1

An adaptive gyroscope based algorithm for temporal gait

analysis

Barry R. Greene, Intel Digital Health Group, Leixlip, Co. Kildare, Ireland and the

TRIL Centre (email: barry.r.greene@intel.com)

Denise McGrath, School of Physiotherapy and Performance Science, University

College Dublin, Ireland (e-mail: denisemcgrath@hotmail.com)

Ross O‟Neill, TRIL Centre, University College Dublin, Ireland

(e-mail: rossoneill@yahoo.com)

Karol J. O‟Donovan, Intel Digital Health Group Leixlip, Co. Kildare, Ireland and the

TRIL Centre

Adrian Burns, Intel Digital Health Group Leixlip, Co. Kildare, Ireland and the TRIL

Centre (e-mail: adrian.burns@intel.com)

Brian Caulfield, CLARITY Centre for Sensor Web Technologies and School of

Physiotherapy and Performance Science, University College Dublin, Ireland

(e-mail: b.caulfield@ucd.ie)

2

An adaptive gyroscope based algorithm for temporal

gait analysis

Barry R. Greene, Denise McGrath, Ross O‟Neill, Karol J. O‟Donovan, Adrian Burns, Brian

Caulfield

Abstract

Body-worn kinematic sensors have been widely proposed as the optimal solution for

portable, low cost, ambulatory monitoring of gait. This study aims to evaluate an

adaptive gyroscope-based algorithm for automated temporal gait analysis using body-

worn wireless gyroscopes. Gyroscope data from nine healthy adult subjects

performing four walks at four different speeds were then compared against data

acquired simultaneously using two force plates and an optical motion capture system.

Data from a poliomyelitis patient, exhibiting pathological gait walking with and

without the aid of a crutch were also compared to the forceplate.

Results show that the mean true error between the adaptive gyroscope algorithm and

force plate was -4.5±14.4 ms and 43.4±6.0 ms for IC and TC points respectively in

healthy subjects. Similarly, the mean true error when data from the polio patient were

compared against the force plate was -75.61±27.53 ms and 99.20±46.00 ms for IC

and TC points respectively. A comparison of the present algorithm against temporal

gait parameters derived from an optical motion analysis system showed good

agreement for nine healthy subjects at four speeds. These results show that the

algorithm reported here could constitute the basis of a robust, portable, low-cost

system for ambulatory monitoring of gait.

3

1. Introduction

Quantitative measurement of gait has been possible for many years using laboratory

based kinematic and kinetic equipment such as marker-based motion capture systems,

force-platforms, and electrophysiological sensors/electrodes. Evaluation of temporal

gait parameters such as stride and swing time and the variability of these parameters

during walking has been shown to be useful in assessing the health of the

neuromuscular system in a variety of populations including the elderly and patients

with specific diseases [10, 12]. Traditional, high-cost laboratory-based methods for

measurement of these parameters involve restrictive environments and long set-up

times. Consequently, an extensive body of research has recently emerged that

examines portable, low-cost, light-weight ambulatory monitoring systems that can

potentially be deployed in a variety of settings such as a doctor‟s office or in the

home. Inertial sensors consisting of either accelerometers [28], gyroscopes [1, 20, 24],

or a combination of sensors [7, 15] have been widely proposed as the optimal solution

for this purpose. The use of gyroscopes for the calculation of temporal parameters is a

particularly attractive solution as, unlike accelerometers, gyroscopes are less sensitive

to the influence of gravity and therefore the signal is less dependent on exact sensor

positioning [27]. Several studies have calculated initial contact (IC) and terminal

contact (TC) times from gyroscopes, and compared their results to either foot-

switches [1, 24] or force-plates and an optical motion capture system [2, 25], with

promising results.

Previous studies used gait event detection algorithms that were applied to the

gyroscope signal to identify the appropriate time-points that corresponded to either

initial or terminal contact. Reliable detection of gait events can be difficult across

4

varying gait-speeds and subject populations. Furthermore, the angular velocity signal

derived from body-worn gyroscopes can be prone to noise and artifact. Han et al. used

an adaptive approach with ankle mounted tri-axial accelerometers to discriminate gait

phases in Parkinson‟s disease patients [11]. Adaptive signal processing approaches

have also found utility in a number of biomedical research areas such as adaptive

noise cancellation [23], heart-beat classification [8] and epileptic seizure prediction

[14]. A temporal gait event detection algorithm that could adapt to the differing

characteristics of the gyroscope signal, for different modes and speeds of human gait

(such as shuffling & fast walking) across varying ages and gait types, as well as being

robust to artifact and noise, could be beneficial.

The purpose of this study was to evaluate the performance of a gyroscope based gait

analysis algorithm that includes adaptive threshold calculation and artifact rejection to

enhance robustness to differing walking characteristics and speeds as well as noise in

the angular velocity signal due to artifact. We sought to test the algorithm against a

force-plate (which is considered the “gold standard” for measuring temporal gait

parameters) and also against an optical motion-capture system (which allows the

capture of a greater number of strides than a force-plate and so can be used to

calculate temporal gait parameters) for the calculation of IC and TC in a group of

healthy volunteers walking at slow, medium and fast pace, as well as while

mimicking shuffling. We also included a patient with Poliomyelitis to test our

algorithm on pathological gait.

5

2. Method

2.1. Experimental set-up

The gait of nine healthy subjects (8M, 1F, mean age: 29.7±3.5) was measured

simultaneously using three gait measurement technologies; body-worn gyroscopes, a

force-plate and an optical motion capture system. The purpose of the study was

explained to all subjects before they provided signed consent to take part in the study.

All speeds were self-selected and not controlled, to minimise the effect of the

experimental protocol on the subjects‟ gait.

Data were recorded whilst each subject performed four walks at four different self-

selected speeds; normal, fast, slow and a mimicked shuffling gait along a 15m

walkway in a motion analysis laboratory. In all, 144 walking trials were completed,

yielding 187 separate IC and 190 TC events for subsequent data analysis.

Data were also recorded from an additional poliomyelitis (polio) patient (M, age 60),

exhibiting pathological gait in both lower limbs. Paralytic poliomyelitis affecting the

left limb requires the patient to wear an above the knee calliper mobility device. He

also has a clubbed left foot and fused left ankle. There is therefore no knee

flexion/extension or ankle dorsi/plantar flexion on the left side. On the right side the

patient walks with a “steppage” gait due to a dropped foot, which causes him to lift

the right foot high, enabling the toes to clear the ground safely. We recorded 22

walking trials (12 with and 10 without the use of a crutch) from this patient. Six trials

were neglected from analysis owing to improper contact with the forceplate and

synchronization errors between acquisition systems.

Fig.1 below illustrates the experimental set-up used to capture the data used in this

study.

6

Insert Fig 1. here

2.2. Force plate data acquisition

Two AMTI (Advanced Mechanical Technology Inc.) force plates were located in the

middle of a 15m walkway. The participants were instructed to walk naturally over the

force plates, at four different speeds. The participants were given an opportunity to

practice their different walking speeds. Subjects wore their own shoes. Force plate

data were acquired at 1kHz. A 10N threshold was applied to the vertical component

of the ground reaction force to detect IC and TC events. This threshold was

considered high enough to avoid erroneous identification of gait events due to noise,

but low enough not to miss them altogether [6]. A step was considered to be valid if

the foot landed clearly within the force platform.

2.3. Optical motion data acquisition

Reference kinematic data were acquired using a CODA optical motion analysis

system (http://www.codamotion.com, Charnwood Dynamics Ltd, Leicestershire, UK).

Two CODA cx1 units were used to acquire data, one placed at either side of the

walkway. The CODA cx1 unit is a commercially available optoelectronic motion

capture system for recording and analyzing human movement. Two CODA infrared

light-emitting diode markers were placed on the left and right foot. Markers were

positioned on the inferior lateral aspect of the heel, and the lateral aspect of the fifth

metatarsal head, on the exterior of the subjects‟ training shoes. The optical kinematic

data were collected at a sampling rate of 200Hz. Kinematic data were analyzed using

the CODAmotion analysis software. Kinematic data from the optical motion capture

system and force plates were synchronized through the analysis software. The IC and

TC times for each trial were calculated manually using the algorithms reported by

7

Hreljac & Marshall [13], referred to here as the Hreljac-Marshall algorithm (HMA).

These algorithms are based on the values of the vertical and horizontal components of

jerk equal to zero, for IC and TC respectively.

2.4. Gyroscope data acquisition

Inertial monitoring unit (IMU) kinematic data was acquired using four SHIMMER

(http://shimmer-research.com/) wireless sensors [4, 17], one each attached to the foot

and shank of the left and right leg. For the purpose of this study, only data from the

sensors on the shank were analysed. Each gyroscope sensor was attached to the

anterior of each shank, oriented to capture movement about the anatomical medio-

lateral axis of the shank of the leg, and secured using surgical tape, half way on the

imaginary line between the Tibial Tuberosity (TT) and the Lateral Malleolus (LM).

Each sensor contained both a tri-axial accelerometer on the baseboard and an add-on

tri-axial gyroscope daughterboard. Each was programmed to sample each axis at a

rate of 102.4Hz. Data were acquired from each body-worn sensor using a custom built

application developed using the BioMOBIUS software development environment

(http://www.biomobius.org) [5, 18]. Gyroscope data were acquired through

BioMOBIUS on a Dell Optiplex GX745 PC, with 2GB RAM and an Intel Core® 2

1.8GHz CPU.

All post-processing and analysis was carried out off-line using the MATLAB®

(http://www.mathworks.com/, Natick, VA, USA) programming environment. The raw

gyroscope data were calibrated to derive the angular velocity vectors with respect to

the sensor unit coordinate axis. A standard calibration procedure [9] was used to

calibrate all gyroscopes used in this study. Before further processing, the raw

gyroscope signal was low pass filtered using zero-phase 5th

order Butterworth filter

with a 5 Hz corner frequency.

8

The BioMOBIUS based acquisition system and the optical motion capture/force plate

system were synchronised using a dedicated trigger output from the CODA motion

analysis system. This trigger is activated at the initiation and deactivated at the

conclusion of a capture from the CODA system. The analog trigger signal was

connected to the analog-to-digital input of a separate SHIMMER device (using an

analog expansion breakout board), sampled at 102.4Hz and transmitted wirelessly via

Bluetooth to the BioMOBIUS acquisition software. Data from the synchronisation

and kinematic SHIMMER devices were simultaneously recorded within

BioMOBIUS.

2.5. Temporal gait parameters

Temporal parameters of gait were derived using an algorithm which adaptively

calculates thresholds in order to determine IC and TC events from the medio-lateral

angular velocity of the shank. A number of authors have proposed methods for

detecting IC and TC points from shank angular velocity signals. The present

algorithm differs from previously reported methods in that it includes adaptive

threshold calculation and artefact rejection to improve robustness to differing walking

characteristics and speeds as well as noise in the angular velocity signal due to

artefact.

In order to ensure the angular velocity signal derived from the gyroscope has the

correct polarity, the „skewness‟ of the signal (a measure of the asymmetry of the

signal) is calculated for each walk. If the skewness of the filtered signal is less than

zero the angular velocity signal is inverted automatically in software to ensure correct

polarity of the signal when applied to the algorithm.

9

2.5.1. Adaptive Threshold calculation

As with previously reported gait analysis algorithms [24, 25] employing gyroscopes,

the IC and TC points are defined relative to the swing phase which is more easily

detected (mid-swing points manifest in the medio-lateral angular velocity signal as

large positive peaks, see Fig. 4 & 5)

A number of adaptive thresholds were employed to ensure accurate and robust

detection of IC and TC points from the medio-lateral shank angular velocity signal

over a range of speeds:

In detecting the mid-swing point for each gait cycle on the medio-lateral angular

velocity signal ( ML ), valid local maximum peaks of the signal should have a

preceding minimum at least th1 deg/s less than the maximum, calculated as

MLth max6.01

In detecting the mid-swing point for each stride from the medio-lateral angular

velocity signal for each leg, valid local maximum peaks of the signal should be

greater than th2 deg/s. th2 is calculated as 0.8 times the mean of all data points greater

than the mean value of the angular velocity signal.

N

i MLMLiNth

12

18.0

where ML is the mean of the medio-lateral angular velocity signal and N is the

number of samples.

For detecting IC points, the local minimum should have a preceding maximum at least

th3 deg/s greater than the local minimum. th3 is calculated as 0.8 times the absolute

value of mean of all points less than the mean angular velocity;

N

i MLMLiNth

13

18.0

10

For detecting TC points, valid local minima should be less than th4. th4 is calculated

as 0.8 times the mean of all points less than the mean value of the angular velocity

signal.

N

i MLMLiNth

14

18.0

For IC detection, valid local minima should be less than th5, defined as the mean

value of the angular velocity signal for that trial ( ML ).

For TC detection, the local minimum should have a preceding maximum at least th6

deg/s greater than the local minimum, defined as 36 2thth .

For mid swing detection, if two max peaks are found within t1 seconds of each other

only the greater max is considered. t1 is defined as 0.5 seconds or sft 5.01 .

Similarly for IC and TC detection following mid-swing detection, only data within ±t2

seconds is considered, defined as 1.5 seconds or sft 5.11 .

Fig. 2 is a flowchart detailing the operation of the adaptive gyroscope algorithm.

Insert Fig. 2 here

2.5.2. Artefact rejection

Temporal gait parameters are calculated from the gait events, IC and TC. An artefact

rejection routine is employed to remove spurious temporal parameters that are

calculated from noisy or artefactual gyroscope data. This routine is also designed to

account for missing and extra IC and TC points detected by the adaptive gyroscope

algorithm. Artefact rejection is based on two strands; examination of temporal

sequence information and examination of times between successive characteristic

points (cycle times)

11

a) Temporal sequence information Once all characteristic points are detected using the present algorithm, each point is

assigned a numerical label 1 to 4, where; 1. right IC, 2. left TC, 3. left IC and 4. right

TC.

A correct gait cycle (if starting on a right IC) will then follow the sequence 1,2,3,4.

This is determined by subtracting each from label the previous label. Those samples

which do not follow this sequence and so do not produce a difference equal to either -

3 or 1 are deemed artefact. Additional checks were also carried out to ensure that a

TC is always preceded by an IC and an IC preceded by a mid-swing point.

b) Gait cycle information The time between adjacent gait characteristic points is calculated for each set of

characteristic points (right IC, left TC, left IC, right TC). This is commonly referred to

as gait cycle time when right IC is the characteristic point used. If the difference

between any successive characteristic point is greater than 2.5 seconds the associated

characteristic point is flagged as artefact. Similarly, if the difference between any

successive IC or TC point is zero seconds, the associated point is flagged as artefact.

Furthermore, any gait parameters with a negative or zero value are also neglected

from analysis.

2.6. Statistical analysis

IC and TC points derived from the body-worn gyroscopes using the algorithm

described here were compared against those derived from the optical motion capture

system (using the HMA) and the force plate using a number of metrics. The true error

is defined as the difference in time (in milliseconds) between the IC or TC time

detected using the force plate, and the time that same point is detected using the

adaptive gyroscope algorithm. The IC and TC characteristic points derived using the

12

present algorithm and the motion capture system were used to calculate the four

temporal gait parameters; stride time, stance time, swing time and step time.

Stride time is defined as the time from IC of one foot to IC of the same foot. The

stance time is defined as the time between IC and TC of the same foot. Similarly

swing time is defined as the time between TC and IC on the same foot. Step time is

calculated as the time between IC on one foot and IC on the other foot. In this study,

the data for left and right feet for each temporal gait parameter were merged as it was

assumed that gait asymmetry would not be an issue in our control population.

Temporal gait parameters derived from shank angular velocity signals were compared

against those obtained from the HMA using the mean percentage error, intraclass

correlation coefficient (ICC(2, k) [26] and the Bland-Altman method [3]. Bland-

Altman plots are shown to illustrate graphically the agreement between temporal

parameters, simultaneously derived using the adaptive gyroscope algorithm and the

HMA.

3. Results

3.1. Force plate results

Table 1 below compares the performance of the adaptive gyroscope-based algorithm

in terms of true error against the force plate in detecting IC and TC points.

Insert Table 1 here

In total, the experiment yielded 187 IC and 190 TC points for comparison. The mean

true error for IC across all speeds was -4.5±14.4 ms while for TC the mean true error

was 43.4±6.0 ms. Fig.3 graphically illustrates the range of true error between IC and

TC point derived automatically from the gyroscope at four different self-selected

speeds against those derived using the force plate.

13

Insert Fig. 3 here

Similarly walking trials recorded from the polio patient yielded 28 IC and 26 TC

points for comparison. The mean true error for IC -75.61±27.53 ms while for TC the

mean true error was 99.20±46.00 ms. For walking trials where the patient made use of

a crutch the mean true error for IC and TC were -80.98±28.29 ms and 106.83±56.62

ms respectively. Similarly, for trials where the patient did not use a crutch the mean

true error for IC and TC were -70.96±27.45 ms and 92.65±36.69 respectively. Fig.4

shows the shank medio-lateral angular velocity signals as well as the force plate

timings (GRF curves) for the polio patient.

Insert Fig. 4 here

3.2. Optical motion capture system results

The mean velocities (taken across all subjects) for each of the walking speeds were as

follows; fast: 1.56 m/s (5.61 km/h); normal: 1.10 m/s (3.95 km/h); slow: 0.65 m/s

(2.36 km/h); shuffle: 0.48 m/s (1.71 km/h).

Table 3 below gives the results for agreement between temporal gait parameters

derived from both adaptive gyroscope algorithm and the HMA at four different

speeds. The results show that there is a good agreement (ICC > 0.80) between both

for each temporal gait parameter.

Insert Table 2 here

Fig.5 is a Bland-Altman plot for stride times calculated from data acquired from the

adaptive gyroscope algorithm and data acquired using the HMA on kinematic data

acquired from the optical motion analysis system during fast walking.

Insert Fig. 5 here

14

4. Discussion

A vital requisite for a portable gait monitoring system is its ability to accurately and

reliably determine when IC and TC occur. In this study, we sought to investigate

whether the temporal parameters obtained from nine healthy subjects, as calculated by

the adaptive gyroscope algorithm using body-worn gyroscopes, corresponded to those

simultaneously obtained from the same subject by two force plates and a marker-

based motion analysis system. In addition we sought to show that the present

algorithm is robust in automatically detecting the gait characteristic points from the

shank inertial sensor data of a patient exhibiting pathological gait.

Adaptive algorithms have recently been employed throughout the literature in

developing robust classification and detection algorithms. The algorithm presented

here applies a similar concept in its ability to automatically adapt to individual gait

features present in the gyroscope signal, through the inclusion of adaptive thresholds.

Previous research by the authors [22] successfully implemented a previously reported

algorithm for gait event detection from body-worn gyroscopes [25]. However, upon

applying this algorithm to our present data set, we experienced difficulties in

identifying gait events across varying speeds and subjects. Whilst this algorithm was

successful in event detection at slow and normal walking speeds, we found it to be

unsuitable for some trials where subjects had been mimicking shuffling or walking

very fast. Furthermore, we found that our implementation of this algorithm was prone

to missing and extra detection of IC and TC points due to noise and artefact that can

appear on the gyroscope derived angular velocity signal. Fig.6 compares the

performance of the adaptive gyroscope algorithm to that of our implementation of the

Salarian et al. algorithm in detecting gait events on two samples of over-ground

walking at two different speeds, from two separate healthy subjects. The graphs on

15

the left hand side show the IC and TC points, as identified by the adaptive algorithm.

The graphs on the right hand side show the IC and TC points as identified by our

implementation of an algorithm previously presented in the literature [25].

Insert Fig. 6 here

Body-worn gyroscopes and associated algorithms to assess gait have been employed

in other studies, with four studies in particular reporting successful validations with

reference systems [2, 15, 24, 25]. For comparison, these studies showed average true

errors (±SD) of -2 ms (±16), -8.7 ms (±12.5), 16.6 ms (±11.9) and -14ms (±23), for IC

events, and -35ms, -2.9 (±26.8), 3.7ms (±26.5) and 23ms (±28), for TC events,

respectively. Using our normal walking speed data for comparison, an average error

of -2.3 ms (±4.87 ms) for IC and 43.61 ms (±15.30 ms) for TC, between the adaptive

gyroscope algorithm and the force plate was found. We can see from these results that

the algorithm presented here compares favourably to previous gait studies using

gyroscopes in the detection of IC. In our study we found the detection of TC to be

biased across all speeds by an average 43.4 ms. A similar finding is observed in

Sabatini et al. [24], who found a bias in the opposite direction of -35 ms for TC

detection from foot worn gyroscopes. However, the low SD of average error for TC

detection across all speeds in this study is encouraging.

A systematic bias in the adaptive gyroscope algorithm‟s detection of TC may be

related to the fact that the point in the angular velocity signal thought to represent TC

may in fact represent a slightly different point in the gait cycle owing to the physical

separation of the shank mounted gyroscope from the toe. In identifying a consistent

bias in the detection of TC, it is possible that calculations can be corrected by taking

the systematic error into account. However, in many cases this would not be

necessary. If this system is ultimately being used to collect baseline data for long-term

16

gait monitoring in the field for example, then any variation/deterioration in gait

patterns will be flagged with respect to baseline measurements, regardless of the

systematic delay in TC identification. Moreover, Miller et al. [19] argue that the

timing of foot-off events may not be as crucial as foot-contact events because the start

and end points of the gait-cycle are defined by foot contact events. The analysis of

stride time is used for the time normalisation of data per gait cycle that is prevalent in

almost all aspects of gait analysis.

In this study, data were analysed from the gyroscope mounted on each shank rather

than data from the gyroscopes mounted on each foot. This was because the signals

from the foot mounted gyroscopes were found to be more prone to noise and artefact,

owing to vibration from heel impact. This finding is in contrast to that found in

previous research by Jasiewicz et al. [15] and so may be a limitation of the sensor

used in this study rather than being generally true for all foot-mounted gyroscopes.

We included a mimicked shuffling gait in this study to test the compliance of the

algorithm to differing gait dynamics. It is of particular interest to note that the

performance of the adaptive algorithm during the mimicked shuffling gait was upheld,

allowing easy detection of gait events, as can be seen in Fig. 6. However, it was

difficult to clearly identify IC and TC during shuffling using the force-plate, as the

increase and decrease of the force exerted on the plate was much more gradual. In

addition, both feet were often simultaneously on the force plate, resulting in a large

amount of unusable steps. This made the comparison difficult, as the force plate no

longer represented the “gold-standard” for this walking speed. Nevertheless, the

adaptive algorithm presented here has been shown to be robust in identifying IC and

TC in this condition, and may be considered a promising method for the calculation of

gait events at very slow speeds that emulate a shuffling gait.

17

A systematic bias in the adaptive gyroscope algorithm‟s detection of IC and TC is

also evident in the mean true error comparison against the forceplate in the polio

patient‟s data. Similar to healthy subjects, TC detection is slightly less accurate than

IC detection, particularly on the left foot where a more irregular gait was exhibited.

However, as previously discussed, systematic errors can be corrected for, and the low

standard deviation in the data is encouraging. As can be seen in figure 3, the adaptive

algorithm successfully extracted IC and TC points from the gyroscope signal, despite

left and right legs producing differing signals. Kotiadis et al. [16] deemed timing

limits of 150ms for early and late heel down detection as acceptable in their system

for control of a drop-foot stimulator. Our results fall well within these limits for

pathological gait.

The second part of our analysis focused on comparing the temporal parameters from a

larger number of steps captured by the optical motion capture system to those

simultaneously captured by the gyroscopes. We used a method that has been

previously described in the literature by Hreljac and Marshall (2000) (HMA) to

extract IC and TC times from kinematic data captured using a marker-based motion

capture system, resulting in absolute average errors of 4.7 and 5.6 ms, respectively,

using a force plate as a reference.

We compared stride, swing, stance and step times as calculated by the adaptive

gyroscope algorithm with the results produced using the HMA. We demonstrated

good to excellent agreement (ICC(2,k) > 0.8) for all parameters, across all speeds, as

can be seen in table 3.

We experienced some of the same problems with the HMA as outlined in O‟Connor

et al. [21]. We were unable to extract IC and TC times automatically as there are

multiple peaks in the acceleration curves and in many cases, the HMA was unable to

18

identify the appropriate peak. The location of the local maximum points had to be

verified manually through visual inspection to ensure the correct points were

identified, by looking at a sagittal view of the markers that could be advanced frame

by frame. This was a laborious process when compared to the ease at which large

amounts of data were automatically processed using the adaptive gyroscope algorithm

described here, with no checking of the data required.

We have assessed the performance of the adaptive gyroscope algorithm reported here

across a range of walking speeds, and we have found that the inclusion of adaptive

threshold calculation for IC and TC determination as well as and artefact rejection has

resulted in good to excellent agreement for temporal parameters, compared to another

kinematic method.

Acknowledgment

This research was completed as part of a wider programme of research within the

TRIL Centre, (Technology Research for Independent Living). The TRIL Centre is a

multi-disciplinary research centre, bringing together researchers from UCD, TCD,

NUIG, Intel, and GE Healthcare, funded by Intel, GE Healthcare and IDA Ireland.

http://www.trilcentre.org. The authors would like to thank Dr. Emer Doheny for

providing useful feedback on the manuscript as well as Mr. Ben Dromey for his help

with graphically detailing the experimental layout.

References

1. Aminian, K., et al., Spatio-temporal parameters of gait measured by an

ambulatory system using miniature gyroscopes. Journal of Biomechanics,

2002. 35(5): p. 689-699.

2. Aminian, K., et al., Evaluation of an ambulatory system for gait analysis in

hip osteoarthritis and after total hip replacement. Gait & Posture, 2004. 20(1):

p. 102-107.

19

3. Bland, J.M. and D.G. Altman, Measuring agreement in method comparison

studies. Statistical Methods in Medical Research, 1999. 8(2): p. 135-160.

4. Burns, A., et al., SHIMMER™ - A Wireless Sensor Platform for Non-invasive

Biomedical Research. IEEE Sensors, 2010. 10(9): p. 1527 - 1534

5. Burns, A., et al. Open Shareable Research Platform for Developing

Interoperable Personal Health Systems. in 1st AMA-IEEE medical technology

conference on individualized medicine. 2010. Washington DC.

6. Catalfamo, P., et al., Detection of gait events using an F-Scan in-shoe pressure

measurement system. Gait & Posture, 2008. 28(3): p. 420-426.

7. Cutti, A., et al., ‘Outwalk’: a protocol for clinical gait analysis based on

inertial and magnetic sensors. Medical and Biological Engineering and

Computing, 2010. 48(1): p. 17-25.

8. de Chazal, P. and R.B. Reilly, A Patient-Adapting Heartbeat Classifier Using

ECG Morphology and Heartbeat Interval Features. IEEE Trans. Biomed.

Eng., 2006. 53(12): p. 2535-2543.

9. Ferraris, F., U. Grimaldi, and M. Parvis, Procedure for effortless in-field

calibration of three-axis rate gyros and accelerometers. Sens. Mater, 1995.

7(5): p. 311-330.

10. Frenkel-Toledo, S., et al., Effect of gait speed on gait rhythmicity in

Parkinson's disease: variability of stride time and swing time respond

differently. J Neuroeng Rehabil., 2005. 31: p. 2-23.

11. Han, J., et al., Adaptive windowing for gait phase discrimination in

Parkinsonian gait using 3-axis acceleration signals. Medical and Biological

Engineering and Computing, 2009. 47(11): p. 1155-1164.

12. Hausdorff, J.M., et al., Altered fractal dynamics of gait: reduced stride-

interval correlations with aging and Huntington's disease. J Appl Physiol,

1997. 82(1): p. 262-269.

13. Hreljac, A. and R.N. Marshall, Algorithms to determine event timing during

normal walking using kinematic data. Journal of Biomechanics, 2000. 33(6):

p. 783-786.

14. Iasemidis, L.D., et al., Adaptive epileptic seizure prediction system. IEEE

Trans. Biomed. Eng., 2003. 50(5): p. 616-627.

15. Jasiewicz, J.M., et al., Gait event detection using linear accelerometers or

angular velocity transducers in able-bodied and spinal-cord injured

individuals. Gait & Posture, 2006. 24(4): p. 502-509.

16. Kotiadis, D., H.J. Hermens, and P.H. Veltink, Inertial Gait Phase Detection

for control of a drop foot stimulator: Inertial sensing for gait phase detection.

Medical Engineering & Physics, 2010. 32(4): p. 287-297.

17. Lorincz, K., et al. Wearable Wireless Sensor Network to Assess Clinical Status

in Patients with Neurological Disorders. in 6th International Symposium on

Information Processing in Sensor Networks. 2007.

18. McGrath, M.J. and T.J. Dishongh, A Common Personal Health Research

Platform -SHIMMER™ and BioMOBIUS™. Intel technology journal, 2009.

13(3): p. 122-147.

19. Miller, A., Gait event detection using a multilayer neural network. Gait &

Posture, 2009. 29(4): p. 542-545.

20. Najafi, B., et al., Measurement of stand-sit and sit-stand transitions using a

miniature gyroscope and its application in fall risk evaluation in the elderly.

IEEE Trans. Biomed. Eng, 2002. 49(8): p. 843-851.

20

21. O'Connor, C.M., et al., Automatic detection of gait events using kinematic

data. Gait & Posture, 2007. 25(3): p. 469-474.

22. O‟Donovan, K.J., et al. SHIMMER: A new tool for temporal Gait analysis. in

Proceedings of IEEE Eng. Med. Biol. 2009. Minneapolis, MN.

23. Rangayyan, R.M., Biomedical Signal Analysis. 2002, New York, NY: IEEE

Press/Wiley.

24. Sabatini, A.M., et al., Assessment of walking features from foot inertial

sensing. IEEE Trans. Biomed. Eng., 2005. 52(3): p. 486-494.

25. Salarian, A., et al., Gait assessment in Parkinson's disease: toward an

ambulatory system for long-term monitoring. IEEE Trans. Biomed. Eng.,

2004. 51(8): p. 1434-1443.

26. Shrout, P.E. and J.L. Fleiss, Intraclass correlations: Uses in assessing rater

reliability. Psychological Bulletin, 1979. 86(2): p. 420-428.

27. Tong, K. and M.H. Granat, A practical gait analysis system using gyroscopes.

Medical Engineering & Physics, 1999. 21(2): p. 87-94.

28. Zijlstra, W. and A.L. Hof, Assessment of spatio-temporal gait parameters

from trunk accelerations during human walking. Gait & Posture, 2003. 18(2):

p. 1-10.

21

Figure captions

Figure 1: Experimental setup to acquire data simultaneously from body-worn

gyroscopes, motion capture system and two force plates.

Figure 2: Flowchart detailing operation of adaptive kinematic algorithm

Figure 3: True error between force plate and adaptive gyroscope algorithm for nine

healthy subjects walking at four different speeds.

Figure 4: Shank medio-lateral angular velocity signals and GRF curves from a polio

patient exhibiting pathological gait in both limbs. Mid-swing, IC and TC points are

marked for each leg. True errors between force-plate and adaptive gyroscope

algorithm in detecting IC and TC points are shown graphically (true errors (in ms) are

as follows; right IC: 16.63, left IC: -74.26, right IC: 75.68, left IC; 148.73).

Figure 5: Bland Altman plot illustrating the agreement between stride times from nine

healthy subjects calculated using the adaptive gyroscope algorithm and those derived

from the HMA using the optical motion capture system during fast walking.

Figure 6: The top two panels show a sample of shank medio-lateral shank angular

velocity signal derived from a gyroscope mounted on the shank while a healthy

subject was mimicking shuffling while the bottom two panels show a sample shank

angular velocity signal from healthy subject while walking fast. The performance of

the adaptive gyroscope algorithm against the Salarian et al. algorithm for both walks

is shown.

22

Figures

Figure 1

23

Start

Gyroscope data

Calibration

LPF

Temporal gait parameters

Low pass filter medio-lateral

angular velocity

Calculate adaptive threshold

parameters from filtered

medio-lateral angular

velocity signal

Data can be obtained using

any mode of kinematic data

acquisition

Artifact rejection

Adaptive threshold calculation

Detect HS and

TO points

Calculate temporal gait parameters

Detect heel-strike and toe-off

points from filtered angular

velocity signals using

adaptive thresholds

Skewness

Correct for orientation of

body-worn sensor

Figure 2

24

-60

-40

-20

0

20

40

60

80

Shuffle Slow Med Fast

Walking speed

Heel str

ike t

rue e

rror

[ms]

-60

-40

-20

0

20

40

60

80

Shuffle Slow Med Fast

Walking speed

True error between forceplate and adaptive gyroscope algorithm

Toe o

ff t

rue e

rror

[ms]

Figure 3

25

Figure 4

26

1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4-0.2

-0.15

-0.1

-0.05

0

0.05

0.1

0.15Agreement between adaptive gyroscope algorihm and HMA stride times during fast walking

Mean of stride times [s]

Diff

ere

nce b

etw

een s

trid

e tim

es [s]

Intraclass correlation

coefficient ICC(2,k) = 0.9571

Figure 5

27

Figure 6

28

Table captions

Table 1: Mean true error (mean TE ± SD) in IC and TC between force plate and the

adaptive gyroscope algorithm for each of the four walking speeds.

Table 2: Mean results for agreement between temporal gait parameters derived from

the adaptive gyroscope algorithm and the HMA (using the optical motion capture

system) at four different speeds. Mean error refers to mean percentage error between

the adaptive gyroscope algorithm and the HMA for each temporal gait parameter.

29

Tables

Shuffle Slow Normal Fast

IC -1.39±36.14 -7.62±8.27 -2.30± 4.87 -6.50±9.15

TC 51.63±25.85 45.28±8.19 43.61±15.30 32.90±14.11

Table 1

30

Shuffle Slow Normal Fast

True

error

[ms]

Error

[%] ICC(2,k)

True

error

[ms]

Error

[%] ICC(2,k)

True

error

[ms]

Error

[%] ICC(2,k)

True

error

[ms]

Error

[%] ICC(2,k)

Stride

time -0.03 3.03 0.96 -4.86 2.23 0.97 -6.35 1.87 0.98 -5.14 2.58 0.97

Stance

time 97.27 10.5 0.86 74.09 9.16 0.8 35.43 8.03 0.92 31.47 6.56 0.93

Swing

time -90.2 29.55 0.83 -89.76 20.93 0.82 -58.04 14.07 0.93 -43.53 14.46 0.94

Step

time -24.31 11.4 0.88 -0.24 5.21 0.91 -6.64 5.59 0.89 -5.74 10.14 0.94

Table 2