Research Article Detection of Lungs Status Using ...downloads.hindawi.com/journals/tswj/2014/182938.pdf · variation of normal lung sounds. In , Tinkelman et al. [ ] suggested a computer

Research ArticleDetection of Lungs Status Using Morphological Complexities ofRespiratory Sounds

Ashok Mondal1 Parthasarathi Bhattacharya2 and Goutam Saha1

1 Department of Electronics and Electrical Communication Engineering Indian Institute of TechnologyKharagpur Kharagpur 721 302 India

2 Institute of Pulmocare and Research Kolkata Kolkata 700 064 India

Correspondence should be addressed to Ashok Mondal ashokrpegmailcom

Received 31 August 2013 Accepted 17 December 2013 Published 6 February 2014

Academic Editors G R Epler and F Varoli

Copyright copy 2014 Ashok Mondal et alThis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Traditionally the clinical diagnosis of a respiratory disease is made from a careful clinical examination including chest auscultationObjective analysis and automatic interpretation of the lung sound based on its physical characters are strongly warranted to assistclinical practice In this paper a new method is proposed to distinguish between the normal and the abnormal subjects using themorphological complexities of the lung sound signals The morphological embedded complexities used in these experiments havebeen calculated in terms of texture information (lacunarity) irregularity index (sample entropy) third order moment (skewness)and fourth order moment (Kurtosis) These features are extracted from a mixed data set of 10 normal and 20 abnormal subjectsand are analyzed using two different classifiers extreme learning machine (ELM) and support vector machine (SVM) networkTheresults are obtained using 5-fold cross-validation The performance of the proposed method is compared with a wavelet analysisbased method The developed algorithm gives a better accuracy of 9286 and sensitivity of 8630 and specificity of 8690 for acomposite feature vector of four morphological indices

1 Introduction

The audio information of respiratory signals is used to findout the pulmonary dysfunctions The diagnostic status of therespiratory system can be assessed by interpreting the audiblecharacteristics of lung sound signals in terms of varyingamplitude intensity and tone quality or modal frequenciesPhysicians examine the lung disease in two ways one isnoninvasive process which includes auscultation pulmonaryfunction test respiratory inductance plethysmograph andphonopneumography technique and the other is invasiveapproach such as chest X-ray or roentgenogram and com-puterized tomography (CT) scan and so forth The invasivediagnostic procedures are expensive time consuming andharmful as in case of X-ray repetition One of the popularnoninvasive approaches is auscultation a stethoscope devicebased technique started after the invention of stethoscopeby French physician Laennec in 1962 [1] It is a most simpleand inexpensive diagnostic tool and widely used by medicalpractitioners However this cost effective and easily handling

appliance is unable to remove interventions produced fromthe surroundings organs namely cardiac sound source andalso form the clinical environment that leads to misdiagnosisof the respiratory diseases

The nonlinear and nonstationary properties of lungsound signal make it difficult to diagnose the lungs statususing only the temporal or spectral characteristics of therespiratory sounds The lung sound (LS) shows a complexdynamics because of the involvement of transmission pathfiltering effect attenuation and its production mechanismwhich is unstable The pathological status of the lungssignifies a morphological deviation of the normal breathsound The pattern complexities of abnormal LS are higherthan that of the normal LS because of the occurrence ofauxiliary signals in case of unhealthy lungs The doctors takehelp of different clinical devices namely modern electronicstethoscope CT scan bronchoscopy and so forth to capturethe various distinctive parameters of the normal as well asabnormal states of the lungs The accuracy of diagnosiswith these diagnostic tools depends on the experience and

Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 182938 9 pageshttpdxdoiorg1011552014182938

2 The Scientific World Journal

knowledge of the physicians and also on the cooperation ofthe patients

With the advances of computer technologies statisticalsignal processing artificial intelligence and pattern recog-nition algorithms lung sound analysis is commenced in anautomated manner The computer aided or microprocessorbased automated tools offer several facilities in terms ofhigh speed large storage capacity and avoid the manualhazards The important intervening step of automated lungsound analysis is the extraction of authentic features thatare inherently correlated with the lungs conditions The finalstage of pattern recognition based lung sound analysis systemis decision making about the underlying disease if any Theaim of feature extraction procedure is to identify the relevantdistinct parameters of the LS signals and to arrange themin vector form that serves as an input during classificationResearchers have developed several feature extraction tech-niques to form the feature vectors based on parametric andnonparametricmethods In parametric techniquemodels theLS signal based on a priori knowledge is in contrast withnonparametric method which characterizes LS signal usinga set of basis functions [2] Many studies have been donein classifying the lung dysfunctions which suggest a varietyof methods based on time frequency and time-frequencydomain analysis [2ndash10]

The first initiative for lung sounds analysis was taken byForgacs et al in the late 1960s [11] After that a number ofresearch works were published in this field In 1973 MurphyJr and Sorensen described a spectral based technique forwheeze sounds analysis [12] and later on Murphy Jr etal introduced a waveform analysis of crackles sounds [13]Dosani and Kraman presented a technique called phono-pneumography in 1983 [14] for measuring the intensityvariation of normal lung sounds In 1991 Tinkelman et al[15] suggested a computer digitized phonopneumographytechnique to detect the airway obstruction in child patientsA classification method has been developed by Cohen andLandsberg to distinguish between normal and abnormalbreath sounds by computing the Mahalanobis distancebetween the known and unknown classes using their meanfeature vectors and covariance matrix of known data [5]Thisis a semiautomated system because breath cycles are selectedmanually and inapplicable for adventitious sounds Sankuret al proposed a technique to discriminate pathologicalfrom normal subjects using autoregressive coefficients and k-nearest neighbour and quadratic classifier in [2] Gavriely etal [4] have analyzed normal lung sounds and characterizedthese sounds by defined some parameters such as ampli-tude frequency and regression lines slopes correspondingto high and low frequency segments They have suggestedthat these parameters are useful for discrimination betweennormal and pathological sounds A neural network basedclassification technique has been proposed by Yeginer et al[6] for distinguishing between normal and abnormal subjectsusing the subphase features like autoregressive coefficientsprediction error and ratio of expiration and inspirationduration The correct classification performance for a smalldatabase is between 70 and 80 The efficiency of thisprocedure degrades due to improper selection of model

order Zheng et al [7] have introduced an algorithm todifferentiate between normal and abnormal lung soundsusing time and frequency domain features combined witha stochastic classification procedure Matsunaga et al havepresented a maximum likelihood approach based methodin distinguishing the adventitious and normal pulmonarysounds [8] This classification technique uses two types ofacoustic modeling methods one is hidden Markov model(HMM) for detection of adventitious sounds and anotherone is microphone dependent model to identify the normalsounds This is a semiautomated processes because theacoustic segments are labeled manually and its performancedegrades for noisy data

Marshall and Boussakta [9] have developed a classifica-tion technique to classify the normal and crackles soundsusing the wavelet features and cross-correlation mathemat-ical tool A wavelet based classification approach has beendeveloped by Kandaswamy et al [10] for discriminatingsix different types of pulmonary sounds by using statisticalfeatures of thewavelet detail coefficients (D1 toD7) producedby Daubechies wavelet of order 8 and artificial neural net-work Another wavelet based analysis has been studied bySello et al [3] to separate healthy subject from pathologicalsubject by characterizing the frequency power distributionof the wavelet coefficients generated by Morlet wavelet fromrespiratory sounds

The difficulties are faced by researchers in analysis of res-piratory sounds that are the interference of lung sounds withheart sounds appearance of different pathological sounds insimilar forms and also the unavailability of the sophisticatedinstrumentation for processing information embedded in thelung sounds The focus of the study is to develop a newmethod for better analysis of lung sounds by exploring thestatistical approaches anddigital signal processing knowledgecombined with pattern recognition algorithms In this papera new technique is proposed to detect the lungs statusnormal and abnormal using the structural complexitiesof LS signals The structural behavioral of the LS signalis parameterized with a number of distinct features suchas sample entropy lacunarity skewness and kurtosis Atwenty-four-dimensional feature vector is formed using theseparametersThe ELMand SVMnetworks are used to evaluatethe efficiency of the developed technique The proposedtechnique gives better performance than the baselinemethod[3]

The rest of this paper is organized as follows Theoret-ical background information on ELM and SVM classifiersis described in Section 2 The methodology to distinguishbetween normal and abnormal status of the lungs is discussedin detail in Section 3 and the database and implementationplatform that are employed in the work are discussed inSection 4 Section 5 depicts the experimental results anddiscusses the efficiency of themethod and conclusion is givenin Section 6

2 Theoretical Background

21 Extreme Learning Machine (ELM) In 2006 Huang et alproposed a high speed and simple learning algorithm to

The Scientific World Journal 3

Input layer Hidden layer Output layer

Inputfeatures Outputs

1

1

1

2

2

3

24

10

Figure 1 Structure of ELM network

remove the drawbacks of conventional learning algorithmsfor a single layer feed forward network (SLFN) [16] Thislearning technique is named as extreme learning machine(ELM) which is thousands times faster than conventionallearning approaches because it avoids the adjustment ofhidden layer parameters (weights and biases) during trainingby choosing them randomly The ELM algorithm trains aSLFN through the three steps which are given in the next

Consider an activation function 119892(119910) which is infinitelydifferentiable hidden nodes and a training set 119878 =

(119910119901 119889119901) 119901 = 1 2 119873 here 119889119901 isin 119877119898 is the output

response for the input sample 119910119901 isin 119877119899

Step 1 Assign hidden layer biases 119887119895 and input weights 119908119895

randomly according to any continuous probability densityfunction 119895 = 1 2

Step 2 Compute the output matrixM of the hidden layer

Step 3 Compute the output weights 120573 = MdaggerD whereD = [1198891 119889119873]

119879 andMdagger is the Moore-Penrose generalizedinverse of the hidden layer matrixH [17 18]

However the classification accuracy of the ELM networkdepends on the number of hidden nodes and the selectionof the activation functions In this work we have chosenradial basis activation function and an ELM network whosehidden layer consists of 10 hidden nodes because it givesbetter accuracy than the other combinationsThe structure ofan ELMnetwork is shown in Figure 1This network is used inour experiments

22 Support Vector Machine (SVM) The support vectormachine (SVM) networkwas proposed by Cortes andVapnikin 1995 as an alternative tool of multilayer feedforward neural

network [19] SVMs are used to solve the classification andregression problems The SVM classifies different patternsthrough the two steps (1) first the training data are mappedto a feature space of high dimension using a nonlinear kernelfunction and (2) after that an optimal hyperplane is con-structed using the method of Lagrange multipliers in orderto separate the individual classes The hyperplane is used todistinguish two linearly separable classes as given by

119889119895 (120596119879119891 (119910119895) + 119887) ge 1 for 119895 = 1 2 119871 (1)

where 119910119895 isin R119899 is 119895th input pattern and 119889119895 isin minus1 1 is thecorresponding output pattern or target for a training dataset119910119895 119889119895

119871

119895=1 119891(sdot) is a nonlinear mapping function

The decision surface of (1) is modified by introducinga nonnegative slack variable 120585 in order to separate twononlinearly separable classes as represented by

119889119895 [120596119879119891 (119910119895) + 119887] ge 1 minus 120585119895 for 119895 = 1 2 119871 (2)

An optimal hyperplane can be obtained by minimizingthe function 119865(120596 120585) with respect to 120596 and 120585119895 and it isexpressed by

119865 (120596 120585) =1

2120596119879120596 + Γ

119871

sum

119895=1

120585119895 (3)

where Γ is the reciprocal of a regularization parameter and itcontrols the tradeoff between complexity of the machine andthe number of nonseparable points [20]

To construct a decision function 120601(119910) (4) for an SVMclassifier it is required to maximize the objective function119876119891(120572)with respect to Lagrange multipliers 120572119895

119871

119895=1 subject to

the two constraints expressed by (6)

120601 (119910) = sign(

119871

sum

119895=1

120572119895119889119895119870119891 (119910 119910119895) + 119887) (4)

119876119891 (120572) =

119871

sum

119895=1

120572119895 minus1

2

119871

sum

119895=1

119871

sum

119894=1

120572119895120572119894119870119891 (119910119895 119910119894) 119889119895119889119894 (5)

subject to

119871

sum

119895=1

119889119895120572119895 = 0 0 le 120572119895 le Γ 119895 = 1 2 119871 (6)

The kernel function 119870119891(119910 119910119895) must satisfy the Mercerrsquoscondition

3 Methodology

31 Enhancement of Lung Sound Signals The recorded lungsounds are contaminated with environmental noise man-made artifacts data recording and processing instrumentsrsquodisturbances and heart sound interference which leadsto an incorrect detection of the lungs conditions In thiswork we have reduced the surrounding noise by recording


0

1

Am

plitu

de

minus1

05

minus05

0 2 4 6 8 10 12 14 16

Sample number times104

(a)

0 2 4 6 8 10 12 14 16

0

1

Sample number

Am

plitu

de

minus1

times104

05

minus05

(b)

Figure 2 (a) It depicts the waveform of a noisy lung sound signal(b) It shows the corresponding differentiation output of the nosynormal LS

the lung sounds in a quite environment and manmadeartifacts are suppressed by placing the stethoscope in aproper way over the recording positions of the subjects Theinstrumental disturbances are removed with a first orderdifferentiation algorithm [21] and the heart sound (HS) noiseis eliminated using a novel algorithm based on empiricalmode decomposition (EMD) technique developed by us andfilled an Indian patent (Ref number 515KOL2011) and thework has been published in [22] The results of the firstorder differentiation algorithm and EMD based techniqueare depicted in Figures 2 and 3 respectively The enhancedlung sound signals are used in the next step to extractthe embedded features for differentiating the normalcy andabnormality

32 Respiratory Cycle Calculation It is seen that inspirationand expiration phase of a lung sound cycle carries significantinformation regarding the lungs conditions Hence one com-plete breathing cycle is required to diagnose the lungsrsquo statususing morphological complexities of the respiratory soundsignals In this study we calculate one complete respiratorycycle using a new algorithmdeveloped by us based onHilberttransform (HT) [23] and published in [24]The algorithm canbe summarized as follows

Step 1 Calculate envelope 119890(119905) of the test signal 119910(119905) usingHT

119890 (119905) = [119910(119905)2+ (119909 (119905))

2]12

(7)

where (sdot) is the Hilbert opertor

Step 2 Smoothing of envelope signal 119890(119905) is done by afinite impulse response (FIR) Butterworth filter of cut-offfrequency 5Hz This operation minimizes the fast vibrationof 119890(119905)

1

0minus1

Am

plitu

de

0 05 1 15 2 25 3 35 4

times104Sample number

(a)

0

10

minus1

Am

plitu

de

05 1 15 2 25 3 35 4


(b)

1

0

minus1Am

plitu

de

0 05 1 15 2 25 3 35 4


(c)

1

0

minus1A

mpl

itude

0 05 1 15 2 25 3 35 4


(d)

1

0

minus1Am

plitu

de

0 05 1 15 2 25 3 35 4


(e)

Figure 3 Graphical results of the EMDbasedHS removal technique(proposed by us) (a) It shows the waveform of a mixed LS signal(20 LS + 80 HS) (b) It shows the waveform of a mixed LS signal(50 LS + 50 HS) (c) It shows the waveform of a mixed LS signal(80 LS + 20 HS) (d) Reconstructed LS signal (e) Residual HSsignal

Step 3 The transition points 119905119901119895=12119899of breathing phases are

determined using the first order derivative of the smoothenenvelope signal 119904

119890(119905) according to the following rule

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895minus1

lt 0

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895

= 0

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895+1

gt 0

(8)

The duration of inspiration or expiration phase is cal-culated by measuring the distance between two consecutiveminima points 119905119901119895

The cycle duration 119862119889 of lung sound

signal is determined by Algorithm 1 The results of the cycledetection algorithm are shown in Figure 4


Require 119905119901119895 = [1199051199011 1199051199012

119905119901119899] 119905119901119895=12119899

is the location of the transition point(1) for 119898 = [1 2 119903] 119903 is the total number of inspiration 119868

119901ℎ or expiration phases 119864119901ℎ do(2) for 119886 = 119868

119901ℎ(119898) to 119864

119901ℎ(119898) do

(3) 119862119889(119886) larr 1

(4) end for(5) end for

Algorithm 1 Calculate 119862119889

0

1

Am

plitu

de

minus10 2 4 6 8 10


(a)

0

1

Am

plitu

de

05

0 2 4 6 8 10


(b)

01

Am

plitu

de

minus10 2 4 6 8 10


(c)

01

Am

plitu

de

minus10 2 4 6 8 10


(d)

01

Am

plitu

de

minus10 2 4 6 8 10 12 14 16


(e)

Figure 4 (a) depicts the waveform of a normal LS signal (b) Itshows the Hilbert envelope and (c) shows the smoothen envelope(d) First order derivative of the smoothen envelope curve (e)The red line defined a respiratory cycle estimated by the distancebetween two consecutive phases

33 Characterizing Parameters of Lung Sounds andFeature Extraction Techniques

331 Essential Parameters of Lung Sounds to Capture the Pat-tern Changes One of themost important steps in respiratorysounds analysis is to identify the relevant inherent propertiesof the lung sound signals These attributes are useful formodeling the healthy and unhealthy conditions of the lungsThe respiratory sounds are complex in nature because of therandomized vibration of the air ways walls and turbulent

flow of gases through the respiratory system The dynamicalcomplexities of the abnormal lung sound are higher thanthat of the normal lung sound because of the presence ofauxiliary sound in abnormal cases Hence morphologicalpatterns of the abnormal lung sounds deviate from that ofthe normal sounds in a certain degree of alignment Thetemporal domain features are not relevant in diagnosing ofrespiratory diseases because of the equivalent resemble forthe both cases of normal and abnormal patients On the otherhand the spectral domain characteristics of the lung soundsdo not meet the requirements of pattern recognition due tothe nonstationary behavior of the signals The respiratorysignals show non-stationarity characteristic because of thechange in lung volume during the breathing process [25]In this work the statistical domain features are explored tomeasure the morphological complexities of the lung soundsignals The feature sets consist of four statistical parameterskurtosis (120594) skewness (120585) lacunarity (120577) and sample entropy(120572) Among these features kurtosis and skewness parameterscan measure the flatness and asymmetric distribution ofthe probability density functions (PDFs) for normal andabnormal lung sound signals as shown in Figures 5(b) and5(d) respectively It is seen from Figure 5 that the flatnessand asymmetric distribution of the abnormal LS are highercompared to normal lung sound Hence the kurtosis andskewness values for abnormal LS must be higher than thatof the normal LS The lacunarity (120577) parameter measures thetexture or heterogeneity information of any objects whichmay be fractal or nonfractal [26] The texture of normal LSis different from that of the abnormal LS because of theirdifferent genesis mechanism Hence the texture informationof the normal and pathological respiratory signals can bemeasured through the lacunarity index The abnormal LSis more heterogeneous and correspondingly gives a higherlacunarity value than the normal LS The sample entropy (120572)is a statistical parameter which can measure the complexityor irregularity in a signal [27] The sample entropy valueincreases with the irregularity property of the signals andvice versa The abnormal LS is irregular or more complexin nature than that of the normal lung sound signal due tothe unstable condition of the respiratory system associatedwith the disease severity Hence abnormal lung sound gives ahigher sample entropy value over the normal one andwe haveverified it in our preliminary work which has been publishedin [24] The complexity of the LS alters with the pathologicalconditions of the lungs The morphological complexity ofthe pathological signals is different from that of the normal


0

01

Am

plitu

de

minus010 2 4 6 8 10


(a)

0 1000 2000 3000 4000 50000

002

004

P(X

)

(b)

0 2 4 6 8 10

1

0

minus1

Am

plitu

de


(c)

0 1000 2000 3000 4000 50000

05

1

P(X

)

X

(d)

Figure 5 (a) and (c) display the normal and abnormal lung sound waveforms (b) and (d) show the probability distribution functions fornormal and abnormal cases

lung sound signals These statistical parameters (120594 120585 120577 120572)

can significantly quantify the morphological changes inpathological and normal respiratory signals Hence thesefeatures may be useful in discriminating the normal andabnormal conditions of the lungs

332 Feature Extraction Techniques In this paper four typesof features (120594 120585 120577 120572) are used to find out the status of thelungs that is normal versus abnormal These features arecomputed in different ways and discussed in the next

Computation of kurtosis (120594) parameter the kurtosisparameter describes the shape of the probability densityfunction in terms of the flatness or peakedness [28] Thekurtosis value is greater for peaked distribution than thatof the flat distribution This parameter is the fourth ordermoment of the distribution and can be defined as

Kurtosis (120594) =119864 [(119910 (119899) minus 120583)

4]

(119864 [(119910 (119899) minus 120583)2])2minus 3 =

1198984

1205904minus 3 (9)

where 119864(sdot) is the expectation operator and 120583 = 119864[119910(119899)] and120590 = radic119864[(119910(119899) minus 120583)

2] are the mean and standard deviation of

the distribution respectively and 119910(119899) is the 119899th sample valueof the lung sound signal

Estimation of skewness (120585) parameter the skewnessparameter measures the asymmetry of the distribution [28]A distribution will be asymmetric when probability densityfunction extends unequally on the left or right sides of

the center point The skewness value is zero for symmetricaldistribution and is positive or negative for asymmetricaldistributionThe parameter is defined as the ratio of the thirdorder moment and the cube of the standard deviation of theprobability distribution and is calculated by

Skewness (120585) =119864 [(119910 (119899) minus 120583)

3]

(119864 [(119910 (119899) minus 120583)2])15

=1198983

1205903 (10)

where 1198983 is the third order moment of the distribution and

119864(sdot) is the expectation operatorCalculation of lacunarity (120577) parameter The concept

of the lacunarity parameter was introduced by Mandelbrotto characterize the fractal objects [29] This parameter candistinguish the objects with the same fractal dimension bymeasuring their texture information The lacunarity featurewas first implemented in the respiratory sound analysis byHadjileontiadis to classify the adventitious lung sounds [30]The following steps are involved in calculating the lacunarityvalue based on gliding box algorithm [26]

Step 1 Computate the box mass 119887ma for a box of length 119897 byplacing it at the origin of the dataset of length119871 and119871 is alwaysgreater than 119897

Step 2 Repeate Step 1 over the entire dataset by sliding thebox with one space to the right direction


Step 3 Calculate the probability distribution 119875(119887ma 119897) of thebox masses by dividing the box masses 119896(119887ma 119897) by the totalnumber of boxes 119861(119897)

Step 4 Estimate of the first (1198721) and second (1198722) momentsof the probability distribution

Step 5 Calculate the lacunarity value for the size 119897 by dividingthe secondmoment by the square of the firstmoment and canbe defined as

120577 (119897) =1198722

11987221

(11)

where 1198721 = sum119887ma119875(119887ma 119897) and 1198722 = sum1198872

ma119875(119887ma 119897)

Computation of sample entropy (120572) parameter sam-ple entropy (SampEnt) is a modified form of approximateentropy (ApEn) was introduced by PincusThe SampEnt toolhas been proposed by Richman and Moorman to reduce thebias caused by self-matching for each template of a data series[27] It is defined as the negative logarithm of the conditionalprobability that two states that are match point-wise for adimension119898within a tolerance 119903 remainmatch in dimension119898 + 1 The SampEnt calculation algorithm consists of theseveral steps that are described next

Step 1 The templates or vectors of size 119898 are formed from agiven time series V(119899)119899=12119873 (in this case V(119899) represents thelung sound signals) as follows

119880119898 (119895) = [V (119895) V (119895 + 1) V (119895 + 119898 minus 1)]

1 ≦ 119896 ≦ 119873 minus 119898 + 1(12)

Step 2 The distance 119889[119880119898(119895) 119880119898(119894)] between the templates119880119898(119895) and 119880119898(119894) is defined as the absolute maximumdifference of their corresponding scalar components andcalculated by

119889 [119880119898 (119895) 119880119898 (119894)] = max119896=01119898minus1

(1003816100381610038161003816V (119895 + 119896) minus V (119894 + 119896)

1003816100381610038161003816)

(13)

Step 3 Counting the number of templates matching fora given template 119880119898(119895) by considering the conditions119889[119880119898(119895) 119880119898(119894)] le 119903 and 119894 = 119895

Step 4 The conditional probability of template matching fora signal having119873

119898(119894) number of templates matching for each

template is computed as

119862119898

(119903) =1

119873 minus 119898

119873minus119898

sum

119894=1

119873119898

(119894)

119873 minus 119898 minus 1 (14)

Step 5 The sample entropy values are calculated by

SampEnt (119898 119903119873) = minus ln[119862119898+1

(119903)

119862119898 (119903)] (15)

where119862119898+1(119903) is the probability that two templateswillmatchfor 119898 + 1 points

4 Experimental Datasets andImplementation Issues

41 Subjects and Data Acquisition The lung sound signalsare recorded from the abnormal as well as normal male andfemale subjects with different types of pulmonary dysfunc-tions Chronic Obstructive Pulmonary Diseases (COPDs)Interstitial Lung Diseases (ILDs) and asthma These record-ings are collected from various resources Audio amp Biosig-nal Processing laboratory IIT Kharagpur and Institute ofPulmocare and Research Kolkata India A total of 120

cycles are collected from 30 recordings of 10 normal and20 abnormal individuals The abnormal lung sounds includewheezes crackles and squawks sounds The lung sounds arerecorded from the anterior suprasternal notch positions ofthe subjects using a single channel data acquisition systemand described in [22] The sound recordings were performedin the sitting position and at relaxing mood of the patientsand stethoscope device was fixed tightly on the recording siteto diminish themanmade artifactsThe acquired LS datawerearranged in 16 bit PCM Mono audio format and stored aslowastwav files at sampling frequency of 8 kHz The recordingshave been done with a large subjects of various age groups

42 Implementation Platform The whole analysis is imple-mented on an ACER-PC with 329GHz Intel core 2 quadCPU and 349GB of RAM The MATLAB (R2008a TheMathworks Inc Natick MA) tool is used for conducting allthe experiments

5 Results and Discussion

The lung sounds data employed in this experiment arereported in Section 4 The datasets consist of pathologicaland normal respiratory sounds These data are validatedthrough three types of medical tests chest X-ray pulmonaryfunction test and high resolution computed tomographic(HRCT) scanThe decision regarding the lungs status that isnormal versus abnormal can be determined in an automatedmanner by using the machine learning properties In thiswork ELM and SVM networks are used to discriminatebetween normal and pathological individuals using theirlung sounds characteristics Both the classifiers use radialbasis function (RBF) for processing the input sample dataThe features are extracted from the 120 breathing cycles of10 normal and 20 abnormal patients The feature datasetis divided into five subsets one subset for testing and theremaining four subsets for training The training and testingprocesses are repeated five times for five individual featuresubsets The classification process runs through two phasestraining and testing During the training period trainingdataset is rendered to classifier to form a generalized modelthat is used in examining the unseen data In testing timethe unknown class feature set is verified by the trained modeland it identifies the test data class The performance of theclassifier is evaluated in terms of percentage of classificationaccuracy (CA) sensitivity (SEN) and specificity (SPE)The values of thesemeasuringmetrics are averaged on 5 trials


Table 1 Performance of ELM and SVM classifier for different sets of the proposed feature vectors

Type of classifier Feature set Feature vector SEN () SPE () CA () Training time (ms) Testing time (ms)ELM Set-1 [120594 120585] 5294 6938 6282 075 057SVM Set-1 [120594 120585] 6156 5971 6071 650 210ELM Set-2 [120594 120585 120577] 8571 8750 9047 083 065SVM Set-2 [120594 120585 120577] 8571 7738 8571 1700 570ELM Set-3 [120594 120585 120577 120572] 8630 8690 9286 420 370SVM Set-3 [120594 120585 120577 120572] 8630 8570 9160 2360 810120594 Kurtosis 120585 Skewness 120577 Lacunarity 120572 Sample Entropy

Table 2 Comparison of performances between the proposed and baseline methods [3]

Method Feature vector Type of classifier SEN () SPE () CA () Training time (ms) Testing time (ms)Proposed [120594 120585 120577 120572] ELM 8630 8690 9286 420 370Method [120594 120585 120577 120572] SVM 8630 8580 9150 2360 810

Baseline [11989125 11989150 11989175] ELM 8047 7971 8766 262 110

Method [4] [11989125 11989150 11989175] SVM 9090 7260 8633 1070 775

SEN sensitivity SPE specificity CA classification accuracy 120594 Kurtosis 120585 Skewness 120577 Lacunarity 120572 sample entropy 11989125 11989150 11989175 quartiles features

of classification procedures for three different feature setsThesemeasuring units are defined by the following equations

CA () =TP + TN

(TN + TP + FN + FP)times 100

SEN () =TP

(TP + FN)times 100

SPE () =TN

(TN + FP)times 100

(16)

The experimental results are shown in Tables 1 and 2These results are obtained by employing the three sets ofcomposite features vectors Set 1 consists of two types offeatures (120594 120585) set 2 comprises of three types of features(120594 120585 120577) and set 3 consists of four types of features (120594 120585 120577 120572)Both the classifiers give better results for set 3 in comparisonwith the remaining two sets (ie set-1 and set-2) becauseof the proper modeling of respiratory sound using higherdimensional feature vector Table 1 shows that set 2 and set3 give much better results than that of set 1 which meanslacunarity and sample entropy features carry more relevantinformation than the skewness and kurtosis parameters Thetraining and testing time increase with the increasing featuredimensions The SVM classifier takes more time than theELM network Unlike the SVM the ELM network avoidsthe tuning of hidden layer biases and input weights TheELM network gives slightly better performance than SVMdue to its universal approximation capability of the targetfunctionThe efficiency of the proposedmethod is justified bycomparing the experimental results with the baselinemethodintroduced by Sello et al [3] The experimental results showthat the developed method gives much improved accuracyover the baseline method and are shown in Table 2 Thereason of poor performance of the existing method is thatit only captures the energy information of the signal interms of frequency quartiles features These features cannot

properly model the respiratory soundsThe quartiles featuresare extracted from the global wavelet spectrum computedby the wavelet transform (WT) However the efficiency ofthe WT depends on the selection of mother wavelet orbasis function The existing technique uses Morlet waveletto generate the wavelet coefficients On the other handthe proposed method captures the structural informationof the lung sound signals in terms of texture informationirregularity index flatness and asymmetric properties of thedistributionThese features are inherently correlated with themorphological characteristics of the respiratory sounds andare capable of mapping the lungs conditions properly Hencethe developed algorithm gives better performance than theexisting technique

6 Conclusion

This paper proposes a new method to detect the normal andabnormal conditions of the lungs in a non-invasive mannerby exploring the inherent morphological characteristics ofthe lung sound signals using ELM network The method isvery fast because ELM network avoids the tuning of inputweights and hidden layer biases by randomly selecting themduring the learning process The efficiency of the algorithmis tested for three combined feature vector setsThe proposedmethod gives a better accuracy of 9286 than the baselinemethod which gives an accuracy of 8766 The proposedmethod is superior in terms of computational complexityand classification accuracy and can be used to develop anautomated diagnostic tool that will assist the doctors indiagnosis the lung status normal versus abnormal

Conflict of Interests

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


References

[1] R Laennec ldquoInvention of the stethoscoperdquo in Acoustics-Hist-orical and Philosophical Development pp 166ndash171 1962

[2] B Sankur Y P Kahya E C Gulert and T Engin ldquoComparisonof AR-based algorithms for respiratory sounds classificationrdquoComputers in Biology and Medicine vol 24 no 1 pp 67ndash761994

[3] S Sello S-K Strambi G de Michele and N Ambrosino ldquoRes-piratory sound analysis in healthy and pathological subjectsa wavelet approachrdquo Biomedical Signal Processing and Controlvol 3 no 3 pp 181ndash191 2008

[4] N Gavriely M Nissan A-H E Rubin and D W CugellldquoSpectral characteristics of chest wall breath sounds in normalsubjectsrdquoThorax vol 50 no 12 pp 1292ndash1300 1995

[5] A Cohen and D Landsberg ldquoAnalysis and automatic classi-fication of breath soundsrdquo IEEE Transactions on BiomedicalEngineering vol 31 no 9 pp 585ndash590 1984

[6] M Yeginer K Ciftci U Cini I Sen G Kilinc and Y P KahyaldquoUsing lung sounds in classification of pulmonary diseasesaccording to respiratory subphasesrdquo in Proceedings of the 26thAnnual International Conference of the IEEE Engineering inMedicine and Biology Society (EMBC rsquo04) vol 1 pp 482ndash485September 2004

[7] H Zheng H Wang L Y Wang and G Yin ldquoLung sound pat-tern analysis for anesthesia monitoringrdquo in Proceedings of theAmerican Control Conference (ACC rsquo05) vol 3 pp 1563ndash1568June 2005

[8] S Matsunaga K Yamauchi M Yamashita and S Miya-hara ldquoClassification between normal and abnormal respiratorysounds based on maximum likelihood approachrdquo in Proceed-ings of the IEEE International Conference on Acoustics Speechand Signal Processing (ICASSP rsquo09) pp 517ndash520 April 2009

[9] A Marshall and S Boussakta ldquoSignal analysis of medicalacoustic sounds with applications to chest medicinerdquo Journal ofthe Franklin Institute vol 344 no 3-4 pp 230ndash242 2007

[10] A Kandaswamy C S Kumar R P Ramanathan S JayaramanandNMalmurugan ldquoNeural classification of lung sounds usingwavelet coefficientsrdquo Computers in Biology and Medicine vol34 no 6 pp 523ndash537 2004

[11] P Forgacs A R Nathoo and H D Richardson ldquoBreathsoundsrdquoThorax vol 26 no 3 pp 288ndash295 1971

[12] R L Murphy Jr and K Sorensen ldquoChest auscultation in thediagnosis of pulmonary asbestosisrdquo Journal of OccupationalMedicine vol 15 no 3 pp 272ndash276 1973

[13] R L H Murphy Jr S K Holford and W C Knowler ldquoVisuallung sound characterization by time expanded wave formanalysisrdquoThe New England Journal of Medicine vol 296 no 17pp 968ndash971 1977

[14] R Dosani and S S Kraman ldquoLung sound intensity variabilityin normal men A contour phonopneumographic studyrdquo Chestvol 83 no 4 pp 628ndash631 1983

[15] D G Tinkelman C Lutz and B Conner ldquoAnalysis of breathsounds in normal and asthmatic children and adults usingComputer Digitized Airway Phonopneumography (CDAP)rdquoRespiratory Medicine vol 85 no 2 pp 125ndash131 1991

[16] G-B Huang Q-Y Zhu and C-K Siew ldquoExtreme learningmachine theory and applicationsrdquoNeurocomputing vol 70 no1ndash3 pp 489ndash501 2006

[17] D Serre Matrices Theory and Applications vol 216 Springer2010

[18] C R Rao and S K Mitra Generalized Inverse of Matrices andIts Applications vol 7 JohnWiley amp Sons New York NY USA1971

[19] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[20] S Haykin Neural Networks and Learning Machines PearsonEducation 3rd edition 2008

[21] Y-C Yeh and W-J Wang ldquoQRS complexes detection for ECGsignal the difference operation methodrdquo Computer Methodsand Programs in Biomedicine vol 91 no 3 pp 245ndash254 2008

[22] A Mondal P S Bhattacharya and G Saha ldquoReduction of heartsound interference from lung sound signals using empiricalmode decomposition techniquerdquo Journal ofMedical Engineeringand Technology vol 35 no 6-7 pp 344ndash353 2011

[23] M Feldman ldquoNonparametric identification of asymmetricnonlinear vibration systems with the Hilbert transformrdquo Jour-nal of Sound and Vibration vol 331 no 14 pp 3386ndash3396 2012

[24] A Mondal P Bhattacharya and G Saha ldquoDiagnosing of thelungs status using morphological anomalies of the signals intransformed domainrdquo in Proceedings of the 4th InternationalConference on Intelligent Human Computer Interaction (IHCIrsquo12) pp 1ndash5 2012

[25] W K Blake Mechanics of Flow-Induced Sound and VibrationAcademic Press Orlando Fla USA 1986

[26] C Allain and M Cloitre ldquoCharacterizing the lacunarity ofrandom and deterministic fractal setsrdquo Physical Review A vol44 no 6 pp 3552ndash3558 1991

[27] H M Al-Angari and A V Sahakian ldquoUse of sample entropyapproach to study heart rate variability in obstructive sleepapnea syndromerdquo IEEETransactions on Biomedical Engineeringvol 54 no 10 pp 1900ndash1904 2007

[28] T-H Kim and H White ldquoOn more robust estimation ofskewness and kurtosisrdquo Finance Research Letters vol 1 no 1pp 56ndash73 2004

[29] G Du and T S Yeo ldquoA novel lacunarity estimation methodapplied to SAR image segmentationrdquo IEEE Transactions onGeoscience and Remote Sensing vol 40 no 12 pp 2687ndash26912002

[30] L J Hadjileontiadis ldquoA texture-based classification of cracklesand squawks using lacunarityrdquo IEEETransactions on BiomedicalEngineering vol 56 no 3 pp 718ndash732 2009

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



knowledge of the physicians and also on the cooperation ofthe patients

With the advances of computer technologies statisticalsignal processing artificial intelligence and pattern recog-nition algorithms lung sound analysis is commenced in anautomated manner The computer aided or microprocessorbased automated tools offer several facilities in terms ofhigh speed large storage capacity and avoid the manualhazards The important intervening step of automated lungsound analysis is the extraction of authentic features thatare inherently correlated with the lungs conditions The finalstage of pattern recognition based lung sound analysis systemis decision making about the underlying disease if any Theaim of feature extraction procedure is to identify the relevantdistinct parameters of the LS signals and to arrange themin vector form that serves as an input during classificationResearchers have developed several feature extraction tech-niques to form the feature vectors based on parametric andnonparametricmethods In parametric techniquemodels theLS signal based on a priori knowledge is in contrast withnonparametric method which characterizes LS signal usinga set of basis functions [2] Many studies have been donein classifying the lung dysfunctions which suggest a varietyof methods based on time frequency and time-frequencydomain analysis [2ndash10]

The first initiative for lung sounds analysis was taken byForgacs et al in the late 1960s [11] After that a number ofresearch works were published in this field In 1973 MurphyJr and Sorensen described a spectral based technique forwheeze sounds analysis [12] and later on Murphy Jr etal introduced a waveform analysis of crackles sounds [13]Dosani and Kraman presented a technique called phono-pneumography in 1983 [14] for measuring the intensityvariation of normal lung sounds In 1991 Tinkelman et al[15] suggested a computer digitized phonopneumographytechnique to detect the airway obstruction in child patientsA classification method has been developed by Cohen andLandsberg to distinguish between normal and abnormalbreath sounds by computing the Mahalanobis distancebetween the known and unknown classes using their meanfeature vectors and covariance matrix of known data [5]Thisis a semiautomated system because breath cycles are selectedmanually and inapplicable for adventitious sounds Sankuret al proposed a technique to discriminate pathologicalfrom normal subjects using autoregressive coefficients and k-nearest neighbour and quadratic classifier in [2] Gavriely etal [4] have analyzed normal lung sounds and characterizedthese sounds by defined some parameters such as ampli-tude frequency and regression lines slopes correspondingto high and low frequency segments They have suggestedthat these parameters are useful for discrimination betweennormal and pathological sounds A neural network basedclassification technique has been proposed by Yeginer et al[6] for distinguishing between normal and abnormal subjectsusing the subphase features like autoregressive coefficientsprediction error and ratio of expiration and inspirationduration The correct classification performance for a smalldatabase is between 70 and 80 The efficiency of thisprocedure degrades due to improper selection of model

order Zheng et al [7] have introduced an algorithm todifferentiate between normal and abnormal lung soundsusing time and frequency domain features combined witha stochastic classification procedure Matsunaga et al havepresented a maximum likelihood approach based methodin distinguishing the adventitious and normal pulmonarysounds [8] This classification technique uses two types ofacoustic modeling methods one is hidden Markov model(HMM) for detection of adventitious sounds and anotherone is microphone dependent model to identify the normalsounds This is a semiautomated processes because theacoustic segments are labeled manually and its performancedegrades for noisy data

Marshall and Boussakta [9] have developed a classifica-tion technique to classify the normal and crackles soundsusing the wavelet features and cross-correlation mathemat-ical tool A wavelet based classification approach has beendeveloped by Kandaswamy et al [10] for discriminatingsix different types of pulmonary sounds by using statisticalfeatures of thewavelet detail coefficients (D1 toD7) producedby Daubechies wavelet of order 8 and artificial neural net-work Another wavelet based analysis has been studied bySello et al [3] to separate healthy subject from pathologicalsubject by characterizing the frequency power distributionof the wavelet coefficients generated by Morlet wavelet fromrespiratory sounds

The difficulties are faced by researchers in analysis of res-piratory sounds that are the interference of lung sounds withheart sounds appearance of different pathological sounds insimilar forms and also the unavailability of the sophisticatedinstrumentation for processing information embedded in thelung sounds The focus of the study is to develop a newmethod for better analysis of lung sounds by exploring thestatistical approaches anddigital signal processing knowledgecombined with pattern recognition algorithms In this papera new technique is proposed to detect the lungs statusnormal and abnormal using the structural complexitiesof LS signals The structural behavioral of the LS signalis parameterized with a number of distinct features suchas sample entropy lacunarity skewness and kurtosis Atwenty-four-dimensional feature vector is formed using theseparametersThe ELMand SVMnetworks are used to evaluatethe efficiency of the developed technique The proposedtechnique gives better performance than the baselinemethod[3]

The rest of this paper is organized as follows Theoret-ical background information on ELM and SVM classifiersis described in Section 2 The methodology to distinguishbetween normal and abnormal status of the lungs is discussedin detail in Section 3 and the database and implementationplatform that are employed in the work are discussed inSection 4 Section 5 depicts the experimental results anddiscusses the efficiency of themethod and conclusion is givenin Section 6

2 Theoretical Background

21 Extreme Learning Machine (ELM) In 2006 Huang et alproposed a high speed and simple learning algorithm to




1

1

1

2

2

3

24

10














119889119895 (120596119879119891 (119910119895) + 119887) ge 1 for 119895 = 1 2 119871 (1)


119871



119889119895 [120596119879119891 (119910119895) + 119887] ge 1 minus 120585119895 for 119895 = 1 2 119871 (2)


119865 (120596 120585) =1

2120596119879120596 + Γ

119871

sum

119895=1

120585119895 (3)



119871

119895=1 subject to


120601 (119910) = sign(

119871

sum

119895=1

120572119895119889119895119870119891 (119910 119910119895) + 119887) (4)

119876119891 (120572) =

119871

sum

119895=1

120572119895 minus1

2

119871

sum

119895=1

119871

sum

119894=1

120572119895120572119894119870119891 (119910119895 119910119894) 119889119895119889119894 (5)

subject to

119871

sum

119895=1

119889119895120572119895 = 0 0 le 120572119895 le Γ 119895 = 1 2 119871 (6)


3 Methodology



0

1

Am

plitu

de

minus1

05

minus05

0 2 4 6 8 10 12 14 16


(a)

0 2 4 6 8 10 12 14 16

0

1

Sample number

Am

plitu

de

minus1

times104

05

minus05

(b)





119890 (119905) = [119910(119905)2+ (119909 (119905))

2]12

(7)



1

0minus1

Am

plitu

de

0 05 1 15 2 25 3 35 4


(a)

0

10

minus1

Am

plitu

de

05 1 15 2 25 3 35 4


(b)

1

0

minus1Am

plitu

de

0 05 1 15 2 25 3 35 4


(c)

1

0

minus1A

mpl

itude

0 05 1 15 2 25 3 35 4


(d)

1

0

minus1Am

plitu

de

0 05 1 15 2 25 3 35 4


(e)





[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895minus1

lt 0

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895

= 0

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895+1

gt 0

(8)





Require 119905119901119895 = [1199051199011 1199051199012

119905119901119899] 119905119901119895=12119899



119901ℎ(119898) to 119864

119901ℎ(119898) do

(3) 119862119889(119886) larr 1



0

1

Am

plitu

de

minus10 2 4 6 8 10


(a)

0

1

Am

plitu

de

05

0 2 4 6 8 10


(b)

01

Am

plitu

de

minus10 2 4 6 8 10


(c)

01

Am

plitu

de

minus10 2 4 6 8 10


(d)

01

Am

plitu

de

minus10 2 4 6 8 10 12 14 16


(e)






0

01

Am

plitu

de

minus010 2 4 6 8 10


(a)

0 1000 2000 3000 4000 50000

002

004

P(X

)

(b)

0 2 4 6 8 10

1

0

minus1

Am

plitu

de


(c)

0 1000 2000 3000 4000 50000

05

1

P(X

)

X

(d)






Kurtosis (120594) =119864 [(119910 (119899) minus 120583)

4]

(119864 [(119910 (119899) minus 120583)2])2minus 3 =

1198984

1205904minus 3 (9)






Skewness (120585) =119864 [(119910 (119899) minus 120583)

3]

(119864 [(119910 (119899) minus 120583)2])15

=1198983

1205903 (10)










120577 (119897) =1198722

11987221

(11)


ma119875(119887ma 119897)



119880119898 (119895) = [V (119895) V (119895 + 1) V (119895 + 119898 minus 1)]

1 ≦ 119896 ≦ 119873 minus 119898 + 1(12)


119889 [119880119898 (119895) 119880119898 (119894)] = max119896=01119898minus1

(1003816100381610038161003816V (119895 + 119896) minus V (119894 + 119896)

1003816100381610038161003816)

(13)





119862119898

(119903) =1

119873 minus 119898

119873minus119898

sum

119894=1

119873119898

(119894)

119873 minus 119898 minus 1 (14)


SampEnt (119898 119903119873) = minus ln[119862119898+1

(119903)

119862119898 (119903)] (15)













Baseline [11989125 11989150 11989175] ELM 8047 7971 8766 262 110

Method [4] [11989125 11989150 11989175] SVM 9090 7260 8633 1070 775



CA () =TP + TN


SEN () =TP

(TP + FN)times 100

SPE () =TN

(TN + FP)times 100

(16)



6 Conclusion





References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












1

1

1

2

2

3

24

10














119889119895 (120596119879119891 (119910119895) + 119887) ge 1 for 119895 = 1 2 119871 (1)


119871



119889119895 [120596119879119891 (119910119895) + 119887] ge 1 minus 120585119895 for 119895 = 1 2 119871 (2)


119865 (120596 120585) =1

2120596119879120596 + Γ

119871

sum

119895=1

120585119895 (3)



119871

119895=1 subject to


120601 (119910) = sign(

119871

sum

119895=1

120572119895119889119895119870119891 (119910 119910119895) + 119887) (4)

119876119891 (120572) =

119871

sum

119895=1

120572119895 minus1

2

119871

sum

119895=1

119871

sum

119894=1

120572119895120572119894119870119891 (119910119895 119910119894) 119889119895119889119894 (5)

subject to

119871

sum

119895=1

119889119895120572119895 = 0 0 le 120572119895 le Γ 119895 = 1 2 119871 (6)


3 Methodology



0

1

Am

plitu

de

minus1

05

minus05

0 2 4 6 8 10 12 14 16


(a)

0 2 4 6 8 10 12 14 16

0

1

Sample number

Am

plitu

de

minus1

times104

05

minus05

(b)





119890 (119905) = [119910(119905)2+ (119909 (119905))

2]12

(7)



1

0minus1

Am

plitu

de

0 05 1 15 2 25 3 35 4


(a)

0

10

minus1

Am

plitu

de

05 1 15 2 25 3 35 4


(b)

1

0

minus1Am

plitu

de

0 05 1 15 2 25 3 35 4


(c)

1

0

minus1A

mpl

itude

0 05 1 15 2 25 3 35 4


(d)

1

0

minus1Am

plitu

de

0 05 1 15 2 25 3 35 4


(e)





[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895minus1

lt 0

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895

= 0

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895+1

gt 0

(8)





Require 119905119901119895 = [1199051199011 1199051199012

119905119901119899] 119905119901119895=12119899



119901ℎ(119898) to 119864

119901ℎ(119898) do

(3) 119862119889(119886) larr 1



0

1

Am

plitu

de

minus10 2 4 6 8 10


(a)

0

1

Am

plitu

de

05

0 2 4 6 8 10


(b)

01

Am

plitu

de

minus10 2 4 6 8 10


(c)

01

Am

plitu

de

minus10 2 4 6 8 10


(d)

01

Am

plitu

de

minus10 2 4 6 8 10 12 14 16


(e)






0

01

Am

plitu

de

minus010 2 4 6 8 10


(a)

0 1000 2000 3000 4000 50000

002

004

P(X

)

(b)

0 2 4 6 8 10

1

0

minus1

Am

plitu

de


(c)

0 1000 2000 3000 4000 50000

05

1

P(X

)

X

(d)






Kurtosis (120594) =119864 [(119910 (119899) minus 120583)

4]

(119864 [(119910 (119899) minus 120583)2])2minus 3 =

1198984

1205904minus 3 (9)






Skewness (120585) =119864 [(119910 (119899) minus 120583)

3]

(119864 [(119910 (119899) minus 120583)2])15

=1198983

1205903 (10)










120577 (119897) =1198722

11987221

(11)


ma119875(119887ma 119897)



119880119898 (119895) = [V (119895) V (119895 + 1) V (119895 + 119898 minus 1)]

1 ≦ 119896 ≦ 119873 minus 119898 + 1(12)


119889 [119880119898 (119895) 119880119898 (119894)] = max119896=01119898minus1

(1003816100381610038161003816V (119895 + 119896) minus V (119894 + 119896)

1003816100381610038161003816)

(13)





119862119898

(119903) =1

119873 minus 119898

119873minus119898

sum

119894=1

119873119898

(119894)

119873 minus 119898 minus 1 (14)


SampEnt (119898 119903119873) = minus ln[119862119898+1

(119903)

119862119898 (119903)] (15)













Baseline [11989125 11989150 11989175] ELM 8047 7971 8766 262 110

Method [4] [11989125 11989150 11989175] SVM 9090 7260 8633 1070 775



CA () =TP + TN


SEN () =TP

(TP + FN)times 100

SPE () =TN

(TN + FP)times 100

(16)



6 Conclusion





References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0

1

Am

plitu

de

minus1

05

minus05

0 2 4 6 8 10 12 14 16


(a)

0 2 4 6 8 10 12 14 16

0

1

Sample number

Am

plitu

de

minus1

times104

05

minus05

(b)





119890 (119905) = [119910(119905)2+ (119909 (119905))

2]12

(7)



1

0minus1

Am

plitu

de

0 05 1 15 2 25 3 35 4


(a)

0

10

minus1

Am

plitu

de

05 1 15 2 25 3 35 4


(b)

1

0

minus1Am

plitu

de

0 05 1 15 2 25 3 35 4


(c)

1

0

minus1A

mpl

itude

0 05 1 15 2 25 3 35 4


(d)

1

0

minus1Am

plitu

de

0 05 1 15 2 25 3 35 4


(e)





[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895minus1

lt 0

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895

= 0

[119889(119867119904

119890(119905))

119889119909]

100381610038161003816100381610038161003816100381610038161003816119905=119905119901119895+1

gt 0

(8)





Require 119905119901119895 = [1199051199011 1199051199012

119905119901119899] 119905119901119895=12119899



119901ℎ(119898) to 119864

119901ℎ(119898) do

(3) 119862119889(119886) larr 1



0

1

Am

plitu

de

minus10 2 4 6 8 10


(a)

0

1

Am

plitu

de

05

0 2 4 6 8 10


(b)

01

Am

plitu

de

minus10 2 4 6 8 10


(c)

01

Am

plitu

de

minus10 2 4 6 8 10


(d)

01

Am

plitu

de

minus10 2 4 6 8 10 12 14 16


(e)






0

01

Am

plitu

de

minus010 2 4 6 8 10


(a)

0 1000 2000 3000 4000 50000

002

004

P(X

)

(b)

0 2 4 6 8 10

1

0

minus1

Am

plitu

de


(c)

0 1000 2000 3000 4000 50000

05

1

P(X

)

X

(d)






Kurtosis (120594) =119864 [(119910 (119899) minus 120583)

4]

(119864 [(119910 (119899) minus 120583)2])2minus 3 =

1198984

1205904minus 3 (9)






Skewness (120585) =119864 [(119910 (119899) minus 120583)

3]

(119864 [(119910 (119899) minus 120583)2])15

=1198983

1205903 (10)










120577 (119897) =1198722

11987221

(11)


ma119875(119887ma 119897)



119880119898 (119895) = [V (119895) V (119895 + 1) V (119895 + 119898 minus 1)]

1 ≦ 119896 ≦ 119873 minus 119898 + 1(12)


119889 [119880119898 (119895) 119880119898 (119894)] = max119896=01119898minus1

(1003816100381610038161003816V (119895 + 119896) minus V (119894 + 119896)

1003816100381610038161003816)

(13)





119862119898

(119903) =1

119873 minus 119898

119873minus119898

sum

119894=1

119873119898

(119894)

119873 minus 119898 minus 1 (14)


SampEnt (119898 119903119873) = minus ln[119862119898+1

(119903)

119862119898 (119903)] (15)













Baseline [11989125 11989150 11989175] ELM 8047 7971 8766 262 110

Method [4] [11989125 11989150 11989175] SVM 9090 7260 8633 1070 775



CA () =TP + TN


SEN () =TP

(TP + FN)times 100

SPE () =TN

(TN + FP)times 100

(16)



6 Conclusion





References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










Require 119905119901119895 = [1199051199011 1199051199012

119905119901119899] 119905119901119895=12119899



119901ℎ(119898) to 119864

119901ℎ(119898) do

(3) 119862119889(119886) larr 1



0

1

Am

plitu

de

minus10 2 4 6 8 10


(a)

0

1

Am

plitu

de

05

0 2 4 6 8 10


(b)

01

Am

plitu

de

minus10 2 4 6 8 10


(c)

01

Am

plitu

de

minus10 2 4 6 8 10


(d)

01

Am

plitu

de

minus10 2 4 6 8 10 12 14 16


(e)






0

01

Am

plitu

de

minus010 2 4 6 8 10


(a)

0 1000 2000 3000 4000 50000

002

004

P(X

)

(b)

0 2 4 6 8 10

1

0

minus1

Am

plitu

de


(c)

0 1000 2000 3000 4000 50000

05

1

P(X

)

X

(d)






Kurtosis (120594) =119864 [(119910 (119899) minus 120583)

4]

(119864 [(119910 (119899) minus 120583)2])2minus 3 =

1198984

1205904minus 3 (9)






Skewness (120585) =119864 [(119910 (119899) minus 120583)

3]

(119864 [(119910 (119899) minus 120583)2])15

=1198983

1205903 (10)










120577 (119897) =1198722

11987221

(11)


ma119875(119887ma 119897)



119880119898 (119895) = [V (119895) V (119895 + 1) V (119895 + 119898 minus 1)]

1 ≦ 119896 ≦ 119873 minus 119898 + 1(12)


119889 [119880119898 (119895) 119880119898 (119894)] = max119896=01119898minus1

(1003816100381610038161003816V (119895 + 119896) minus V (119894 + 119896)

1003816100381610038161003816)

(13)





119862119898

(119903) =1

119873 minus 119898

119873minus119898

sum

119894=1

119873119898

(119894)

119873 minus 119898 minus 1 (14)


SampEnt (119898 119903119873) = minus ln[119862119898+1

(119903)

119862119898 (119903)] (15)













Baseline [11989125 11989150 11989175] ELM 8047 7971 8766 262 110

Method [4] [11989125 11989150 11989175] SVM 9090 7260 8633 1070 775



CA () =TP + TN


SEN () =TP

(TP + FN)times 100

SPE () =TN

(TN + FP)times 100

(16)



6 Conclusion





References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0

01

Am

plitu

de

minus010 2 4 6 8 10


(a)

0 1000 2000 3000 4000 50000

002

004

P(X

)

(b)

0 2 4 6 8 10

1

0

minus1

Am

plitu

de


(c)

0 1000 2000 3000 4000 50000

05

1

P(X

)

X

(d)






Kurtosis (120594) =119864 [(119910 (119899) minus 120583)

4]

(119864 [(119910 (119899) minus 120583)2])2minus 3 =

1198984

1205904minus 3 (9)






Skewness (120585) =119864 [(119910 (119899) minus 120583)

3]

(119864 [(119910 (119899) minus 120583)2])15

=1198983

1205903 (10)










120577 (119897) =1198722

11987221

(11)


ma119875(119887ma 119897)



119880119898 (119895) = [V (119895) V (119895 + 1) V (119895 + 119898 minus 1)]

1 ≦ 119896 ≦ 119873 minus 119898 + 1(12)


119889 [119880119898 (119895) 119880119898 (119894)] = max119896=01119898minus1

(1003816100381610038161003816V (119895 + 119896) minus V (119894 + 119896)

1003816100381610038161003816)

(13)





119862119898

(119903) =1

119873 minus 119898

119873minus119898

sum

119894=1

119873119898

(119894)

119873 minus 119898 minus 1 (14)


SampEnt (119898 119903119873) = minus ln[119862119898+1

(119903)

119862119898 (119903)] (15)













Baseline [11989125 11989150 11989175] ELM 8047 7971 8766 262 110

Method [4] [11989125 11989150 11989175] SVM 9090 7260 8633 1070 775



CA () =TP + TN


SEN () =TP

(TP + FN)times 100

SPE () =TN

(TN + FP)times 100

(16)



6 Conclusion





References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













120577 (119897) =1198722

11987221

(11)


ma119875(119887ma 119897)



119880119898 (119895) = [V (119895) V (119895 + 1) V (119895 + 119898 minus 1)]

1 ≦ 119896 ≦ 119873 minus 119898 + 1(12)


119889 [119880119898 (119895) 119880119898 (119894)] = max119896=01119898minus1

(1003816100381610038161003816V (119895 + 119896) minus V (119894 + 119896)

1003816100381610038161003816)

(13)





119862119898

(119903) =1

119873 minus 119898

119873minus119898

sum

119894=1

119873119898

(119894)

119873 minus 119898 minus 1 (14)


SampEnt (119898 119903119873) = minus ln[119862119898+1

(119903)

119862119898 (119903)] (15)













Baseline [11989125 11989150 11989175] ELM 8047 7971 8766 262 110

Method [4] [11989125 11989150 11989175] SVM 9090 7260 8633 1070 775



CA () =TP + TN


SEN () =TP

(TP + FN)times 100

SPE () =TN

(TN + FP)times 100

(16)



6 Conclusion





References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation














Baseline [11989125 11989150 11989175] ELM 8047 7971 8766 262 110

Method [4] [11989125 11989150 11989175] SVM 9090 7260 8633 1070 775



CA () =TP + TN


SEN () =TP

(TP + FN)times 100

SPE () =TN

(TN + FP)times 100

(16)



6 Conclusion





References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










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









Documents

Research Article Detection of Lungs Status Using ...downloads.hindawi.com/journals/tswj/2014/182938.pdf · variation of normal lung sounds. In , Tinkelman et al. [ ] suggested a computer