Grasps Recognition and Evaluation of Stroke Patients for Supporting Rehabilitation Therapy
-
Upload
transumanisti -
Category
Documents
-
view
5 -
download
0
Transcript of Grasps Recognition and Evaluation of Stroke Patients for Supporting Rehabilitation Therapy
Research ArticleGrasps Recognition and Evaluation of Stroke Patients forSupporting Rehabilitation Therapy
Beatriz Leon1 Angelo Basteris1 Francesco Infarinato2 Patrizio Sale2 Sharon Nijenhuis3
Gerdienke Prange3 and Farshid Amirabdollahian1
1 Adaptive Systems Research Group at the School of Computer Science University of Hertfordshire Hatfield Hertfordshire AL10 9ABUK
2 IRCCS San Raffaele Pisana Via di Val Cannuta 247 00166 Roma Italy3 Roessingh Research and Development Roessinghsbleekweg 33b 7522 AH Enschede The Netherlands
Correspondence should be addressed to Beatriz Leon bleonhertsacuk
Received 27 June 2014 Accepted 18 August 2014 Published 2 September 2014
Academic Editor Giorgio Ferriero
Copyright copy 2014 Beatriz Leon et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Stroke survivors often suffer impairments on their wrist and hand Robot-mediated rehabilitation techniques have been proposedas a way to enhance conventional therapy based on intensive repeated movements Amongst the set of activities of daily livinggrasping is one of the most recurrent Our aim is to incorporate the detection of grasps in the machine-mediated rehabilitationframework so that they can be incorporated into interactive therapeutic games In this study we developed and tested a methodbased on support vector machines for recognizing various grasp postures wearing a passive exoskeleton for hand and wristrehabilitation after stroke The experiment was conducted with ten healthy subjects and eight stroke patients performing thegrasping gestures The method was tested in terms of accuracy and robustness with respect to intersubjectsrsquo variability anddifferences between different grasps Our results show reliable recognition while also indicating that the recognition accuracy canbe used to assess the patientsrsquo ability to consistently repeat the gestures Additionally a grasp quality measure was proposed tomeasure the capabilities of the stroke patients to perform grasp postures in a similar way than healthy people These two measurescan be potentially used as complementary measures to other upper limb motion tests
1 Introduction
Stroke survivors are often unable to perform the fine motorcontrol activities which are required during activities ofdaily living amongst which grasping is one of the mostrecurrent Robots represent an appealing tool for exercise-based approach in neurorehabilitation due to their ability todeliver repetitive training
Previous studies indicate that the inclusion of robotrehabilitation training improves short- and long-term motorcontrol of the impaired upper limb of patients after a stroke[1 2] However evidence of the transfer of robotic trainingeffects to activities in daily life is limited [3] Therefore theinclusion of functional tasks such as grasping objects is vitalto increase the practice and improvement in such activitiesfor stroke rehabilitation [4 5] A next step in this direction
is to incorporate the detection of various grasps in the robotrehabilitation frameworks and to evaluate their quality inorder to detect any improvement in the rehabilitation process
The SCRIPT (Supervised Care ampRehabilitation InvolvingPersonal Telerobotics) project aims at delivering an affordablesystem for home-based rehabilitation of the hand and wristfor stroke survivors [6] A passive exoskeleton (Figure 1)has been developed within the project in order to facilitatethe patientsrsquo fingers and wrist extension Self-administeredtraining at home is performed through repetitive interactionwith games based on functional exercises (rehabilitationgames) to enhance engagement The games are controlledby wearing the orthosis and performing several arm andhand movements The challenge is to use the limited set ofsensors provided by the device to recognize the various grasppostures and incorporate them into the games
Hindawi Publishing CorporationBioMed Research InternationalVolume 2014 Article ID 318016 14 pageshttpdxdoiorg1011552014318016
2 BioMed Research International
Elastic cord
Finger cap
Leaf spring fittedwith bending sensors
Figure 1 SCRIPT passive orthosis showing the bending sensors andleaf springs
In previouswork [7] we implemented and tested differentmethods for recognizing four different grasp postures per-formed while wearing the SCRIPT orthosis We found thatwith the support vector machines (SVM) method we couldachieve an overall accuracy of more than 90 with smallcomputational time SVM has been successfully applied tothe classification of a variety of biomedical conditions [8ndash10] and more specifically to study different aspects of strokepatients including the classification of carotid artery plaques[11] the study of dietary patterns [12] or using readings of ashoe-based sensor to identify sitting standing and walkingpostures [13] However few studies have been reported onthe classification of grasp postures Tavakolan et al [14] usedSVM for pattern recognition of surface electromyographysignals of four forearm muscles in order to classify eighthand gestures On the other hand Puthenveettil et al [15]used linear discriminant analysis to classify hand preshapesin poststroke patients using data from theCyberGlove In thisstudy the SVM approach was selected to perform the gesturerecognition
We conducted an investigation in a group of eightpoststroke patients to determine the accuracy of the gesturerecognition wearing the SCRIPT device using the SVMapproach and compared it with the samples obtained fora group of ten healthy subjects and their relationship tothe level of hand impairment The goal was to assess thegrasp recognition with this particular user group in orderto confirm its suitability prior to deployment in a largerscale clinical evaluation of the prototype orthotic deviceand the SCRIPT system Successful grasp recognition willprovide a more versatile set of gestures to the therapeutichuman-machine interaction system that acts as a mediumto support home-based rehabilitation After a given graspposture is recognized the aim is to evaluate the quality of thegrasp in order to measure the patient progress throughoutthe rehabilitation process In this study we propose a graspquality measure that can be calculated with the availablesensor readings of the orthosisThesemeasures are comparedwith the results of other commonly used upper limb motiontests such as the action research arm test (ARAT) [16]
Table 1 Details of healthy participants
Id Country Age Gender Dominant1 UK 27 Male Right2 UK 32 Male Right3 UK 33 Male Right4 UK 30 Female Right5 UK 32 Female Right6 Italy 47 Female Left7 Italy 50 Male Right8 Italy 52 Female Right9 Italy 50 Male Left10 Italy 51 Female Left
2 Materials and Methods
21 Device The SCRIPT passive exoskeleton [17] is a passivedevicewhich applies external extension torques on the fingersthrough five leaf springs which are connected to the fingercaps through an elastic cord (Figure 1) The elastic cordenables the fingers freedom of movement relative to the leafspring and also allows adjusting the level of extension supportprovided by the device The leaf springs are fitted with com-mercial resistive flex sensors [18] which measure their flexionwith a resolution of 1 degree However the deflection of theleaf spring is not the actual flexion angle of the finger butthe two quantities are related by a monotonically increasingfunction [17] Subjects are free to laterally abductadduct thefingers and oppose the thumb but these movements are notsupported nor sensed by the orthosis The orthosis measuresonly overall finger flexion in a range from 0 to 90 degrees andwrist flexion and extension in a range from 90 to minus45 degrees
It is important to note that this passive orthosis asmany of those in the field of rehabilitation robotics may notconform to the conventional definition of a robot althoughconsisting of sensors passive actuators and a decision mak-ing component However our research is applicable in thefield of robot-assisted training and rehabilitation roboticswhere sensors are utilized towards benchmarking motor per-formance aiding a meaningful two-way interaction betweenthe human and the machine
22 Study Participants Two groups of participants wererecruited for this study a group of healthy subjects anda group of stroke patients In the first group ten healthysubjects with no previous injuries of fingers hand or wristvolunteered to participate in this study (Table 1) This studywas carried out at the University of Hertfordshire UKapproved by the university ethics committee (Ethics protocolnumber COMSTUH00008) and at the IRCCS San RaffaelePisana (Rome Italy) Participants were recruited amongfaculty staff by advertising on an internal mailing list
Thegroupof stroke patientswere selectedwith the criteriaused in the SCRIPT project in order to enable them theuse of the orthosis [19] They were patients with a unilateralischemic or hemorrhagic stroke between 6 months and 5years ago They had limitations in arm and hand function
BioMed Research International 3
while being able to actively flex the elbow by at least 15∘ and toactively flex their fingers by a quarter of their passive rangeThey also had the ability to understand and follow instruc-tions For stroke patients individually fitted orthoses wereusedWith these criteria a total of eight patientswere selected(Table 2) One patient was recruited at the RehabilitationCentreHet Roessingh (Enschede theNetherlands) and sevenat the IRCCS San Raffaele Pisana (Rome Italy) In both casesthe experiment was part of an on-going clinical study forwhich ethical approval had been obtained at the respectivelocal ethics committees
The action research arm test (ARAT) [16] and Fugl-Meyerassessment (FM) [20 21] have been used as quantitativemeasures to evaluate the arm motor recovery after strokeThe ARAT consists of four sections (A) grasp (B) grip (C)pinch and (D) gross arm movement The FM assessment forthe upper extremity consists on a scale of 66 points dividedin three subsections proximal (shoulder and elbow) distal(wrist and hand) and coordination (test of tremor dysmetriaand time) Details of the results of these tests for the strokepatients can also be seen in Table 2
23 Grasp Gestures In recent literature several activitiesof daily living have been identified as the most importantto train after stroke [4 22 23] They include eating withcutlery drinking holding objects while walking takingmoney from purse openclose clothing combing hair andknob manipulation We have selected in agreement withhealthcare professionals three grasp gestures needed in orderto perform these tasks (shown in Figure 2) Two are classifiedas precision grips the tripod (the thumb opposes the indexand middle finger) and the lateral grasp (the thumb holdsan object against the side of the index finger) The third onethe cylindrical grasp is classified as a power grasp (all fingersmake contact with the object) Keller et al [24] identifiedthe tripod and the lateral grasps as the most frequentlyused prehensile patterns for static and dynamic graspingrespectively The relaxed posture of the hand was used as theforth gesture in order to enable the recognition when thesubjects were not performing any grasp
Two of the selected gestures are evaluated in the ARATthe tripod grasp is tested in section A (grasp subtest) andthe cylindrical grasp in section B (grip subtest) of ARAT Itis expected that the recognition of the gestures is related tothe ability to move the hand measured by ARAT
24 Grasp RecognitionMethod Theproblem of hand posturedetection has been previously approached using vision-based[25 26] or glove-based [25 27ndash29] methods dependingon the constraints of the specific applications In our casegiven the bulk of the device vision-based approaches forthe recognition of the hand postures are not suitable as thehand is practically occluded and therefore the recognitionshould be based on the sensory readings for each fingerprovided by the exoskeleton sensors In a previous work wecompared several glove-based approaches to recognize grasppostures performing experiments on five healthy participantswearing the SCRIPT passive exoskeleton [7] We compared
three methods one based on the statistics of the flexion dataanother based on neural networks and finally one based onsupport vector machines (SVM) We found that with thelast method we could achieve an overall accuracy of morethan 90with small computational time (lt60ms)Thereforein this study we used the SVM approach to determine theaccuracy of recognition for healthy participants and strokepatients wearing the SCRIPT device
SVM is a popular machine learning technique for classi-fication A support vector machine is a supervised learningclassifier that constructs a set of hyperplanes in a high-dimensional space (support vectors) that are used to classifythe data A good separation is achieved by the hyperplanethat has the largest distance to the nearest training data pointof any class The hyperplanes are found solving the followingoptimization problem [30]
min120596119887120576
1
2
120596
119879
120596 + 119862
119897
sum
119894=1
120576
119894
subject to 119910
119894(120596
119879
120593 (119909
119894) + 119887 ge 1 minus 120576
119894) 120576
119894ge 0
(1)
where (119909
119894 119910
119894) | 119909
119894isin 119877
119901
119910
119894isin minus1 1 are the training
set of 119897 instance-label pairs 119909119894is 119901-dimensional real vector
120596 the normal vector of the hyperplane and 119862 gt 0 thepenalty parameter of the error term The training vectors 119909
119894
are mapped into a 119901-dimensional spaces by the function 120593
and in order to create nonlinear classifiersa kernel functionis used In our work we used a radial basis function (RBF) asthe kernel function given that it can handle the case when therelation between class labels and attributes is nonlinear [31]It is defined as
119870(119909
119894 119909
119895) = exp (minus120574100381710038171003817
1003817
1003817
119909
119894minus119909
119895
1003817
1003817
1003817
1003817
1003817
2
) 120574 gt 0 (2)
where 120574 is the kernel parameter Therefore two parametersare needed 119862and 120574 In order to find the best values forthese parameters we used a V-fold cross-validation techniquedividing the training set for one subject into V subsets of equalsize and testing the classifier on the remaining V minus 1 subsetsIn this work Vwas taken as 5 the value of the cost parameter119862 was varied as 119862 = 2
119909 119909 = minus5 5 and the value of thekernel parameter 120574 was varied as 120574 = 2
119910 119910 = minus4 0Thevalues that gave the highest validation accuracy were 119862 = 4
and 120574 = 1The method was implemented in Python using the LIB-
SVM package (httpwwwcsientuedutwsimcjlinlibsvm)The flexion angles were normalized in the range from 0 to 1(corresponding to 0 to 90 degrees) and the selected error tostop the training phase was set to 0001
25 Grasp Evaluation Method In the field of robotics manygrasp quality measures have been developed that allow thecomparison of different aspects of the robotic grasp [32] In[33] the most common robot grasp quality measures havebeen adapted to the evaluation of the grasp of the humanhand From this set of quality measures the one proposed by[34] whichmeasures how close a given grasp is to a referenceposture is the only one that can be used with the sensor
4 BioMed Research International
Table2Detailsof
thes
troke
patie
nts
IdCou
ntry
Age
(yrs)
Times
ince
acutee
vent
Gender
Affected
Dom
inant
Stroke
type
ARA
T(m
ax57
points)
FM(m
ax66
points)
AB
CD
Total
Proxim
alDistal
Coo
rdination
Total
1Holland
6913
mon
ths
Male
Left
Left
Ischem
ia18
1017
853
2916
449
2Ita
ly73
7mon
ths
Female
Left
Right
Ischem
ia12
812
941
2812
343
3Ita
ly88
6mon
ths
Male
Left
Right
Ischem
ia12
86
632
2712
443
4Ita
ly83
7mon
ths
Male
Left
Right
Ischem
ia6
40
313
2811
342
5Ita
ly85
6mon
ths
Male
Right
Right
Ischem
ia12
812
941
2812
343
6Ita
ly72
6mon
ths
Male
Left
Right
Ischem
ia18
1212
951
2823
354
7Ita
ly81
7mon
ths
Female
Left
Right
Ischem
ia18
1212
951
3524
362
8Ita
ly69
7mon
ths
Male
Left
Right
Ischem
ia12
86
632
2612
341
ARA
T=arm
scores
ofthea
ctionresearch
arm
test
FM=arm
scores
oftheF
ugl-M
eyer
motor
assessment
BioMed Research International 5
(a) Tripod grasp (b) Lateral grasp
(c) Cylindrical grasp (d) Relaxed position
Figure 2 Selected gestures to be recognized performed wearing the SCRIPT device
information provided by the SCRIPT orthosisThis index hasbeen adapted for this study as follows
119866119876 = 1 minus
119899
sum
119894=1
120596
119894(
119910
119894minus 119886
119894
119877
119894
)
2
(3)
where 119899 is the number of hand joints 120596119894is a weight factor
119910
119894is the current finger flexion angle and 119877
119894is the joint angle
range used to normalize the index calculated as themaximumbetween the reference posture 119886
119894and either the upper or lower
angle limit The index has to be maximized so that the graspis optimal when all joints are at the reference posture havinga quality measure of one and it goes to zero when all its jointsare at their maximum angle limits
The reference posture 119886
119894is taken in this case as the
one performed by the healthy subjects This measure thenenables the evaluation of how far is a poststroke patient fromperforming a grasp in away similar to a healthy subject How-ever the postures obtained from healthy subjects performingeach gesture are likely to have variations between subjectsespecially when some of the fingers were not playing an activerole in the grasp (eg the ring and little fingers in the tripodgrasp) In order to consider this variance we have includeda weight factor 120596
119894that will give high scores to fingers whose
postures have small standard deviation and vice versa The
corresponding weights for each finger and a given gesture canbe calculated as
119908
119894119892=
120572
119892
120590
119894119892
(4)
where 119894 and 119892 are the given finger and gesture 120590119894119892
is thestandard deviation of the finger flexion over all subjects forthe given finger and grasp and 120572
119892is calculated as
120572
119892= 5
5
prod
119894=1
120590
119894119892
times ((prod120590
119894119892)
119894=1234
+ (prod120590
119894119892)
119894=1345
+ (prod120590
119894119892)
119894=1245
+ (prod120590
119894119892)
119894=1235
+(prod120590
119894119892)
119894=2345
)
minus1
(5)
26 Experimental Protocol The participants took part inone session lasting half an hour conducted by a researcheror a therapist They were asked to wear a SCRIPT passiveorthosis on the impaired hand or one of their hands (in thecase of healthy participants) while sitting in front of a PCSubsequently they were instructed to mimic the picture of
6 BioMed Research International
a gesture shown on the screen (Figure 2) The participantthen confirmed that heshe achieved the desired gesture andthen pressing a button on the screen the flexion angles ofthe gesturewere saved After confirmation theywere asked torelax the hand and press a button At that moment the anglesof the relaxed posture were also saved
Each subject performed 6 repetitions of each gesture ina pseudorandom sequence resulting in the capture of 24gestures Data were then postprocessed by Python ad hocapplications The results of the classification system werecalculated using the following values for each gesture 119894
(i) True positives (TP) gestures correctly classified asgesture 119894
(ii) False negatives (FN) gestures 119894 incorrectly classifiedas other gestures
(iii) False positives (FP) other gestures incorrectly classi-fied as gesture 119894
(iv) True negatives (TN) other gestures correctly classi-fied as nongesture 119894
These are commonly used to evaluate the sensitivity andspecificity of a clinical test in its ability to confirmor refute thepresence of a diseaseThe sensitivity (true positive rate) refersto the ability of the test to correctly identify those patientswith the disease and the specificity (true negative rate) refersto the ability to correctly identify those patients without thedisease [35] In this study the outcome of the test is not binarytherefore we calculated the performance measures focusingon each gesture We used the accuracy (ability to correctlyidentify positive and negatives) true positive rate (ability tocorrectly identify the positives) and false positive rate (lackof ability to correctly classify the negatives) The last oneis measuring the opposite of the specificity as it enables usto focus on evaluating the recognition performance for agiven gesture instead of looking at the classification of othergestures They are calculated as follows
Accuracy = TP + TNTotal testing data
True positive rate (TPR) = TPTP + FN
False positive rate (FPR) = FPTN + FP
(6)
27 Data Analysis The data acquired for each subject wasdivided into two sets for training and testing purposes Asthe patients will need to perform the training procedure eachtime before starting the games (or the games will not beable to reliably recognize their gestures) the less numberof samples required to train the model the better In [7]it has been shown that a high accuracy can be achievedwith 4 training samples thus allowing very short calibrationtime and making this approach suitable for home-basedrehabilitationThen a training set using 4 samples per gesturewas used to train the model Results were considered takinginto account all the possible permutations of 4 trainingsamples
The overall results of the recognition performance aresummarized as median and interquartile range (using boxplots) differentiated between the participantrsquos type (healthyor stroke patient) The information provided by the differentperformance measures is compared using a Pearson corre-lation in order to determine if we can rely on one measure-ment for the recognition assessment We also compared thevariability of the flexion angles over all subjects performingthe different grasp gestures and the results of the accuracy ofgesture recognition per participant
Additionally the results of the proposed grasp qualitymeasure are also presented summarized using box plotsshowing the variation between the participantrsquos type thedifferent gestures and the intervariability between the par-ticipants In order to assess how these three parametersinfluenced the results of the grasp quality measure a multipleregression analysis was performed This type of analysiscan be used to consider multiple independent variablescalculating the least square estimates for a data set [36] Usingthis analysis we can devise our model using the followingequation
119866119876
119894= 119887
0+ 119887
1Participant type
119894
+ 119887
2Grasp type
119894
+ 119887
3Subject
119894
+ 120576
119894
(7)
where 119887119900represents the constant and 120576 is the modelling error
In this case predictors are classification variables with morethan two categories therefore dummy coding is required toinclude them in the regression equation The technique todo this coding is to create an independent variable for eachindependent category except one as a dichotomyThe omittedvariable provides a baseline for comparison while avoidingmulticollinearity [36] In this analysis we used ldquohealthyparticipantsrdquo ldquorelaxed posturerdquo and ldquosubject 1 (healthy)rdquo asour base line for each one of our predictors The ldquoEnterrdquomethod was used in order to force the model to consider allvariables as significant variables in the model
It is intuitive to assume that the level of impairment of thepatients should be correlated with their ability to consistentlyperform the gestures in a similar way (recognition accuracy)and similar to the ones performed by healthy subjects (GQmeasure) As the ARAT and FM tests are common waysto evaluate the level of capabilities of the upper limb wecorrelated the accuracies of recognition and the grasp qualitywith the results of these test using the Pearson coefficientTheIBM SPSS statistical package for Windows version 210 wasused to perform the analysis of the data
3 Results
This section presents the results of the experiments in twoparts the results of the recognition of the different gesturesperformed by healthy subjects and stroke patients and theresults of the evaluation of those grasps
31 Grasp Recognition Results The results of the true positiverate of the training and testing phases of the experimentsare presented in Table 3 The number of gestures correctly
BioMed Research International 7
ParticipantsStroke patientsHealthy subjects
Accu
racy
()
100
80
60
40
20
0
(a)
ParticipantsStroke patientsHealthy subjects
TPR
()
100
80
60
40
20
0
(b)
FPR
()
100
80
60
40
20
0
ParticipantsStroke patientsHealthy subjects
(c)
Figure 3 Performance measures evaluating the recognition of grasp postures (a) accuracy (b) true positive rate and (c) false positive rateResults are presented discriminated between healthy and stroke participants
Table 3 Training time and true positive rate () between thetraining and testing phases
Meantraining time
(ms)
Meantraining TPR
()
Mean testingTPR ()
Healthysubjects
1 1800 plusmn 06 58 plusmn 14 87 plusmn 22 1733 plusmn 06 100 plusmn 0 100 plusmn 03 2200 plusmn 07 84 plusmn 14 91 plusmn 04 1933 plusmn 05 79 plusmn 25 98 plusmn 35 1733 plusmn 06 84 plusmn 14 91 plusmn 06 1867 plusmn 05 58 plusmn 10 83 plusmn 47 2200 plusmn 14 63 plusmn 9 85 plusmn 48 2267 plusmn 16 98 plusmn 8 97 plusmn 49 2200 plusmn 06 31 plusmn 9 91 plusmn 110 1867 plusmn 05 84 plusmn 16 90 plusmn 2
1980 plusmn 05 74 plusmn 24 91 plusmn 6
Strokepatients
1 2600 plusmn 07 100 plusmn 0 99 plusmn 22 2200 plusmn 04 61 plusmn 15 91 plusmn 43 2200 plusmn 04 38 plusmn 6 59 plusmn 44 1933 plusmn 07 38 plusmn 11 56 plusmn 45 2467 plusmn 06 38 plusmn 6 58 plusmn 46 2067 plusmn 06 33 plusmn 6 68 plusmn 57 2067 plusmn 03 47 plusmn 19 93 plusmn 28 2133 plusmn 04 43 plusmn 26 78 plusmn 3
2208 plusmn 06 50 plusmn 25 75 plusmn 17
recognized greatly increased from the training to the testingphase It can clearly be seen that the testing overall recogni-tion performance of gestures performed by the stroke patientsis lower than the ones performed by the healthy subjects asit was expected but still they produced a high percentageof true positive recognized gestures (TPR mean of 75)The computational time taken for training the model was onaverage 198ms for healthy subjects and 221ms for strokepatients
In order to evaluate inmore detail the recognition resultsthe selected performance measures are shown in Figure 3Ideally the accuracy and true positive rate should be closeto 100 and the false positive rate close to zero In this casethe median accuracy of recognition of gestures performed byhealthy subjects was over 95 and 87 for stroke patientsThe true positive rate showed lower values with respect to theaccuracy median values for healthy subjects of over 91 andfor stroke patients of over 75 The healthy subjects showedmedian values of false positive rates of 2 and 8 for strokepatients
The different performance measures can also provideinformation about the specific recognition of each gestureSpecifically the true positive rate refers to the ability ofthe method to correctly identify a specific gesture whilstthe false positive rate refers to the percentage of gestureswrongly classified Figure 4 shows the accuracy true positiverate and false positive rate for each gesture The difficultyof recognition of the different grasp gestures is different forthe healthy subjects or stroke patients The relax postureis a clearly distinctive gesture therefore showing the bestaccuracies best TPR and FPR values close to zero Forhealthy subjects the tripod and cylindrical gestures were themost difficult to be recognized and the most misrecognizedFor stroke patients the three gestures presented similardifficulties to be recognized and the tripod and lateral graspwere the most misrecognized
Table 4 presents the Pearson correlation values of thedifferent performancemeasures for the given conditionsTheaccuracy is inversely correlated with the false negative rateover all conditions (correlation coefficient C lt minus097 withstatistical significance) and also highly correlated with theTPR (C gt 099) The different measures of performance pro-vide specific information on the performance of recognitionbut the accuracy could be selected as the overall measureof recognition performance Therefore for the followinganalysis the accuracy will be used
It is expected that the grasping capabilities of the par-ticipants affect the recognition of hand postures especiallyin the case of stroke patients Therefore we also studied
8 BioMed Research InternationalAc
cura
cy (
)
100
80
60
40
20
0
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
Accu
racy
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(a)
TPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
TPR
()
100
80
60
40
20
0
Part
icip
ants
Part
icip
ants
Hea
lthy
subj
ects
Stro
ke p
atie
nts
(b)
FPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
FPR
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(c)
Figure 4 Performance measures for each gesture evaluating the recognition of hand postures (a) accuracy (b) true positive rate and (c)false positive rate
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
2 BioMed Research International
Elastic cord
Finger cap
Leaf spring fittedwith bending sensors
Figure 1 SCRIPT passive orthosis showing the bending sensors andleaf springs
In previouswork [7] we implemented and tested differentmethods for recognizing four different grasp postures per-formed while wearing the SCRIPT orthosis We found thatwith the support vector machines (SVM) method we couldachieve an overall accuracy of more than 90 with smallcomputational time SVM has been successfully applied tothe classification of a variety of biomedical conditions [8ndash10] and more specifically to study different aspects of strokepatients including the classification of carotid artery plaques[11] the study of dietary patterns [12] or using readings of ashoe-based sensor to identify sitting standing and walkingpostures [13] However few studies have been reported onthe classification of grasp postures Tavakolan et al [14] usedSVM for pattern recognition of surface electromyographysignals of four forearm muscles in order to classify eighthand gestures On the other hand Puthenveettil et al [15]used linear discriminant analysis to classify hand preshapesin poststroke patients using data from theCyberGlove In thisstudy the SVM approach was selected to perform the gesturerecognition
We conducted an investigation in a group of eightpoststroke patients to determine the accuracy of the gesturerecognition wearing the SCRIPT device using the SVMapproach and compared it with the samples obtained fora group of ten healthy subjects and their relationship tothe level of hand impairment The goal was to assess thegrasp recognition with this particular user group in orderto confirm its suitability prior to deployment in a largerscale clinical evaluation of the prototype orthotic deviceand the SCRIPT system Successful grasp recognition willprovide a more versatile set of gestures to the therapeutichuman-machine interaction system that acts as a mediumto support home-based rehabilitation After a given graspposture is recognized the aim is to evaluate the quality of thegrasp in order to measure the patient progress throughoutthe rehabilitation process In this study we propose a graspquality measure that can be calculated with the availablesensor readings of the orthosisThesemeasures are comparedwith the results of other commonly used upper limb motiontests such as the action research arm test (ARAT) [16]
Table 1 Details of healthy participants
Id Country Age Gender Dominant1 UK 27 Male Right2 UK 32 Male Right3 UK 33 Male Right4 UK 30 Female Right5 UK 32 Female Right6 Italy 47 Female Left7 Italy 50 Male Right8 Italy 52 Female Right9 Italy 50 Male Left10 Italy 51 Female Left
2 Materials and Methods
21 Device The SCRIPT passive exoskeleton [17] is a passivedevicewhich applies external extension torques on the fingersthrough five leaf springs which are connected to the fingercaps through an elastic cord (Figure 1) The elastic cordenables the fingers freedom of movement relative to the leafspring and also allows adjusting the level of extension supportprovided by the device The leaf springs are fitted with com-mercial resistive flex sensors [18] which measure their flexionwith a resolution of 1 degree However the deflection of theleaf spring is not the actual flexion angle of the finger butthe two quantities are related by a monotonically increasingfunction [17] Subjects are free to laterally abductadduct thefingers and oppose the thumb but these movements are notsupported nor sensed by the orthosis The orthosis measuresonly overall finger flexion in a range from 0 to 90 degrees andwrist flexion and extension in a range from 90 to minus45 degrees
It is important to note that this passive orthosis asmany of those in the field of rehabilitation robotics may notconform to the conventional definition of a robot althoughconsisting of sensors passive actuators and a decision mak-ing component However our research is applicable in thefield of robot-assisted training and rehabilitation roboticswhere sensors are utilized towards benchmarking motor per-formance aiding a meaningful two-way interaction betweenthe human and the machine
22 Study Participants Two groups of participants wererecruited for this study a group of healthy subjects anda group of stroke patients In the first group ten healthysubjects with no previous injuries of fingers hand or wristvolunteered to participate in this study (Table 1) This studywas carried out at the University of Hertfordshire UKapproved by the university ethics committee (Ethics protocolnumber COMSTUH00008) and at the IRCCS San RaffaelePisana (Rome Italy) Participants were recruited amongfaculty staff by advertising on an internal mailing list
Thegroupof stroke patientswere selectedwith the criteriaused in the SCRIPT project in order to enable them theuse of the orthosis [19] They were patients with a unilateralischemic or hemorrhagic stroke between 6 months and 5years ago They had limitations in arm and hand function
BioMed Research International 3
while being able to actively flex the elbow by at least 15∘ and toactively flex their fingers by a quarter of their passive rangeThey also had the ability to understand and follow instruc-tions For stroke patients individually fitted orthoses wereusedWith these criteria a total of eight patientswere selected(Table 2) One patient was recruited at the RehabilitationCentreHet Roessingh (Enschede theNetherlands) and sevenat the IRCCS San Raffaele Pisana (Rome Italy) In both casesthe experiment was part of an on-going clinical study forwhich ethical approval had been obtained at the respectivelocal ethics committees
The action research arm test (ARAT) [16] and Fugl-Meyerassessment (FM) [20 21] have been used as quantitativemeasures to evaluate the arm motor recovery after strokeThe ARAT consists of four sections (A) grasp (B) grip (C)pinch and (D) gross arm movement The FM assessment forthe upper extremity consists on a scale of 66 points dividedin three subsections proximal (shoulder and elbow) distal(wrist and hand) and coordination (test of tremor dysmetriaand time) Details of the results of these tests for the strokepatients can also be seen in Table 2
23 Grasp Gestures In recent literature several activitiesof daily living have been identified as the most importantto train after stroke [4 22 23] They include eating withcutlery drinking holding objects while walking takingmoney from purse openclose clothing combing hair andknob manipulation We have selected in agreement withhealthcare professionals three grasp gestures needed in orderto perform these tasks (shown in Figure 2) Two are classifiedas precision grips the tripod (the thumb opposes the indexand middle finger) and the lateral grasp (the thumb holdsan object against the side of the index finger) The third onethe cylindrical grasp is classified as a power grasp (all fingersmake contact with the object) Keller et al [24] identifiedthe tripod and the lateral grasps as the most frequentlyused prehensile patterns for static and dynamic graspingrespectively The relaxed posture of the hand was used as theforth gesture in order to enable the recognition when thesubjects were not performing any grasp
Two of the selected gestures are evaluated in the ARATthe tripod grasp is tested in section A (grasp subtest) andthe cylindrical grasp in section B (grip subtest) of ARAT Itis expected that the recognition of the gestures is related tothe ability to move the hand measured by ARAT
24 Grasp RecognitionMethod Theproblem of hand posturedetection has been previously approached using vision-based[25 26] or glove-based [25 27ndash29] methods dependingon the constraints of the specific applications In our casegiven the bulk of the device vision-based approaches forthe recognition of the hand postures are not suitable as thehand is practically occluded and therefore the recognitionshould be based on the sensory readings for each fingerprovided by the exoskeleton sensors In a previous work wecompared several glove-based approaches to recognize grasppostures performing experiments on five healthy participantswearing the SCRIPT passive exoskeleton [7] We compared
three methods one based on the statistics of the flexion dataanother based on neural networks and finally one based onsupport vector machines (SVM) We found that with thelast method we could achieve an overall accuracy of morethan 90with small computational time (lt60ms)Thereforein this study we used the SVM approach to determine theaccuracy of recognition for healthy participants and strokepatients wearing the SCRIPT device
SVM is a popular machine learning technique for classi-fication A support vector machine is a supervised learningclassifier that constructs a set of hyperplanes in a high-dimensional space (support vectors) that are used to classifythe data A good separation is achieved by the hyperplanethat has the largest distance to the nearest training data pointof any class The hyperplanes are found solving the followingoptimization problem [30]
min120596119887120576
1
2
120596
119879
120596 + 119862
119897
sum
119894=1
120576
119894
subject to 119910
119894(120596
119879
120593 (119909
119894) + 119887 ge 1 minus 120576
119894) 120576
119894ge 0
(1)
where (119909
119894 119910
119894) | 119909
119894isin 119877
119901
119910
119894isin minus1 1 are the training
set of 119897 instance-label pairs 119909119894is 119901-dimensional real vector
120596 the normal vector of the hyperplane and 119862 gt 0 thepenalty parameter of the error term The training vectors 119909
119894
are mapped into a 119901-dimensional spaces by the function 120593
and in order to create nonlinear classifiersa kernel functionis used In our work we used a radial basis function (RBF) asthe kernel function given that it can handle the case when therelation between class labels and attributes is nonlinear [31]It is defined as
119870(119909
119894 119909
119895) = exp (minus120574100381710038171003817
1003817
1003817
119909
119894minus119909
119895
1003817
1003817
1003817
1003817
1003817
2
) 120574 gt 0 (2)
where 120574 is the kernel parameter Therefore two parametersare needed 119862and 120574 In order to find the best values forthese parameters we used a V-fold cross-validation techniquedividing the training set for one subject into V subsets of equalsize and testing the classifier on the remaining V minus 1 subsetsIn this work Vwas taken as 5 the value of the cost parameter119862 was varied as 119862 = 2
119909 119909 = minus5 5 and the value of thekernel parameter 120574 was varied as 120574 = 2
119910 119910 = minus4 0Thevalues that gave the highest validation accuracy were 119862 = 4
and 120574 = 1The method was implemented in Python using the LIB-
SVM package (httpwwwcsientuedutwsimcjlinlibsvm)The flexion angles were normalized in the range from 0 to 1(corresponding to 0 to 90 degrees) and the selected error tostop the training phase was set to 0001
25 Grasp Evaluation Method In the field of robotics manygrasp quality measures have been developed that allow thecomparison of different aspects of the robotic grasp [32] In[33] the most common robot grasp quality measures havebeen adapted to the evaluation of the grasp of the humanhand From this set of quality measures the one proposed by[34] whichmeasures how close a given grasp is to a referenceposture is the only one that can be used with the sensor
4 BioMed Research International
Table2Detailsof
thes
troke
patie
nts
IdCou
ntry
Age
(yrs)
Times
ince
acutee
vent
Gender
Affected
Dom
inant
Stroke
type
ARA
T(m
ax57
points)
FM(m
ax66
points)
AB
CD
Total
Proxim
alDistal
Coo
rdination
Total
1Holland
6913
mon
ths
Male
Left
Left
Ischem
ia18
1017
853
2916
449
2Ita
ly73
7mon
ths
Female
Left
Right
Ischem
ia12
812
941
2812
343
3Ita
ly88
6mon
ths
Male
Left
Right
Ischem
ia12
86
632
2712
443
4Ita
ly83
7mon
ths
Male
Left
Right
Ischem
ia6
40
313
2811
342
5Ita
ly85
6mon
ths
Male
Right
Right
Ischem
ia12
812
941
2812
343
6Ita
ly72
6mon
ths
Male
Left
Right
Ischem
ia18
1212
951
2823
354
7Ita
ly81
7mon
ths
Female
Left
Right
Ischem
ia18
1212
951
3524
362
8Ita
ly69
7mon
ths
Male
Left
Right
Ischem
ia12
86
632
2612
341
ARA
T=arm
scores
ofthea
ctionresearch
arm
test
FM=arm
scores
oftheF
ugl-M
eyer
motor
assessment
BioMed Research International 5
(a) Tripod grasp (b) Lateral grasp
(c) Cylindrical grasp (d) Relaxed position
Figure 2 Selected gestures to be recognized performed wearing the SCRIPT device
information provided by the SCRIPT orthosisThis index hasbeen adapted for this study as follows
119866119876 = 1 minus
119899
sum
119894=1
120596
119894(
119910
119894minus 119886
119894
119877
119894
)
2
(3)
where 119899 is the number of hand joints 120596119894is a weight factor
119910
119894is the current finger flexion angle and 119877
119894is the joint angle
range used to normalize the index calculated as themaximumbetween the reference posture 119886
119894and either the upper or lower
angle limit The index has to be maximized so that the graspis optimal when all joints are at the reference posture havinga quality measure of one and it goes to zero when all its jointsare at their maximum angle limits
The reference posture 119886
119894is taken in this case as the
one performed by the healthy subjects This measure thenenables the evaluation of how far is a poststroke patient fromperforming a grasp in away similar to a healthy subject How-ever the postures obtained from healthy subjects performingeach gesture are likely to have variations between subjectsespecially when some of the fingers were not playing an activerole in the grasp (eg the ring and little fingers in the tripodgrasp) In order to consider this variance we have includeda weight factor 120596
119894that will give high scores to fingers whose
postures have small standard deviation and vice versa The
corresponding weights for each finger and a given gesture canbe calculated as
119908
119894119892=
120572
119892
120590
119894119892
(4)
where 119894 and 119892 are the given finger and gesture 120590119894119892
is thestandard deviation of the finger flexion over all subjects forthe given finger and grasp and 120572
119892is calculated as
120572
119892= 5
5
prod
119894=1
120590
119894119892
times ((prod120590
119894119892)
119894=1234
+ (prod120590
119894119892)
119894=1345
+ (prod120590
119894119892)
119894=1245
+ (prod120590
119894119892)
119894=1235
+(prod120590
119894119892)
119894=2345
)
minus1
(5)
26 Experimental Protocol The participants took part inone session lasting half an hour conducted by a researcheror a therapist They were asked to wear a SCRIPT passiveorthosis on the impaired hand or one of their hands (in thecase of healthy participants) while sitting in front of a PCSubsequently they were instructed to mimic the picture of
6 BioMed Research International
a gesture shown on the screen (Figure 2) The participantthen confirmed that heshe achieved the desired gesture andthen pressing a button on the screen the flexion angles ofthe gesturewere saved After confirmation theywere asked torelax the hand and press a button At that moment the anglesof the relaxed posture were also saved
Each subject performed 6 repetitions of each gesture ina pseudorandom sequence resulting in the capture of 24gestures Data were then postprocessed by Python ad hocapplications The results of the classification system werecalculated using the following values for each gesture 119894
(i) True positives (TP) gestures correctly classified asgesture 119894
(ii) False negatives (FN) gestures 119894 incorrectly classifiedas other gestures
(iii) False positives (FP) other gestures incorrectly classi-fied as gesture 119894
(iv) True negatives (TN) other gestures correctly classi-fied as nongesture 119894
These are commonly used to evaluate the sensitivity andspecificity of a clinical test in its ability to confirmor refute thepresence of a diseaseThe sensitivity (true positive rate) refersto the ability of the test to correctly identify those patientswith the disease and the specificity (true negative rate) refersto the ability to correctly identify those patients without thedisease [35] In this study the outcome of the test is not binarytherefore we calculated the performance measures focusingon each gesture We used the accuracy (ability to correctlyidentify positive and negatives) true positive rate (ability tocorrectly identify the positives) and false positive rate (lackof ability to correctly classify the negatives) The last oneis measuring the opposite of the specificity as it enables usto focus on evaluating the recognition performance for agiven gesture instead of looking at the classification of othergestures They are calculated as follows
Accuracy = TP + TNTotal testing data
True positive rate (TPR) = TPTP + FN
False positive rate (FPR) = FPTN + FP
(6)
27 Data Analysis The data acquired for each subject wasdivided into two sets for training and testing purposes Asthe patients will need to perform the training procedure eachtime before starting the games (or the games will not beable to reliably recognize their gestures) the less numberof samples required to train the model the better In [7]it has been shown that a high accuracy can be achievedwith 4 training samples thus allowing very short calibrationtime and making this approach suitable for home-basedrehabilitationThen a training set using 4 samples per gesturewas used to train the model Results were considered takinginto account all the possible permutations of 4 trainingsamples
The overall results of the recognition performance aresummarized as median and interquartile range (using boxplots) differentiated between the participantrsquos type (healthyor stroke patient) The information provided by the differentperformance measures is compared using a Pearson corre-lation in order to determine if we can rely on one measure-ment for the recognition assessment We also compared thevariability of the flexion angles over all subjects performingthe different grasp gestures and the results of the accuracy ofgesture recognition per participant
Additionally the results of the proposed grasp qualitymeasure are also presented summarized using box plotsshowing the variation between the participantrsquos type thedifferent gestures and the intervariability between the par-ticipants In order to assess how these three parametersinfluenced the results of the grasp quality measure a multipleregression analysis was performed This type of analysiscan be used to consider multiple independent variablescalculating the least square estimates for a data set [36] Usingthis analysis we can devise our model using the followingequation
119866119876
119894= 119887
0+ 119887
1Participant type
119894
+ 119887
2Grasp type
119894
+ 119887
3Subject
119894
+ 120576
119894
(7)
where 119887119900represents the constant and 120576 is the modelling error
In this case predictors are classification variables with morethan two categories therefore dummy coding is required toinclude them in the regression equation The technique todo this coding is to create an independent variable for eachindependent category except one as a dichotomyThe omittedvariable provides a baseline for comparison while avoidingmulticollinearity [36] In this analysis we used ldquohealthyparticipantsrdquo ldquorelaxed posturerdquo and ldquosubject 1 (healthy)rdquo asour base line for each one of our predictors The ldquoEnterrdquomethod was used in order to force the model to consider allvariables as significant variables in the model
It is intuitive to assume that the level of impairment of thepatients should be correlated with their ability to consistentlyperform the gestures in a similar way (recognition accuracy)and similar to the ones performed by healthy subjects (GQmeasure) As the ARAT and FM tests are common waysto evaluate the level of capabilities of the upper limb wecorrelated the accuracies of recognition and the grasp qualitywith the results of these test using the Pearson coefficientTheIBM SPSS statistical package for Windows version 210 wasused to perform the analysis of the data
3 Results
This section presents the results of the experiments in twoparts the results of the recognition of the different gesturesperformed by healthy subjects and stroke patients and theresults of the evaluation of those grasps
31 Grasp Recognition Results The results of the true positiverate of the training and testing phases of the experimentsare presented in Table 3 The number of gestures correctly
BioMed Research International 7
ParticipantsStroke patientsHealthy subjects
Accu
racy
()
100
80
60
40
20
0
(a)
ParticipantsStroke patientsHealthy subjects
TPR
()
100
80
60
40
20
0
(b)
FPR
()
100
80
60
40
20
0
ParticipantsStroke patientsHealthy subjects
(c)
Figure 3 Performance measures evaluating the recognition of grasp postures (a) accuracy (b) true positive rate and (c) false positive rateResults are presented discriminated between healthy and stroke participants
Table 3 Training time and true positive rate () between thetraining and testing phases
Meantraining time
(ms)
Meantraining TPR
()
Mean testingTPR ()
Healthysubjects
1 1800 plusmn 06 58 plusmn 14 87 plusmn 22 1733 plusmn 06 100 plusmn 0 100 plusmn 03 2200 plusmn 07 84 plusmn 14 91 plusmn 04 1933 plusmn 05 79 plusmn 25 98 plusmn 35 1733 plusmn 06 84 plusmn 14 91 plusmn 06 1867 plusmn 05 58 plusmn 10 83 plusmn 47 2200 plusmn 14 63 plusmn 9 85 plusmn 48 2267 plusmn 16 98 plusmn 8 97 plusmn 49 2200 plusmn 06 31 plusmn 9 91 plusmn 110 1867 plusmn 05 84 plusmn 16 90 plusmn 2
1980 plusmn 05 74 plusmn 24 91 plusmn 6
Strokepatients
1 2600 plusmn 07 100 plusmn 0 99 plusmn 22 2200 plusmn 04 61 plusmn 15 91 plusmn 43 2200 plusmn 04 38 plusmn 6 59 plusmn 44 1933 plusmn 07 38 plusmn 11 56 plusmn 45 2467 plusmn 06 38 plusmn 6 58 plusmn 46 2067 plusmn 06 33 plusmn 6 68 plusmn 57 2067 plusmn 03 47 plusmn 19 93 plusmn 28 2133 plusmn 04 43 plusmn 26 78 plusmn 3
2208 plusmn 06 50 plusmn 25 75 plusmn 17
recognized greatly increased from the training to the testingphase It can clearly be seen that the testing overall recogni-tion performance of gestures performed by the stroke patientsis lower than the ones performed by the healthy subjects asit was expected but still they produced a high percentageof true positive recognized gestures (TPR mean of 75)The computational time taken for training the model was onaverage 198ms for healthy subjects and 221ms for strokepatients
In order to evaluate inmore detail the recognition resultsthe selected performance measures are shown in Figure 3Ideally the accuracy and true positive rate should be closeto 100 and the false positive rate close to zero In this casethe median accuracy of recognition of gestures performed byhealthy subjects was over 95 and 87 for stroke patientsThe true positive rate showed lower values with respect to theaccuracy median values for healthy subjects of over 91 andfor stroke patients of over 75 The healthy subjects showedmedian values of false positive rates of 2 and 8 for strokepatients
The different performance measures can also provideinformation about the specific recognition of each gestureSpecifically the true positive rate refers to the ability ofthe method to correctly identify a specific gesture whilstthe false positive rate refers to the percentage of gestureswrongly classified Figure 4 shows the accuracy true positiverate and false positive rate for each gesture The difficultyof recognition of the different grasp gestures is different forthe healthy subjects or stroke patients The relax postureis a clearly distinctive gesture therefore showing the bestaccuracies best TPR and FPR values close to zero Forhealthy subjects the tripod and cylindrical gestures were themost difficult to be recognized and the most misrecognizedFor stroke patients the three gestures presented similardifficulties to be recognized and the tripod and lateral graspwere the most misrecognized
Table 4 presents the Pearson correlation values of thedifferent performancemeasures for the given conditionsTheaccuracy is inversely correlated with the false negative rateover all conditions (correlation coefficient C lt minus097 withstatistical significance) and also highly correlated with theTPR (C gt 099) The different measures of performance pro-vide specific information on the performance of recognitionbut the accuracy could be selected as the overall measureof recognition performance Therefore for the followinganalysis the accuracy will be used
It is expected that the grasping capabilities of the par-ticipants affect the recognition of hand postures especiallyin the case of stroke patients Therefore we also studied
8 BioMed Research InternationalAc
cura
cy (
)
100
80
60
40
20
0
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
Accu
racy
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(a)
TPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
TPR
()
100
80
60
40
20
0
Part
icip
ants
Part
icip
ants
Hea
lthy
subj
ects
Stro
ke p
atie
nts
(b)
FPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
FPR
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(c)
Figure 4 Performance measures for each gesture evaluating the recognition of hand postures (a) accuracy (b) true positive rate and (c)false positive rate
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
BioMed Research International 3
while being able to actively flex the elbow by at least 15∘ and toactively flex their fingers by a quarter of their passive rangeThey also had the ability to understand and follow instruc-tions For stroke patients individually fitted orthoses wereusedWith these criteria a total of eight patientswere selected(Table 2) One patient was recruited at the RehabilitationCentreHet Roessingh (Enschede theNetherlands) and sevenat the IRCCS San Raffaele Pisana (Rome Italy) In both casesthe experiment was part of an on-going clinical study forwhich ethical approval had been obtained at the respectivelocal ethics committees
The action research arm test (ARAT) [16] and Fugl-Meyerassessment (FM) [20 21] have been used as quantitativemeasures to evaluate the arm motor recovery after strokeThe ARAT consists of four sections (A) grasp (B) grip (C)pinch and (D) gross arm movement The FM assessment forthe upper extremity consists on a scale of 66 points dividedin three subsections proximal (shoulder and elbow) distal(wrist and hand) and coordination (test of tremor dysmetriaand time) Details of the results of these tests for the strokepatients can also be seen in Table 2
23 Grasp Gestures In recent literature several activitiesof daily living have been identified as the most importantto train after stroke [4 22 23] They include eating withcutlery drinking holding objects while walking takingmoney from purse openclose clothing combing hair andknob manipulation We have selected in agreement withhealthcare professionals three grasp gestures needed in orderto perform these tasks (shown in Figure 2) Two are classifiedas precision grips the tripod (the thumb opposes the indexand middle finger) and the lateral grasp (the thumb holdsan object against the side of the index finger) The third onethe cylindrical grasp is classified as a power grasp (all fingersmake contact with the object) Keller et al [24] identifiedthe tripod and the lateral grasps as the most frequentlyused prehensile patterns for static and dynamic graspingrespectively The relaxed posture of the hand was used as theforth gesture in order to enable the recognition when thesubjects were not performing any grasp
Two of the selected gestures are evaluated in the ARATthe tripod grasp is tested in section A (grasp subtest) andthe cylindrical grasp in section B (grip subtest) of ARAT Itis expected that the recognition of the gestures is related tothe ability to move the hand measured by ARAT
24 Grasp RecognitionMethod Theproblem of hand posturedetection has been previously approached using vision-based[25 26] or glove-based [25 27ndash29] methods dependingon the constraints of the specific applications In our casegiven the bulk of the device vision-based approaches forthe recognition of the hand postures are not suitable as thehand is practically occluded and therefore the recognitionshould be based on the sensory readings for each fingerprovided by the exoskeleton sensors In a previous work wecompared several glove-based approaches to recognize grasppostures performing experiments on five healthy participantswearing the SCRIPT passive exoskeleton [7] We compared
three methods one based on the statistics of the flexion dataanother based on neural networks and finally one based onsupport vector machines (SVM) We found that with thelast method we could achieve an overall accuracy of morethan 90with small computational time (lt60ms)Thereforein this study we used the SVM approach to determine theaccuracy of recognition for healthy participants and strokepatients wearing the SCRIPT device
SVM is a popular machine learning technique for classi-fication A support vector machine is a supervised learningclassifier that constructs a set of hyperplanes in a high-dimensional space (support vectors) that are used to classifythe data A good separation is achieved by the hyperplanethat has the largest distance to the nearest training data pointof any class The hyperplanes are found solving the followingoptimization problem [30]
min120596119887120576
1
2
120596
119879
120596 + 119862
119897
sum
119894=1
120576
119894
subject to 119910
119894(120596
119879
120593 (119909
119894) + 119887 ge 1 minus 120576
119894) 120576
119894ge 0
(1)
where (119909
119894 119910
119894) | 119909
119894isin 119877
119901
119910
119894isin minus1 1 are the training
set of 119897 instance-label pairs 119909119894is 119901-dimensional real vector
120596 the normal vector of the hyperplane and 119862 gt 0 thepenalty parameter of the error term The training vectors 119909
119894
are mapped into a 119901-dimensional spaces by the function 120593
and in order to create nonlinear classifiersa kernel functionis used In our work we used a radial basis function (RBF) asthe kernel function given that it can handle the case when therelation between class labels and attributes is nonlinear [31]It is defined as
119870(119909
119894 119909
119895) = exp (minus120574100381710038171003817
1003817
1003817
119909
119894minus119909
119895
1003817
1003817
1003817
1003817
1003817
2
) 120574 gt 0 (2)
where 120574 is the kernel parameter Therefore two parametersare needed 119862and 120574 In order to find the best values forthese parameters we used a V-fold cross-validation techniquedividing the training set for one subject into V subsets of equalsize and testing the classifier on the remaining V minus 1 subsetsIn this work Vwas taken as 5 the value of the cost parameter119862 was varied as 119862 = 2
119909 119909 = minus5 5 and the value of thekernel parameter 120574 was varied as 120574 = 2
119910 119910 = minus4 0Thevalues that gave the highest validation accuracy were 119862 = 4
and 120574 = 1The method was implemented in Python using the LIB-
SVM package (httpwwwcsientuedutwsimcjlinlibsvm)The flexion angles were normalized in the range from 0 to 1(corresponding to 0 to 90 degrees) and the selected error tostop the training phase was set to 0001
25 Grasp Evaluation Method In the field of robotics manygrasp quality measures have been developed that allow thecomparison of different aspects of the robotic grasp [32] In[33] the most common robot grasp quality measures havebeen adapted to the evaluation of the grasp of the humanhand From this set of quality measures the one proposed by[34] whichmeasures how close a given grasp is to a referenceposture is the only one that can be used with the sensor
4 BioMed Research International
Table2Detailsof
thes
troke
patie
nts
IdCou
ntry
Age
(yrs)
Times
ince
acutee
vent
Gender
Affected
Dom
inant
Stroke
type
ARA
T(m
ax57
points)
FM(m
ax66
points)
AB
CD
Total
Proxim
alDistal
Coo
rdination
Total
1Holland
6913
mon
ths
Male
Left
Left
Ischem
ia18
1017
853
2916
449
2Ita
ly73
7mon
ths
Female
Left
Right
Ischem
ia12
812
941
2812
343
3Ita
ly88
6mon
ths
Male
Left
Right
Ischem
ia12
86
632
2712
443
4Ita
ly83
7mon
ths
Male
Left
Right
Ischem
ia6
40
313
2811
342
5Ita
ly85
6mon
ths
Male
Right
Right
Ischem
ia12
812
941
2812
343
6Ita
ly72
6mon
ths
Male
Left
Right
Ischem
ia18
1212
951
2823
354
7Ita
ly81
7mon
ths
Female
Left
Right
Ischem
ia18
1212
951
3524
362
8Ita
ly69
7mon
ths
Male
Left
Right
Ischem
ia12
86
632
2612
341
ARA
T=arm
scores
ofthea
ctionresearch
arm
test
FM=arm
scores
oftheF
ugl-M
eyer
motor
assessment
BioMed Research International 5
(a) Tripod grasp (b) Lateral grasp
(c) Cylindrical grasp (d) Relaxed position
Figure 2 Selected gestures to be recognized performed wearing the SCRIPT device
information provided by the SCRIPT orthosisThis index hasbeen adapted for this study as follows
119866119876 = 1 minus
119899
sum
119894=1
120596
119894(
119910
119894minus 119886
119894
119877
119894
)
2
(3)
where 119899 is the number of hand joints 120596119894is a weight factor
119910
119894is the current finger flexion angle and 119877
119894is the joint angle
range used to normalize the index calculated as themaximumbetween the reference posture 119886
119894and either the upper or lower
angle limit The index has to be maximized so that the graspis optimal when all joints are at the reference posture havinga quality measure of one and it goes to zero when all its jointsare at their maximum angle limits
The reference posture 119886
119894is taken in this case as the
one performed by the healthy subjects This measure thenenables the evaluation of how far is a poststroke patient fromperforming a grasp in away similar to a healthy subject How-ever the postures obtained from healthy subjects performingeach gesture are likely to have variations between subjectsespecially when some of the fingers were not playing an activerole in the grasp (eg the ring and little fingers in the tripodgrasp) In order to consider this variance we have includeda weight factor 120596
119894that will give high scores to fingers whose
postures have small standard deviation and vice versa The
corresponding weights for each finger and a given gesture canbe calculated as
119908
119894119892=
120572
119892
120590
119894119892
(4)
where 119894 and 119892 are the given finger and gesture 120590119894119892
is thestandard deviation of the finger flexion over all subjects forthe given finger and grasp and 120572
119892is calculated as
120572
119892= 5
5
prod
119894=1
120590
119894119892
times ((prod120590
119894119892)
119894=1234
+ (prod120590
119894119892)
119894=1345
+ (prod120590
119894119892)
119894=1245
+ (prod120590
119894119892)
119894=1235
+(prod120590
119894119892)
119894=2345
)
minus1
(5)
26 Experimental Protocol The participants took part inone session lasting half an hour conducted by a researcheror a therapist They were asked to wear a SCRIPT passiveorthosis on the impaired hand or one of their hands (in thecase of healthy participants) while sitting in front of a PCSubsequently they were instructed to mimic the picture of
6 BioMed Research International
a gesture shown on the screen (Figure 2) The participantthen confirmed that heshe achieved the desired gesture andthen pressing a button on the screen the flexion angles ofthe gesturewere saved After confirmation theywere asked torelax the hand and press a button At that moment the anglesof the relaxed posture were also saved
Each subject performed 6 repetitions of each gesture ina pseudorandom sequence resulting in the capture of 24gestures Data were then postprocessed by Python ad hocapplications The results of the classification system werecalculated using the following values for each gesture 119894
(i) True positives (TP) gestures correctly classified asgesture 119894
(ii) False negatives (FN) gestures 119894 incorrectly classifiedas other gestures
(iii) False positives (FP) other gestures incorrectly classi-fied as gesture 119894
(iv) True negatives (TN) other gestures correctly classi-fied as nongesture 119894
These are commonly used to evaluate the sensitivity andspecificity of a clinical test in its ability to confirmor refute thepresence of a diseaseThe sensitivity (true positive rate) refersto the ability of the test to correctly identify those patientswith the disease and the specificity (true negative rate) refersto the ability to correctly identify those patients without thedisease [35] In this study the outcome of the test is not binarytherefore we calculated the performance measures focusingon each gesture We used the accuracy (ability to correctlyidentify positive and negatives) true positive rate (ability tocorrectly identify the positives) and false positive rate (lackof ability to correctly classify the negatives) The last oneis measuring the opposite of the specificity as it enables usto focus on evaluating the recognition performance for agiven gesture instead of looking at the classification of othergestures They are calculated as follows
Accuracy = TP + TNTotal testing data
True positive rate (TPR) = TPTP + FN
False positive rate (FPR) = FPTN + FP
(6)
27 Data Analysis The data acquired for each subject wasdivided into two sets for training and testing purposes Asthe patients will need to perform the training procedure eachtime before starting the games (or the games will not beable to reliably recognize their gestures) the less numberof samples required to train the model the better In [7]it has been shown that a high accuracy can be achievedwith 4 training samples thus allowing very short calibrationtime and making this approach suitable for home-basedrehabilitationThen a training set using 4 samples per gesturewas used to train the model Results were considered takinginto account all the possible permutations of 4 trainingsamples
The overall results of the recognition performance aresummarized as median and interquartile range (using boxplots) differentiated between the participantrsquos type (healthyor stroke patient) The information provided by the differentperformance measures is compared using a Pearson corre-lation in order to determine if we can rely on one measure-ment for the recognition assessment We also compared thevariability of the flexion angles over all subjects performingthe different grasp gestures and the results of the accuracy ofgesture recognition per participant
Additionally the results of the proposed grasp qualitymeasure are also presented summarized using box plotsshowing the variation between the participantrsquos type thedifferent gestures and the intervariability between the par-ticipants In order to assess how these three parametersinfluenced the results of the grasp quality measure a multipleregression analysis was performed This type of analysiscan be used to consider multiple independent variablescalculating the least square estimates for a data set [36] Usingthis analysis we can devise our model using the followingequation
119866119876
119894= 119887
0+ 119887
1Participant type
119894
+ 119887
2Grasp type
119894
+ 119887
3Subject
119894
+ 120576
119894
(7)
where 119887119900represents the constant and 120576 is the modelling error
In this case predictors are classification variables with morethan two categories therefore dummy coding is required toinclude them in the regression equation The technique todo this coding is to create an independent variable for eachindependent category except one as a dichotomyThe omittedvariable provides a baseline for comparison while avoidingmulticollinearity [36] In this analysis we used ldquohealthyparticipantsrdquo ldquorelaxed posturerdquo and ldquosubject 1 (healthy)rdquo asour base line for each one of our predictors The ldquoEnterrdquomethod was used in order to force the model to consider allvariables as significant variables in the model
It is intuitive to assume that the level of impairment of thepatients should be correlated with their ability to consistentlyperform the gestures in a similar way (recognition accuracy)and similar to the ones performed by healthy subjects (GQmeasure) As the ARAT and FM tests are common waysto evaluate the level of capabilities of the upper limb wecorrelated the accuracies of recognition and the grasp qualitywith the results of these test using the Pearson coefficientTheIBM SPSS statistical package for Windows version 210 wasused to perform the analysis of the data
3 Results
This section presents the results of the experiments in twoparts the results of the recognition of the different gesturesperformed by healthy subjects and stroke patients and theresults of the evaluation of those grasps
31 Grasp Recognition Results The results of the true positiverate of the training and testing phases of the experimentsare presented in Table 3 The number of gestures correctly
BioMed Research International 7
ParticipantsStroke patientsHealthy subjects
Accu
racy
()
100
80
60
40
20
0
(a)
ParticipantsStroke patientsHealthy subjects
TPR
()
100
80
60
40
20
0
(b)
FPR
()
100
80
60
40
20
0
ParticipantsStroke patientsHealthy subjects
(c)
Figure 3 Performance measures evaluating the recognition of grasp postures (a) accuracy (b) true positive rate and (c) false positive rateResults are presented discriminated between healthy and stroke participants
Table 3 Training time and true positive rate () between thetraining and testing phases
Meantraining time
(ms)
Meantraining TPR
()
Mean testingTPR ()
Healthysubjects
1 1800 plusmn 06 58 plusmn 14 87 plusmn 22 1733 plusmn 06 100 plusmn 0 100 plusmn 03 2200 plusmn 07 84 plusmn 14 91 plusmn 04 1933 plusmn 05 79 plusmn 25 98 plusmn 35 1733 plusmn 06 84 plusmn 14 91 plusmn 06 1867 plusmn 05 58 plusmn 10 83 plusmn 47 2200 plusmn 14 63 plusmn 9 85 plusmn 48 2267 plusmn 16 98 plusmn 8 97 plusmn 49 2200 plusmn 06 31 plusmn 9 91 plusmn 110 1867 plusmn 05 84 plusmn 16 90 plusmn 2
1980 plusmn 05 74 plusmn 24 91 plusmn 6
Strokepatients
1 2600 plusmn 07 100 plusmn 0 99 plusmn 22 2200 plusmn 04 61 plusmn 15 91 plusmn 43 2200 plusmn 04 38 plusmn 6 59 plusmn 44 1933 plusmn 07 38 plusmn 11 56 plusmn 45 2467 plusmn 06 38 plusmn 6 58 plusmn 46 2067 plusmn 06 33 plusmn 6 68 plusmn 57 2067 plusmn 03 47 plusmn 19 93 plusmn 28 2133 plusmn 04 43 plusmn 26 78 plusmn 3
2208 plusmn 06 50 plusmn 25 75 plusmn 17
recognized greatly increased from the training to the testingphase It can clearly be seen that the testing overall recogni-tion performance of gestures performed by the stroke patientsis lower than the ones performed by the healthy subjects asit was expected but still they produced a high percentageof true positive recognized gestures (TPR mean of 75)The computational time taken for training the model was onaverage 198ms for healthy subjects and 221ms for strokepatients
In order to evaluate inmore detail the recognition resultsthe selected performance measures are shown in Figure 3Ideally the accuracy and true positive rate should be closeto 100 and the false positive rate close to zero In this casethe median accuracy of recognition of gestures performed byhealthy subjects was over 95 and 87 for stroke patientsThe true positive rate showed lower values with respect to theaccuracy median values for healthy subjects of over 91 andfor stroke patients of over 75 The healthy subjects showedmedian values of false positive rates of 2 and 8 for strokepatients
The different performance measures can also provideinformation about the specific recognition of each gestureSpecifically the true positive rate refers to the ability ofthe method to correctly identify a specific gesture whilstthe false positive rate refers to the percentage of gestureswrongly classified Figure 4 shows the accuracy true positiverate and false positive rate for each gesture The difficultyof recognition of the different grasp gestures is different forthe healthy subjects or stroke patients The relax postureis a clearly distinctive gesture therefore showing the bestaccuracies best TPR and FPR values close to zero Forhealthy subjects the tripod and cylindrical gestures were themost difficult to be recognized and the most misrecognizedFor stroke patients the three gestures presented similardifficulties to be recognized and the tripod and lateral graspwere the most misrecognized
Table 4 presents the Pearson correlation values of thedifferent performancemeasures for the given conditionsTheaccuracy is inversely correlated with the false negative rateover all conditions (correlation coefficient C lt minus097 withstatistical significance) and also highly correlated with theTPR (C gt 099) The different measures of performance pro-vide specific information on the performance of recognitionbut the accuracy could be selected as the overall measureof recognition performance Therefore for the followinganalysis the accuracy will be used
It is expected that the grasping capabilities of the par-ticipants affect the recognition of hand postures especiallyin the case of stroke patients Therefore we also studied
8 BioMed Research InternationalAc
cura
cy (
)
100
80
60
40
20
0
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
Accu
racy
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(a)
TPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
TPR
()
100
80
60
40
20
0
Part
icip
ants
Part
icip
ants
Hea
lthy
subj
ects
Stro
ke p
atie
nts
(b)
FPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
FPR
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(c)
Figure 4 Performance measures for each gesture evaluating the recognition of hand postures (a) accuracy (b) true positive rate and (c)false positive rate
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
4 BioMed Research International
Table2Detailsof
thes
troke
patie
nts
IdCou
ntry
Age
(yrs)
Times
ince
acutee
vent
Gender
Affected
Dom
inant
Stroke
type
ARA
T(m
ax57
points)
FM(m
ax66
points)
AB
CD
Total
Proxim
alDistal
Coo
rdination
Total
1Holland
6913
mon
ths
Male
Left
Left
Ischem
ia18
1017
853
2916
449
2Ita
ly73
7mon
ths
Female
Left
Right
Ischem
ia12
812
941
2812
343
3Ita
ly88
6mon
ths
Male
Left
Right
Ischem
ia12
86
632
2712
443
4Ita
ly83
7mon
ths
Male
Left
Right
Ischem
ia6
40
313
2811
342
5Ita
ly85
6mon
ths
Male
Right
Right
Ischem
ia12
812
941
2812
343
6Ita
ly72
6mon
ths
Male
Left
Right
Ischem
ia18
1212
951
2823
354
7Ita
ly81
7mon
ths
Female
Left
Right
Ischem
ia18
1212
951
3524
362
8Ita
ly69
7mon
ths
Male
Left
Right
Ischem
ia12
86
632
2612
341
ARA
T=arm
scores
ofthea
ctionresearch
arm
test
FM=arm
scores
oftheF
ugl-M
eyer
motor
assessment
BioMed Research International 5
(a) Tripod grasp (b) Lateral grasp
(c) Cylindrical grasp (d) Relaxed position
Figure 2 Selected gestures to be recognized performed wearing the SCRIPT device
information provided by the SCRIPT orthosisThis index hasbeen adapted for this study as follows
119866119876 = 1 minus
119899
sum
119894=1
120596
119894(
119910
119894minus 119886
119894
119877
119894
)
2
(3)
where 119899 is the number of hand joints 120596119894is a weight factor
119910
119894is the current finger flexion angle and 119877
119894is the joint angle
range used to normalize the index calculated as themaximumbetween the reference posture 119886
119894and either the upper or lower
angle limit The index has to be maximized so that the graspis optimal when all joints are at the reference posture havinga quality measure of one and it goes to zero when all its jointsare at their maximum angle limits
The reference posture 119886
119894is taken in this case as the
one performed by the healthy subjects This measure thenenables the evaluation of how far is a poststroke patient fromperforming a grasp in away similar to a healthy subject How-ever the postures obtained from healthy subjects performingeach gesture are likely to have variations between subjectsespecially when some of the fingers were not playing an activerole in the grasp (eg the ring and little fingers in the tripodgrasp) In order to consider this variance we have includeda weight factor 120596
119894that will give high scores to fingers whose
postures have small standard deviation and vice versa The
corresponding weights for each finger and a given gesture canbe calculated as
119908
119894119892=
120572
119892
120590
119894119892
(4)
where 119894 and 119892 are the given finger and gesture 120590119894119892
is thestandard deviation of the finger flexion over all subjects forthe given finger and grasp and 120572
119892is calculated as
120572
119892= 5
5
prod
119894=1
120590
119894119892
times ((prod120590
119894119892)
119894=1234
+ (prod120590
119894119892)
119894=1345
+ (prod120590
119894119892)
119894=1245
+ (prod120590
119894119892)
119894=1235
+(prod120590
119894119892)
119894=2345
)
minus1
(5)
26 Experimental Protocol The participants took part inone session lasting half an hour conducted by a researcheror a therapist They were asked to wear a SCRIPT passiveorthosis on the impaired hand or one of their hands (in thecase of healthy participants) while sitting in front of a PCSubsequently they were instructed to mimic the picture of
6 BioMed Research International
a gesture shown on the screen (Figure 2) The participantthen confirmed that heshe achieved the desired gesture andthen pressing a button on the screen the flexion angles ofthe gesturewere saved After confirmation theywere asked torelax the hand and press a button At that moment the anglesof the relaxed posture were also saved
Each subject performed 6 repetitions of each gesture ina pseudorandom sequence resulting in the capture of 24gestures Data were then postprocessed by Python ad hocapplications The results of the classification system werecalculated using the following values for each gesture 119894
(i) True positives (TP) gestures correctly classified asgesture 119894
(ii) False negatives (FN) gestures 119894 incorrectly classifiedas other gestures
(iii) False positives (FP) other gestures incorrectly classi-fied as gesture 119894
(iv) True negatives (TN) other gestures correctly classi-fied as nongesture 119894
These are commonly used to evaluate the sensitivity andspecificity of a clinical test in its ability to confirmor refute thepresence of a diseaseThe sensitivity (true positive rate) refersto the ability of the test to correctly identify those patientswith the disease and the specificity (true negative rate) refersto the ability to correctly identify those patients without thedisease [35] In this study the outcome of the test is not binarytherefore we calculated the performance measures focusingon each gesture We used the accuracy (ability to correctlyidentify positive and negatives) true positive rate (ability tocorrectly identify the positives) and false positive rate (lackof ability to correctly classify the negatives) The last oneis measuring the opposite of the specificity as it enables usto focus on evaluating the recognition performance for agiven gesture instead of looking at the classification of othergestures They are calculated as follows
Accuracy = TP + TNTotal testing data
True positive rate (TPR) = TPTP + FN
False positive rate (FPR) = FPTN + FP
(6)
27 Data Analysis The data acquired for each subject wasdivided into two sets for training and testing purposes Asthe patients will need to perform the training procedure eachtime before starting the games (or the games will not beable to reliably recognize their gestures) the less numberof samples required to train the model the better In [7]it has been shown that a high accuracy can be achievedwith 4 training samples thus allowing very short calibrationtime and making this approach suitable for home-basedrehabilitationThen a training set using 4 samples per gesturewas used to train the model Results were considered takinginto account all the possible permutations of 4 trainingsamples
The overall results of the recognition performance aresummarized as median and interquartile range (using boxplots) differentiated between the participantrsquos type (healthyor stroke patient) The information provided by the differentperformance measures is compared using a Pearson corre-lation in order to determine if we can rely on one measure-ment for the recognition assessment We also compared thevariability of the flexion angles over all subjects performingthe different grasp gestures and the results of the accuracy ofgesture recognition per participant
Additionally the results of the proposed grasp qualitymeasure are also presented summarized using box plotsshowing the variation between the participantrsquos type thedifferent gestures and the intervariability between the par-ticipants In order to assess how these three parametersinfluenced the results of the grasp quality measure a multipleregression analysis was performed This type of analysiscan be used to consider multiple independent variablescalculating the least square estimates for a data set [36] Usingthis analysis we can devise our model using the followingequation
119866119876
119894= 119887
0+ 119887
1Participant type
119894
+ 119887
2Grasp type
119894
+ 119887
3Subject
119894
+ 120576
119894
(7)
where 119887119900represents the constant and 120576 is the modelling error
In this case predictors are classification variables with morethan two categories therefore dummy coding is required toinclude them in the regression equation The technique todo this coding is to create an independent variable for eachindependent category except one as a dichotomyThe omittedvariable provides a baseline for comparison while avoidingmulticollinearity [36] In this analysis we used ldquohealthyparticipantsrdquo ldquorelaxed posturerdquo and ldquosubject 1 (healthy)rdquo asour base line for each one of our predictors The ldquoEnterrdquomethod was used in order to force the model to consider allvariables as significant variables in the model
It is intuitive to assume that the level of impairment of thepatients should be correlated with their ability to consistentlyperform the gestures in a similar way (recognition accuracy)and similar to the ones performed by healthy subjects (GQmeasure) As the ARAT and FM tests are common waysto evaluate the level of capabilities of the upper limb wecorrelated the accuracies of recognition and the grasp qualitywith the results of these test using the Pearson coefficientTheIBM SPSS statistical package for Windows version 210 wasused to perform the analysis of the data
3 Results
This section presents the results of the experiments in twoparts the results of the recognition of the different gesturesperformed by healthy subjects and stroke patients and theresults of the evaluation of those grasps
31 Grasp Recognition Results The results of the true positiverate of the training and testing phases of the experimentsare presented in Table 3 The number of gestures correctly
BioMed Research International 7
ParticipantsStroke patientsHealthy subjects
Accu
racy
()
100
80
60
40
20
0
(a)
ParticipantsStroke patientsHealthy subjects
TPR
()
100
80
60
40
20
0
(b)
FPR
()
100
80
60
40
20
0
ParticipantsStroke patientsHealthy subjects
(c)
Figure 3 Performance measures evaluating the recognition of grasp postures (a) accuracy (b) true positive rate and (c) false positive rateResults are presented discriminated between healthy and stroke participants
Table 3 Training time and true positive rate () between thetraining and testing phases
Meantraining time
(ms)
Meantraining TPR
()
Mean testingTPR ()
Healthysubjects
1 1800 plusmn 06 58 plusmn 14 87 plusmn 22 1733 plusmn 06 100 plusmn 0 100 plusmn 03 2200 plusmn 07 84 plusmn 14 91 plusmn 04 1933 plusmn 05 79 plusmn 25 98 plusmn 35 1733 plusmn 06 84 plusmn 14 91 plusmn 06 1867 plusmn 05 58 plusmn 10 83 plusmn 47 2200 plusmn 14 63 plusmn 9 85 plusmn 48 2267 plusmn 16 98 plusmn 8 97 plusmn 49 2200 plusmn 06 31 plusmn 9 91 plusmn 110 1867 plusmn 05 84 plusmn 16 90 plusmn 2
1980 plusmn 05 74 plusmn 24 91 plusmn 6
Strokepatients
1 2600 plusmn 07 100 plusmn 0 99 plusmn 22 2200 plusmn 04 61 plusmn 15 91 plusmn 43 2200 plusmn 04 38 plusmn 6 59 plusmn 44 1933 plusmn 07 38 plusmn 11 56 plusmn 45 2467 plusmn 06 38 plusmn 6 58 plusmn 46 2067 plusmn 06 33 plusmn 6 68 plusmn 57 2067 plusmn 03 47 plusmn 19 93 plusmn 28 2133 plusmn 04 43 plusmn 26 78 plusmn 3
2208 plusmn 06 50 plusmn 25 75 plusmn 17
recognized greatly increased from the training to the testingphase It can clearly be seen that the testing overall recogni-tion performance of gestures performed by the stroke patientsis lower than the ones performed by the healthy subjects asit was expected but still they produced a high percentageof true positive recognized gestures (TPR mean of 75)The computational time taken for training the model was onaverage 198ms for healthy subjects and 221ms for strokepatients
In order to evaluate inmore detail the recognition resultsthe selected performance measures are shown in Figure 3Ideally the accuracy and true positive rate should be closeto 100 and the false positive rate close to zero In this casethe median accuracy of recognition of gestures performed byhealthy subjects was over 95 and 87 for stroke patientsThe true positive rate showed lower values with respect to theaccuracy median values for healthy subjects of over 91 andfor stroke patients of over 75 The healthy subjects showedmedian values of false positive rates of 2 and 8 for strokepatients
The different performance measures can also provideinformation about the specific recognition of each gestureSpecifically the true positive rate refers to the ability ofthe method to correctly identify a specific gesture whilstthe false positive rate refers to the percentage of gestureswrongly classified Figure 4 shows the accuracy true positiverate and false positive rate for each gesture The difficultyof recognition of the different grasp gestures is different forthe healthy subjects or stroke patients The relax postureis a clearly distinctive gesture therefore showing the bestaccuracies best TPR and FPR values close to zero Forhealthy subjects the tripod and cylindrical gestures were themost difficult to be recognized and the most misrecognizedFor stroke patients the three gestures presented similardifficulties to be recognized and the tripod and lateral graspwere the most misrecognized
Table 4 presents the Pearson correlation values of thedifferent performancemeasures for the given conditionsTheaccuracy is inversely correlated with the false negative rateover all conditions (correlation coefficient C lt minus097 withstatistical significance) and also highly correlated with theTPR (C gt 099) The different measures of performance pro-vide specific information on the performance of recognitionbut the accuracy could be selected as the overall measureof recognition performance Therefore for the followinganalysis the accuracy will be used
It is expected that the grasping capabilities of the par-ticipants affect the recognition of hand postures especiallyin the case of stroke patients Therefore we also studied
8 BioMed Research InternationalAc
cura
cy (
)
100
80
60
40
20
0
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
Accu
racy
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(a)
TPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
TPR
()
100
80
60
40
20
0
Part
icip
ants
Part
icip
ants
Hea
lthy
subj
ects
Stro
ke p
atie
nts
(b)
FPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
FPR
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(c)
Figure 4 Performance measures for each gesture evaluating the recognition of hand postures (a) accuracy (b) true positive rate and (c)false positive rate
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
BioMed Research International 5
(a) Tripod grasp (b) Lateral grasp
(c) Cylindrical grasp (d) Relaxed position
Figure 2 Selected gestures to be recognized performed wearing the SCRIPT device
information provided by the SCRIPT orthosisThis index hasbeen adapted for this study as follows
119866119876 = 1 minus
119899
sum
119894=1
120596
119894(
119910
119894minus 119886
119894
119877
119894
)
2
(3)
where 119899 is the number of hand joints 120596119894is a weight factor
119910
119894is the current finger flexion angle and 119877
119894is the joint angle
range used to normalize the index calculated as themaximumbetween the reference posture 119886
119894and either the upper or lower
angle limit The index has to be maximized so that the graspis optimal when all joints are at the reference posture havinga quality measure of one and it goes to zero when all its jointsare at their maximum angle limits
The reference posture 119886
119894is taken in this case as the
one performed by the healthy subjects This measure thenenables the evaluation of how far is a poststroke patient fromperforming a grasp in away similar to a healthy subject How-ever the postures obtained from healthy subjects performingeach gesture are likely to have variations between subjectsespecially when some of the fingers were not playing an activerole in the grasp (eg the ring and little fingers in the tripodgrasp) In order to consider this variance we have includeda weight factor 120596
119894that will give high scores to fingers whose
postures have small standard deviation and vice versa The
corresponding weights for each finger and a given gesture canbe calculated as
119908
119894119892=
120572
119892
120590
119894119892
(4)
where 119894 and 119892 are the given finger and gesture 120590119894119892
is thestandard deviation of the finger flexion over all subjects forthe given finger and grasp and 120572
119892is calculated as
120572
119892= 5
5
prod
119894=1
120590
119894119892
times ((prod120590
119894119892)
119894=1234
+ (prod120590
119894119892)
119894=1345
+ (prod120590
119894119892)
119894=1245
+ (prod120590
119894119892)
119894=1235
+(prod120590
119894119892)
119894=2345
)
minus1
(5)
26 Experimental Protocol The participants took part inone session lasting half an hour conducted by a researcheror a therapist They were asked to wear a SCRIPT passiveorthosis on the impaired hand or one of their hands (in thecase of healthy participants) while sitting in front of a PCSubsequently they were instructed to mimic the picture of
6 BioMed Research International
a gesture shown on the screen (Figure 2) The participantthen confirmed that heshe achieved the desired gesture andthen pressing a button on the screen the flexion angles ofthe gesturewere saved After confirmation theywere asked torelax the hand and press a button At that moment the anglesof the relaxed posture were also saved
Each subject performed 6 repetitions of each gesture ina pseudorandom sequence resulting in the capture of 24gestures Data were then postprocessed by Python ad hocapplications The results of the classification system werecalculated using the following values for each gesture 119894
(i) True positives (TP) gestures correctly classified asgesture 119894
(ii) False negatives (FN) gestures 119894 incorrectly classifiedas other gestures
(iii) False positives (FP) other gestures incorrectly classi-fied as gesture 119894
(iv) True negatives (TN) other gestures correctly classi-fied as nongesture 119894
These are commonly used to evaluate the sensitivity andspecificity of a clinical test in its ability to confirmor refute thepresence of a diseaseThe sensitivity (true positive rate) refersto the ability of the test to correctly identify those patientswith the disease and the specificity (true negative rate) refersto the ability to correctly identify those patients without thedisease [35] In this study the outcome of the test is not binarytherefore we calculated the performance measures focusingon each gesture We used the accuracy (ability to correctlyidentify positive and negatives) true positive rate (ability tocorrectly identify the positives) and false positive rate (lackof ability to correctly classify the negatives) The last oneis measuring the opposite of the specificity as it enables usto focus on evaluating the recognition performance for agiven gesture instead of looking at the classification of othergestures They are calculated as follows
Accuracy = TP + TNTotal testing data
True positive rate (TPR) = TPTP + FN
False positive rate (FPR) = FPTN + FP
(6)
27 Data Analysis The data acquired for each subject wasdivided into two sets for training and testing purposes Asthe patients will need to perform the training procedure eachtime before starting the games (or the games will not beable to reliably recognize their gestures) the less numberof samples required to train the model the better In [7]it has been shown that a high accuracy can be achievedwith 4 training samples thus allowing very short calibrationtime and making this approach suitable for home-basedrehabilitationThen a training set using 4 samples per gesturewas used to train the model Results were considered takinginto account all the possible permutations of 4 trainingsamples
The overall results of the recognition performance aresummarized as median and interquartile range (using boxplots) differentiated between the participantrsquos type (healthyor stroke patient) The information provided by the differentperformance measures is compared using a Pearson corre-lation in order to determine if we can rely on one measure-ment for the recognition assessment We also compared thevariability of the flexion angles over all subjects performingthe different grasp gestures and the results of the accuracy ofgesture recognition per participant
Additionally the results of the proposed grasp qualitymeasure are also presented summarized using box plotsshowing the variation between the participantrsquos type thedifferent gestures and the intervariability between the par-ticipants In order to assess how these three parametersinfluenced the results of the grasp quality measure a multipleregression analysis was performed This type of analysiscan be used to consider multiple independent variablescalculating the least square estimates for a data set [36] Usingthis analysis we can devise our model using the followingequation
119866119876
119894= 119887
0+ 119887
1Participant type
119894
+ 119887
2Grasp type
119894
+ 119887
3Subject
119894
+ 120576
119894
(7)
where 119887119900represents the constant and 120576 is the modelling error
In this case predictors are classification variables with morethan two categories therefore dummy coding is required toinclude them in the regression equation The technique todo this coding is to create an independent variable for eachindependent category except one as a dichotomyThe omittedvariable provides a baseline for comparison while avoidingmulticollinearity [36] In this analysis we used ldquohealthyparticipantsrdquo ldquorelaxed posturerdquo and ldquosubject 1 (healthy)rdquo asour base line for each one of our predictors The ldquoEnterrdquomethod was used in order to force the model to consider allvariables as significant variables in the model
It is intuitive to assume that the level of impairment of thepatients should be correlated with their ability to consistentlyperform the gestures in a similar way (recognition accuracy)and similar to the ones performed by healthy subjects (GQmeasure) As the ARAT and FM tests are common waysto evaluate the level of capabilities of the upper limb wecorrelated the accuracies of recognition and the grasp qualitywith the results of these test using the Pearson coefficientTheIBM SPSS statistical package for Windows version 210 wasused to perform the analysis of the data
3 Results
This section presents the results of the experiments in twoparts the results of the recognition of the different gesturesperformed by healthy subjects and stroke patients and theresults of the evaluation of those grasps
31 Grasp Recognition Results The results of the true positiverate of the training and testing phases of the experimentsare presented in Table 3 The number of gestures correctly
BioMed Research International 7
ParticipantsStroke patientsHealthy subjects
Accu
racy
()
100
80
60
40
20
0
(a)
ParticipantsStroke patientsHealthy subjects
TPR
()
100
80
60
40
20
0
(b)
FPR
()
100
80
60
40
20
0
ParticipantsStroke patientsHealthy subjects
(c)
Figure 3 Performance measures evaluating the recognition of grasp postures (a) accuracy (b) true positive rate and (c) false positive rateResults are presented discriminated between healthy and stroke participants
Table 3 Training time and true positive rate () between thetraining and testing phases
Meantraining time
(ms)
Meantraining TPR
()
Mean testingTPR ()
Healthysubjects
1 1800 plusmn 06 58 plusmn 14 87 plusmn 22 1733 plusmn 06 100 plusmn 0 100 plusmn 03 2200 plusmn 07 84 plusmn 14 91 plusmn 04 1933 plusmn 05 79 plusmn 25 98 plusmn 35 1733 plusmn 06 84 plusmn 14 91 plusmn 06 1867 plusmn 05 58 plusmn 10 83 plusmn 47 2200 plusmn 14 63 plusmn 9 85 plusmn 48 2267 plusmn 16 98 plusmn 8 97 plusmn 49 2200 plusmn 06 31 plusmn 9 91 plusmn 110 1867 plusmn 05 84 plusmn 16 90 plusmn 2
1980 plusmn 05 74 plusmn 24 91 plusmn 6
Strokepatients
1 2600 plusmn 07 100 plusmn 0 99 plusmn 22 2200 plusmn 04 61 plusmn 15 91 plusmn 43 2200 plusmn 04 38 plusmn 6 59 plusmn 44 1933 plusmn 07 38 plusmn 11 56 plusmn 45 2467 plusmn 06 38 plusmn 6 58 plusmn 46 2067 plusmn 06 33 plusmn 6 68 plusmn 57 2067 plusmn 03 47 plusmn 19 93 plusmn 28 2133 plusmn 04 43 plusmn 26 78 plusmn 3
2208 plusmn 06 50 plusmn 25 75 plusmn 17
recognized greatly increased from the training to the testingphase It can clearly be seen that the testing overall recogni-tion performance of gestures performed by the stroke patientsis lower than the ones performed by the healthy subjects asit was expected but still they produced a high percentageof true positive recognized gestures (TPR mean of 75)The computational time taken for training the model was onaverage 198ms for healthy subjects and 221ms for strokepatients
In order to evaluate inmore detail the recognition resultsthe selected performance measures are shown in Figure 3Ideally the accuracy and true positive rate should be closeto 100 and the false positive rate close to zero In this casethe median accuracy of recognition of gestures performed byhealthy subjects was over 95 and 87 for stroke patientsThe true positive rate showed lower values with respect to theaccuracy median values for healthy subjects of over 91 andfor stroke patients of over 75 The healthy subjects showedmedian values of false positive rates of 2 and 8 for strokepatients
The different performance measures can also provideinformation about the specific recognition of each gestureSpecifically the true positive rate refers to the ability ofthe method to correctly identify a specific gesture whilstthe false positive rate refers to the percentage of gestureswrongly classified Figure 4 shows the accuracy true positiverate and false positive rate for each gesture The difficultyof recognition of the different grasp gestures is different forthe healthy subjects or stroke patients The relax postureis a clearly distinctive gesture therefore showing the bestaccuracies best TPR and FPR values close to zero Forhealthy subjects the tripod and cylindrical gestures were themost difficult to be recognized and the most misrecognizedFor stroke patients the three gestures presented similardifficulties to be recognized and the tripod and lateral graspwere the most misrecognized
Table 4 presents the Pearson correlation values of thedifferent performancemeasures for the given conditionsTheaccuracy is inversely correlated with the false negative rateover all conditions (correlation coefficient C lt minus097 withstatistical significance) and also highly correlated with theTPR (C gt 099) The different measures of performance pro-vide specific information on the performance of recognitionbut the accuracy could be selected as the overall measureof recognition performance Therefore for the followinganalysis the accuracy will be used
It is expected that the grasping capabilities of the par-ticipants affect the recognition of hand postures especiallyin the case of stroke patients Therefore we also studied
8 BioMed Research InternationalAc
cura
cy (
)
100
80
60
40
20
0
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
Accu
racy
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(a)
TPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
TPR
()
100
80
60
40
20
0
Part
icip
ants
Part
icip
ants
Hea
lthy
subj
ects
Stro
ke p
atie
nts
(b)
FPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
FPR
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(c)
Figure 4 Performance measures for each gesture evaluating the recognition of hand postures (a) accuracy (b) true positive rate and (c)false positive rate
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
6 BioMed Research International
a gesture shown on the screen (Figure 2) The participantthen confirmed that heshe achieved the desired gesture andthen pressing a button on the screen the flexion angles ofthe gesturewere saved After confirmation theywere asked torelax the hand and press a button At that moment the anglesof the relaxed posture were also saved
Each subject performed 6 repetitions of each gesture ina pseudorandom sequence resulting in the capture of 24gestures Data were then postprocessed by Python ad hocapplications The results of the classification system werecalculated using the following values for each gesture 119894
(i) True positives (TP) gestures correctly classified asgesture 119894
(ii) False negatives (FN) gestures 119894 incorrectly classifiedas other gestures
(iii) False positives (FP) other gestures incorrectly classi-fied as gesture 119894
(iv) True negatives (TN) other gestures correctly classi-fied as nongesture 119894
These are commonly used to evaluate the sensitivity andspecificity of a clinical test in its ability to confirmor refute thepresence of a diseaseThe sensitivity (true positive rate) refersto the ability of the test to correctly identify those patientswith the disease and the specificity (true negative rate) refersto the ability to correctly identify those patients without thedisease [35] In this study the outcome of the test is not binarytherefore we calculated the performance measures focusingon each gesture We used the accuracy (ability to correctlyidentify positive and negatives) true positive rate (ability tocorrectly identify the positives) and false positive rate (lackof ability to correctly classify the negatives) The last oneis measuring the opposite of the specificity as it enables usto focus on evaluating the recognition performance for agiven gesture instead of looking at the classification of othergestures They are calculated as follows
Accuracy = TP + TNTotal testing data
True positive rate (TPR) = TPTP + FN
False positive rate (FPR) = FPTN + FP
(6)
27 Data Analysis The data acquired for each subject wasdivided into two sets for training and testing purposes Asthe patients will need to perform the training procedure eachtime before starting the games (or the games will not beable to reliably recognize their gestures) the less numberof samples required to train the model the better In [7]it has been shown that a high accuracy can be achievedwith 4 training samples thus allowing very short calibrationtime and making this approach suitable for home-basedrehabilitationThen a training set using 4 samples per gesturewas used to train the model Results were considered takinginto account all the possible permutations of 4 trainingsamples
The overall results of the recognition performance aresummarized as median and interquartile range (using boxplots) differentiated between the participantrsquos type (healthyor stroke patient) The information provided by the differentperformance measures is compared using a Pearson corre-lation in order to determine if we can rely on one measure-ment for the recognition assessment We also compared thevariability of the flexion angles over all subjects performingthe different grasp gestures and the results of the accuracy ofgesture recognition per participant
Additionally the results of the proposed grasp qualitymeasure are also presented summarized using box plotsshowing the variation between the participantrsquos type thedifferent gestures and the intervariability between the par-ticipants In order to assess how these three parametersinfluenced the results of the grasp quality measure a multipleregression analysis was performed This type of analysiscan be used to consider multiple independent variablescalculating the least square estimates for a data set [36] Usingthis analysis we can devise our model using the followingequation
119866119876
119894= 119887
0+ 119887
1Participant type
119894
+ 119887
2Grasp type
119894
+ 119887
3Subject
119894
+ 120576
119894
(7)
where 119887119900represents the constant and 120576 is the modelling error
In this case predictors are classification variables with morethan two categories therefore dummy coding is required toinclude them in the regression equation The technique todo this coding is to create an independent variable for eachindependent category except one as a dichotomyThe omittedvariable provides a baseline for comparison while avoidingmulticollinearity [36] In this analysis we used ldquohealthyparticipantsrdquo ldquorelaxed posturerdquo and ldquosubject 1 (healthy)rdquo asour base line for each one of our predictors The ldquoEnterrdquomethod was used in order to force the model to consider allvariables as significant variables in the model
It is intuitive to assume that the level of impairment of thepatients should be correlated with their ability to consistentlyperform the gestures in a similar way (recognition accuracy)and similar to the ones performed by healthy subjects (GQmeasure) As the ARAT and FM tests are common waysto evaluate the level of capabilities of the upper limb wecorrelated the accuracies of recognition and the grasp qualitywith the results of these test using the Pearson coefficientTheIBM SPSS statistical package for Windows version 210 wasused to perform the analysis of the data
3 Results
This section presents the results of the experiments in twoparts the results of the recognition of the different gesturesperformed by healthy subjects and stroke patients and theresults of the evaluation of those grasps
31 Grasp Recognition Results The results of the true positiverate of the training and testing phases of the experimentsare presented in Table 3 The number of gestures correctly
BioMed Research International 7
ParticipantsStroke patientsHealthy subjects
Accu
racy
()
100
80
60
40
20
0
(a)
ParticipantsStroke patientsHealthy subjects
TPR
()
100
80
60
40
20
0
(b)
FPR
()
100
80
60
40
20
0
ParticipantsStroke patientsHealthy subjects
(c)
Figure 3 Performance measures evaluating the recognition of grasp postures (a) accuracy (b) true positive rate and (c) false positive rateResults are presented discriminated between healthy and stroke participants
Table 3 Training time and true positive rate () between thetraining and testing phases
Meantraining time
(ms)
Meantraining TPR
()
Mean testingTPR ()
Healthysubjects
1 1800 plusmn 06 58 plusmn 14 87 plusmn 22 1733 plusmn 06 100 plusmn 0 100 plusmn 03 2200 plusmn 07 84 plusmn 14 91 plusmn 04 1933 plusmn 05 79 plusmn 25 98 plusmn 35 1733 plusmn 06 84 plusmn 14 91 plusmn 06 1867 plusmn 05 58 plusmn 10 83 plusmn 47 2200 plusmn 14 63 plusmn 9 85 plusmn 48 2267 plusmn 16 98 plusmn 8 97 plusmn 49 2200 plusmn 06 31 plusmn 9 91 plusmn 110 1867 plusmn 05 84 plusmn 16 90 plusmn 2
1980 plusmn 05 74 plusmn 24 91 plusmn 6
Strokepatients
1 2600 plusmn 07 100 plusmn 0 99 plusmn 22 2200 plusmn 04 61 plusmn 15 91 plusmn 43 2200 plusmn 04 38 plusmn 6 59 plusmn 44 1933 plusmn 07 38 plusmn 11 56 plusmn 45 2467 plusmn 06 38 plusmn 6 58 plusmn 46 2067 plusmn 06 33 plusmn 6 68 plusmn 57 2067 plusmn 03 47 plusmn 19 93 plusmn 28 2133 plusmn 04 43 plusmn 26 78 plusmn 3
2208 plusmn 06 50 plusmn 25 75 plusmn 17
recognized greatly increased from the training to the testingphase It can clearly be seen that the testing overall recogni-tion performance of gestures performed by the stroke patientsis lower than the ones performed by the healthy subjects asit was expected but still they produced a high percentageof true positive recognized gestures (TPR mean of 75)The computational time taken for training the model was onaverage 198ms for healthy subjects and 221ms for strokepatients
In order to evaluate inmore detail the recognition resultsthe selected performance measures are shown in Figure 3Ideally the accuracy and true positive rate should be closeto 100 and the false positive rate close to zero In this casethe median accuracy of recognition of gestures performed byhealthy subjects was over 95 and 87 for stroke patientsThe true positive rate showed lower values with respect to theaccuracy median values for healthy subjects of over 91 andfor stroke patients of over 75 The healthy subjects showedmedian values of false positive rates of 2 and 8 for strokepatients
The different performance measures can also provideinformation about the specific recognition of each gestureSpecifically the true positive rate refers to the ability ofthe method to correctly identify a specific gesture whilstthe false positive rate refers to the percentage of gestureswrongly classified Figure 4 shows the accuracy true positiverate and false positive rate for each gesture The difficultyof recognition of the different grasp gestures is different forthe healthy subjects or stroke patients The relax postureis a clearly distinctive gesture therefore showing the bestaccuracies best TPR and FPR values close to zero Forhealthy subjects the tripod and cylindrical gestures were themost difficult to be recognized and the most misrecognizedFor stroke patients the three gestures presented similardifficulties to be recognized and the tripod and lateral graspwere the most misrecognized
Table 4 presents the Pearson correlation values of thedifferent performancemeasures for the given conditionsTheaccuracy is inversely correlated with the false negative rateover all conditions (correlation coefficient C lt minus097 withstatistical significance) and also highly correlated with theTPR (C gt 099) The different measures of performance pro-vide specific information on the performance of recognitionbut the accuracy could be selected as the overall measureof recognition performance Therefore for the followinganalysis the accuracy will be used
It is expected that the grasping capabilities of the par-ticipants affect the recognition of hand postures especiallyin the case of stroke patients Therefore we also studied
8 BioMed Research InternationalAc
cura
cy (
)
100
80
60
40
20
0
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
Accu
racy
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(a)
TPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
TPR
()
100
80
60
40
20
0
Part
icip
ants
Part
icip
ants
Hea
lthy
subj
ects
Stro
ke p
atie
nts
(b)
FPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
FPR
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(c)
Figure 4 Performance measures for each gesture evaluating the recognition of hand postures (a) accuracy (b) true positive rate and (c)false positive rate
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
BioMed Research International 7
ParticipantsStroke patientsHealthy subjects
Accu
racy
()
100
80
60
40
20
0
(a)
ParticipantsStroke patientsHealthy subjects
TPR
()
100
80
60
40
20
0
(b)
FPR
()
100
80
60
40
20
0
ParticipantsStroke patientsHealthy subjects
(c)
Figure 3 Performance measures evaluating the recognition of grasp postures (a) accuracy (b) true positive rate and (c) false positive rateResults are presented discriminated between healthy and stroke participants
Table 3 Training time and true positive rate () between thetraining and testing phases
Meantraining time
(ms)
Meantraining TPR
()
Mean testingTPR ()
Healthysubjects
1 1800 plusmn 06 58 plusmn 14 87 plusmn 22 1733 plusmn 06 100 plusmn 0 100 plusmn 03 2200 plusmn 07 84 plusmn 14 91 plusmn 04 1933 plusmn 05 79 plusmn 25 98 plusmn 35 1733 plusmn 06 84 plusmn 14 91 plusmn 06 1867 plusmn 05 58 plusmn 10 83 plusmn 47 2200 plusmn 14 63 plusmn 9 85 plusmn 48 2267 plusmn 16 98 plusmn 8 97 plusmn 49 2200 plusmn 06 31 plusmn 9 91 plusmn 110 1867 plusmn 05 84 plusmn 16 90 plusmn 2
1980 plusmn 05 74 plusmn 24 91 plusmn 6
Strokepatients
1 2600 plusmn 07 100 plusmn 0 99 plusmn 22 2200 plusmn 04 61 plusmn 15 91 plusmn 43 2200 plusmn 04 38 plusmn 6 59 plusmn 44 1933 plusmn 07 38 plusmn 11 56 plusmn 45 2467 plusmn 06 38 plusmn 6 58 plusmn 46 2067 plusmn 06 33 plusmn 6 68 plusmn 57 2067 plusmn 03 47 plusmn 19 93 plusmn 28 2133 plusmn 04 43 plusmn 26 78 plusmn 3
2208 plusmn 06 50 plusmn 25 75 plusmn 17
recognized greatly increased from the training to the testingphase It can clearly be seen that the testing overall recogni-tion performance of gestures performed by the stroke patientsis lower than the ones performed by the healthy subjects asit was expected but still they produced a high percentageof true positive recognized gestures (TPR mean of 75)The computational time taken for training the model was onaverage 198ms for healthy subjects and 221ms for strokepatients
In order to evaluate inmore detail the recognition resultsthe selected performance measures are shown in Figure 3Ideally the accuracy and true positive rate should be closeto 100 and the false positive rate close to zero In this casethe median accuracy of recognition of gestures performed byhealthy subjects was over 95 and 87 for stroke patientsThe true positive rate showed lower values with respect to theaccuracy median values for healthy subjects of over 91 andfor stroke patients of over 75 The healthy subjects showedmedian values of false positive rates of 2 and 8 for strokepatients
The different performance measures can also provideinformation about the specific recognition of each gestureSpecifically the true positive rate refers to the ability ofthe method to correctly identify a specific gesture whilstthe false positive rate refers to the percentage of gestureswrongly classified Figure 4 shows the accuracy true positiverate and false positive rate for each gesture The difficultyof recognition of the different grasp gestures is different forthe healthy subjects or stroke patients The relax postureis a clearly distinctive gesture therefore showing the bestaccuracies best TPR and FPR values close to zero Forhealthy subjects the tripod and cylindrical gestures were themost difficult to be recognized and the most misrecognizedFor stroke patients the three gestures presented similardifficulties to be recognized and the tripod and lateral graspwere the most misrecognized
Table 4 presents the Pearson correlation values of thedifferent performancemeasures for the given conditionsTheaccuracy is inversely correlated with the false negative rateover all conditions (correlation coefficient C lt minus097 withstatistical significance) and also highly correlated with theTPR (C gt 099) The different measures of performance pro-vide specific information on the performance of recognitionbut the accuracy could be selected as the overall measureof recognition performance Therefore for the followinganalysis the accuracy will be used
It is expected that the grasping capabilities of the par-ticipants affect the recognition of hand postures especiallyin the case of stroke patients Therefore we also studied
8 BioMed Research InternationalAc
cura
cy (
)
100
80
60
40
20
0
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
Accu
racy
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(a)
TPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
TPR
()
100
80
60
40
20
0
Part
icip
ants
Part
icip
ants
Hea
lthy
subj
ects
Stro
ke p
atie
nts
(b)
FPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
FPR
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(c)
Figure 4 Performance measures for each gesture evaluating the recognition of hand postures (a) accuracy (b) true positive rate and (c)false positive rate
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
8 BioMed Research InternationalAc
cura
cy (
)
100
80
60
40
20
0
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
Accu
racy
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(a)
TPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
GestureLateralCylindricalTripodRelaxed
TPR
()
100
80
60
40
20
0
Part
icip
ants
Part
icip
ants
Hea
lthy
subj
ects
Stro
ke p
atie
nts
(b)
FPR
()
100
80
60
40
20
0
GestureLateralCylindricalTripodRelaxed
Part
icip
ants
Hea
lthy
subj
ects
GestureLateralCylindricalTripodRelaxed
FPR
()
100
80
60
40
20
0
Part
icip
ants
Stro
ke p
atie
nts
(c)
Figure 4 Performance measures for each gesture evaluating the recognition of hand postures (a) accuracy (b) true positive rate and (c)false positive rate
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
BioMed Research International 9
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
Flex
ion
(deg
)
10000
7500
5000
2500
000
ParticipantsStroke patientsHealthy subjects
fingerSmallRing RingMiddleIndexThumb
fingerSmallMiddleIndexThumb
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
Figure 5 Variability of the different finger flexion values produced by all subjects while performing the various gestures
Table 4 Correlation between performance measures
Performance measures Participant Pearson correlation
Accuracy versus TPR Healthy subject 0999lowast
Stroke patient 1000lowast
Accuracy versus FPR Healthy subject minus0973lowast
Stroke patient minus0997lowastlowastCorrelation is significant at the 001 level (2-tailed)
the variability of the finger flexion to produce the differentgestures and their impact on the recognition
Figure 5 shows a summary of the variability of theflexion angles over all subjects performing the differentgrasp gestures As expected the relaxed posture is the mostconsistent over all gestures For healthy subjects the ringfinger and little finger are the ones with greater variation asthey can freely move and are not actively participating inthe tripod grasp The ranges of variation for stroke patientsare not similar to the healthy subjects As the patients havedifferent levels of impairment the flexion of each finger hashigher ranges across stroke patients with several outliersexcept in the case of the ring finger and little finger wherethe variation was quite small (max 45 degrees)This might be
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
10 BioMed Research International
Subject10987654321
Accu
racy
()
100
80
60
40
20
0
Subject10987654321
Participants
Stroke patientsHealthy subjects
Figure 6 Results of the accuracy of gesture recognition per participant The left graph presents the results for healthy participants and theright one for stroke patients
ARAT600050004000300020001000
Mea
n ac
cura
cy (
)
10000
8000
6000
4000
2000
000
Figure 7 Association between accuracy of recognition and theresults of the ARAT
due to the limited mobility of these fingers and therefore thepatients had to rely on the thumb index finger and middlefinger to grasp the objects
The results of the accuracies of recognition of the differentgestures per participant are presented in Figure 6 Thevariation of accuracies of healthy subjects (between 89 and100) was smaller than the ones of stroke patients (between72 and 100)
The correlation results of the recognition accuracies withthe results of the arm motor recovery tests are presentedin Table 5 The higher positive correlations are between the
Table 5 Correlation between accuracy and common arm motorrecovery tests
Test compared withmean accuracy Pearson correlation Sig
ARAT total 0651 0040ARAT A 0630 0047ARAT B 0532 0087ARAT C 0680 0032ARAT D 0502 0103FM total 0461 0125FM proximal 0481 0114FM distal 0386 0172FM coordination 0141 0370
accuracies and the ARAT test especially to sections A (graspsubtest) and C (pinch subtest) This is not surprising giventhat the ARAT specifically assesses dexterity while FM isa much broader measure of motor impairment Figure 7presents accuracy and ARAT score values for each subject tovisually show the high association
32 Grasp Evaluation Results The results of the evaluation ofeach of the grasps using the proposed qualitymeasure GQ arepresented in Figure 8Thedifference between the participantrsquostype and the grasp performed are shown in Figure 8(a)The healthy subjects presented smaller variations than thestroke patients and their median grasp quality was higherThe higher variation was presented while the participants
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
BioMed Research International 11G
Q
100080060040020000100080060040020000
100080060040020000100080060040020000
GQ
GQ
GQ
Ges
ture
Ges
ture
Rela
xed
Trip
od
Cylin
dric
alLa
tera
l
ParticipantsStroke patientsHealthy subjects
ParticipantsStroke patientsHealthy subjects
(a)
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
GQ
100
080
060
040
020
000
Participants
Stroke patientsHealthy subjects
Subject10987654321
Subject10987654321
Ges
ture
Rela
xed
Trip
odCy
lindr
ical
Late
ral
(b)
Figure 8 Results of the grasp quality (GQ) for healthy subjects and stroke patients per gesture (a) summary of the results and (b) resultsshowing variability per participant
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
12 BioMed Research International
Table 6 Correlation between the grasp quality measure (GQ) and common arm motor recovery tests
Test compared with GQ Relax posture Tripod grasp Cylindrical grasp Lateral graspPearson correlation Sig Pearson correlation Sig Pearson correlation Sig Pearson correlation Sig
ARAT total minus0365 0187 minus0387 0172 minus0247 0278 0078 0427ARAT A minus0375 0180 minus0443 0136 minus0247 0278 0008 0492ARAT B minus0139 0371 minus0214 0305 minus0007 0494 0209 0310ARAT C minus0527 0090 minus0502 0102 minus0434 0141 minus0055 0448ARAT D minus0077 0429 minus0051 0452 0020 0482 0347 0200FM total minus0117 0391 minus0133 0376 minus0005 0496 minus0037 0465FM proximal 0130 0380 minus0083 0422 minus0002 0498 minus0071 0434FM distal minus0042 0461 minus0086 0419 0044 0459 0023 0479FM coordination minus0638lowast 0044 minus0660lowast 0038 minus0574 0069 minus0441 0137lowastCorrelation is significant at the 005 level (1-tailed)ARAT = arm scores of the action research arm test FM = arm scores of the Fugl-Meyer motor assessment
Table 7 Multiple regression results
Model 119877 119877 square Adjusted 119877 square Std Error of the Estimate Change statistics119877 Square change 119865 change df1 df2 Sig 119865 change
1 0780 0609 0597 004365 0609 53469 20 687 0000
performed the lateral grasp the grasp quality varied 24(075ndash099) for healthy patients and 28 (067ndash095) forstroke patients
The variability between different participants is shownin Figure 8(b) The grasp performed by healthy participantsgot a measure of quality above 085 for the relaxed tripodand cylindrical grasps However performing the lateral grasppresented a higher variability especially for subjects 6 and10 who obtained grasp qualities as low as 057 and 073respectively Stroke patients presented a higher variabilitybetween subjects but in general there was a consistent graspquality per subject and gesture (maximum 17)mdashexceptpatient 1 performing the cylindrical grasp who showed avariation of 25 (066ndash091)
The correlation results of the grasp quality with the resultsof the arm motor recovery tests are presented in Table 6 Ingeneral there are no significant correlations between the testsand the results obtained from the quality measure except fora negative correlation with the Fugl-Meyer coordination testperforming the relax posture and the tripod grasp
In order to assess how the quality measure GQ was influ-enced by the type of participant (healthy subject versus strokepatient) the gesture and the specific subject performing thegrasps a multiple regression analysis was performed Theresults of the analysis are presented in Table 7 The 119877 valueof 078 shows that the predictors are a good estimation of thequality measureThe 1198772 indicates that the predictors accountfor 609 of the variation of the grasp quality measure andas the adjusted 1198772 is similar to it it means that the model isable to generalize Also the standard error 004365 indicatesthatmost samplemeans from the estimated values are similarto the quality measure means The 119865 change is an importantvalue as it indicates that this model causes 1198772 to change fromzero to 0609 which is significant with a probability less than0001
4 Discussion
In this study we examined the performance of a techniquebased on support vector machines for the recognition ofhand gestures using the finger flexionextension anglescomparing a group of healthy subjects with a group ofpoststroke patients The results for stroke patients in generalshow lower accuracies lower true positive rates and greaterfalse positive rates than the ones performed by the healthysubjectsThis is due to the fact that patientrsquos impairment afterstroke affects their ability to reproduce gestures with smallvariations and in a repeatable way However we hypothesizedthat the accuracy for gesture recognition for stroke patientscan provide an insight into patientsrsquo ability to consistentlyreproduce gestures This was corroborated with the highcorrelation between the recognition accuracies and the scoresof the ARAT (119862 = 0651 for ARAT total) showing thatthe accuracy of recognition can be used as a measure ofthe grasping capabilities for the impaired hand Thereforethis method shows potency to be used to detect changesrelated to lack of ability to reproduce gestures in a consistentway
The results of the grasp quality measure showed thatthe healthy subjects presented smaller variations than thestroke patients and their grasp quality is higher showingthat the measure can be used to assess the capabilities ofstroke patients to perform grasp postures in a similar waythan healthy people However the variation between healthysubjects is higher than we expected This can be due tothe mechanical design of the orthosis as the readings ofthe sensors could vary from orthosis to orthosis dependingon the level of tension of the elastic cords With a moreaccurate device as the currently developed next version ofthe SCRIPT orthosis it is expected that the variability ofthe healthy postures will be reduced which will give more
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
BioMed Research International 13
consistent high quality values for the healthy participants andtherefore would be more reliable for the evaluation of strokepatient grasps Also a broader study with a larger numberof healthy subjects and stroke patients could give a moreaccurate measure of the quality of the grasp
Also the lack of correlation between the grasp qualityGQ and the results of the arm motor recovery tests couldindicate that this measure is perhaps providing differentinformation about the patients indicating specifically theirlevel of ability to perform each of the different gestures whichis not specifically measured by the common tests The resultsof the multiple regression analysis showed that the type ofparticipant (healthy subject or stroke patient) the grasp andthe specific subject can account for 61 of the grasp qualityvariability which indicate presence of other indicators thathave not been considered in our model We hypothesizethat these variations could be due to differences influencedby gender the hand performing the gestures (dominant ornondominant) or the level of disability of the stroke patientsFuture work is needed to study the influences of these factorswhen more participants and more accurate readings areavailable
The technique presented in this paper will be used forthe recognition of hand postures needed for the activitiesof daily life while patients playing rehabilitation games withthe SCRIPT orthosis over a period of six weeks With theseresults more exhaustive analysis can be performed whichcould provide insights into improvements on the ability toperform the grasp postures over time influencing the overallmotor performance
5 Conclusion
The results obtained with this study show that a techniquebased on support vector machines can be used to recognizedifferent grasp gestures for stroke patients with a validaccuracy while playing rehabilitation games wearing a spe-cially designed exoskeleton for their rehabilitation This willallow the training of various grasp postures to improve theperformance of these postures needed for several activitiesof daily living Moreover we showed that the accuracy ofrecognition can be used to assess the ability of the strokepatients to consistently repeat the gestures and the proposedgrasp quality GQ can be used to measure the capabilities ofthe stroke patients to perform grasp postures in a similar waythan healthy people These two measures could be used ascomplementary measures to the other upper limb motiontests such as ARAT and they can be potentially appliedto evaluate the grasp performance of patients using otherorthosis able to measure finger flexion
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work has been partly funded by the European Commu-nity Seventh Framework Programme under Grant agreementno FP7-ICT-288698 (SCRIPT) The authors are gratefulto the SCRIPT consortium for their ongoing commitmentand dedication to the project and to the participants thatvoluntarily agreed to participate in this study They wouldalso like to thank Professor Vittorio Sanguineti for hisvaluable advice of the classification method and ValentinaLombardi and Daniele Galafate for their help on performingthe experiments on stroke patients of the IRCCS San RaffaelePisana Italy
References
[1] G B Prange M J A Jannink C G M Groothuis-OudshoornH J Hermens and M J Ijzerman ldquoSystematic review of theeffect of robot-aided therapy on recovery of the hemiparetic armafter strokerdquo Journal of Rehabilitation Research amp Developmentvol 43 no 2 pp 171ndash183 2006
[2] G Kwakkel B J Kollen and H I Krebs ldquoEffects of robot-assisted therapy on upper limb recovery after stroke a system-atic reviewrdquo Neurorehabilitation and Neural Repair vol 22 no2 pp 111ndash121 2008
[3] J Mehrholz T Platz J Kugler andM Pohl ldquoElectromechanicaland robot-assisted arm training for improving arm functionand activities of daily living after strokerdquo Stroke vol 40 no 5pp e392ndashe393 2009
[4] A A A Timmermans H A M Seelen R D Willmann etal ldquoArm and hand skills training preferences after strokerdquoDisability and Rehabilitation vol 31 no 16 pp 1344ndash1352 2009
[5] A A Timmermans H A Seelen R D Willmann and HKingma ldquoTechnology-assisted training of arm-hand skills instroke concepts on reacquisition ofmotor control and therapistguidelines for rehabilitation technology designrdquo Journal ofNeuroEngineering and Rehabilitation vol 6 no 1 article 1 2009
[6] G B Prange H J Hermens J Schafer et al ldquoSCRIPT tele-robotics at home-functional architecture and clinical appli-cationrdquo in Proceedings of the 6th International Symposiumon E-Health Services and Technologies (EHST rsquo12) GenevaSwitzerland 2012
[7] B Leon A Basteris and F Amirabdollahian ldquoComparingrecognition methods to identify different types of grasp forhand rehabilitationrdquo in Proceedings of the 7th International Con-ference on Advances in Computer-Human Interactions (ACHI14) pp 109ndash114 Barcelona Spain 2014
[8] N Foubert AMMcKee R A Goubran and F Knoefel ldquoLyingand sitting posture recognition and transition detection using apressure sensor arrayrdquo in Proceedings of the IEEE Symposium onMedical Measurements and Applications (MeMeA rsquo12) pp 1ndash6May 2012
[9] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005
[10] M A Oskoei and H Huosheng ldquoSupport vector machine-based classification scheme for myoelectric control applied toupper limbrdquo IEEE Transactions on Biomedical Engineering vol55 pp 1956ndash1965 2008
[11] U R Acharya S V Sree M M R Krishnan et al ldquoAtheroscle-rotic risk stratification strategy for carotid arteries using
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007
14 BioMed Research International
texture-based featuresrdquo Ultrasound in Medicine amp Biology vol38 no 6 pp 899ndash915 2012
[12] C-M Kastorini G Papadakis H J Milionis et al ldquoCompara-tive analysis of a-priori and a-posteriori dietary patterns usingstate-of-the-art classification algorithms a casecase-controlstudyrdquo Artificial Intelligence in Medicine vol 59 pp 175ndash1832013
[13] G D Fulk and E Sazonov ldquoUsing sensors tomeasure activity inpeople with strokerdquo Topics in Stroke Rehabilitation vol 18 no6 pp 746ndash757 2011
[14] M Tavakolan Z G Xiao and C Menon ldquoA preliminaryinvestigation assessing the viability of classifying hand posturesin seniorsrdquo BioMedical Engineering Online vol 10 article 792011
[15] S Puthenveettil G Fluet Q Qinyin and S AdamovichldquoClassification of hand preshaping in persons with stroke usingLinear Discriminant Analysisrdquo in Proceedings of the AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (EMBC rsquo12) pp 4563ndash4566 2012
[16] N Yozbatiran L Der-Yeghiaian and S C Cramer ldquoA stan-dardized approach to performing the action research arm testrdquoNeurorehabilitation and Neural Repair vol 22 no 1 pp 78ndash902008
[17] S Ates J Lobo-Prat P Lammertse H van der Kooij and AH Stienen ldquoSCRIPT Passive Orthosis Design and technicalevaluation of the wrist and hand orthosis for rehabilitationtraining at homerdquo in Proceedings of the International Conferenceon Rehabilitation Robotics pp 1ndash6 Jun 2013
[18] S Electronics Flex Sensor 22rdquo httpswwwsparkfuncomproducts10264
[19] S Nijenhuis G Prange J Schafer J Rietman and J BuurkeldquoFeasibility of a personalized armhand training system foruse at home after stroke results so farrdquo in Proceedings ofthe International NeuroRehabilitation Symposium (INRS rsquo13)Zurich Switzerland 2013
[20] A R Fugl Meyer L Jaasko and I Leyman ldquoThe post strokehemiplegic patient I a method for evaluation of physicalperformancerdquo Scandinavian Journal of Rehabilitation Medicinevol 7 no 1 pp 13ndash31 1975
[21] D J Gladstone C J Danells and S E Black ldquoThe Fugl-meyerassessment of motor recovery after stroke a critical review ofits measurement propertiesrdquo Neurorehabilitation and NeuralRepair vol 16 no 3 pp 232ndash240 2002
[22] D J Magermans E K J Chadwick H E J Veeger and F CT van der Helm ldquoRequirements for upper extremity motionsduring activities of daily livingrdquo Clinical Biomechanics vol 20no 6 pp 591ndash599 2005
[23] C J van Andel N Wolterbeek C A M Doorenbosch DVeeger and J Harlaar ldquoComplete 3D kinematics of upperextremity functional tasksrdquo Gait and Posture vol 27 no 1 pp120ndash127 2008
[24] A D Keller C L Taylor and V Zahm Studies to Determinethe Functional Requirements for Hand and Arm ProsthesisDepartment of Engineering University of California 1947
[25] A Chaudhary J L Raheja K Das and S Raheja ldquoIntelligentapproaches to interact with machines using hand gesturerecognition in natural way a surveyrdquo International Journal ofComputer Science amp Engineering Survey (IJCSES) vol 2 2011
[26] L Lamberti and F Camastra ldquoHandy a real-time three colorglove-based gesture recognizer with learning vector quantiza-tionrdquoExpert SystemswithApplications vol 39 no 12 pp 10489ndash10494 2012
[27] J Joseph and J LaViola A Survey of Hand Posture and GestureRecognition Techniques and Technology BrownUniversity 1999
[28] X Deyou ldquoA neural network approach for hand gesturerecognition in virtual reality driving training system of SPGrdquoin Proceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 519ndash522 August 2006
[29] R Palm and B Hiev ldquoGrasp recognition by time-clusteringfuzzy modeling and Hidden Markov Models (HMM)mdashacomparative studyrdquo in Proceedings of the IEEE InternationalConference on Fuzzy Systems (FUZZ 08) pp 599ndash605 HongKong June 2008
[30] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systems andTechnology vol 2 no 3 article 27 2011
[31] C W Hsu C C Chang and C J Lin ldquoA practical guide tosupport vector classificationrdquo 2010
[32] R Suarez M Roa and J Cornella ldquoGrasp quality measuresrdquoTech Rep Technical University of Catalonia 2006
[33] B Leon J L Sancho-Bru N J Jarque-Bou A Morales and MA Roa ldquoEvaluation of human prehension using grasp qualitymeasuresrdquo International Journal of Advanced Robotic Systemsvol 9 article 116 2012
[34] A Liegeois ldquoAutomatic supervisory control of the configurationand behavior of multibody mechanismsrdquo IEEE Transactions onSystems Man and Cybernetics vol 7 no 12 pp 868ndash871 1977
[35] A G Lalkhen and AMcCluskey ldquoClinical tests Sensitivity andspecificityrdquo Continuing Education in Anaesthesia Critical Careand Pain vol 8 no 6 pp 221ndash223 2008
[36] F Amirabdollahian R Loureiro E Gradwell C CollinWHar-win and G Johnson ldquoMultivariate analysis of the Fugl-Meyeroutcome measures assessing the effectiveness of GENTLESrobot-mediated stroke therapyrdquo Journal of NeuroEngineeringand Rehabilitation vol 4 article 4 2007