Research Article Effect of Wireless Channels on Detection...

Research ArticleEffect of Wireless Channels on Detection and Classification ofAsthma Attacks in Wireless Remote Health Monitoring Systems

Orobah Al-Momani and Khaled M Gharaibeh

Yarmouk University Irbid 21163 Jordan

Correspondence should be addressed to Khaled M Gharaibeh kmgharaiyuedujo

Received 9 September 2013 Revised 22 December 2013 Accepted 26 December 2013 Published 10 February 2014

Academic Editor Velio Macellari

Copyright copy 2014 O Al-Momani and K M Gharaibeh This is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

This paper aims to study the performance of support vector machine (SVM) classification in detecting asthma attacks in a wirelessremotemonitoring scenarioThe effect ofwireless channels on decisionmaking of the SVMclassifier is studied in order to determinethe channel conditions under which transmission is not recommended from a clinical point of view The simulation results showthat the performance of the SVM classification algorithm in detecting asthma attacks is highly influenced by themobility of the userwhere Doppler effects are manifested The results also show that SVM classifiers outperform other methods used for classificationof cough signals such as the hidden markov model (HMM) based classifier specially when wireless channel impairments areconsidered

1 Introduction

The application of wireless telemedicine in monitoringpatients is seen as a useful and potentially powerful toolto help patients seek medical treatment [1 2] In remotemonitoring systems of asthmatic patients cough signals (aswell as other body vital signs) are collected by sensorsattached to the patient and sent to a PC located in a hospitalusing wireless technology [3ndash7] A classification algorithmis then applied to the received signals in order to decidethe state of the patient Doctor intervention can then bedone by calling the patient to take precautions take a certainmedication or to go to a health care centre for propertreatment

One of the main purposes of such systems is to helppatients take precautions in the case of asthma attacks [4]An asthma attack is a sudden worsening of asthma symptomscaused by an exposure to allergens or irritants such asinhaling dry and cold air and certain allergens such as petspollen dust and smokeThe symptoms of asthma attackmayvary in severity and duration from person to person Themain symptom that indicates an asthma attack is coughingwhich has different frequency and strength from regularcoughing [4] Other signs of Asthma attack include headache

blue colour in skin difficulty in talking and difficulty inbreathing When these signs of asthma attack are noticed apatient should immediately seek medical treatment in orderto prevent severe asthma attacks which may cause death [5]

With wireless transmission channel impairments such asamplitude variations time dispersion and Doppler effectsresult in degradation of the receiver ability to recover thetransmitted signalThismeans that the classification accuracyof medical signals transmitted over wireless channels will bedeteriorated as a result of errors in the received data Unlikevoice data medical signals are more influenced by channelimpairments because these signals have low bandwidths andare very sensitive to channel impairments [8] Therefore itis important to relate the classification accuracy and hencediagnosability of body signs to channel parameters in orderto identify the channel conditions under which transmissionis not recommended from a clinical point of view

This problem has been investigated in the literature wherethe objective was to determine the transmission conditionswhere diagnosability of received signals is possible [8ndash11]For example it was shown in [8] that successful transmissionof ECG is highly influenced by the mobility of the trans-mitter where it was shown that the receiver BER exceedsthe acceptable limit required for correct classification and

Hindawi Publishing CorporationInternational Journal of Telemedicine and ApplicationsVolume 2014 Article ID 816369 8 pageshttpdxdoiorg1011552014816369

2 International Journal of Telemedicine and Applications

diagnosability of ECG signals at speeds above 50KmHrIn [9] a wireless telemedicine system for transmission ofphotoplethysmography (PPG) signals was analyzed where itwas shown that successful transmission of medical data overthe tested channels requires a receiver bit error rate of lessthan 10minus7

In this paper support vector machines (SVMrsquos) are usedfor classification of cough signals transmitted through a wire-less channel in order to make decision about the occurrenceof asthma attacks of asthmatic patients who are free to moveSVM classifiers have extensively been used in classificationof speech-like signals and were shown to provide higherclassification capability for classification of cough than othertechniques [12 13]

The performance of the SVM classification algorithmin detection of asthma attacks is evaluated by computingthe probability of correct classification at different channelconditions (channel models) and is compared to the per-formance of hidden markov model (HMM) classifier [14]The classification accuracy is related to the channel SNR inorder to determine the channel conditions under which theclassifier produces acceptable results for detection of asthmaattacks

2 Wireless Remote Health MonitoringSystem for Asthma Patients

Figure 1 shows a system model for a wireless remote healthmonitoring system which can be used for detection ofasthma attacks of patients who are free to move In thissystem cough data are captured by a microphone and thentransmitted through a wireless channel At the receiver sidethe received signal is demodulated and then entered into afeature extraction block Extracted features of the receivedsignal are used by the classification algorithm which makesa decision about the existence of an asthma attack

One of the main challenges in the design of wire-less remote health monitoring system is the reliability ofthe communication channel [13] Communication channelsintroduce impairments to the transmitted signal which resultin the inability of the receiver to recover the original signalcorrectly Hence errors introduced by the wireless channelimpact signal classification and result in wrong interpretationof medical data

There are two main types of signal degradation intro-duced by wireless channels the first is attenuation andrandom variation of signal amplitude and the second isdistortion of the signal spectrum Signal attenuation resultsfrom the degradation of the signal power level over distancewhile random variation of signal amplitude results fromchannel noise and multipath fading effects Noise effects aremodelled by additive white Gaussian noise (AWGN) with apower spectral density that depends on the channel signal-to-noise ratio (SNR)

With AWGN channel model white Gaussian noise isadded to the transmitted signal based on a specified SNRtherefore the received signal can be expressed as

119903 (119905) = 119904 (119905) + 119899 (119905) (1)

where 119904(119905) is transmitted signal and 119899(119905) is a noise signalThe noise signal is assumed to be statistically independentof 119904(119905) stationary Gaussian noise process with zero meanand two-sided PSD of 119873

0

2WattsHz The AWGN model isparticularly simple to use in the detection of signals and in thedesign of optimum receiver in most communication systems

Random variations of signal amplitude due to multipathfading effects are usually modelled by the Rayleigh channelmodel [14 15] In a Rayleigh channel model signals fromdifferent paths having different phases and similar signalstrengths are received to produce a Rayleigh distributedsignal amplitude The received signal from a Rayleigh fadingchannel is modelled as [13 15]

119903 (119905) = ℎ (119905) 119904 (119905) + 119899 (119905) (2)

where ℎ(119905) is the fading amplitude which has a Rayleighprobability density function (PDF) given by

119891ℎ

(ℎ) =

ℎ

1205902

119890minusℎ

22120590

2

ℎ ge 0 (3)

where 1205902 is the average received power and ℎ is the signalmagnitude

Distortion of signal spectrum is usually attributed totwo main phenomena the first is multipath delay andthe second is the Doppler effects Multipath delay resultsfrom multipath phenomena in wireless channels where thereceived signal consists of multiple reflected signal compo-nents which have different delays Delay distortion resultsin intersymbol interference (ISI) where successive symbolsinterfere and have significant effects on transmitted signalswhen their bandwidth is larger than the coherence bandwidthof the fading channel (frequency selective channels) [16]TheDoppler spread is a measure a spectral broadening causedby time varying nature of the channel Doppler spread isusually related to the mobility of either the receiver or thetransmitter and is defined as the frequency range in whichthe frequency of the received signal changes due to Dopplereffects Doppler shift leads to signal distortion which causeserrors in the received signalTheDoppler power spectrum fora narrowband fast fading channel is modelled as [16]

119878 (119891) =

1

120587119891119863

radic1 minus (119891119891119863

)2

10038161003816100381610038161198911003816100381610038161003816le 119891119863

(4)

where 119891119863

is the maximum Doppler shift introduced by thechannel which is linearly related to the speed of either thetransmitter or the receiver The Doppler spectrum specifiesthe frequency spectrum of the Rayleigh fading signal andhence its autocorrelation function

Since medical signals usually have smaller bandwidthsthan the channel coherence bandwidth the delay spreadcan be neglected On the other hand small variations insignal spectrum due to Doppler spread can cause significantdistortion to the transmitted signal when the signal band-width is small Therefore in the simulations that will followdelay spread will be neglected and the performance of theclassification system will be tested under different values ofthe Doppler spread

International Journal of Telemedicine and Applications 3

Microphone (cough signal)

Remote PC feature

extraction and (classification)

(digitization

and modulation)

(demodulation) Channel

AWGN and Rayleigh

Wireless RxWireless Tx

Figure 1 Block diagram of a wireless monitoring system for asthma attack detection

3 Support Vector Machine (SVM)Classification

SVM classifiers have extensively been used in classificationof speech-like signals and were shown to provide higherclassification capability for classification of cough than othertechniques [12]

SVM is a classification algorithm that performs a classi-fication task by constructing a hyperplane in a multidimen-sional space that separates data into two different categoriesAn SVM classifier consists of 119871 training points where eachinput x

119894

has 119863 dimensions and belongs to one of two classes1198671

and 1198672

which correspond to 119910119894

= minus1 or +1 where 119910119894

denotes the classifier output Training data can be representedin the form x

119894

119910119894

where 119894 = 1 2 119871 and x isin 119877119863 [17]If the data is linearly separable then classification of a datapoint x is done bymaximizing themargin separating the twohyperplanes which results in a decision rule of the form [18]

Decide Class 1 if sign (119871

sum

119894=1

120572119894

x119894

119910119894

sdot x + 119887 ) = 1

Decide Class 2 if sign(119871

sum

119894=1

120572119894

x119894

119910119894

sdot x + 119887 ) = minus1

(5)

where sign(sdot) is the sign function and 120572119894

are Lagrangemultipliers and are referred to as the support vectors (SV)[18]

When data is not linearly inseparable a kernel function isused to map data to higher dimension such that the resultingdata is linearly separable [18] The original formulation ofthe SVM classifier remains the same except that every dotproduct in (5) is replaced by a nonlinear kernel functionTherefore with using a kernel function the decision rule in(5) is reformulated as [18]


sum

119894=1

120572119894

119910119894

119870(x119894

x) + 119887 ) = 1


sum

119894=1

120572119894

119910119894

119870(x119894

x) + 119887 ) = minus1

(6)

where 119870(x119894

x) is the kernel function which has differentforms depending on the application The most commonkernel functions are the polynomial the radial basis function(RBF) and the sigmoid kernel functions [19]

The parameters of the SVM classification algorithm(including the parameters of the kernel function) can befound using cross-validation of the available training datawhere the parameters that give the best classification accuracyare selected [17]

Tree binaryclassifier

Feature extraction(MFCC)

Training algorithm

Database(classes)

Test signal

Feature extraction (MFCC)

Training data (sound signals)

C1 C2 Cnmiddot middot middot

SpeechRegular coughAsthma attackNormal asthmaArtifacts

Figure 2 Classifier model

Frameblocking

Windowing FFT Coughsignal

MFCC Mel-ScaleDCT Log (∙)

Figure 3 Block diagram of feature extraction process using MFCC[20]

4 Asthma Attack Detection UsingSVM Classification

Figure 3 shows the SVM classification model used in thispaper for detection of asthma attacks The classification pro-cess involves extraction of features generation of a classifierdata base from a training set and decisionmaking for a giventest signal

41 Feature Extraction In the classification model inFigure 2 features of cough signals are extracted and thenused by the classification algorithm in order to reducethe dimensionality of the classification problem Featureextraction can be done using different methods such asFourier transform and Wavelet transform-based methods[21]

One of the common Fourier transform-based methods isthe Mel Frequency Cepstral Coefficient (MFCC) techniquewhich has been used extensively in feature extraction ofspeech-like signals [22] MFCC is a type of parametricrepresentation of speech-like signals and has been extensivelyused in speech recognition because of its ability to capturerelevant information of human speech [20] MFCC featureextraction process is known for its robustness against time


varying nature of speech signals which is a desired featurewhen processing cough data from asthmatic patients

MFCCs are the coefficients that represent the short-termpower spectrum of a signal obtained from linear cosinetransform of the log power spectrum on a nonlinear ldquoMelrdquoscale of frequency [22] The process of calculating MFCCs isillustrated in Figure 3 Digitized cough data is first processedby the framing block which performs segmentation of thecough signal samples into 119873 frames Windowing is appliedto the resulting frames before being spectrally analyzed usingthe fast Fourier transform (FFT) The power spectrum ismapped onto the Mel scale using the approximation [20]

119891119898

= 2595 log10

(1 +

119891

700

) (7)

where 119891 (Hz) is the normal frequency and 119891119898

is the Melfrequency MFCCs are calculated from the discrete cosinetransform (DCT) of the log of the signal spectrum which isequivalent to passing the Mel scaled log spectrum through abank of triangular shaped bandpass filters distributed alonga Mel scaled frequency band of interest [14]

42 MultiClass Classification of Cough Signals The initialform of SVM classifier is binary as clear from (1) and (2)where the output of the learned function is either positiveor negative Asthma attack detection requires classificationof cough signals into five classes normal cough coughrelated to asthma attack cough related to a regular asthmaepisode speech and artefacts (sounds from patient dailyactivity including drinking laughing clearing through andcrying) To generate multiclass SVMs from binary SVMsa number of methods have been proposed in the literature[23ndash25] One of the most popular methods is the binary treesupport vectormachine (BTSVM)which combines SVM andbinary trees BTSVM decomposes an N-class problem intoNndash1 subproblem each separating a pair of classes with SVMclassifier BTSVM has a number of characteristics such aslower number of binary classifiers and faster decision speed[23]

Figure 5 shows a tree-based classification process ofcough signals The classifier uses four stages of classificationfor classifying a sound signal into the five classes mentionedabove The training set for the tree classifier is generatedby dividing the database into two disjoint groups at eachclassification stage

43 Kernel Function Selection Given the diversity of thesound signals incorporated in the training set of the classifiertraining data constitutes a linearly inseparable databaseTherefore an RBF kernel function is used as in (2) TheRBF kernel has less hyperparameters and less numericalcomplexity than other kernel functions [17] The RBF kernelfunction is described by [18]

119870(x119894

x119895

) = 119890(minus|x119894minusx119895|

22120590

2)

(8)

The parameter 120590 determines the area of influence of thesupport vector (SV) over the data space Large value of 120590 will

Table 1 MFCC parameters used in feature extraction of coughsignals

Parameter ValueNumber of MFCCcoefficient 20

Window Hamming119908 (119899) = 054 minus 046 cos(2119899(119873 minus 1))

Window length (N) 256Frame size 25msFFT size 256Feature vector dimension 69

allow a SV to have a strong influence over a large area andreduces the SV but reducing the SV results in increased error[17]

5 Simulation and Verification

In the model shown in Figure 6 test signals are transmittedthrough the wireless channel after performing analogue todigital (AD) conversion and digital modulation At thereceiver the signal is converted back to the analogue domainafter digital demodulation and then applied to the classifierto make a decision about the existence of asthma attacks

The performance of the classification system under wire-less channel impairments is quantified by computing theprobability of correct classification (119875

119888

) (the classificationrate) The probability of correct classification (119875

119888

) is definedas the ratio of the number of correctly classified samples andthe total number of the samples [19]

119875119888

() =Numbr of correctly classified samples

Total number of samplestimes 100

(9)

51 Data Collection The classification system was testedusing recordings of cough signals from real asthmaticpatients The recordings were obtained from 18 patients atthe King Abdullah University Hospital (KAUH) in Jordanin order to develop the training set of the classifier (Datawas obtained after getting the consent of patients throughthe hospital procedure and policies) The causes of coughin these patients were as follows 10 patients had coughcaused by asthma and 8 patients had cough caused by anasthma attack All recording were sampled at frequency11025Hz and were divided into samples where each patientproduces at least 5 samples or 6 samples On the other hand144 samples from normal subjects including normal coughspeech and artefacts (drink laughter clearing through andcrying) were obtained from httpwwwfreesoundorg [26]Table 3 shows the five types of sound signals used for trainingthe classifiersThe recordings were digitized at a frequency of11025 samplessecond and encoded at 16 bits per sample

Table 1 shows the parameters used for feature extractionof cough signals using MFCC as discussed in the previous


1000 2000 3000 4000 5000 6000 7000 8000 9000

4

3

2

1

0

0

minus1

minus2

minus3

minus4

times104

(a)

0 200 400 600 800 1000 1200 1400

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

minus6

MFCC

(b)

Figure 4 (a) A cough signal of an asthmatic patient and (b) its MFCC

Table 2 Fading channel parameters

Modulation orderM 8Fading sequence length (N) 180019

120593119899

120579119899

Uniformly distributed randomvariables

120593119899

= 120579119899

= 2lowast pi lowast rand (1119873) minus piDoppler shift 0 10 100HzSNR 0ndash35 dB

Sound signal

Classifier 1

Classifier 2

Artifact

Classifier 3

Speech Cough

Non-Artifact

Normal cough Asthma cough

Classifier 4

Asthma Asthma attack

Figure 5 Tree SVM classifier for detection of asthma attacks

section Figures 4(a) and 4(b) show a cough signal and itsMFCC features

52 Wireless Channel Models Two channel models are con-sidered an AWGN channel model and a Rayleigh fading

Unknown cough from patient

Channel(AWGN or Rayleigh)

Digital modulation (16 QAM)

AD

Digital demodulation

(16 QAM) DA Classifier

Decision

Figure 6 Block diagram of simulation model

channel model With AWGN channel noise is simply addedto the transmitted signal using the Matlab function ldquoawgnrdquofunction where SNR can be specified A Rayleigh fadingchannel model is generated using the sum-of-sinusoidsmethod in which complex Gaussian noise with a powerspectral density (PSD) that is equal to the Doppler powerspectrum in (8) is approximated by a finite sum of weightedsinusoids as [27]

ℎ (119905) = radic2

119872

119872

sum

119899=1

cos (120596119889

119905 cos120572119899

+ 120593119899

)

+ 119895radic2

119872

119872

sum

119899=1

cos (120596119889

119905 sin120572119899

+ 120593119899

)

(10)

where 120572119899

= (2120587119899 minus 120587 minus 120579119899

)4119872 119899 = 1 2 119872 120596119889

is the maximum angular Doppler frequency 120593119899

and 120579119899

are independent random variables uniformly distributed on[minus120587 120587] for all 119899 With this formulation the fading amplitude(|ℎ(119905)|) approximates a Rayleigh random variable with a PDFas in (7)The parameters of the channel model in (10) used inthe simulations which will follow are shown in Table 2


0 2 4 6 8 10 12 14 16 18

SNR

1

095

09

085

08

075

07

065

06

055

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

Figure 7 Probability of correct classification in AWGN channelversus SNR lowastSVM and ∘HMM

Table 3 Sound signals used in generating the training set of theclassifier

Class Number of samplesArtefacts 48Speech 48Normal cough 48Asthma cough 48Asthma attack 48Total 240

In the simulation model shown in Figure 6 test signalsare transmitted through both channel models at differentSNRs and Doppler shifts and the number of correctly clas-sified samples is counted The minimum SNR required foran acceptable 119875

119888

is determined for both channel modelsFurthermore in the case of Rayleigh fading channel themaximum Doppler frequency that results in acceptable clas-sification rate is determined at a given SNR

53 Simulation Results Using the AWGNchannelmodel theaverage classification rate of the SVM classifier was computedat different SNRs and compared to the classification rateobtained from using the hidden markov models approach in[14] Figure 7 shows 119875

119888

versus SNR for the SVM classifier andcompared to that of an HMM classifier

Under a Rayleigh channel model 119875119888

is calculated atdifferent SNRs and different values of the Doppler shift ofthe channel Figures 8(a)ndash8(c) show119875

119888

versus SNR atDopplershifts of 0 10 and 100Hz of the SVM classifier and comparedto the HMM-based classifier

54 Discussion In the case of AWGNchannel the simulationresults show that a maximum 119875

119888

of 90 is obtained for theSVM classifier at SNR = 16 dB and 86 for the HMM-basedclassifier at SNR = 17 dB The results show that the SVMclassifier outperforms the HMM classifier at all SNRs

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

(a)

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

045

(b)

0 5 10 15 20 25 30 35

SNR

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

045

(c)

Figure 8 Probability of correct classification in Rayleigh fadingchannel versus SNR (a) 119891

119863

= 0Hz (b) 119891119863

= 10Hz and (c) 119891119863

=100Hz lowastSVM and ∘HMM


In Rayleigh fading channel the simulation results showthat the SVM classifier outperforms the HMM classifier atall Doppler frequency shifts From Figures 8(a)ndash8(c) themaximum 119875

119888

of both classifiers is obtained at SNRs above33 dB which is about 14 dB higher than the case of using anAWGN channel This means that SNR needs to be increasedin order to overcome the loss in 119875

119888

due to the mobilityof the transmitter In general the simulation results showthat the SVM classifier outperforms the HMM classifierwhen channel impairments are considered The maximumdifference in 119875

119888

between the two classifiers is about 17 at119891119863

= 100Hz and SNR = 15 dB

6 Conclusion

In this paper classification of cough signals in a wirelessremote health monitoring scenario using SVM classificationalgorithm has been analyzed The reliability of the classifica-tion system has been tested under different wireless channelmodels (AWGN and Rayleigh fading) where it has beenshown that channel impairments have a significant effecton the accuracy of the classification system The simulationresults show that SVM classification is capable of detectingasthma attacks from cough signals transmitted through awireless communication channel provided that the channelSNR is increased to overcome the channel impairmentscaused by noise amplitude variations due to fading and themobility of the wireless transmitter at the patientrsquos side Thesimulation results also show that SVM classifier providesbetter classification accuracy than the HMM-based classifierunder the same channel conditions The results presented inthis paper can be used in the design of remote healthmonitor-ing systems for asthmatic patients where system parameterscan be designed considering the minimum requirements forchannel SNR required for achieving the desired classificationaccuracy

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

The authors would like to thank Dr Nedal Shawagfeh andDr Alaa Ayasra from King Abdullah University Hospital(KAUH) for their help in providing the cough data

References

[1] B Woodward R Istepanian and C Richards ldquoDesign of atelemedicine system using a mobile telephonerdquo IEEE Transac-tions on Information Technology in Biomedicine vol 5 no 1 pp13ndash15 2001

[2] A Abidoye N Azeez A Adesina and K Agbele ldquoUsingWearable Sensors for Remote Healthcare Monitoring SystemrdquoJournal of Sensor Technology vol 1 pp 22ndash28 2011

[3] J Finkelestein and R H Friedman ldquoHome asthma tele-monitoring systemrdquo in Proceedings of IEEE Annual Northeast

Bioengineering Conference pp 103ndash104 Storrs Conn USA2000

[4] J Finkelstein and R Friedman ldquoTelemedicine system to sup-port asthma self-managementrdquo in Proceedings of the IEEEEMBS International Conference on Information TechnologyApplications in Biomedicine pp 164ndash167 Arlington Va USA2000

[5] H Chu C Chang Z Hui and J Tsai ldquoA ubiquitous warningsystem for asthma-inducementrdquo in Proceedings of the IEEEInternational Conference on Sensor Networks Ubiquitous andTrustworthy Computing pp 186ndash191 Taichung Taiwan 2006

[6] M Okubo Y Imai T Ishikawa T Hayasaka and S UenoldquoRespiration monitoring for home-care patients of respiratorydiseases with therapeutic aidsrdquo in Proceedings of the of the 4thEuropean Conference of the International Federation for Medicaland Biological Engineering vol 22 pp 1117ndash1120 AntwerpBelgium November 2008

[7] A Bumatay R Chan K Lauher et al ldquoCoupled mobile phoneplatform with peak flow meter enables real-time lung functionassessmentrdquo IEEE Sensors Journal vol 12 no 3 pp 685ndash6912012

[8] R Habib ldquoModeling of GSM mobile telemedical systemrdquo inProceedings of the IEEE Annual Engineering in Medicine andBiology Society Conference pp 926ndash930 Hong Kong 1998

[9] R Istepanian and M Brien ldquoModeling of photo plethysmog-raphy mobile telemedicine systemrdquo in Proceedings of the IEEEInternational Conference EMBS pp 987ndash990 Chicago Ill USA1997

[10] J R Gallego A Hernandez-Solana M Canales J Lafuente AValdovinos and J Fernandez-Navajas ldquoPerformance analysis ofmultiplexed medical data transmission for mobile emergencycare over theUMTS channelrdquo IEEETransactions on InformationTechnology in Biomedicine vol 9 no 1 pp 13ndash22 2005

[11] httpwwwwebmdcom[12] J-C Wang J-F Wang C-B Lin K-T Jian and W Kuok

ldquoContent-based audio classification using support vectormachines and independent component analysisrdquo in Proceedingsof the IEEE International Conference on Pattern Recognition(ICPR rsquo06) pp 157ndash160 Hong Kong August 2006

[13] J Chol and M Zhou ldquoRecent advances in wireless sensor net-works for healthmonitoringrdquo International Journal of IntelligentControl and Systems vol 15 pp 49ndash58 2010

[14] M Sergio S S Birring I D Pavord and D H EvansldquoDetection of cough signals in continuous audio recordingsusing hiddenMarkovmodelsrdquo IEEETransactions on BiomedicalEngineering vol 53 no 6 pp 1078ndash1083 2006

[15] B Sklar ldquoRayleigh fading channels in mobile digital communi-cation systemsrdquo IEEE Communications Magazine vol 35 no 7pp 90ndash100 1997

[16] S Popa N Draghiciu and R Reiz ldquoFading types in wirelesscommunications systemsrdquo Journal of Electrical and ElectronicsEngineering vol 1 no 1 pp 233ndash237 2008

[17] C Burges A Tutorial on Support Vector Machines For PatternRecognition Kluwer Academic Publishers Boston Mass USA1998

[18] D Srivstain and L Bhambu ldquoData classification using supportvector machinerdquo Journal of Theoretical and Applied InformationTechnology vol 12 no 1 pp 1960ndash1971 2009

[19] C Hsu C Chang and C Lin ldquoA Practical Guide to SupportVector Classificationrdquo httpwwwcsientuedutw


[20] A Bala A Kumber and N Birla ldquoVoice command recognitionsystem based on MFCC and DWTrdquo International Journal ofEngineering Science and Technology vol 2 pp 7335ndash7342 2010

[21] HChatrzarrinAArcelus RGoubran andFKnoefel ldquoFeatureextraction for the differentiation of dry and wet cough soundsrdquoin Proceedings of the IEEE International Symposium on MedicalMeasurements and Applications (MeMeA rsquo11) pp 162ndash166 BariItaly May 2011

[22] R Hasan M Jamil and G Rabbain ldquoSpeaker identificationusing Mel frequency cepstral coefficientsrdquo in Proceedings ofthe 3rd International Conference on Electrical amp ComputerEngineering (ICECE rsquo04) pp 565ndash568 Dhaka Bangladesh2004

[23] H Hoa T An and T Dat ldquoSemi-supervised tree support vectormachine for online cough recognitionrdquo in Proceedings of theInternational Speech Communication Association (ISCA rsquo11) pp1637ndash1640 Florence Italy 2011

[24] G Sun Z Wang and M Wang ldquoA new multi-classificationmethod based on binary tree support vector machinerdquo inProceedings of the IEEE International Conference on InnovativeComputing Information and Control Liaoning China June2008

[25] M Pal ldquoMulticlass approaches for support vector machinebased land cover classificationrdquo in Proceedings of the 8th AnnualInternational Conference Map India 2005 httparxivorgftparxivpapers080208022411pdf

[26] httpwwwfreesoundorg[27] N Kostov ldquoMobile radio channels modelling in MATLABrdquo

Radio Engineering vol 12 pp 12ndash16 2003

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



diagnosability of ECG signals at speeds above 50KmHrIn [9] a wireless telemedicine system for transmission ofphotoplethysmography (PPG) signals was analyzed where itwas shown that successful transmission of medical data overthe tested channels requires a receiver bit error rate of lessthan 10minus7

In this paper support vector machines (SVMrsquos) are usedfor classification of cough signals transmitted through a wire-less channel in order to make decision about the occurrenceof asthma attacks of asthmatic patients who are free to moveSVM classifiers have extensively been used in classificationof speech-like signals and were shown to provide higherclassification capability for classification of cough than othertechniques [12 13]

The performance of the SVM classification algorithmin detection of asthma attacks is evaluated by computingthe probability of correct classification at different channelconditions (channel models) and is compared to the per-formance of hidden markov model (HMM) classifier [14]The classification accuracy is related to the channel SNR inorder to determine the channel conditions under which theclassifier produces acceptable results for detection of asthmaattacks

2 Wireless Remote Health MonitoringSystem for Asthma Patients

Figure 1 shows a system model for a wireless remote healthmonitoring system which can be used for detection ofasthma attacks of patients who are free to move In thissystem cough data are captured by a microphone and thentransmitted through a wireless channel At the receiver sidethe received signal is demodulated and then entered into afeature extraction block Extracted features of the receivedsignal are used by the classification algorithm which makesa decision about the existence of an asthma attack

One of the main challenges in the design of wire-less remote health monitoring system is the reliability ofthe communication channel [13] Communication channelsintroduce impairments to the transmitted signal which resultin the inability of the receiver to recover the original signalcorrectly Hence errors introduced by the wireless channelimpact signal classification and result in wrong interpretationof medical data

There are two main types of signal degradation intro-duced by wireless channels the first is attenuation andrandom variation of signal amplitude and the second isdistortion of the signal spectrum Signal attenuation resultsfrom the degradation of the signal power level over distancewhile random variation of signal amplitude results fromchannel noise and multipath fading effects Noise effects aremodelled by additive white Gaussian noise (AWGN) with apower spectral density that depends on the channel signal-to-noise ratio (SNR)

With AWGN channel model white Gaussian noise isadded to the transmitted signal based on a specified SNRtherefore the received signal can be expressed as

119903 (119905) = 119904 (119905) + 119899 (119905) (1)

where 119904(119905) is transmitted signal and 119899(119905) is a noise signalThe noise signal is assumed to be statistically independentof 119904(119905) stationary Gaussian noise process with zero meanand two-sided PSD of 119873

0

2WattsHz The AWGN model isparticularly simple to use in the detection of signals and in thedesign of optimum receiver in most communication systems

Random variations of signal amplitude due to multipathfading effects are usually modelled by the Rayleigh channelmodel [14 15] In a Rayleigh channel model signals fromdifferent paths having different phases and similar signalstrengths are received to produce a Rayleigh distributedsignal amplitude The received signal from a Rayleigh fadingchannel is modelled as [13 15]

119903 (119905) = ℎ (119905) 119904 (119905) + 119899 (119905) (2)

where ℎ(119905) is the fading amplitude which has a Rayleighprobability density function (PDF) given by

119891ℎ

(ℎ) =

ℎ

1205902

119890minusℎ

22120590

2

ℎ ge 0 (3)

where 1205902 is the average received power and ℎ is the signalmagnitude

Distortion of signal spectrum is usually attributed totwo main phenomena the first is multipath delay andthe second is the Doppler effects Multipath delay resultsfrom multipath phenomena in wireless channels where thereceived signal consists of multiple reflected signal compo-nents which have different delays Delay distortion resultsin intersymbol interference (ISI) where successive symbolsinterfere and have significant effects on transmitted signalswhen their bandwidth is larger than the coherence bandwidthof the fading channel (frequency selective channels) [16]TheDoppler spread is a measure a spectral broadening causedby time varying nature of the channel Doppler spread isusually related to the mobility of either the receiver or thetransmitter and is defined as the frequency range in whichthe frequency of the received signal changes due to Dopplereffects Doppler shift leads to signal distortion which causeserrors in the received signalTheDoppler power spectrum fora narrowband fast fading channel is modelled as [16]

119878 (119891) =

1

120587119891119863

radic1 minus (119891119891119863

)2

10038161003816100381610038161198911003816100381610038161003816le 119891119863

(4)

where 119891119863

is the maximum Doppler shift introduced by thechannel which is linearly related to the speed of either thetransmitter or the receiver The Doppler spectrum specifiesthe frequency spectrum of the Rayleigh fading signal andhence its autocorrelation function

Since medical signals usually have smaller bandwidthsthan the channel coherence bandwidth the delay spreadcan be neglected On the other hand small variations insignal spectrum due to Doppler spread can cause significantdistortion to the transmitted signal when the signal band-width is small Therefore in the simulations that will followdelay spread will be neglected and the performance of theclassification system will be tested under different values ofthe Doppler spread



Remote PC feature


(digitization

and modulation)


AWGN and Rayleigh






119894


and 1198672


= minus1 or +1 where 119910119894


119894

119910119894



sum

119894=1

120572119894

x119894

119910119894

sdot x + 119887 ) = 1


sum

119894=1

120572119894

x119894

119910119894

sdot x + 119887 ) = minus1

(5)





sum

119894=1

120572119894

119910119894

119870(x119894

x) + 119887 ) = 1


sum

119894=1

120572119894

119910119894

119870(x119894

x) + 119887 ) = minus1

(6)

where 119870(x119894





Training algorithm

Database(classes)

Test signal






Frameblocking











119891119898

= 2595 log10

(1 +

119891

700

) (7)






119870(x119894

x119895

) = 119890(minus|x119894minusx119895|

22120590

2)

(8)










119888


119888


119875119888



(9)




1000 2000 3000 4000 5000 6000 7000 8000 9000

4

3

2

1

0

0

minus1

minus2

minus3

minus4

times104

(a)

0 200 400 600 800 1000 1200 1400

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

minus6

MFCC

(b)




120593119899

120579119899


120593119899

= 120579119899


Sound signal

Classifier 1

Classifier 2

Artifact

Classifier 3

Speech Cough

Non-Artifact


Classifier 4








AD



Decision



ℎ (119905) = radic2

119872

119872

sum

119899=1

cos (120596119889

119905 cos120572119899

+ 120593119899

)

+ 119895radic2

119872

119872

sum

119899=1

cos (120596119889

119905 sin120572119899

+ 120593119899

)

(10)

where 120572119899

= (2120587119899 minus 120587 minus 120579119899

)4119872 119899 = 1 2 119872 120596119889


and 120579119899



0 2 4 6 8 10 12 14 16 18

SNR

1

095

09

085

08

075

07

065

06

055

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate





119888



119888




119888



119888


095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

(a)

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

045

(b)

0 5 10 15 20 25 30 35

SNR

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

045

(c)


119863

= 0Hz (b) 119891119863

= 10Hz and (c) 119891119863




119888


119888


119888



6 Conclusion




Acknowledgments


References
































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Remote PC feature


(digitization

and modulation)


AWGN and Rayleigh






119894


and 1198672


= minus1 or +1 where 119910119894


119894

119910119894



sum

119894=1

120572119894

x119894

119910119894

sdot x + 119887 ) = 1


sum

119894=1

120572119894

x119894

119910119894

sdot x + 119887 ) = minus1

(5)





sum

119894=1

120572119894

119910119894

119870(x119894

x) + 119887 ) = 1


sum

119894=1

120572119894

119910119894

119870(x119894

x) + 119887 ) = minus1

(6)

where 119870(x119894





Training algorithm

Database(classes)

Test signal






Frameblocking











119891119898

= 2595 log10

(1 +

119891

700

) (7)






119870(x119894

x119895

) = 119890(minus|x119894minusx119895|

22120590

2)

(8)










119888


119888


119875119888



(9)




1000 2000 3000 4000 5000 6000 7000 8000 9000

4

3

2

1

0

0

minus1

minus2

minus3

minus4

times104

(a)

0 200 400 600 800 1000 1200 1400

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

minus6

MFCC

(b)




120593119899

120579119899


120593119899

= 120579119899


Sound signal

Classifier 1

Classifier 2

Artifact

Classifier 3

Speech Cough

Non-Artifact


Classifier 4








AD



Decision



ℎ (119905) = radic2

119872

119872

sum

119899=1

cos (120596119889

119905 cos120572119899

+ 120593119899

)

+ 119895radic2

119872

119872

sum

119899=1

cos (120596119889

119905 sin120572119899

+ 120593119899

)

(10)

where 120572119899

= (2120587119899 minus 120587 minus 120579119899

)4119872 119899 = 1 2 119872 120596119889


and 120579119899



0 2 4 6 8 10 12 14 16 18

SNR

1

095

09

085

08

075

07

065

06

055

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate





119888



119888




119888



119888


095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

(a)

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

045

(b)

0 5 10 15 20 25 30 35

SNR

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

045

(c)


119863

= 0Hz (b) 119891119863

= 10Hz and (c) 119891119863




119888


119888


119888



6 Conclusion




Acknowledgments


References
































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












119891119898

= 2595 log10

(1 +

119891

700

) (7)






119870(x119894

x119895

) = 119890(minus|x119894minusx119895|

22120590

2)

(8)










119888


119888


119875119888



(9)




1000 2000 3000 4000 5000 6000 7000 8000 9000

4

3

2

1

0

0

minus1

minus2

minus3

minus4

times104

(a)

0 200 400 600 800 1000 1200 1400

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

minus6

MFCC

(b)




120593119899

120579119899


120593119899

= 120579119899


Sound signal

Classifier 1

Classifier 2

Artifact

Classifier 3

Speech Cough

Non-Artifact


Classifier 4








AD



Decision



ℎ (119905) = radic2

119872

119872

sum

119899=1

cos (120596119889

119905 cos120572119899

+ 120593119899

)

+ 119895radic2

119872

119872

sum

119899=1

cos (120596119889

119905 sin120572119899

+ 120593119899

)

(10)

where 120572119899

= (2120587119899 minus 120587 minus 120579119899

)4119872 119899 = 1 2 119872 120596119889


and 120579119899



0 2 4 6 8 10 12 14 16 18

SNR

1

095

09

085

08

075

07

065

06

055

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate





119888



119888




119888



119888


095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

(a)

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

045

(b)

0 5 10 15 20 25 30 35

SNR

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

045

(c)


119863

= 0Hz (b) 119891119863

= 10Hz and (c) 119891119863




119888


119888


119888



6 Conclusion




Acknowledgments


References
































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










1000 2000 3000 4000 5000 6000 7000 8000 9000

4

3

2

1

0

0

minus1

minus2

minus3

minus4

times104

(a)

0 200 400 600 800 1000 1200 1400

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

minus6

MFCC

(b)




120593119899

120579119899


120593119899

= 120579119899


Sound signal

Classifier 1

Classifier 2

Artifact

Classifier 3

Speech Cough

Non-Artifact


Classifier 4








AD



Decision



ℎ (119905) = radic2

119872

119872

sum

119899=1

cos (120596119889

119905 cos120572119899

+ 120593119899

)

+ 119895radic2

119872

119872

sum

119899=1

cos (120596119889

119905 sin120572119899

+ 120593119899

)

(10)

where 120572119899

= (2120587119899 minus 120587 minus 120579119899

)4119872 119899 = 1 2 119872 120596119889


and 120579119899



0 2 4 6 8 10 12 14 16 18

SNR

1

095

09

085

08

075

07

065

06

055

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate





119888



119888




119888



119888


095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

(a)

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

045

(b)

0 5 10 15 20 25 30 35

SNR

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

045

(c)


119863

= 0Hz (b) 119891119863

= 10Hz and (c) 119891119863




119888


119888


119888



6 Conclusion




Acknowledgments


References
































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0 2 4 6 8 10 12 14 16 18

SNR

1

095

09

085

08

075

07

065

06

055

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate





119888



119888




119888



119888


095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

(a)

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

0 5 10 15 20 25 30 35

SNR

045

(b)

0 5 10 15 20 25 30 35

SNR

095

09

085

08

075

07

065

06

055

05

Prob

abili

ty o

f cor

rect

clas

sifica

tion

rate

045

(c)


119863

= 0Hz (b) 119891119863

= 10Hz and (c) 119891119863




119888


119888


119888



6 Conclusion




Acknowledgments


References
































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











119888


119888


119888



6 Conclusion




Acknowledgments


References
































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation




















RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Research Article Effect of Wireless Channels on Detection...

Documents

Transcript of Research Article Effect of Wireless Channels on Detection...