applied sciences

Article Classification of Sound Signal Using Multiple Features

Yaseen, Gui-Young Son and Soonil Kwon *

Department of Digital Contents, Sejong University, Seoul 05006, Korea; [email protected] (Y.); [email protected] (G.-Y.S.) * Correspondence: [email protected]; Tel.: +82-2-3408-3847

 Received: 10 October 2018; Accepted: 20 November 2018; Published: 22 November 2018 

Abstract: Cardiac disorders are critical and must be diagnosed in the early stage using routine examination with high precision. Cardiac auscultation is a technique to analyze and listen to heart sound using electronic , an electronic stethoscope is a device which provides the digital recording of the heart sound called (PCG). This PCG signal carries useful information about the functionality and status of the heart and hence several signal processing and machine learning technique can be applied to study and diagnose heart disorders. Based on PCG signal, the heart sound signal can be classified to two main categories i.e., normal and abnormal categories. We have created database of 5 categories of heart sound signal (PCG signals) from various sources which contains one normal and 4 are abnormal categories. This study proposes an improved, automatic classification algorithm for cardiac disorder by heart sound signal. We extract features from phonocardiogram signal and then process those features using machine learning techniques for classification. In features extraction, we have used Mel Frequency Cepstral Coefficient (MFCCs) and Discrete Wavelets Transform (DWT) features from the heart sound signal, and for learning and classification we have used support vector machine (SVM), deep neural network (DNN) and centroid displacement based k nearest neighbor. To improve the results and classification accuracy, we have combined MFCCs and DWT features for training and classification using SVM and DWT. From our experiments it has been clear that results can be greatly improved when Mel Frequency Cepstral Coefficient and Discrete Wavelets Transform features are fused together and used for classification via support vector machine, deep neural network and k-neareast neighbor(KNN). The methodology discussed in this paper can be used to diagnose heart disorders in patients up to 97% accuracy. The code and dataset can be accessed at “https://github.com/yaseen21khan/Classification-of-Heart- Sound-Signal-Using-Multiple-Features-/blob/master/README.md”.

Keywords: heart sound signal classification; discrete wavelets transform; mel frequency cepstral coefficient

1. Introduction Cardiovascular system is a perpetual source of data that sanctions soothsaying or distinguishing among cardiovascular diseases. External constrains can lead to cardiac diseases that can cause sudden [1]. Cardiovascular diseases are censorious and must be detected with no time delay [2]. Heart diseases may be identified by elucidating the cardiac sound data. The heart sound signal characteristics may vary with respect to different kinds of heart diseases. A huge difference in the pattern can be found between a normal heart sound signal and abnormal heart sound signal as their PCG signal varies from each other with respect to time, amplitude, intensity, homogeneity, spectral content, etc [3].

Appl. Sci. 2018, 8, 2344; doi:10.3390/app8122344 www.mdpi.com/journal/applsci Appl. Sci. 2018, 8, 2344 2 of 14 Appl. Sci. 2018, 8, x FOR PEER REVIEW 2 of 14

CardiovascularCardiovascular auscultationauscultation isis thethe principalprincipal yetyet simplesimple diagnosingdiagnosing methodmethod usedused toto assessassess andand analyzeanalyze thethe operationoperation andand functionalityfunctionality ofof thethe heart.heart. ItIt isis aa techniquetechnique ofof listeninglistening toto heartheart soundsound withwith stethoscope.stethoscope. TheThe mainmain sourcesource ofof thethe generationgeneration ofof heartheart soundsound isis duedue toto unsteadyunsteady momentmoment ofof bloodblood calledcalled bloodblood turbulence.turbulence. The openingopening andand closingclosing ofof atrioventricularatrioventricular values,values, mitralmitral andand tricuspidtricuspid valuesvalues causescauses differentialdifferential bloodblood pressurepressure and and high high acceleration acceleration and and retardation retardation of of blood blood flow flow[4]. [4]. TheThe auscultationauscultation ofof cardiaccardiac disordersdisorders isis carriedcarried outout usingusing anan electronicelectronic stethoscopestethoscope whichwhich isis indeedindeed costcost effectiveeffective and and non-invasive non-invasive approach approach [3]. [3]. The The digital digital recording recording of heart of sound heart withsound the with help ofthe electronic help of stethoscopeelectronic stethoscope is called PCG is called [4]. PCG [4]. OnceOnce thethe heartheart soundsound isis obtained,obtained, itit cancan bebe classifiedclassified viavia computercomputer aidedaided softwaresoftware techniques,techniques, thesethese techniquestechniques needneed moremore accuratelyaccurately defineddefined heartheart soundsound cyclecycle forfor featurefeature extractionextraction process [[3].3]. ManyMany differentdifferent automated automated classifying classifying approaches approaches have have been used,been includingused, including artificial artificial neural network neural (ANN)network and (ANN) hidden and Markov hidden model Markov (HMM). model Multilayer (HMM). Multilayer perceptron-back perceptron-back propagation propagation (MLP-BP), Wavelet(MLP-BP), coefficients Wavelet andcoefficients neural networks and neural were networks used for were classification used for of classification heart sound signal.of heart But sound they havesignal. often But resultedthey have in often lower resulted accuracy in due lowe tor segmentationaccuracy due errorto segmentation [2]. error [2]. ToTo achieveachieve better better accuracy, accuracy, in in this this work, work, we we propose propose SVM, SVM, centroid centroid displacement displacement based based KNN KNN and DNNand DNN based based classification classification algorithm algorithm using discreteusing di waveletsscrete wavelets transform transform (DWT) and(DWT) (MFCC) and features.(MFCC) Thefeatures. recognition The recognition accuracy canaccuracy be increased can be increased for both noisyfor both and noisy clear and speech clear signals, speechusing signals, discrete using waveletsdiscrete wavelets transform transform fused together fused together with MFCCs with MFCCs features features [5,6]. We[5,6 extracted]. We extracted MFCCs MFCCs features features and discreteand discrete wavelets wavelets transform transform features features for heart for sound heart signalssound (normalsignals and(normal abnormal) and abnormal) and classified and themclassified using them support using vector support machine, vector deep machine, neural deep network neural and network centroid and displacement centroid displacement based KNN, thebased highest KNN, accuracy the highest we achievedaccuracy sowe far achieved is 97.9%. so far is 97.9%. TheThe beginningbeginning ofof thisthis paperpaper describesdescribes somesome backgroundbackground knowledgeknowledge aboutabout heartheart soundsound signalsignal processingprocessing andand categoriescategories ofof heartheart disordersdisorders inin briefbrief detail,detail, afterwardsafterwards thethe methodologymethodology ofof featuresfeatures (MFCCs(MFCCs andand DWT)DWT) andand classifiersclassifiers wewe usedused isis discussed.discussed. InIn thethe experimentexperiment sectionsection wewe talktalk aboutabout classificationclassification ofof features,features, detaildetail aboutabout classifiers,classifiers, thethe toolstools wewe used,used, andand databasedatabase inin detail.detail. InIn thethe end, wewe havehave discussed discussed our our results results and and performance performance evaluation evaluation using using accuracy accuracy and and averaged averaged score F1 score and theand last the paragraph last paragraph concludes concludes the paper. the paper.

2.2. BackgroundBackground TheThe humanhuman heart heart is is comprised comprised of of four four chambers. chambers. Two Two of themof them are are called called atrias atrias and and they they make make the upperthe upper portion portion of the of heart the heart while while remaining remaining two chambers two chambers are called are called ventricles ventricles and they and make they lowermake portionlower portion of heart, of andheart, blood and entersblood theenters heart the through heart through atrias andatrias exit and through exit through ventricles. ventricles. Normal Normal heart soundsheart sounds (Figure (Figure1a is an 1a example is an example of normal of heartnormal sound heart signal) sound are signal) generated are generated by closing by and closing opening and ofopening the valves of the of valves heart. Theof heart. heart The sound heart signals sound produced, signals produced, are directly are relateddirectly to related opening to opening and closing and ofclosing the valves, of the blood valves, flow blood and flow viscosity. and viscosity. Therefore Th duringerefore exercise during or exercise other activities or other whichactivities increases which theincreases ,the heart blood rate, flow blood through flow the through valves isthe increased valves is and increased hence heartand hence sound heart signal’s sound intensity signal’s is increasedintensity is and increased during and during (low shock blood (low flow) blood the sound flow) signal’sthe sound intensity signal’s is intensity decreased is decreased [2]. Figure [2].1f alsoFigure shows 1 f also the shows spectrum the ofspectrum a PCG signal. of a PCG signal.

(a) (b)

Figure 1. Cont. Appl. Sci. 2018, 8, 2344 3 of 14 Appl. Sci. 2018, 8, x FOR PEER REVIEW 3 of 14

(c) (d)

(e) (f)

Figure 1.1. ((aa)) AA normal normal heart heart sound sound signal; signal; (b )( Murmurb) Murmur in in systole (MVP); (MVP); (c) Mitral (c) Mitral Regurgitation Regurgitation (MR); ((MR);d) Mitral (d) Mitral Stenosis(MS); (MS); (e) Aortic (e) Stenosis (AS); ((AS);f) Spectrum (f) Spectrum of a PCG of a signal. PCG signal.

The movements of heart valves generate audible sounds of frequency range lower than 2 kHz, commonly referrefer to to as as “lub-dub” “lub-dub” The The “lub” “lub” is the is firstthe portionfirst portion in the in heart the soundheart signalsound and signal is denoted and is bydenoted (S1), it by is generated(S1), it is generated from the closingfrom the of closing mitral and of mitral tricuspid and valve. tricuspid One valve. complete One cardiac complete cycle cardiac is the heartcycle is sound the heart signal sound wave signal which wave starts which from S1starts and from ends S1 to and the startends of to next the S1start and of itnext is described S1 and it asis onedescribed heartbeat. as one The heartbeat. duration, The pitch, duration, and shape pitch, of and the heartshape sound of the showsheart sound us detail shows about us thedetail different about conditionsthe different of conditions heart [4]. Closing of heart of [4]. mitral Closing valves of is mitral followed valves by closingis followed of tricuspid by closing valve of andtricuspid usually valve the delayand usually between the this delay operation between is this 20 to operation 30 ms. Due is 20 to to the 30 left ms. Due to the contraction left ventricle first, contraction the first, componentthe mitral valve occurs component first followed occurs by tricuspidfirst followed valve by component in the component signal. If the in duration the signal. between If the theseduration two between sound components these two sound is in betweencomponents 100 is to in 200 be ms,tween it is 100 called to 200 a split ms, it and is called its frequency a split and range its liesfrequency from 40 range to 200 lies Hz from and 40 it isHz considered to 200 Hz fataland it if is the considered delay is above fatal 30if the ms delay [7]. “Dub”, is above is generated30 ms [7]. (the“Dub”, second is generated heart sound (the component second heart denoted sound bycompon “”),ent when denoted the aortic by “S2”), vales andwhen pulmonary the aortic valvesvales and are closed.pulmonary S1 is valves usually are of closed. longer S1 time is usually period of and longer lower time frequency period thanand lower S2 which frequency has shorter than S2 duration which buthas highershorter frequencyduration but ranges higher (50 tofrequency 250 Hz). ranges A2 is another(50 to 250 heart Hz). sound A2 is signal another component heart sound generated signal duecomponent to movement generated of Aortic due valveto movement and P2 is of another Aortic sound and signalP2 is componentanother heart generated sound duesignal to movementcomponent ofgenerated pulmonary due valve. to movement of pulmonary is higher valve. compared Aortic to pressure pulmonary is higher pressure compared hence the to A2pulmonary component pressure appears hence before the P2 A2 in thecomponent heart sound appears signal. before The presenceP2 in the of heart cardiac sound diseases signal. such The as pulmonarypresence of stenosiscardiac diseases and atrial such septal as pulmonary defect can bestenos identifiedis and atrial by analyzing septal defect closely can the be split identified between by aorticanalyzing component closely the A2 andsplit pulmonarybetween aortic component component P2 andA2 theirand pulmonary intensities. component During deep P2 emotionaland their andintensities. excitement During state, deep the emotional duration betweenand excitement A2 and state, P2 may the be duration longer than between usual, A2 causing and P2 long may and be widelonger splitting than usual, in between causing the long components. and wide This spli statetting is in considered between fatalthe components. if the delay between This state these is componentconsidered fatal is longer if the than delay 30 between ms [7]. these component is longer than 30 ms [7]. Beside “lub-dub”, some noisy signals may may be present in the called murmurs. Murmurs, bothboth normalnormal and and fatal, fatal, are are continuous continuous vibrations vibrations produced produced due due to the to irregular the irregular flow offlow blood of inblood cardiovascular in cardiovascular system. system. Murmurs Murmurs can be of can two be types of two “normal types murmurs” “normal murmurs” and “abnormal and “abnormal murmurs”, normalmurmurs”, murmurs normal are murmurs usually present are usually in heart present sound signalin heart component sound signal of infants, component children of and infants, adults duringchildren exercise and adults also during in women exercise (during also pregnancy), in women (during this type pregnancy), of murmur this can type be detected of murmur in the can first be heartdetected sound. in the Abnormal first heart murmurs sound. Abnormal are usually murm presenturs in are heart usually patience present which in indicateheart patience a heart which valve defectindicate called a heart stenosis valve (squeezed defect called heart stenosis valves) and(squeezed regurgitation heart (bleedingvalves) and heart regurgitation valves) [7]. Murmurs(bleeding heart valves) [7]. Murmurs are the unusual sound present in heart sound cycle of an abnormal heart sound which indicates some abnormalities. Based on the murmur position in the heart cycle, they Appl. Sci. 2018, 8, 2344 4 of 14 are the unusual sound present in heart sound cycle of an abnormal heart sound which indicates some abnormalities. Based on the murmur position in the heart cycle, they can be called systolic murmurs, diastolic murmurs and [4]. These murmurs sounds and clicks are a key points to identify [7]. Murmurs in the heart sound can be discovered using stethoscope, or phonocardiography. Murmurs are classified as continuous murmurs, systolic murmurs and diastolic murmurs. Systolic murmurs are present during systole and they are generated during the ventricles contraction (ventricular ejection). In the heart sound component they are present between S1 and S2 hence are called as systolic murmurs. Based on their types, these systolic murmurs can be called as either ejection murmurs (, pulmonary stenosis, or aortic stenosis) Figure1e, or regurgitant murmurs (ventricular septal defect, tricuspid regurgitation, mitral regurgitation or ) Figure1c. Diastolic murmurs are created during (after systole), when the ventricles relax. Diastolic murmurs are present between the second and first heart sound portion, this type of murmur is usually due to mitral stenosis (MS) or aortic regurgitation (AR) by as shown in Figure1d. Mitral valve prolapse (MVP) is a disease where the murmur sound is present in between the systole section as shown in Figure1b. The murmur of AR is high pitched and the murmur of AS is low pitched. In mitral regurgitation (MR), systolic component S1 is either soft or buried or absent, and the diastolic component S2 is widely split. In mitral stenosis, the murmur is low pitched and rumbling and is present along the diastolic component. In mitral valve prolapse, the murmur can be found throughout S1compnent. The sound signal of VSD and MR are mostly alike [2,7]. Summarizing the above discussion, we can see that heart sound signal can be acquired using and electronic stethoscope, in each complete cycle of heart sound signal we have S1– intervals, and S4 are rare heart sounds and are not normally audible but can be shown on the graphical recording i.e., phonocardiogram [4]. S3 and S4 intervals are called murmur sound and the heart sound signal which carries murmurs are called abnormal heart sounds and can be classified according to murmur position in the signal. In this study automatic classification of heart sound signal is carried out using five categories, one normal category and four abnormal categories. These abnormal categories are aortic stenosis (AS), mitral regurgitation (MR), mitral stenosis (MS) and MVP (murmur exist in the systole interval). Figure1 shows graphical representation of theses heart sound signals categorically. The classification of heart sound signal can be carried out by several machine learning classifiers available for biomedical signal processing. One of them is SVM which can be used for classification of heart sounds. SVM is machine learning, classification and recognition technique that totally works on statistical learning and theorems, the classification accuracy of SVM is much more efficient than conventional classification techniques [8]. Another classifier that recently gained attention is DNN, DNN acoustic models have high performance in speech processing as well as other bio medical signal processing [9]. In this paper, our study shows that these classifiers (SVM and DNN also centroid displacement based KNN) have good performance for heart sound signal classification. The performance of Convolutional Neural Network is optimized for image classification, so this method is not suitable for the feature set we used in this paper. We also used Recurrent Neural Network in the pilot test, but our DNN method showed better results [10,11].

3. Methodology The heart sound signal contains useful data about the working and health status of the heart, the heart sound signal can be processed using signal processing approach to diagnose various heart diseases before the condition of the heart get worse. Hence, various signal processing techniques can be applied to process the PCG signal. The stages involved in processing and diagnosing the heart sound signal are, the acquisition of heart sound signal, noise removal, sampling the PCG signal at a specific frequency rate, feature extraction, training and classification. As shown in the following Figure2. Appl. Sci. 2018, 8, 2344 5 of 14 Appl. Sci. 2018, 8, x FOR PEER REVIEW 5 of 14

(a) (b)

Figure 2.2. (a) ProposedProposed heartheart soundsound signalsignal classificationclassification algorithmalgorithm using SVM;SVM; ((b)) ProposedProposed heartheart soundsound signalsignal classificationclassification algorithmalgorithm usingusing DNN.DNN.

From heart sound sound signal signal we we extract extract two two different different types types of offeatures: features: MFCCs MFCCs and andDWT. DWT. To Toclassify classify these these features features we weused used SVM, SVM, centroid centroid displacement displacement based based KNN KNN and andDNN DNN as classifiers. as classifiers. For Forperformance performance evaluation evaluation of features of features using using these these classifiers, classifiers, we we carried carried out out our our experiment andand evaluatedevaluated the averaged F1 F1 score score and and accuracy accuracy for for all all of of the the three three features features (MFCCs, (MFCCs, DWT DWT and and a acombination combination of of these) these) using using the the three three clas classifiessifies (SVM, (SVM, DNN DNN and and centroid centroid based based KNN). KNN).

3.1.3.1. Mel Frequency CepstralCepstral CoefficientsCoefficients (MFCCs)(MFCCs) We extractedextracted features features as as a seta set of measured,of measured, non-redundant non-redundant and derivedand derived values values from heartfrom signalheart whichsignal haswhich been has used been in ourused study. in our As study. we used As two we typesused oftwo features, types of MFCCs, features, DWT MFCCs, and the DWT combination and the ofcombination both, MFCCs of both, features MFCCs are enormously features are used enormously in signal processingused in signal and processing recognition. and They recognition. were first calculatedThey were and first used calculated in speech and analysis used byin speech Davis and analysis Mermelstein by Davis in theand 1980’s. Mermelstein Human canin the recognize 1980’s. smallHuman changes can recognize in pitch ofsmall sound changes signal in at pitch lower of frequencies sound signal and at thelower MFCC frequencies scale is linearand the for MFCC those frequenciesscale is linear having for those range frequencies <1 kHz. Thehaving power range aspect <1 kHz. of MFCCs The power is that aspect they of are MFCCs able to is represent that they theare signalable to information represent the efficiently. signal information The MFCCs efficientl featurey. areThe almost MFCCs similar feature to are the almost log filter similar bank to energies the log andfilter include bank energies the mel and scale include which makesthe mel it scale an exact which copy makes on smaller it an exact scale copy of what on smaller humans scale understand, of what thehumans frequency understand, in mel scalethe frequency is given by in Equationmel scale(1): is given by Equation (1): f Mel ( f ) = 2595 log (1+ ) (1) Mel (f) = 2595 log (1 + 700 ) (1) In MFCC algorithm, using Hamming window an audio signal is reshaped to smaller windows In MFCC algorithm, using Hamming window an audio signal is reshaped to smaller windows hence splitting of signal to frames is done. Using Fast Fourier Transform, spectrum is computed for hence splitting of signal to frames is done. Using Fast Fourier Transform, spectrum is computed for each frame, using filter bank each spectrum is being weighted. Finally using Logarithm and Discrete each frame, using filter bank each spectrum is being weighted. Finally using Logarithm and Discrete Cosine Transform, the MFCC vector is computed [12,13]. Compared to other features, MFCCs have Cosine Transform, the MFCC vector is computed [12,13]. Compared to other features, MFCCs have good performance against noisy signal when used in speech signal processing hence can be used in good performance against noisy signal when used in speech signal processing hence can be used in biomedical signal processing [12]. During feature extraction, we resampled each signal frequency biomedical signal processing [12]. During feature extraction, we resampled each signal frequency to to 8 KHz, the length of the each signal’s feature extracted is 19 and the length of frame in each 8 KHz, the length of the each signal’s feature extracted is 19 and the length of frame in each sample is sample is 240 and the step size is 80. The following diagram represents the extraction of MFCCs 240 and the step size is 80. The following diagram represents the extraction of MFCCs features features Figure3a. Figure 3a. Appl.Appl. Sci. Sci.2018 2018, 8,, 8 2344, x FOR PEER REVIEW 6 of6 of 14 14

(a) (b)

FigureFigure 3. 3.(a ()a MFCCs) MFCCs computation; computation; (b ()b Decomposition) Decomposition tree tree of of wavelets wavelets (3rd (3rd level). level).

3.2.3.2. Discrete Discrete Wavelets Wavelets Transform Transform (DWT) (DWT) DWTDWT is is mathematical mathematical tool tool for decomposingfor decomposing data inda ata top in downa top fashion. down fashion. DWT represent DWT represent a function a infunction terms of in a terms rough of overall a rough form, overall and aform, wide and range a wi ofde details. range Despiteof details. of theDespite requirement of the requirement and type ofand function type of i.e., function signals imagesi.e., signals etc. DWTimages offers etc. sublimeDWT offers methodology sublime methodology and technique and for representingtechnique for therepresenting amount of the detail amount present of detail [14]. Wavelets present [14]. perform Wave andlets offerperform scale and based offer analysis scale based for a analysis given data. for a Agiven wide rangedata. A of wide applications range of and appl usageications has been and foundusage forhas these been wavelets found for including these wavelets signal processing, including mathematicssignal processing, and numerical mathematics analysis and and, numerical for its better analysis performance and, for its in signalsbetter performance and image processing in signals it and is consideredimage processing an alternative it is to considered Fast Fourier an Transform alternative as DWT to Fast provide Fourier time-frequency Transform representation as DWT provide [15]. Whentime-frequency there is a need representation for processing [15]. and When analyzing there non is stationarya need for tool, processing DWT can beand used analyzing [16]. Study non showsstationary that discrete tool, DWT wavelets can be transform used [16]. have Study high performanceshows that discrete in speech wavelets signal processing transform sohave far andhigh especiallyperformance when in combined speech signal with MFCCs,processing the so performance far and especially is boosted when further combined [5]. with MFCCs, the performanceDWT can beis boosted defined further as a short [5]. wave which carries its energy and information condensed in time. They areDWT limited can be in spacedefined and as possessing a short wave an average which carries value approaching its energy and to zero.information Each signal condensed contains in twotime. components They are limited i.e., low in frequency space and and possessing high frequency an average components. value approaching Low frequency to zero. components Each signal mostlycontains contain two maximumcomponents information i.e., low frequency part, and highand frequencyhigh frequency components components. convey quality.Low frequency That is whycomponents in DWT estimation mostly contain and details maximum are being information rated. The part, estimation and high is of frequency the high-scale, components low-frequency convey componentsquality. That of theis why signal in and DWT the Detailsestimation are theand low-scale, details are high-frequency being rated. components. The estimation In Figure is of3 b,the ithigh-scale, is shown that low-frequency down sampling components a given signal of canthe be signal divided and into itsthe componentDetails are at levelthe onelow-scale, it has cA1,high-frequency cD1 approximated components. and detailed In Figure components 3b, it is shown and at that level down two sampling it has cA2 a andgiven cD2 signal estimated can be anddivided detailed into components. its component Down at level sampling one it has ensures cA1, cD1 the absenceapproximated of redundancy and detailed in a givencomponents data [17 and]. Mathematically,at level two it DWThas cA2 can and be expressedcD2 estimated in the and following detailed Equation components. (2) as: Down sampling ensures the absence of redundancy in a given data [17].Z ∞Mathematically,1 t DWT− 2xy can be expressed in the following ( ) = ( ) ( ) Equation (2) as: DWT x, y Z t p ϕ x dt (2) −∞ |2x| 2 DWT (x, y) = 𝑍(𝑡) 𝜑 ( ) dt (2) The above equation can be used to extract more useful|| data from audio signal using DWT [18]. In caseThe of DWTabove the equation length ofcan feature be used vector to extract is 24. more useful data from audio signal using DWT [18]. In case of DWT the length of feature vector is 24. 3.3. Support Vector Machine (SVM) 3.3.Its Support working Vector is based Machine on Statistics(SVM) Learning Theory (SLT) and structure risk minimization principal (SRMP), it is considered a unique method in machine learning for signal processing. It has unique Its working is based on Statistics Learning Theory (SLT) and structure risk minimization performance in solving nonlinear learning problems and sample learning problems [19,20]. Based on principal (SRMP), it is considered a unique method in machine learning for signal processing. It has optimization technology SVM is a unique tool in solving machine learning challenges. It has been first unique performance in solving nonlinear learning problems and sample learning problems [19,20]. used by Vapnik in 1990’s and now days they are used in pattern recognition in image processing and Based on optimization technology SVM is a unique tool in solving machine learning challenges. It has been first used by Vapnik in 1990’s and now days they are used in pattern recognition in image Appl. Sci. 2018, 8, 2344 7 of 14 signal processing. Based on SLT and SRMP, it can solve many complex structures, process big data appropriate to over learning [20]. The use of SVM has been started as hyper plane classifier and is useful in a situation where data is required to be separated linearly [21]. d Suppose having sample Si and its sorts Ti, expressed as (Si,Ti), S ∈ R ,Ti ∈ (1, −1), i = 1, ... , N. 2 The d is dimension number of input space. For standard SVM, its classifying border or edge is ||w|| . Making the classifying borer or edge to maximum is equal to making the ||w||2 to minimum. Therefore, the optimization problem of making the classifying edge to maximum can be expressed in the follow quadratic programming problem:

min 1 N wbε J (w, ε) = kwk2 + C ε (3) 2 ∑i=1 i Subject to  T  Ti w ϕ(Si) + b ≥ 1 − εI i = 1, ..., N (4) εi ≥ 0, i = 1, ..., N where εi ≥ 0 (i = 1, ... , N ) is flexible variable, which ensures classification validity under linear non-separable case, and parameter C is a positive real constant which determines penalties to approximation errors, a higher value of C is similar as to assigning a higher penalty to errors. The function ϕ is a non-linear mapping function, by which the non-linear problem can be mapped as linear problem of a high dimension space. In the transforming space, we can get optimal hyper plane. The SVM algorithm can discern right classification to sample by solving above quadratic programming problem [21]. We choose nonlinear kernel (quadratic SVM) model for our experiments, in that case SVM needs another parameter gamma along with C to be optimized for better results. Both of these values greatly affect the classification decision of SVM, larger values of gamma brings classifier closer to training accuracy and lower values make it away from it [22]. Similarly lower values of C are more suitable in bringing decision function closer to training accuracy and vice versa. In our experiments we achieved good training accuracy using C values in range of 0.0003 to 0.0001 and gamma value of log103.

3.4. Deep Neural Network (DNN) DNN is a new AI algorithm for machine learning and has been widely used in many research and developing fields. DNN is inspired bio natural object’s visual mechanism where it have many layers of neural network, it process data from their light carrying organs to center of the brain, replicating that mechanism, in DNN the process is done layer by layer where the information are carried out from one layer to another, sequentially extracting edge features, shape feature, part feature, and eventually forms abstract concept [23]. A DNN is a network of neurons in which they are grouped together in multiple layers, those layers are interconnected together in several sequential layers and neurons accepts the neuron activation from previous layer and calculate weighted sum multiply it with neuron and process the output to next layer and so on. The neurons in each layer of network combined execute a complex nonlinear mapping from the beginning to the last. This mapping causes the network to learn from neuron back propagation techniques using weighted sum of neurons. Figure4a shows a simple neural network where there is an input to collection of many hidden layers and a final output. Appl.Appl. Sci. Sci.2018 2018, 8, ,8 2344, x FOR PEER REVIEW 88 of of 14 14

(a) (b)

FigureFigure 4.4. (a()a Neural) Neural network network composed composed of many of many interrelated interrelated neurons; neurons; (b) Data (collectionb) Data collectionprocedure procedureoutline. outline.

ForFor our our DNN DNN model model we we kept kept batch batch size size equal equal to to 100, 100, three three hidden hidden layers layers and and 2000 2000 epochs epochs for for eacheach training training phase. phase. WeWe inputinput thethe datadata toto layers layers in in chunks chunks and and the the total total numbers numbers of of chunks chunks were were calculatedcalculated from from the the total total number number of of features features length length divided divided by by 40, 40, as as we we selected selected the the chunk chunk size size to to be be 4040 for for optimum optimum results. results. 4. Experiments 4. Experiments We collected and arranged data in the form of database, which consist of two sets of data: We collected and arranged data in the form of database, which consist of two sets of data: normal set and abnormal set, the whole data in database (normal and abnormal) have five categories, normal set and abnormal set, the whole data in database (normal and abnormal) have five one normal category (N) and four abnormal categories, the four abnormal categories were: aortic categories, one normal category (N) and four abnormal categories, the four abnormal categories stenosis (AS) mitral stenosis (MS) mitral regurgitation (MR) mitral valve prolapse (MVP), details of were: aortic stenosis (AS) mitral stenosis (MS) mitral regurgitation (MR) mitral valve prolapse data is shown in Table1, there complete detail (w.r.t to category) were discussed in background section. (MVP), details of data is shown in Table 1, there complete detail (w.r.t to category) were discussed in The total numbers of audio files were 1000 for normal and abnormal categories (200 audio files/per background section. The total numbers of audio files were 1000 for normal and abnormal categories category), the files are in .wav format. Our data collection procedure outline is shown in Figure4b, (200 audio files/per category), the files are in .wav format. Our data collection procedure outline is in which we first selected cardiac disorder category, collected data then applied data standardization shown in Figure 4b, in which we first selected cardiac disorder category, collected data then applied after filtering [AppendixA]. data standardization after filtering [Appendix A]. Table 1. Dataset detail. Table 1. Dataset detail. Classes Number of Samples Per Class Classes Number of Samples Per Class NormalNormal N N200 200 AS AS200 200 MR 200 Abnormal MR 200 Abnormal MS 200 MS MVP200 200 MVP 200 Total 1000 Total 1000

AfterAfter gatheringgathering thethe data,data, the experiments we we perfor performed,med, can can be be categorized categorized into into nine nine types, types, as asdiscussed discussed in in methodology methodology section, section, we we did did first first 3 3ex experimentsperiments using using SVM SVM classifier classifier and and three three types types of offeatures: features: MFCCs, MFCCs, DWT DWT and and MFCCs MFCCs + + DWT DWT (fused (fused fe features),atures), second threethree experimentsexperiments usingusing DNN DNN andand last last 3 3 experiments experiments using using centroid centroid displacement displacement based based KNN, KNN, each each of of these these features features were were trained trained andand classified classified using using each each classifier classifier individually. individually. InIn casecase ofof centroid centroid displacement displacementbased basedKNN KNN we we choosechoose the the value value of of k k = = 2.5 2.5 and and that that worked worked well well for for our our dataset dataset training. training. [ 24[24].]. ForFor training training andand classificationclassification of MFCCs MFCCs featur featureses using using SVM SVM the the parameters parameters and and arguments arguments we weused used while while extracting extracting MFCCs MFCCs features features were were as follows: as follows: sampling sampling frequency frequency 8000 Hz, 8000 feature Hz, feature vector vectorhaving having length length 19 and 19 andframe frame size sizein sample in sample 240 240 with with step step size size of of 80.Similarly, 80.Similarly, forfor trainingtraining andand classificationclassification of of DWT DWT features features using using SVM SVM the parametersthe parameters and argumentsand arguments we used we whileused extractingwhile extracting DWT featuresDWT features were as were follows: as follows: sampling sampling frequency frequency 8000 Hz, 8000 and Hz, length and of length feature of vector feature 24. vector Using 24. fivefold Using crossfivefold validation cross validation technique, technique, we classified we ourclassified data using our data SVM using and DNN.SVM and From DNN. dataset, From 20% dataset, of sample 20% dataof sample in each data class in waseach used class for was testing used for and testing eighty and percent eighty of percent sample of data sample in each data class in each for training, class for Appl. Sci. 2018, 8, 2344 9 of 14 Appl. Sci. 2018, 8, x FOR PEER REVIEW 9 of 14 training,the performance the performance of each machine of each learning machine algorithm learning withalgorithm different with features different were features recorded were with recorded respect withto their respect given to accuracies. their given accuracies. Referring to Figure 55,, thethe datadata trainingtraining phasephase hashas beenbeen shown,shown, wherewhere datadata andand featuresfeatures areare distributed clearly clearly shows shows that that in in case case of ofDWT, DWT, the the data data is spread is spread more more across across its region its region in plan, inplan, as it representsas it represents both bothlower lower and higher and higher frequencies frequencies on its on axes its axeshence hence better better performance performance evaluation. evaluation.

(a) (b)

(c)

Figure 5. (a) MFCCs Features data pattern; ( b) DWT Features data pattern; ( c) DWT plus MFCCs Features data pattern.

Figure 66 showsshows confusionconfusion matricesmatrices forfor featuresfeatures typetype andand thethe trainingtraining model,model, asas explainedexplained inin thethe footer section ofof thisthis figure,figure, a;a; d;d; gg belongsbelongs toto MFCCsMFCCs features, features, b, b, e, e, h h belongs belongs to to DWT DWT features features and and c, c, f, f,i belongsi belongs to to combined combined features features (MFCCs (MFCCs plus plus DWT). DWT). The heart sound signal preprocessing and classification is performed by three different methods and classifiers. In case of any speech and signal processing, study shows that features extracted with wavelets transform have greater performance and when those features are used in combination with MFCCs, the accuracy goes even higher [5,6,25]. Appl. Sci.Appl. 2018 Sci., 8, 2018x FOR, 8, 2344PEER REVIEW 1010 of 14of 14

Figure 6. Cont. Appl. Sci. 2018, 8, 2344 11 of 14 Appl. Sci. 2018, 8, x FOR PEER REVIEW 11 of 14

Figure 6. (a) Confusion matrixFigure for 6. MFCCs(a) Confusion Features matrix and Centroid for MFCCs displacement Features based and KNNCentroid classifier; displacement based KNN classifier; (b) Confusion matrix for DWT Features and Centroid displacement based KNN classifier; (c) Confusion (b) Confusion matrix for DWT Features and Centroid displacement based KNN classifier; (c) matrix for MFCCs plus DWT Features and Centroid displacement based KNN classifier; (d) Confusion Confusion matrix for MFCCs plus DWT Features and Centroid displacement based KNN classifier; matrix for MFCCs Features and DNN classifier; (e) Confusion matrix for DWT Features and DNN classifier; (f) Confusion matrix(d) Confusion for MFCCs plusmatrix DWT for Features MFCCs and Features DNN classifier; and (DNNg) Confusion classifier; matrix (e) Confusion matrix for DWT for MFCCs Features andFeatures SVM classifier; and DNN (h) Confusion classifier; matrix (f) forConfusion DWT Features matrix and for SVM MFCCs classifier; plus DWT Features and DNN (i) Confusion matrix for MFCCsclassifier; plus (g DWT) Confusion Features andmatrix SVM for classifier. MFCCs Features and SVM classifier; (h) Confusion matrix for DWT Features and SVM classifier; (i) Confusion matrix for MFCCs plus DWT Features and SVM 5. Results and Discussion classifier. In our experiments we classified our extracted features from heart sound signal database (including normal and abnormalThe heart category) sound using signal three differentpreprocessing classifiers, and SVM, clas DNNsification and centroid is performed by three different based KNN. Table2 showsmethods the baseline and classifiers. results obtained In case using of MFCCs any speech features and and theirsignal corresponding processing, study shows that features accuracies. In order toextracted improve the with accuracy, wavelets we found transform out some have optimum greater values performance for feature extractionand when those features are used in such as sampling frequencycombination and feature with vector MFCCs, length the on accura whichcy the goes resulting even higher accuracies [5,6,25]. were higher than before and the results are shown in Table3. For better accuracy, new features (DWT) were taken into consideration for our experiments, after experimenting and training our data with DWT we 5. Results and Discussion have achieved results shown in Table4. We can see from Tables3 and4 that there is improvement in the accuracy but we consideredIn our to combineexperiments those we features classified together our and extracte train usingd features SVM and from DNN, heart sound signal database after performing experiments(including and normal finding and some abnormal optimal values category) for sampling using three frequency, different we classifiers, arrived SVM, DNN and centroid at the conclusion thatbased we have KNN. much betterTable output2 shows than usingthe baseline those features result separately,s obtained and using we can MFCCs features and their see the results from Tablecorresponding5. Unfortunately accuracies. DNN couldn’t In order perform to improve well as the compared accuracy, to SVM we found and the out some optimum values for reason for that is we have very limited amount of data available and we know that DNN requires feature extraction such as sampling frequency and feature vector length on which the resulting thousands of files and features for training and classification, so in case if sufficient amount of data accuracies were higher than before and the results are shown in Table 3. For better accuracy, new were available, DNN could have performed higher than SVM . The highest accuracyfeatures we achieved (DWT) from were those taken experiments into consideration was from fused for our (combined) experiments, features after we experimenting and training classified; these combinedour featuresdata with have DWT never we been have used achieved for classification results shown of heart in sound Table signal 4. We before. can see from Table 3 and 4 that Tables3–6 summarizesthere the output, is improvement and classification in the in accuracy terms of accuracybut we consid averagedered f1 to score combine sensitivity those features together and train and specificity of our experimentusing SVM in and detail. DNN, Several after experiments performing were experiment performed withs and MFCCs finding and some DWT optimal values for sampling with variant frequencies,frequency, variant Cepstral we arrived coefficient at the and conclusion their corresponding that we have changing much accuracy better output was than using those features recorded. From our experimentsseparately, Tables and 3we–5 summarizecan see the the results last and from final Table results 5. of Unfortunately MFCCs, DWT and DNN couldn’t perform well as combined features (MFCCscompared and DWT) to SVM respectively. and the reason for that is we have very limited amount of data available and we know that DNN requires thousands of files and features for training and classification, so in case if Table 2. Baseline results with MFCCs features. sufficient amount of data were available, DNN could have performed higher than SVM ClassifierThe highest accuracy we Features achieved from those Accuracyexperiments was from fused (combined) features SVMwe classified; these combined MFCC features have never been 87.2% used for classification of heart sound signal DNNbefore. Tables 3–6 summarizes MFCC the output, and classification 82.3% in terms of accuracy averaged f1 score Centroid displacement base KkNN MFCC 72.2% sensitivity and specificity of our experiment in detail. Several experiments were performed with MFCCs and DWT with variant frequencies, variant Cepstral coefficient and their corresponding changing accuracy was recorded. From our experiments Tables 3–5 summarize the last and final results of MFCCs, DWT and combined features (MFCCs and DWT) respectively.

Table 2. Baseline results with MFCCs features.

Classifier Features Accuracy SVM MFCC 87.2% Appl. Sci. 2018, 8, 2344 12 of 14

Table 3. Improved Result based on MFCCs Features.

KNN MFCCs SVM DNN (Centroid Displacement Based) Accuracy 80.2% 91.6% 86.5% Average F1 Score 92.5% 86.0% 96.9%

Table 4. Result based on DWT Features.

KNN DWT SVM DNN (Centroid Displacement Based) Accuracy 91.8% 92.3% 87.8% Average F1 Score 98.3% 99.1% 95.6%

Table 5. Result based on MFCCs and DWT Features.

KNN MFCCs + DWT SVM DNN (Centroid Displacement Based) Accuracy 97.4% 97.9% 92.1% Average F1 Score 99.2% 99.7% 98.3%

Table 6. Sensitivity and Specificity calculated for each classifier performance based on the Figure6 confusion matrix.

Classifiers Features Sensitivity Specificity MFCCs 81.98% 93.5% Centroid displacement based KNN DWT 92.0% 97.9% MFCCs + DWT 97.6% 98.8% MFCCs 86.8% 95.1% DNN DWT 91.6% 97.4% MFCCs + DWT 94.5% 98.2% MFCCs 87.3% 96.6% SVM DWT 92.3% 98.4% MFCCs + DWT 98.2% 99.4%

Depending on the size of dataset and type of features in the classification experiment different results can be achieved with variant accuracies. Using fivefold cross validation the evaluation results are achieved as shown in Table5. It is observed that the overall maximum accuracy of 97.9% (F1 score of 99.7%) has been achieved using combine features of discrete wavelet transform and MFCCs, and training them via SVM classifier. In case of centroid displacement based KNN the highest accuracy achieved is 97.4% (F1 score of 99.2%). Similarly, in case of DNN, the maximum accuracy of 89.3% has been achieved when we used MFCCs Features combined with DWT features, keeping the parameters for all the features same as mentioned in above paragraph, we trained our deep neural network with features (MFCCs and DWT), in deep neural network we kept batch size of 100, three hidden layers and one output layer and 2000 epochs. In case of DNN too, we can observe from our results table that using combined features of MFCCs and DWT, we can get higher results by efficiently utilizing the information present in the heart sound signal. The improved accuracy is due to usage of completely different domain signals (MFCCs and DWT) together. Appl. Sci. 2018, 8, 2344 13 of 14

6. Conclusions PCG signals carries information about the functioning of heart valves during heartbeat, hence these signals are very important in diagnosing heart problems at early stage. To detect heart problems with great precision, the new features (MFCCs and DWT) were used and have been proposed in this study to detect 4 types of abnormal heart diseases category and one normal category. The results that have been obtained shows clear performance superiority and can help in the early detection of heart disease during auscultation examination. This work can be greatly improved if the dataset is prepared on a larger scale and also if the data features are managed in a more innovative way, then the performance can be more fruitful. Some new features can also be introduced to asses and analyses heart sound signal more accurately.

Author Contributions: The authors contributed equally to this work. Conceptualization, visualization, writing and preparing the manuscript is done by Y., reviewing, supervision, project administration, validation and funding was done by S.K., Software related work, data creation, project administration was the task of G.Y.S. Funding: This research was funded by the Ministry of SMEs and Startups (No. S2442956). The APC was funded by Sejong University. Acknowledgments: This work was supported by the Ministry of SMEs and Startups (No. S2442956), the development of intelligent stethoscope sounds analysis engine and interactive smart stethoscope service platform. Conflicts of Interest: The authors declare no conflict of interest.

Appendix A The data was collected from random sources, such as books (Auscultation skills CD, Heart sound made easy) and websites (48 different websites provided the data including Washington, Texas, 3M, and Michigan and so on). After conducting efficient cross check test, we excluded files having extreme noise, data was sampled to 8000 Hz frequency rate and was converted to mono channel, 3 period heart sound signal, data sampling, conversion and editing was done in cool edit software.

References

1. Bolea, J.; Laguna, P.; Caiani, E.G.; Almeida, R. Heart rate and ventricular repolarization variabilities interactions modification by microgravity simulation during head-down bed rest test. In Proceedings of the 2013 IEEE 26th International Symposium onComputer-Based Medical Systems (CBMS), Porto, Portugal, 22–23 June 2013; IEEE: New York, NY, USA, 2013; pp. 552–553. 2. Kwak, C.; Kwon, O.-W. Cardiac disorder classification by heart sound signals using murmur likelihood and hidden markov model state likelihood. IET Signal Process. 2012, 6, 326–334. [CrossRef] 3. Mondal, A.; Kumar, A.K.; Bhattacharya, P.; Saha, G. Boundary estimation of cardiac events s1 and s2 based on hilbert transform and adaptive thresholding approach. In Proceedings of the 2013 Indian Conference on Medical Informatics and Telemedicine (ICMIT), Kharagpur, India, 28–30 March 2013; IEEE: New York, NY, USA, 2013; pp. 43–47. 4. Randhawa, S.K.; Singh, M. Classification of heart sound signals using multi-modal features. Procedia Computer Science 2015, 58, 165–171. [CrossRef] 5. Anusuya, M.; Katti, S. Comparison of different speech feature extraction techniques with and without wavelet transform to kannada speech recognition. Int. J. Comput. Appl. 2011, 26, 19–24. [CrossRef] 6. Turner, C.; Joseph, A. A wavelet packet and mel-frequency cepstral coefficients-based feature extraction method for speaker identification. Procedia Comput. Sci. 2015, 61, 416–421. [CrossRef] 7. Thiyagaraja, S.R.; Dantu, R.; Shrestha, P.L.; Chitnis, A.; Thompson, M.A.; Anumandla, P.T.; Sarma, T.; Dantu, S. A novel heart-mobile interface for detection and classification of heart sounds. Biomed. Signal Process. Control 2018, 45, 313–324. [CrossRef] 8. Wu, J.-B.; Zhou, S.; Wu, Z.; Wu, X.-M. Research on the method of characteristic extraction and classification of phonocardiogram. In Proceedings of the 2012 International Conference on Systems and Informatics (ICSAI), Yantai, China, 19–20 May 2012; IEEE: New York, NY, USA, 2012; pp. 1732–1735. Appl. Sci. 2018, 8, 2344 14 of 14

9. Maas, A.L.; Qi, P.; Xie, Z.; Hannun, A.Y.; Lengerich, C.T.; Jurafsky, D.; Ng, A.Y. Building dnn acoustic models for large vocabulary speech recognition. Comput. Speech Lang. 2017, 41, 195–213. [CrossRef] 10. Nanni, L.; Aguiar, R.L.; Costa, Y.M.; Brahnam, S.; Silla, C.N., Jr.; Brattin, R.L.; Zhao, Z. Bird and whale species identification using sound images. IET Comput. Vis. 2017, 12, 178–184. [CrossRef] 11. Nanni, L.; Costa, Y.M.; Aguiar, R.L.; Silla, C.N., Jr.; Brahnam, S. Ensemble of deep learning, visual and acoustic features for music genre classification. J. New Music Res. 2018, 1–15. [CrossRef] 12. Nair, A.P.; Krishnan, S.; Saquib, Z. Mfcc based noise reduction in asr using kalman filtering. In Proceedings of the Conference on Advances in Signal Processing (CASP), Pune, India, 9–11 June 2016; IEEE: New York, NY, USA, 2016; pp. 474–478. 13. Almisreb, A.A.; Abidin, A.F.; Tahir, N.M. Comparison of speech features for arabic phonemes recognition system based malay speaker. In Proceedings of the 2014 IEEE Conference on Systems, Process and Control (ICSPC), Kuala Lumpur, Malaysia, 12–14 December 2014; IEEE: New York, NY, USA, 2014; pp. 79–83. 14. DeRose, E.J.S.T.D.; Salesin, D.H. Wavelets for computer graphics: A primer part 1 y. Way 1995, 6, 1. 15. Ghazali, K.H.; Mansor, M.F.; Mustafa, M.M.; Hussain, A. Feature extraction technique using discrete wavelet transform for image classification. In Proceedings of the 5th Student Conference on Research and Development SCOReD 2007, Selangor, Malaysia, 11–12 December 2007; IEEE: New York, NY, USA, 2007; pp. 1–4. 16. Fahmy, M.M. Palmprint recognition based on mel frequency cepstral coefficients feature extraction. Ain Shams Eng. J. 2010, 1, 39–47. [CrossRef] 17. Sharma, R.; Kesarwani, A.; Mathur, P.M. A novel compression algorithm using dwt. In Proceedings of the 2014 Annual IEEE India Conference (INDICON), Pune, India, 11–13 December 2014; IEEE: New York, NY, USA, 2014; pp. 1–4. 18. Li, M.; Chen, W.; Zhang, T. Classification of epilepsy eeg signals using dwt-based envelope analysis and neural network ensemble. Biomed. Signal Process. Control 2017, 31, 357–365. [CrossRef] 19. Zhong, Y.-C.; Li, F. Model identification study on micro robot mobile in liquid based on support vector machine. In Proceedings of the 3rd IEEE International Conference onNano/Micro Engineered and Molecular Systems, NEMS 2008, Sanya, China, 6–9 January 2008; IEEE: New York, NY, USA, 2008; pp. 55–59. 20. Lu, P.; Xu, D.-P.; Liu, Y.-B. Study of fault diagnosis model based on multi-class wavelet support vector machines. In Proceedings of the 2005 International Conference on Machine Learning and Cybernetics, Guangzhou, China, 18–21 August 2005; IEEE: New York, NY, USA; pp. 4319–4321. 21. Yang, K.-H.; Zhao, L.-L. Application of the improved support vector machine on vehicle recognition. In Proceedings of the 2008 International Conference on Machine Learning and Cybernetics, Kunming, China, 12–15 July 2008; IEEE: New York, NY, USA, 2008; pp. 2785–2789. 22. Chapelle, O.; Vapnik, V.; Bousquet, O.; Mukherjee, S. Choosing multiple parameters for support vector machines. Mach. Learn. 2002, 46, 131–159. [CrossRef] 23. Yi, H.; Sun, S.; Duan, X.; Chen, Z. A study on deep neural networks framework. In Proceedings of the 2016 IEEE Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC), Xi’an, China, 3–5 October 2016; IEEE: New York, NY, USA, 2016; pp. 1519–1522. 24. Nguyen, B.P.; Tay, W.-L.; Chui, C.-K. Robust biometric recognition from palm depth images for gloved hands. IEEE Trans. Hum.-Mach. Syst. 2015, 45, 799–804. [CrossRef] 25. Janse, P.V.; Magre, S.B.; Kurzekar, P.K.; Deshmukh, R. A comparative study between mfcc and dwt feature extraction technique. Int. J. Eng. Res. Technol. 2014, 3, 3124–3127.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).