A Novel Method for Recognition of Bioradiolocation Signal Breathing Patterns for Noncontact...

Hindawi Publishing CorporationInternational Journal of Antennas and PropagationVolume 2013 Article ID 969603 8 pageshttpdxdoiorg1011552013969603

Research ArticleA Novel Method for Recognition of BioradiolocationSignal Breathing Patterns for Noncontact Screening ofSleep Apnea Syndrome

Maksim Alekhin1 Lesya Anishchenko1 Alexander Tataraidze12 Sergey Ivashov1

Vladimir Parashin2 Lyudmila Korostovtseva3 Yurii Sviryaev3 and Alexey Bogomolov4

1 Remote Sensing Laboratory Bauman Moscow State Technical University Moscow 105005 Russia2 Department of Biomedical Engineering Bauman Moscow State Technical University Moscow 105005 Russia3 Sleep Laboratory Almazov Federal Heart Blood and Endocrinology Centre Saint Petersburg 197341 Russia4 State Research and Testing Institute of Military Medicine of the Ministry of Defense of Russia Moscow 127083 Russia

Correspondence should be addressed to Maksim Alekhin alexxxmakcnarodru

Received 10 March 2013 Accepted 9 June 2013

Academic Editor Francesco Soldovieri

Copyright copy 2013 Maksim Alekhin et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited

Anovelmethod for recognition of breathing patterns of bioradiolocation signals breathing patterns (BSBP) in the task of noncontactscreening of sleep apnea syndrome (SAS) is proposed and implemented on the base of wavelet transform (WT) and neuralnetwork (NNW) applications Selection of the optimal parameters of WT includes determination of the proper level of waveletdecomposition and the best basis for feature extraction using modified entropy criterion Selection of the optimal properties ofNNWincludes defining the best number of hidden neurons and learning algorithm for the chosenNNWtopologyThe effectivenessof the proposed approach is tested on clinically verified database of BRL signals corresponding to the three classes of breathingpatterns obstructive sleep apnea (OSA) central sleep apnea (CSA) normal calm sleeping (NCS)without sleep-disordered breathing(SDB) episodes

1 Introduction

One of the critical areas of sleep medicine is implementationof novel technical approaches for remote vital signs monitor-ing [1] particularly in screening of sleep disordered breathing(SDB) which is character for sleep apnea syndrome (SAS) [2]Early detection and proper classification of SDB episodes arean important aspect of SAS treatment strategy planning andtaking of opportune preventive measures in clinical practice[3]

Bioradiolocation (BRL) [4] is a modern remote sens-ing technique allowing to perform noncontact vital signsmonitoring of living objects [5] on the base of analysisof specific biometric modulation in reflected radiolocationsignal During tidal breathing process the modulation ismostly determined by reciprocating displacements of skinsurface in abdominal and thoracic areas of chest wall due

to periodic contractions of respiratory muscles [6] Thereliability and effectiveness of BRL technology application innoncontact respiratory monitoring [7] and remote screeningof SAS [8] on the base of breathing pattern analysis wereconvincingly demonstrated

Synthesis of intellectual radar data processing systemsfor target tracking localization and recognition is an up todate task of modern cybernetics Methods and algorithmsfor automated recognition of patterns in nonstationary BRLsignals are in demand in complex ergatic and biomedicalsystems [9]

Practically improvement of performance of patternrecognition can be achieved by applying one of the followingapproaches or their combinations firstly by optimizingdecision rules secondly by applying more efficient featureextraction methods [10] Decision rules implemented on thebase of neural networks (NNW) technology are considered to

2 International Journal of Antennas and Propagation

Data preprocessing

(filtering smoothing and Z-normalization)

Feature extraction

(discrete fast wavelet transform)

Classification of patterns

(neural network classifier)

OSA CSA NCS

0 16 32 48 64 80 96 112 128t

I t

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16j

10

08

06

04

02

00

Vj

10

05

minus00

minus05

minus10

Figure 1 Scheme of the proposed method for recognition of BPBS on the base of WT and NNW

be promising in practical applications of pattern recognitiontasks [11] In turn wavelet transform (WT) is a perspectivemathematical apparatus which is especially demanded informing of attribute space of nonstationary multicomponentsignals with noise for which high-frequency components ofshort duration and extended low-frequency components aretypical [12]

The aim of this study is development of a novel methodfor recognition of breathing patterns of bioradiolocationsignals (BPBS) in the task of noncontact screening of SASwith automated selection of optimal parameters of WTand NNW The proposed approach is tested in recognitionof clinically verified BPBS corresponding to the followingclasses obstructive sleep apnea (OSA) central sleep apnea(CSA) normal calm sleeping (NCS) without SDB

2 Structure of the Proposed PatternRecognition Method

The proposed method for recognition of BPBS in noncontactscreening of SAS includes the followingmain steps (Figure 1)

(i) data preprocessing procedure (including high-passand low-pass Butterworth filtering resamplingsmoothing with five-point moving average filter andZ-normalization)

(ii) feature extraction (forming of BPBS attribute vectorsconsisting of mean-squared values of WT detailedcoefficients of each BRL signal quadrature)

(iii) classification of patterns (applying the fixed numberof resulting components of BPBS attribute vectors tothe inputs of NNW classifiermdashRumelhart multilayerperceptron (MLP) [13]mdashas an output comes the valuecorresponding to one of the target classes OSA CSAand NCS)

For feature extraction an attribute space of absolute valuesof WT detailed coefficients for both BRL signal quadratureswas constructed Thereby each component of BPBS attributevector on the chosen wavelet decomposition level can becalculated as follows

119881119895 =radic(119889119876119895 )2+ (119889119868119895)2 (1)

International Journal of Antennas and Propagation 3

Initialization of parameters

Select the optimal decomposition level by OLD criterion

Set a new type of wavelet basis for decomposition

Select a class of patterns

Select a signal pattern for analysis

I-quadrature of BRL signal Q-quadrature of BRL signal

Z-normalization of signals Z-normalization of signals

Wavelet decomposition Wavelet decomposition

Form attribute vectors of selected patterns

Calculate entropy estimates for attribute vectors

Yes Anotherpattern

Form an array of entropyestimates values for

all signal patterns fromthe selected class

No

Yes Anotherbasis

No

Yes Anotherclass

No

Calculate the MEC value for the selected wavelet basis

Form an arrayof MEC criterion valuesfor all types of analyzed

wavelet bases

Select the best wavelet basis by MEC criterion

Obtain the optimal parameters of waveletdecomposition of BRL signal patterns

Calculate an averaged entropy estimate for the class

Form an array of averagedentropy estimates valuesfor all classes at selected

wavelet basis

Figure 2 Scheme of the algorithm for automated selection of the optimal parameters of WT for improvement of BPBS feature extraction

where119881119895 is a resulting component of BPBS attribute vector 119895is current number of a resulting component of BPBS attributevector 119889119876119895 is the detailed wavelet coefficient for119876-quadratureof BRL signal 119889119868119895 is the detailed wavelet coefficient for 119868-quadrature of BRL signal

For improvement of the performance of the proposedBSBP recognition method (Figure 1) procedures for both

automated selection of the optimalWTparameters (Figure 2)and NNW properties (Figure 3) were developed

3 Optimization Criterions

31 Criterion of the Optimal Level of Wavelet DecompositionThe estimate 119891119898 of the upper frequency limit for the range in



Selection of the optimal parameters ofwavelet decomposition for

forming of attribute vectors of patterns

Synthesis of a minimally complex neural network

Add a neuron of the network hidden layer

Yes

No

Yes

No

Repeattraining

Form an array of networkperformance indicators

values for the set numberof hidden neurons

Calculate averaged network performance indicators

Adda hiddenneuron

Form an array of averagednetwork performance

indicators values for eachnumber of hidden neurons

Find the maximum averaged recognition accuracy value

Select the optimal network structure by MRA criterion

Set a new type of network learning algorithm

Training of the network for pattern recognition

Repeattraining

Form an array of networkperformance indicatorsvalues for the set typeof learning algorithm

Define the priority of network performance indicators

Anothertype of learning

algorithm

Find the maximum averaged recognition accuracy

Select the best neural network realization

Obtain the optimal parameters of waveletdecomposition and neural network classifier


Yes

No

Yes

No


indicators values for alltypes of learning algorithms

Figure 3 Scheme of the algorithm for automated selection of the optimal parameters of NNW for improvement of BPBS classification


which most of quasi-stationary signal energy is concentrated(the upper limit of normal breathing frequency range) isconsidered to be known for BRL signals [4] The maximumpossible frequency value for the registered signal is calculatedin accordance withNyquist theorem (is half the sampling rate119891119904) Then the optimal decomposition level (ODL) of WT foran analyzed BRL signal can be calculated from the relation[14]

119871119900 = [log2 (119891119904

2119891119898

)] + 1 = minus [log2 (2119891119898Δ119905)] + 1 (2)

where 119891119898 is the upper limit of frequency band in which themost of signal energy is concentrated 119891119904 is sampling rate Δ119905is sampling period

Thus further decomposition of analyzed BRL signals tothe levels exceeding the threshold of ODL is not effective ascalculated detailed coefficients ofWTwill not be informativein the aspect of effective feature extraction for BPBS attributespace constructing

32 Criterion of the Optimal Basis of Wavelet Transform Forselection of the optimal basis ofWT from the class of orthog-onal wavelets with compact support a modified entropybased criterion (MEC) is proposed which is calculated on thebase of logarithm energy entropy estimation in the task ofclassification of BPBS

119864119900 = minusradic1

119862119873(

119862

sum

119896=1

(

119873

sum

119894=1

(

119870

sum

119895=1

ln(radic(119889119876119895 )2+ (119889119868119895)2))))

997888rarr min(3)

where 119864119900 is estimate of MEC 119862 is number of classes ofpatterns119873 is number of patterns in each class119870 is number ofresulting components in attribute vectors ln(1198892119895) is logarithmenergy entropy estimate 119889119876119895 is the detailed wavelet coefficientfor 119876-quadrature of BRL signal 119889119868119895 is the detailed waveletcoefficient for 119868-quadrature of BRL signal

Selection of the optimal basis of WT for effective BPBSattribute space forming should be performed using meansquared values of detailed wavelet coefficients of each BRLsignal quadrature for MEC calculations

33 Criterion of the Optimal Number of Hidden Neurons Forselection of the optimal number of hidden neurons of MLPfor BPBS recognition the mean recognition accuracy (MRA)criterion is proposed Varying in each NNW operation testthe number of hidden neurons an estimate of MRA iscalculated as follows

119860119900 =1

119872119861(

119872

sum

119898=1

(

119861

sum

119887=1

119862119887)) 997888rarr max (4)

where 119860119900 is estimate of MRA 119872 is number of operationtests 119861 is number of analyzed wavelet basis119862119887 is recognitionaccuracy for analyzed wavelet basis

Selection of the optimal number of hidden neurons ofMLPwith application ofMRA criterion should be performedusing wavelet bases with such ordinal indexes for which theminimal values of MEC are achieved on training data set

4 Optimization Algorithm

41 Structure of the Optimization Algorithm The proposedalgorithm for automated selection of optimal parameters ofWT and NNW for improving the performance of BPBSrecognition consists of the two main steps In the first stepfor informative feature extraction applying WT initially ageneral class of wavelets is defined then a set of wavelet baseswith ordinal indexes for wavelet families from the generalclass is formed the optimal level of wavelet decomposition isdetermined and finally the optimal wavelet basis is selectedon the base of MEC In the second step for improvingNNW operation performance after preliminary estimationof number of hidden neurons such their optimal amount isfound for which the best MRA value is achieved on trainingdata set and then the best NNW learning algorithm isselected

42 Procedure for Selection of the Optimal Parameters ofWavelet Transform The proposed procedure for selection ofthe optimal parameters of WT includes the following mainsteps (Figure 2)

(i) setting initial parameters characterized by the ana-lyzed BRL signal properties (the maximum possiblefrequency of signal sampling rate and others)

(ii) defining general class of wavelets based on desir-able properties for BRL signal processing (orthog-onal wavelets with compact supportmdashproper time-frequency localization and available fast algorithmsfor implementing WT)

(iii) forming a set of bases with ordinal indexes from wa-velet families (Dobeshi Coiflet and Symlet familiesmdashthe properties of orthogonality regularity symme-try and approximating power of scaling functionsdepend on the ordinal index of the wavelet basis)

(iv) calculating the optimal level of wavelet decomposi-tion (estimation of ODL threshold on the base of themaximum possible frequency of the analyzed BRLsignal and sampling rate values)

(v) selecting the optimal wavelet basis (estimation ofMEC on the base of log energy entropy value for thetask of BPBS classification)

43 Procedure for Selection of the Optimal Parameters of Neu-ralNetwork Theproposed procedure for selection of optimalparameters of NNW classifier on the base of MLP withnonlinear activation function includes the following mainsteps (Figure 3)

(i) setting initial parameters characterized by the condi-tions of the classification task (dimensions of input


and output vectors recognition accuracy error val-ues and others)

(ii) setting the number of hidden layers (one layermdashtheminimal complexity availability of wide range of opti-mization and learning algorithms)

(iii) estimating the upper boundary limit of hiddenneurons number (in accordance with Kolmogorov-Hecht-Nielsen theorem [15] for MLP)

(iv) choosing structural optimization method (simpleiterative algorithm of adding extra neurons)

(v) defining the optimal number of hidden neurons(calculation of MRA during several operation tests ofNNW on a training data set)

(vi) selecting the best NNW learning algorithm (on thebase of such performance indicators as recognitionaccuracy time needed computational complexityand others)

5 Experimental Study

For the test and optimization of the proposed methodsand algorithms for automated recognition of BPBS theclinically verified database of BRL signals for subjects withSAS collected during simultaneous registration of full-nightpolysomnography (PSG) acquired in the Sleep Laboratoryof Almazov Federal Heart Blood and Endocrinology Centre(Figure 4) was used [8] The sample included 7 subjects(4 males and 3 females aged 43ndash62 years with body massindex of 216ndash577) depending on severity of SAS 4 severe1 moderate 1 mild 1 normal The full-night PSG recordswere collected with Embla N7000 system SimultaneouslyBioRascan multifrequency BRL system with a continuous-wave signal and step frequency modulation (developed atRemote Sensing Laboratory of Bauman Moscow State Tech-nical University) was used applying the operating frequencyrange 36ndash40GHz The internal clock of BRL and PSGsystems were synchronized for further verification [8]

For forming BPBS attribute vectors both BRL signalquadratures were used according to (1) with the same lengthof 128 counts corresponding to 128 seconds satisfying the rec-ommendations for screening of SAS [2] In correspondencewith (2) for BRL signals at sampling rate of 119891119904 = 100Hz withmaximum breathing frequency [16] not exceeding the valueof 119891119898 = 10Hz the OLD value cames to 119871119900 = 3 providing 16components in the structure of BPBS attribute vectors

The experimental data set included three classes of BPBS(Figure 5) obstructive sleep apnea (OSA) central sleep apnea(CSA) normal calm sleeping without SDB (NCS)

The experimental data set included 240 realizations ofBPBS related to the three classes (OSA CSA and NCS) inthe following proportion

(i) 90 patterns (30 in each class)mdashtraining set

(ii) 30 patterns (10 in each class)mdashvalidation set

(iii) 120 patterns (40 in each class)mdashtest set

Figure 4 Scheme of the experiment on simultaneous registrationof BRL and PSG data during sleep

Table 1 Selection of the optimal wavelet basis for BPBS featureextraction

Wavelet basis Entropy estimates1198641 1198642 1198643 119864119900

DBSH13 minus31840 minus70530 minus35636 minus49188SMLT13 minus32790 minus80104 minus44964 minus56313CFLT5 minus39672 minus71158 minus33537 minus50866Where 1198641 1198642 and 1198643 are averaged entropy estimates for the three classes ofBPBS 119864119900 is MEC estimates

6 Results and Discussion

61 Selection of the Optimal Wavelet Basis The optimal wa-velet basis for constructing BPBS attribute space was selectedfrom orthogonal wavelets with compact support includingthe following ones available in MATLAB [12]

(i) Dobeshimdash16 bases (ordinal indexes from 1 to 16)

(ii) Symletmdash14 bases (ordinal indexes from 4 to 17)

(iii) Coifletmdash5 bases (ordinal indexes from 1 to 5)

For eachwavelet basis BPBS attribute vectorswere formed onthe base of mean-squared values of WT detailed coefficientsof each BRL signal quadrature according to (1) The resultsof the calculations of MEC estimates for the best bases fromeach family are given in Table 1

Thus wavelet basis Symlet 13 should be considered thebest according to MEC but for the analysis of effectivenessand consistency of the proposed entropy criterion itself thebest bases for Dobeshi and Coiflet wavelet families were alsosubsequently considered

62 Selection of the Optimal Number of Hidden NeuronsAfter forming of attribute vectors of BPBS applying selectedwavelet bases (Dobeshi 13 Symlet 13 and Coiflet 5) the train-ing of NNW classifier was 10 times independently performedfor each case varying the number of hidden neurons from 1 to


0 20 40 60 80 100 120 140 180160 200

1005

0minus05

minus10

t (s)

S(t)

ru

CSA

(a)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

OSA

(b)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

NCS

(c)

Figure 5 Typical BRL signals for the three classes of analyzed patterns (OSA CSA and NCS)

Table 2 Selection of the optimal number of hidden neurons forNNW

Number of neurons Recognition accuracy values119860DBSH13 119860SMLT13 119860CFLT5 119860Max 119860119900

1 0360 0455 0479 0479 04312 0561 0620 0562 0620 05813 0732 0675 0746 0746 07184 0730 0793 0734 0793 07525 0789 0823 0756 0823 07896 0768 0818 0808 0818 07987 0799 0842 0814 0842 08188 0821 0825 0815 0825 08209 0836 0850 0836 0850 084110 0813 0848 0833 0848 083111 0820 0839 0812 0839 082412 0821 0831 0825 0831 082613 0836 0823 0827 0836 082914 0788 0820 0818 0820 080915 0824 0782 0794 0824 0800Where119860DBSH13119860SMLT13 and119860CFLT5 are mean BPBS recognition accuracyvalues for corresponding wavelet bases (Dobeshi 13 Symlet 13 and Coiflet 5)at the set number of hidden neurons of NNW 119860Max is the maximum BPBSrecognition accuracy value 119860119900 is MRA estimate

15 MRA values for the three wavelet bases at varied numberof hidden neurons are given in Table 2

Thus according to MRA criterion the optimal numberof hidden neurons of NNW should be considered equal to9 which also corresponds to the upper boundary limit esti-mated from Kolmogorov-Hecht-Nielsen theorem for MLP[15] Herewith the best absolute value of recognition accuracywas also achieved for wavelet basis Symlet 13The estimate (2)of MEC should also be considered effective and consistent asvarying the number of hidden neurons from 1 to 15 at meanrecognition accuracy not less than 75 for each analyzedwavelet the basis ofWTwith the optimalMEC value allowedto achieve absolutely the best value of recognition accuracyvalue

63 Selection of the Best Learning Algorithm Previously fordefining the optimal number of hidden neurons scaled con-jugate gradient algorithm was used for training as it is con-sidered to be the most multipurpose for feed forward NNWtopologies [13] Subsequently to select the best learning algo-rithm the following gradient based variants implemented inMATLAB were considered [11]

(i) first-order gradient (FOG) algorithmsmdashresilientbackpropagation (RPA)

(ii) conjugate gradient (CG) algorithmsmdashFletcher-Reeves (FRA) Polak-Ribiere (PRA) Beale-Powell(BPA) Moller (MA)

(iii) quasi-Newton (QN) algorithmsmdashLevenberg-Mar-quardt (LMA) Broyden-Fletcher-Goldfarb-Shanno(BFGSA) Battiti (BA)

Selection of the best learning algorithm was performed forMLP with one hidden layer with 9 hidden neurons withapplication of wavelet basis Symlet 13 on the 3rd level ofwavelet decomposition in forming of BPBS attribute vectors

Each time after 10 independent NNW operation testson the test data set the following averaged performanceindicators for learning algorithms were calculated meanrecognition accuracy number of epochs needed for trainingand time spent Cross-validation criterion [11] was used todetermine the moment to stop training The results in orderof priority of the algorithms are given in Table 3

Thus LMA learning algorithm should be considered thebest in training of NNW for BPBS recognition in the task ofnoncontact SAS screening providing the mean accuracy notless than 84 with type II error not exceeding 8 for SDBepisodes The best absolute recognition accuracy of 867was also achieved for NNW realization with LMA learningalgorithm (Table 4)

7 Conclusion

A novel method for of BPBS recognition in the task ofnoncontact SAS screening with automated selection of opti-mal parameters of WT and NNW was developed The BPBS


Table 3 Selection of the best NNW learning algorithm

Algorithm Performance indicatorsRank Title Accuracy Epochs Time (ms)I LMA 0842 12 493II MA 0826 41 801III BA 0811 39 936IV RP 0802 29 592V BPA 0798 25 757VI BFGSA 0794 23 811VII PRA 0793 21 683VIII FRA 0754 26 648

Table 4 Efficiency of BPBS recognition

True OSA CSA NCS 120573err

OSA 36 0 3 69CSA 1 35 2 72NCS 6 4 33 208120572err 152 95 104 867

attribute space was formed on the base of mean-squaredvalues of wavelet detailed coefficients of each BRL signalquadrature MLP with one hidden layer and nonlinearactivation function was used as a classifier The proposedmethod was tested on clinically verified database of BRLsignals related to the three classes of BPBS OSA CSA NCS

In accordance with proposed MEC and ODL criterionsbasis Symlet 13 from the general class of wavelets withcompact support on the 3rd level of wavelet decompositionwas considered to be the best one for feature extractionThe estimate of MEC itself should be considered effectiveand consistent Calculation of MRA criterion revealed thatthe optimal number of NNW hidden neurons is equal to9 which also corresponded to the upper boundary limitestimated from Kolmogorov-Hecht-Nielsen theorem LMAshould be considered the best training algorithm for theproposed classifier providing the mean recognition accuracynot less than 84 with type II error not exceeding 8

Acknowledgments

The research was performed in the framework of the ldquoActiveand Passive Microwaves for Security and Subsurface imaging(AMISS)rdquo EU 7th Framework Marie Curie Actions IRSESProject (PIRSES-GA-2010-269157) and supported by thegrants of the Ministry of Education and Science of RussianFederation and Russian Foundation for Basic Research

References

[1] F Soldovieri I Catapano L Crocco L Anishchenko and SIvashov ldquoA feasibility study for life signs monitoring via acontinuous-wave radarrdquo International Journal of Antennas andPropagation vol 2012 Article ID 420178 5 pages 2012

[2] American Sleep Disorders Association Standards of Prac-tice Committee ldquoPractice parameters for the indications for

polysomnography and related proceduresrdquo Sleep vol 20 no 6pp 406ndash422 1997

[3] L Korostovtseva Y Sviryaev N Zvartau A Konradi andA Kalinkin ldquoPrognosis and cardiovascular morbidity andmortality in prospective study of hypertensive patients withobstructive sleep apnea syndrome in St Petersburg RussiardquoMedical Science Monitor vol 17 no 3 pp 146ndash153 2011

[4] A S Bugaev V V Chapursky S I Ivashov V V Razevig AP Sheyko and I A Vasilyev ldquoThrough wall sensing of humanbreathing and heart beating by monochromatic radarrdquo in Pro-ceedings of the 10th International Conference Ground PenetratingRadar (GPR rsquo04) pp 291ndash294 Delft The Netherlands June2004

[5] S Ivashov V Razevig A Sheyko and I Vasilyev ldquoDetection ofhuman breathing and heartbeat by remote radarrdquo in Proceedingsof the Progress in Electromagnetics Research Symposium (PIERSrsquo04) pp 663ndash666 Pisa Italy March 2004

[6] L Liu Z Liu and B E Barrowes ldquoThrough-wall bio-radiolocation with UWB impulse radar observation simula-tion and signal extractionrdquo IEEE Journal of Selected Topics inApplied Earth Observations and Remote Sensing vol 4 no 4pp 791ndash798 2011

[7] M Alekhin L Anishchenko A Tataraidze S Ivashov VParashin andADyachenko ldquoComparison of bio-radiolocationand respiratory plethysmography signals in time and frequencydomains on the base of cross correlation and spectral analysisrdquoInternational Journal of Antennas and Propagation vol 2013Article ID 410692 6 pages 2013

[8] M Alekhin L Anishchenko A Zhuravlev et al ldquoEstimationof information value of diagnostic data obtained by bioradiolo-cation pneumography in non-contact screening of sleep apneasyndromerdquo Biomedical Engineering vol 47 no 2 pp 96ndash992013

[9] S Stergiopoulos Advanced Signal Processing Handbook Theoryand Implementation for Radar Sonar and Medical Imaging RealTime Systems CRC Press New York NY USA 2000

[10] C Bishop Pattern Recognition and Machine Training SpringerNew York NY USA 2006

[11] I Nabney Netlab Algorithms for Pattern Recognition SpringerNew York NY USA 2001

[12] M Weeks Digital Signal Processing Using MATLAB andWavelets Jones amp Bartlett Training Boston Mass USA 2ndedition 2011

[13] S Haykin Neural Networks A Comprehensive FoundationPrentice Hall New York NY USA 2nd edition 1998

[14] R L Kirlin and R M Dizaji ldquoOptimum sampling frequencyin wavelet based signal compressionrdquo in Proceedings of theCanadian Conference on Electrical and Computer Egineering(CCECE rsquo00) pp 426ndash429 May 2000

[15] R Hecht-Nielsen ldquoKolmogorovrsquos Mapping Neural NetworkExistence Theoremrdquo in Proceedings of the 3rd InternationalConference on Neural Networks pp 11ndash14 1987

[16] D A Korchagina M D Alekhin and L N Anishchenko ldquoBio-radiolocation method at chest wall motion analysis during tidalbreathingrdquo in Proceedings of the 7th European Radar Conference(EuRAD rsquo10) pp 475ndash478 Paris France October 2010

Submit your manuscripts athttpwwwhindawicom

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2013

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2013Part I


DistributedSensor Networks


ISRN Signal Processing



Mechanical Engineering

Advances in

Modelling amp Simulation in EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2013

Advances inOptoElectronics


Volume 2013

ISRN Sensor Networks


VLSI Design


Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawi Publishing Corporation httpwwwhindawicom Volume 2013

The Scientific World Journal

ISRN Robotics



Antennas andPropagation


ISRN Electronics



thinspJournalthinspofthinsp

Sensors


Active and Passive Electronic Components

Chemical EngineeringInternational Journal of



Electrical and Computer Engineering

Journal of

ISRN Civil Engineering


Advances inAcoustics ampVibration



Data preprocessing

(filtering smoothing and Z-normalization)

Feature extraction

(discrete fast wavelet transform)

Classification of patterns

(neural network classifier)

OSA CSA NCS

0 16 32 48 64 80 96 112 128t

I t

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16j

10

08

06

04

02

00

Vj

10

05

minus00

minus05

minus10

Figure 1 Scheme of the proposed method for recognition of BPBS on the base of WT and NNW

be promising in practical applications of pattern recognitiontasks [11] In turn wavelet transform (WT) is a perspectivemathematical apparatus which is especially demanded informing of attribute space of nonstationary multicomponentsignals with noise for which high-frequency components ofshort duration and extended low-frequency components aretypical [12]

The aim of this study is development of a novel methodfor recognition of breathing patterns of bioradiolocationsignals (BPBS) in the task of noncontact screening of SASwith automated selection of optimal parameters of WTand NNW The proposed approach is tested in recognitionof clinically verified BPBS corresponding to the followingclasses obstructive sleep apnea (OSA) central sleep apnea(CSA) normal calm sleeping (NCS) without SDB

2 Structure of the Proposed PatternRecognition Method

The proposed method for recognition of BPBS in noncontactscreening of SAS includes the followingmain steps (Figure 1)

(i) data preprocessing procedure (including high-passand low-pass Butterworth filtering resamplingsmoothing with five-point moving average filter andZ-normalization)

(ii) feature extraction (forming of BPBS attribute vectorsconsisting of mean-squared values of WT detailedcoefficients of each BRL signal quadrature)

(iii) classification of patterns (applying the fixed numberof resulting components of BPBS attribute vectors tothe inputs of NNW classifiermdashRumelhart multilayerperceptron (MLP) [13]mdashas an output comes the valuecorresponding to one of the target classes OSA CSAand NCS)

For feature extraction an attribute space of absolute valuesof WT detailed coefficients for both BRL signal quadratureswas constructed Thereby each component of BPBS attributevector on the chosen wavelet decomposition level can becalculated as follows

119881119895 =radic(119889119876119895 )2+ (119889119868119895)2 (1)












Yes Anotherpattern



No

Yes Anotherbasis

No

Yes Anotherclass

No



wavelet bases





wavelet basis













Yes

No

Yes

No

Repeattraining




Adda hiddenneuron







Repeattraining




algorithm





Yes

No

Yes

No






119871119900 = [log2 (119891119904

2119891119898

)] + 1 = minus [log2 (2119891119898Δ119905)] + 1 (2)




119864119900 = minusradic1

119862119873(

119862

sum

119896=1

(

119873

sum

119894=1

(

119870

sum

119895=1

ln(radic(119889119876119895 )2+ (119889119868119895)2))))

997888rarr min(3)




119860119900 =1

119872119861(

119872

sum

119898=1

(

119861

sum

119887=1

119862119887)) 997888rarr max (4)









































0 20 40 60 80 100 120 140 180160 200

1005

0minus05

minus10

t (s)

S(t)

ru

CSA

(a)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

OSA

(b)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

NCS

(c)














7 Conclusion







OSA 36 0 3 69CSA 1 35 2 72NCS 6 4 33 208120572err 152 95 104 867



Acknowledgments


References




















Journal of



RotatingMachinery


Volume 2013Part I








Advances in




Volume 2013



VLSI Design




ISRN Robotics





ISRN Electronics




Sensors







Journal of
















Yes Anotherpattern



No

Yes Anotherbasis

No

Yes Anotherclass

No



wavelet bases





wavelet basis













Yes

No

Yes

No

Repeattraining




Adda hiddenneuron







Repeattraining




algorithm





Yes

No

Yes

No






119871119900 = [log2 (119891119904

2119891119898

)] + 1 = minus [log2 (2119891119898Δ119905)] + 1 (2)




119864119900 = minusradic1

119862119873(

119862

sum

119896=1

(

119873

sum

119894=1

(

119870

sum

119895=1

ln(radic(119889119876119895 )2+ (119889119868119895)2))))

997888rarr min(3)




119860119900 =1

119872119861(

119872

sum

119898=1

(

119861

sum

119887=1

119862119887)) 997888rarr max (4)









































0 20 40 60 80 100 120 140 180160 200

1005

0minus05

minus10

t (s)

S(t)

ru

CSA

(a)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

OSA

(b)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

NCS

(c)














7 Conclusion







OSA 36 0 3 69CSA 1 35 2 72NCS 6 4 33 208120572err 152 95 104 867



Acknowledgments


References




















Journal of



RotatingMachinery


Volume 2013Part I








Advances in




Volume 2013



VLSI Design




ISRN Robotics





ISRN Electronics




Sensors







Journal of











Yes

No

Yes

No

Repeattraining




Adda hiddenneuron







Repeattraining




algorithm





Yes

No

Yes

No






119871119900 = [log2 (119891119904

2119891119898

)] + 1 = minus [log2 (2119891119898Δ119905)] + 1 (2)




119864119900 = minusradic1

119862119873(

119862

sum

119896=1

(

119873

sum

119894=1

(

119870

sum

119895=1

ln(radic(119889119876119895 )2+ (119889119868119895)2))))

997888rarr min(3)




119860119900 =1

119872119861(

119872

sum

119898=1

(

119861

sum

119887=1

119862119887)) 997888rarr max (4)









































0 20 40 60 80 100 120 140 180160 200

1005

0minus05

minus10

t (s)

S(t)

ru

CSA

(a)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

OSA

(b)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

NCS

(c)














7 Conclusion







OSA 36 0 3 69CSA 1 35 2 72NCS 6 4 33 208120572err 152 95 104 867



Acknowledgments


References




















Journal of



RotatingMachinery


Volume 2013Part I








Advances in




Volume 2013



VLSI Design




ISRN Robotics





ISRN Electronics




Sensors







Journal of







119871119900 = [log2 (119891119904

2119891119898

)] + 1 = minus [log2 (2119891119898Δ119905)] + 1 (2)




119864119900 = minusradic1

119862119873(

119862

sum

119896=1

(

119873

sum

119894=1

(

119870

sum

119895=1

ln(radic(119889119876119895 )2+ (119889119868119895)2))))

997888rarr min(3)




119860119900 =1

119872119861(

119872

sum

119898=1

(

119861

sum

119887=1

119862119887)) 997888rarr max (4)









































0 20 40 60 80 100 120 140 180160 200

1005

0minus05

minus10

t (s)

S(t)

ru

CSA

(a)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

OSA

(b)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

NCS

(c)














7 Conclusion







OSA 36 0 3 69CSA 1 35 2 72NCS 6 4 33 208120572err 152 95 104 867



Acknowledgments


References




















Journal of



RotatingMachinery


Volume 2013Part I








Advances in




Volume 2013



VLSI Design




ISRN Robotics





ISRN Electronics




Sensors







Journal of

































0 20 40 60 80 100 120 140 180160 200

1005

0minus05

minus10

t (s)

S(t)

ru

CSA

(a)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

OSA

(b)

1005

0minus05

minus100 20 40 60 80 100 120 140 180160 200

t (s)

S(t)

ru

NCS

(c)














7 Conclusion







OSA 36 0 3 69CSA 1 35 2 72NCS 6 4 33 208120572err 152 95 104 867



Acknowledgments


References




















Journal of



RotatingMachinery


Volume 2013Part I








Advances in




Volume 2013



VLSI Design




ISRN Robotics





ISRN Electronics




Sensors







Journal of










OSA 36 0 3 69CSA 1 35 2 72NCS 6 4 33 208120572err 152 95 104 867



Acknowledgments


References




















Journal of



RotatingMachinery


Volume 2013Part I








Advances in




Volume 2013



VLSI Design




ISRN Robotics





ISRN Electronics




Sensors







Journal of







Journal of



RotatingMachinery


Volume 2013Part I








Advances in




Volume 2013



VLSI Design




ISRN Robotics





ISRN Electronics




Sensors







Journal of





A Novel Method for Recognition of Bioradiolocation Signal Breathing Patterns for Noncontact...

Documents

Transcript of A Novel Method for Recognition of Bioradiolocation Signal Breathing Patterns for Noncontact...