a hybrid expert system approach for telemonitoring of vocal fold pathology

14
Applied Soft Computing 13 (2013) 4148–4161 Contents lists available at ScienceDirect Applied Soft Computing j ourna l ho me page: www.elsevier.com/locate /asoc A hybrid expert system approach for telemonitoring of vocal fold pathology M. Hariharan a,, Kemal Polat b , R. Sindhu c , Sazali Yaacob a a School of Mechatronic Engineering, Universiti Malaysia Perlis (UniMAP), Campus Pauh Putra, Perlis 02600, Malaysia b Department of Electrical and Electronics Engineering, Faculty of Engineering and Architecture, Abant Izzet Baysal University, 14280 Bolu, Turkey c School of Microelectronic Engineering, Universiti Malaysia Perlis (UniMAP), Campus Pauh Putra, Perlis 02600, Malaysia a r t i c l e i n f o Article history: Received 24 August 2012 Received in revised form 8 March 2013 Accepted 3 June 2013 Available online 28 June 2013 Keywords: Vocal fold pathology Feature extraction Feature weighting Compressed voice samples Classification a b s t r a c t Acoustical parameters extracted from the recorded voice samples are actively pursued for accurate detec- tion of vocal fold pathology. Most of the system for detection of vocal fold pathology uses high quality voice samples. This paper proposes a hybrid expert system approach to detect vocal fold pathology using the compressed/low quality voice samples which includes feature extraction using wavelet packet transform, clustering based feature weighting and classification. In order to improve the robustness and discrimination ability of the wavelet packet transform based features (raw features), we propose clustering based feature weighting methods including k-means clustering (KMC), fuzzy c-means (FCM) clustering and subtractive clustering (SBC). We have investigated the effectiveness of raw and weighted features (obtained after applying feature weighting methods) using four different classifiers: Least Square Support Vector Machine (LS-SVM) with radial basis kernel, k-means nearest neighbor (kNN) classifier, probabilistic neural network (PNN) and classification and regression tree (CART). The proposed hybrid expert system approach gives a promising classification accuracy of 100% using the feature weighting methods and also it has potential application in remote detection of vocal fold pathology. © 2013 Elsevier B.V. All rights reserved. 1. Introduction People who have trouble in using their voices are about 7.5 mil- lion approximately in the United States [1]. Due to the nature of job, unhealthy social habits and voice abuse, people suffering from vocal fold pathology have been increasing dramatically. Any prob- lem in the vocal folds causes involuntary movements of one or more muscles of the larynx which results in hoarseness and even- tually reduces the voice quality [2,3]. Hence, voice can be a reliable source to investigate the vocal fold pathologies. Medical profession- als use subjective techniques or invasive methods such as the direct inspection of the vocal folds and the observation of the vocal folds by endoscopic instruments to detect vocal fold pathology. These techniques require costly resources such as special light sources, endoscopic instruments and specialized video-camera equipments, besides they are expensive, risky, time consuming, annoying for patients [2,3]. In order to circumvent these problems, non-invasive methods have been developed to help the medical professionals to detect vocal fold pathology. Corresponding author. Tel.: +60 134622469; fax: +60 49885167. E-mail addresses: [email protected], [email protected] (M. Hariharan). With the proliferation in signal processing techniques, voice signal can be used for the detection of vocal fold pathology and its quantitative information plays a prominent role to understand the process of vocal fold pathology formation. In the last three decades, several research works have been carried out on the automatic detection and classification of vocal fold pathologies by means of acoustic analysis, parametric and non-parametric feature extraction methods, automatic pattern recognition or statistical methods [2–14]. A large amount of acoustic parameters have been proposed and its effectiveness has been proven by experimental researches. The important parameters are pitch, jitter, shimmer, harmonics-to-noise, normalized noise energy, Mel-frequency cep- stral coefficients (MFCCs) wavelet/wavelet packet based features, non-linear dynamic analysis (approximation entropy, correlation dimension, Lyapunaov exponent etc.) and higher order spectral (HOS) analysis [2,4–9,15–18]. Most of the earlier systems developed for detecting vocal fold pathology use high quality voice samples, which requires 5–10 MB of storage space. The high quality voice samples need to be compressed, in order to store them occupying the minimal stor- age space and to share/transmit the recorded high quality voice samples among different research laboratories over telecommu- nication networks. Hence, reliable methods are to be developed to detect the vocal fold pathology using the voice samples which are compressed in mp3 format at different bit rates (160, 96, 64, 1568-4946/$ see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.asoc.2013.06.004

Upload: sazali

Post on 25-Dec-2016

219 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: A hybrid expert system approach for telemonitoring of vocal fold pathology

Av

Ma

b

c

a

ARRAA

KVFFCC

1

ljvlmtsaibtebpmd

w

1h

Applied Soft Computing 13 (2013) 4148–4161

Contents lists available at ScienceDirect

Applied Soft Computing

j ourna l ho me page: www.elsev ier .com/ locate /asoc

hybrid expert system approach for telemonitoring ofocal fold pathology

. Hariharana,∗, Kemal Polatb, R. Sindhuc, Sazali Yaacoba

School of Mechatronic Engineering, Universiti Malaysia Perlis (UniMAP), Campus Pauh Putra, Perlis 02600, MalaysiaDepartment of Electrical and Electronics Engineering, Faculty of Engineering and Architecture, Abant Izzet Baysal University, 14280 Bolu, TurkeySchool of Microelectronic Engineering, Universiti Malaysia Perlis (UniMAP), Campus Pauh Putra, Perlis 02600, Malaysia

r t i c l e i n f o

rticle history:eceived 24 August 2012eceived in revised form 8 March 2013ccepted 3 June 2013vailable online 28 June 2013

eywords:ocal fold pathology

a b s t r a c t

Acoustical parameters extracted from the recorded voice samples are actively pursued for accurate detec-tion of vocal fold pathology. Most of the system for detection of vocal fold pathology uses high qualityvoice samples. This paper proposes a hybrid expert system approach to detect vocal fold pathologyusing the compressed/low quality voice samples which includes feature extraction using wavelet packettransform, clustering based feature weighting and classification. In order to improve the robustnessand discrimination ability of the wavelet packet transform based features (raw features), we proposeclustering based feature weighting methods including k-means clustering (KMC), fuzzy c-means (FCM)

eature extractioneature weightingompressed voice sampleslassification

clustering and subtractive clustering (SBC). We have investigated the effectiveness of raw and weightedfeatures (obtained after applying feature weighting methods) using four different classifiers: Least SquareSupport Vector Machine (LS-SVM) with radial basis kernel, k-means nearest neighbor (kNN) classifier,probabilistic neural network (PNN) and classification and regression tree (CART). The proposed hybridexpert system approach gives a promising classification accuracy of 100% using the feature weightingmethods and also it has potential application in remote detection of vocal fold pathology.

. Introduction

People who have trouble in using their voices are about 7.5 mil-ion approximately in the United States [1]. Due to the nature ofob, unhealthy social habits and voice abuse, people suffering fromocal fold pathology have been increasing dramatically. Any prob-em in the vocal folds causes involuntary movements of one or

ore muscles of the larynx which results in hoarseness and even-ually reduces the voice quality [2,3]. Hence, voice can be a reliableource to investigate the vocal fold pathologies. Medical profession-ls use subjective techniques or invasive methods such as the directnspection of the vocal folds and the observation of the vocal foldsy endoscopic instruments to detect vocal fold pathology. Theseechniques require costly resources such as special light sources,ndoscopic instruments and specialized video-camera equipments,esides they are expensive, risky, time consuming, annoying foratients [2,3]. In order to circumvent these problems, non-invasive

ethods have been developed to help the medical professionals to

etect vocal fold pathology.

∗ Corresponding author. Tel.: +60 134622469; fax: +60 49885167.E-mail addresses: [email protected],

[email protected] (M. Hariharan).

568-4946/$ – see front matter © 2013 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.asoc.2013.06.004

© 2013 Elsevier B.V. All rights reserved.

With the proliferation in signal processing techniques, voicesignal can be used for the detection of vocal fold pathology andits quantitative information plays a prominent role to understandthe process of vocal fold pathology formation. In the last threedecades, several research works have been carried out on theautomatic detection and classification of vocal fold pathologies bymeans of acoustic analysis, parametric and non-parametric featureextraction methods, automatic pattern recognition or statisticalmethods [2–14]. A large amount of acoustic parameters have beenproposed and its effectiveness has been proven by experimentalresearches. The important parameters are pitch, jitter, shimmer,harmonics-to-noise, normalized noise energy, Mel-frequency cep-stral coefficients (MFCCs) wavelet/wavelet packet based features,non-linear dynamic analysis (approximation entropy, correlationdimension, Lyapunaov exponent etc.) and higher order spectral(HOS) analysis [2,4–9,15–18].

Most of the earlier systems developed for detecting vocal foldpathology use high quality voice samples, which requires 5–10 MBof storage space. The high quality voice samples need to becompressed, in order to store them occupying the minimal stor-age space and to share/transmit the recorded high quality voice

samples among different research laboratories over telecommu-nication networks. Hence, reliable methods are to be developedto detect the vocal fold pathology using the voice samples whichare compressed in mp3 format at different bit rates (160, 96, 64,
Page 2: A hybrid expert system approach for telemonitoring of vocal fold pathology

ft Com

4gcuMNtoccuccMompciws

mwtliTacaicWhta1tpotfrfIpnwwLaoarl[ltcsamcor

M. Hariharan et al. / Applied So

8, 24, and 8 kb/s). In [19], researchers have pioneered to investi-ate the effect of the Moving Pictures Expert Group (MPEG) audioompression on the automatic detection of vocal fold pathologysing a wide variety of acoustic parameters extracted from theulti-Dimensional Voice Program (MDVP) of Kay Elemetrics Corp.icolas Saenz Lechon et al. have proposed a system to inves-

igate the effects of audio compression in automatic detectionf vocal fold pathologies [20]. They used Mel-frequency cepstraloefficients along with their derivatives and noise parameters toharacterize the voice signals. The classification was performedsing Gaussian mixture model and support vector machine. Theyoncluded that there was no performance degradation for theompressed voice samples with binary rates of above 64 kbps.oran et al. [21,22] have developed a system for the detection

f vocal fold pathologies using voice samples which were trans-itted over a telephone line. They employed classical acoustic

erturbation parameters and used a simple linear discriminantlassifier to classify the voice samples into normal or patholog-cal ones. The performance of the system was only around 85%,

hich was very less compared to results using high quality voiceamples.

From the previous works, it has been observed that the perfor-ance of the system degrades severely when the voice samplesere compressed by using bit rates lower than 64 kbps using

he conventional acoustical parameters since the compression atower bit rates diminishes the fidelity and introduces some signif-cant alterations in harmonic structure of the voice samples [19].his paper focuses on the development of hybrid expert systempproach for efficient detection of vocal fold pathology using theompressed voice samples. The proposed hybrid expert systempproach has the following stages: voice compression using Audac-ty 2.0.1 [23], feature extraction using wavelet packet transform,lustering based feature weighting methods and classification.

avelet packet transform based features were chosen since theyave been successfully applied by many researchers not only inhe area of automatic detection of vocal fold pathologies [5–7] butlso in various applications [24–26]. They achieved accuracy up-to00% approximately under different experimental conditions andheir results highly depend on the optimal selection of waveletacket based features using feature selection/transformation andptimization techniques. In this paper, three clustering based fea-ure weighting methods were proposed such as k-means clustering,uzzy c-means clustering and subtractive clustering to improve theobustness and discrimination ability of the wavelet packet trans-orm based features. Voice samples from Massachusetts Eye and Earnfirmary (MEEI) voice disorders database [27] and MAPACI speechathology database [28] were used, in order to gauge the robust-ess and independence of the algorithms to the database. In thisork, the efficiency of the wavelet packet transform based raw andeighted features of compressed voice samples were tested using

S-SVM, kNN, PNN and CART. LS-SVM offers high classificationccuracy in the previous works and is easy to handle compared tother classification techniques [5,7,16,29]. kNN classifier has somedvantages like easy to understand and implement, no training isequired, robust to the search space, can be updated online at veryittle cost and few parameters to tune such as distance metric and k30]. PNN is useful for classification problems since its high speed ofearning, simple learning rule, and single pass training comparedo multi-layer perceptrons [11,12,31]. CART is a non-parametriclassifier and provides easily comprehensible decision strategies,calable to large problems, can handle large number of variables,nd insensitive to outliers [32–34]. Two schemes of data validation

ethods were used (Conventional Validation – ConV and 10-fold

ross validation – CrossV), in order to demonstrate the consistencyf the classification results. Two projection based dimensionalityeduction methods such as principal component analysis (PCA)

puting 13 (2013) 4148–4161 4149

and linear discriminant analysis (LDA) were used to form a smalluncorrelated feature set.

The organization of the paper is as follows: the dataset used inthis paper is explained in Section 2. A detailed description of themethodology adopted in this paper is described in Section 3 whichincludes feature extraction, feature weighting methods, dimen-sionality reduction methods and classifiers. Section 4 explains theinterpretation of the results and discussion. Section 5 concludes thepaper.

2. Database

In this paper, two databases were used such as MEEI voice dis-order database and MAPACI speech pathology database. MEEI voicedisorder database contains 53 normal and 657 pathological voicesamples. The voice samples were the sustained phonation of thevowel /ah/ (1–3 s) long and reading (12 s) of the “Rainbow Passage”from patients with normal voices and a wide variety of organic, neu-rological, traumatic, and psychogenic voice disorders in differentstages. All the voice samples were collected in a controlled envi-ronment and sampled with a sampling rate of 25 or 50 kHz and16 bits of resolution. A subset of 226 voices samples of sustainedphonation of the vowel /ah/ (173 pathological + 53 normal) wereused according to [35]. In MAPACI speech pathology database, allthe voice samples were recorded using a Senheiser headset micro-phone at 44,100 Hz during the lifetime project of MAPACI (2003).This database consists of 24 male voice samples (12 normal + 12pathological) and 24 female voice samples (12 normal + 12 patho-logical). The details of recordings (Range, average and standarddeviation in years) used in this paper were tabulated in Table 1.Compressed voice datasets were created by compressing the highquality voice samples with a bit rate of 8 kbps, an output samplingrate of 24 kHz and 16-bit resolution using Audacity 2.0.1. A bit rateof 8 kbps was chosen since the harmonic structure of voice sam-ples was severely affected and it is a challenge for researchers topropose robust features for efficient detection of vocal fold pathol-ogy using such compressed voice samples. Figs. 1 and 2 showthe uncompressed and compressed pathological and normal voicesamples (MEEI voice disorder database and MAPACI speech pathol-ogy database) respectively. From the Figs, it can be observed thata normal voice sample has regular/periodic oscillation due to thesymmetrical vibration of the vocal fold and a pathological voicesample has irregular/aperiodic oscillations due to the asymmetri-cal vibration of the vocal fold. It can also be seen from the figuresthat the structure and quality of the original audio signals have beenseverely affected after compression.

3. Methodology

The block diagram of the proposed study was shown in Fig. 3.Analysis of voice samples was carried out by means of frame-based(short-time basis) analysis [2,3]. In frame-based analysis, voicesamples were segmented into frames of 40 ms long using a Ham-ming window with 50% overlap. From each recording of the MEEIvoice samples, 8 and 26 central frames from pathological and nor-mal voices samples were selected respectively among the ones thatbelong to the most stationary portion of the sustained speech sig-nal as it is proposed in [36,37]. This selection yielded 1384 framesof pathological voice and another 1378 frames of normal voice.Similarly, for the MAPACI voice samples, 65 central frames wereselected and this selection yielded 1560 frames of pathological andnormal voice samples. Then the each frame was parameterized by

means of wavelet packet transform with energy and entropy fea-tures. Data pre-processing step is one of the most important stagesin pattern recognition problems. In order to improve the robust-ness and discrimination ability of the extracted features, clustering
Page 3: A hybrid expert system approach for telemonitoring of vocal fold pathology

4150 M. Hariharan et al. / Applied Soft Computing 13 (2013) 4148–4161

Table 1Details of the recordings used in this study.

MEEI voice disorder database

Subjects Range (years) Average (years) Standard deviation (years)

Male Female Male Female Male Female Male Female

MEEI voice disorder databaseNormal 21 32 26–58 22–52 38.8 34.2 8.49 7.87Pathologic 70 103 26–58 21–51 41.7 37.6 9.38 8.19

MAPACI speech pathology database

Normal 12 12 20–37 24–3Pathologic 12 12 27–68 20–6

Fig. 1. Plots of uncompressed patholo

Fig. 2. Plots of compressed patholog

9 24.7 32.1 5.6 5.53 49.5 40.8 13.9 15.7

gical and normal voice samples.

ical and normal voice samples.

Page 4: A hybrid expert system approach for telemonitoring of vocal fold pathology

M. Hariharan et al. / Applied Soft Computing 13 (2013) 4148–4161 4151

gram

bspde

3

bfmdtbqpdbdp3eEw

E

E

wwcFu

Fig. 3. Overall block dia

ased feature weighting methods were proposed. In the followingections, feature extraction using wavelet packet transform, datare-processing using clustering based feature weighting methods,imensionality reduction using PCA and LDA, and classifiers werexplained.

.1. Wavelet packet transform based feature extraction

In the last decades, wavelet and wavelet packet transform haveeen applied for the analysis of voice samples to detect the vocalold pathology, since they provide robust representation of nor-

al and pathological voice samples in both spatial and frequencyomain. Basically, wavelet and wavelet packet transform analyzeshe signal by decomposing into approximation (lower frequencyand, via low-pass filtering) and detail coefficients (higher fre-uency band, via high-pass filtering). In wavelet decompositionrocedure, only lower frequency bands were used for consecutiveecompositions but in wavelet packet decomposition procedure,oth lower and higher frequency bands were used for furtherecompositions. In this paper, all the voice samples were decom-osed into 5 levels using wavelet packet transform and yields2 wavelet packet subbands. Energy and entropy features werextracted from 32 wavelet packet subband coefficients as shown inqs. (1) and (2). Features extracted by using the following equationsere called as ‘raw features’.

nergym = log

⎛⎜⎜⎜⎜⎝

N∑i=1

(Cim,n)

2

N

⎞⎟⎟⎟⎟⎠

, where m = 1,2,· · ·, M,

n = 0,1,· · ·, 2M − 1 (1)

ntropym = −N∑

i=1

Cim,n log(Ci

m,n), where m = 1,2,· · ·, M,

n = 0,1,· · ·, 2M − 1 (2)

here m represents the number of decomposition level, C is the

avelet packet coefficient, N is the number of wavelet packet

oefficients at each level and n represents wavelet packet node.eatures extracted from all the wavelet packet subbands weresed since the effect of pathological factor in the voice samples

of the proposed study.

cannot be detected from specific wavelet subbands or frequencyrange. Considering the fact that pathological voice signals containmore rapid variations and the popularity of Daubechies waveletsin speech applications, the mother wavelet function was chosen tobe ‘db4’ in our study. The order was chosen to be low, to model thetransients and rapid variations in a signal efficiently [5,7,14,38].

3.2. Data pre-processing using clustering based feature weighting

Data preprocessing methods are employed prior to classifica-tion process and they are widely applied for noise removing, outlierdetection, data normalization, data transformation and missingvalue imputation [39–42]. There are many data preprocessingmethods available in the literature based on clustering based fea-ture weighting [39–42]. Clustering algorithms are normally usedin many fields which include machine learning, data grouping,document retrieval, image analysis and pattern recognition etc.In pattern classification problems, the performance of the learnedmodels highly depends on the relevant and robust features. In orderto improve the robustness of the features and the classificationaccuracy, clustering algorithms were used as feature weightingmethod to transform the extracted non-linearly separable featuresto linearly separable features [39–42]. Clustering algorithms areused not only to study the similarity or dissimilarity of the featuresbut also useful for compression and reduction of the features.

In this work, k-means clustering, fuzzy c-means clustering andsubtractive clustering were used as a clustering algorithm in theprocess of feature weighting. The working of clustering based fea-ture weighting is summarized (Fig. 4) follows: Firstly, the clustercenters of each feature belonging to dataset using k-means clus-tering or fuzzy c-means clustering or subtractive clustering werefound. Next, the ratios of means of features to their centers werecalculated. Finally, these ratios were multiplied with each respec-tive feature. For more information about these feature weightingalgorithms, the readers are referred to Refs. [39–42].

3.3. Dimensionality reduction

In this paper, we have extracted 32 energy and entropy featuresfrom the compressed audio data. In order to map the original large

feature space (32 energy and entropy) into reduced feature space,there are two feature transformation approaches such as sequen-tial methods and projection methods. Sequential methods do notremove the feature redundancy completely since the final solution
Page 5: A hybrid expert system approach for telemonitoring of vocal fold pathology

4152 M. Hariharan et al. / Applied Soft Computing 13 (2013) 4148–4161

The ca lculation of cluste r ce nters of

each feature in dataset usingk-means

clus ter ing orfuz zy c-means

clus ter ingorsubtrac tiv e clus ter ing

The computation of the rat ios of means

of features to their centers (the calculated

values are the multiplying coefficients)

These ratios were multiplied with

each respective feature in dataset

Fig. 4. The block diagram of proposed clustering based feature weighting method.

Table 2Details of training and testing set.

Number of frames Type of validation MEEI database MAPACI speech pathology database

MEEI Database (1384pathological + 1378 normal)

CrossV Feature vectors are divided randomly into10 sets and training is repeated for 10times

Feature vectors are divided randomly into10 sets and training is repeated for 10times

= 1933

dtIoo

3a

mIiitPt[

FS

MAPACI speech pathologydatabase (1560pathological + 1560 normal)

ConV Training

epends on the extracted feature set. In projection methods, projec-ion is performed by a linear or non-linear transformation [38,43].n this paper, two projection based dimensionality reduction meth-ds (PCA and LDA) were used for reducing the original dimensionf the feature space without the loss of classification accuracy.

.3.1. Dimensionality reduction with principal componentnalysis

PCA is a well-known linear projection method and one of theost popular techniques for dimensionality reduction [38,43,44].

t is used to transform the original large set of correlated featuresnto small set of uncorrelated features (principal components). PCAs an unsupervised feature transformation method since it performs

he vector projection without any knowledge of their class labels.CA reveals about the hidden information from the original fea-ure space by maximizing the variance of the projected vectors38,43,44]. The steps are summarized as follows:

ig. 5. The class distribution of raw and weighted PCA entropy features according to theBC.

(70%) Testing = 829(30%) Training = 2184 (70%) Testing = 936(30%)

1. Estimate the covariance matrix from the original feature set.2. Perform the eigenvalue decomposition to compute the eigenvec-

tors/eigenvalues of covariance matrix and sort the eigenvaluesin descending order.

3. Form the transformation matrix using the respective eigenvec-tors.

4. Transform the original larger feature space into new featurespace using transformation matrix.

5. From the transformed matrix, only the features of eigenvaluegreater than ‘1’ were selected and forms the small set of uncorre-lated features. After applying PCA, we obtained a reduced featureset with 10 uncorrelated features.

3.3.2. Dimensionality reduction with linear discriminant analysisLDA seeks to minimize the distances among the vectors belong-

ing to the same class and to maximize the distances among theclass centers. LDA utilizes supervised projection method and is to

principal components 1, 5 and 10. (a) Raw PCA features, (b) KMC, (c) FCM and (d)

Page 6: A hybrid expert system approach for telemonitoring of vocal fold pathology

M.

Hariharan

et al.

/ A

pplied Soft

Computing

13 (2013)

4148–4161

4153

Table 3Classification accuracies for MEEI database [energy features, 10-fold cross validation and conventional validation].

Classifiers Energy features Raw features Weighted features

KMC FCM SBC

SE SP ACC SE SP ACC SE SP ACC SE SP ACC

10-fold cross validationLS-SVM

32 features 93.02 ± 0.33 91.49 ± 0.22 92.24 ± 0.24 99.77 ± 0.09 99.52 ± 0.06 99.64 ± 0.06 100 ± 0.00 99.93 ± 0.00 99.96 ± 0.00 100 ± 0.00 99.61 ± 0.05 99.80 ± 0.03PCA 92.06 ± 0.10 90.78 ± 0.24 91.41 ± 0.15 99.12 ± 0.12 99.43 ± 0.07 99.28 ± 0.05 98.36 ± 0.09 97.90 ± 0.12 98.13 ± 0.07 99.90 ± 0.04 98.32 ± 0.06 99.10 ± 0.04LDA 90.65 ± 0.08 89.03 ± 0.05 89.83 ± 0.05 99.85 ± 0.02 99.78 ± 0.00 99.82 ± 0.01 99.64 ± 0.00 99.64 ± 0.00 99.64 ± 0.00 98.96 ± 0.05 98.13 ± 0.00 98.54 ± 0.02

PNN 32 features 90.62 ± 2.46 88.59 ± 0.47 89.54 ± 1.34 98.21 ± 2.38 99.35 ± 0.17 98.74 ± 1.24 99.99 ± 0.05 99.88 ± 0.06 99.93 ± 0.03 98.80 ± 2.37 98.12 ± 0.35 98.42 ± 1.14PCA 89.97 ± 0.96 88.62 ± 1.29 89.28 ± 1.07 98.53 ± 0.20 98.19 ± 0.20 98.36 ± 0.16 100.00 ± 0.00 99.65 ± 0.15 99.82 ± 0.07 97.13 ± 1.60 97.68 ± 0.34 97.39 ± 0.84LDA 90.71 ± 0.05 89.03 ± 0.03 89.85 ± 0.04 99.83 ± 0.04 99.77 ± 0.05 99.80 ± 0.03 99.64 ± 0.00 99.70 ± 0.05 99.67 ± 0.03 99.10 ± 0.20 98.28 ± 0.05 98.69 ± 0.09

KNN 32 features 91.96 ± 0.56 87.89 ± 0.43 89.82 ± 0.28 100.00 ± 0.00 99.78 ± 0.12 99.89 ± 0.06 100.00 ± 0.00 99.62 ± 0.15 99.81 ± 0.08 99.82 ± 0.18 97.09 ± 0.80 98.41 ± 0.37PCA 90.44 ± 1.03 87.70 ± 0.62 89.02 ± 0.77 100.00 ± 0.00 99.45 ± 0.24 99.72 ± 0.12 100.00 ± 0.00 99.26 ± 0.33 99.63 ± 0.17 98.46 ± 0.36 96.74 ± 0.55 97.58 ± 0.24LDA 88.31 ± 2.83 86.20 ± 1.87 87.22 ± 2.33 99.81 ± 0.10 99.67 ± 0.06 99.74 ± 0.05 99.68 ± 0.15 99.54 ± 0.04 99.61 ± 0.07 98.59 ± 0.27 98.23 ± 0.18 98.41 ± 0.19

CART 32 features 87.71 ± 0.28 86.27 ± 0.42 86.97 ± 0.20 99.73 ± 0.09 99.58 ± 0.09 99.65 ± 0.08 99.71 ± 0.09 99.46 ± 0.21 99.59 ± 0.12 99.73 ± 0.06 99.78 ± 0.05 99.75 ± 0.04PCA 84.46 ± 0.81 83.37 ± 0.86 83.91 ± 0.81 99.61 ± 0.11 99.29 ± 0.16 99.45 ± 0.08 99.55 ± 0.06 99.29 ± 0.08 99.42 ± 0.04 94.60 ± 0.37 94.42 ± 0.90 94.50 ± 0.46LDA 84.80 ± 0.57 83.41 ± 0.31 84.09 ± 0.41 99.70 ± 0.09 99.61 ± 0.04 99.65 ± 0.04 99.62 ± 0.13 99.49 ± 0.09 99.56 ± 0.04 98.32 ± 0.23 98.10 ± 0.17 98.21 ± 0.15

Conventional validationLS-SVM

32 features 92.62 ± 0.87 90.47 ± 1.97 91.51 ± 1.36 99.54 ± 0.26 99.64 ± 0.28 99.59 ± 0.17 99.44 ± 0.26 98.73 ± 0.66 99.08 ± 0.39 100 ± 0.00 99.54 ± 0.29 99.77 ± 0.14PCA 92.01 ± 0.98 90.29 ± 1.33 91.12 ± 0.87 99.28 ± 0.49 99.01 ± 0.59 99.14 ± 0.35 100 ± 0.00 99.95 ± 0.10 99.98 ± 0.05 99.85 ± 0.20 98.83 ± 0.45 99.34 ± 0.22LDA 90.22 ± 1.55 89.05 ± 1.21 89.61 ± 1.18 99.83 ± 0.20 99.73 ± 0.14 99.78 ± 0.15 99.61 ± 0.17 99.64 ± 0.21 99.63 ± 0.14 99.27 ± 0.38 98.07 ± 0.62 98.66 ± 0.31

PNN 32 features 90.11 ± 2.62 87.95 ± 0.63 88.94 ± 1.37 99.95 ± 0.15 99.84 ± 0.12 99.90 ± 0.07 99.95 ± 0.14 99.82 ± 0.08 99.89 ± 0.07 98.66 ± 2.80 97.80 ± 0.56 98.17 ± 1.33PCA 89.22 ± 0.90 88.00 ± 1.34 88.57 ± 0.99 100.00 ± 0.00 99.67 ± 0.13 99.83 ± 0.07 100.00 ± 0.00 99.53 ± 0.12 99.76 ± 0.06 96.91 ± 1.75 97.33 ± 0.36 97.10 ± 0.93LDA 91.14 ± 0.47 89.27 ± 0.40 90.17 ± 0.35 99.84 ± 0.03 99.76 ± 0.05 99.80 ± 0.03 99.65 ± 0.06 99.70 ± 0.11 99.68 ± 0.05 99.09 ± 0.26 98.25 ± 0.18 98.66 ± 0.11

KNN 32 features 91.75 ± 0.52 87.23 ± 0.58 89.35 ± 0.36 100.00 ± 0.00 99.70 ± 0.13 99.85 ± 0.06 100.00 ± 0.00 99.58 ± 0.20 99.79 ± 0.10 98.92 ± 0.89 96.50 ± 0.77 97.68 ± 0.65PCA 90.81 ± 1.28 87.15 ± 0.57 88.88 ± 0.77 100.00 ± 0.00 99.51 ± 0.33 99.75 ± 0.17 100.00 ± 0.00 99.28 ± 0.49 99.64 ± 0.25 98.11 ± 0.41 96.23 ± 0.49 97.15 ± 0.27LDA 89.97 ± 2.20 86.91 ± 1.18 88.35 ± 1.59 99.93 ± 0.13 99.56 ± 0.29 99.74 ± 0.15 99.89 ± 0.17 99.35 ± 0.41 99.62 ± 0.20 98.50 ± 0.38 98.21 ± 0.14 98.35 ± 0.1832 features 87.07 ± 0.45 85.95 ± 0.55 86.48 ± 0.21 99.64 ± 0.10 99.66 ± 0.11 99.65 ± 0.07 99.64 ± 0.07 99.62 ± 0.09 99.63 ± 0.07 99.67 ± 0.13 99.62 ± 0.16 99.64 ± 0.09

CART PCA 84.23 ± 0.52 82.89 ± 0.48 83.50 ± 0.27 99.50 ± 0.11 99.36 ± 0.12 99.43 ± 0.10 99.35 ± 0.13 99.14 ± 0.13 99.24 ± 0.08 94.16 ± 0.43 93.71 ± 0.42 93.92 ± 0.31LDA 85.21 ± 0.46 84.27 ± 0.24 84.71 ± 0.23 99.75 ± 0.08 99.64 ± 0.05 99.69 ± 0.04 99.72 ± 0.11 99.47 ± 0.10 99.59 ± 0.06 98.41 ± 0.12 97.96 ± 0.10 98.18 ± 0.10

Page 7: A hybrid expert system approach for telemonitoring of vocal fold pathology

4154

M.

Hariharan

et al.

/ A

pplied Soft

Computing

13 (2013)

4148–4161

Table 4Classification accuracies for MAPACI database [energy features, 10-fold cross validation and conventional validation].

Classifiers Energy features Raw features Weighted features

KMC FCM SBC

SE SP ACC SE SP ACC SE SP ACC SE SP ACC

10-fold cross validationLS-SVM

32 features 94.02 ± 0.18 90.85 ± 0.29 92.38 ± 0.18 99.62 ± 0.09 99.55 ± 0.08 99.58 ± 0.04 99.06 ± 0.05 98.84 ± 0.10 98.95 ± 0.05 99.94 ± 0.00 99.97 ± 0.03 99.96 ± 0.02PCA 91.39 ± 0.20 86.51 ± 0.16 88.79 ± 0.10 98.36 ± 0.10 97.96 ± 0.11 98.16 ± 0.07 99.83 ± 0.04 99.34 ± 0.04 99.59 ± 0.03 99.94 ± 0.02 99.20 ± 0.04 99.57 ± 0.03LDA 82.69 ± 0.11 73.08 ± 0.06 77.06 ± 0.07 99.00 ± 0.03 99.67 ± 0.02 99.33 ± 0.02 98.40 ± 0.05 98.76 ± 0.03 98.58 ± 0.03 99.61 ± 0.00 99.27 ± 0.03 99.44 ± 0.02

PNN 32 features 92.45 ± 1.73 88.40 ± 0.39 90.30 ± 0.67 99.90 ± 0.13 99.99 ± 0.02 99.95 ± 0.06 99.92 ± 0.14 99.96 ± 0.08 99.94 ± 0.07 99.83 ± 0.28 99.65 ± 0.05 99.74 ± 0.13PCA 92.61 ± 2.78 85.37 ± 1.97 88.54 ± 0.90 99.97 ± 0.03 99.04 ± 0.80 99.50 ± 0.41 99.93 ± 0.10 98.59 ± 1.12 99.24 ± 0.56 99.96 ± 0.05 97.75 ± 1.64 98.82 ± 0.86LDA 84.05 ± 0.45 72.22 ± 0.28 76.88 ± 0.08 98.84 ± 0.17 99.67 ± 0.10 99.25 ± 0.06 98.30 ± 0.05 98.85 ± 0.07 98.57 ± 0.02 99.67 ± 0.07 99.21 ± 0.10 99.44 ± 0.02

KNN 32 features 93.62 ± 1.54 87.51 ± 0.75 90.32 ± 0.62 99.96 ± 0.03 99.99 ± 0.03 99.97 ± 0.02 99.94 ± 0.02 99.91 ± 0.05 99.93 ± 0.03 99.96 ± 0.05 99.55 ± 0.22 99.76 ± 0.10PCA 92.14 ± 1.78 86.56 ± 0.50 89.13 ± 0.84 99.99 ± 0.03 99.44 ± 0.16 99.71 ± 0.08 99.96 ± 0.08 99.19 ± 0.13 99.57 ± 0.07 99.93 ± 0.05 98.23 ± 0.44 99.06 ± 0.22LDA 73.32 ± 4.07 69.76 ± 1.84 71.35 ± 2.81 99.09 ± 0.04 99.30 ± 0.16 99.20 ± 0.08 98.07 ± 0.19 98.40 ± 0.38 98.23 ± 0.26 99.45 ± 0.27 99.20 ± 0.13 99.32 ± 0.20

CART 32 features 83.44 ± 0.78 82.70 ± 0.73 83.06 ± 0.62 99.29 ± 0.15 99.21 ± 0.19 99.25 ± 0.11 99.73 ± 0.11 99.77 ± 0.10 99.75 ± 0.08 99.50 ± 0.13 99.06 ± 0.10 99.28 ± 0.07PCA 79.84 ± 0.82 79.06 ± 0.69 79.44 ± 0.68 99.15 ± 0.18 98.95 ± 0.25 99.05 ± 0.18 98.92 ± 0.12 98.77 ± 0.22 98.84 ± 0.09 95.94 ± 0.33 95.53 ± 0.42 95.73 ± 0.19LDA 68.91 ± 0.42 67.40 ± 0.35 68.12 ± 0.34 98.98 ± 0.07 99.17 ± 0.09 99.07 ± 0.04 97.81 ± 0.18 97.71 ± 0.21 97.76 ± 0.15 99.08 ± 0.08 99.10 ± 0.04 99.09 ± 0.04

Conventional validationLS-SVM

32 features 93.30 ± 1.01 90.43 ± 1.50 91.78 ± 0.55 99.53 ± 0.39 99.38 ± 0.27 99.46 ± 0.18 99.85 ± 0.18 100 ± 0.00 99.43 ± 0.09 99.81 ± 0.19 99.94 ± 0.10 99.87 ± 0.12PCA 90.73 ± 1.81 85.59 ± 0.70 87.96 ± 0.72 98.31 ± 0.77 97.79 ± 0.58 98.04 ± 0.54 99.83 ± 0.22 99.21 ± 0.26 99.52 ± 0.19 99.87 ± 0.11 98.92 ± 0.53 99.39 ± 0.30LDA 82.27 ± 0.80 73.24 ± 0.63 77.02 ± 0.64 99.00 ± 0.34 99.70 ± 0.27 99.35 ± 0.22 98.09 ± 0.42 98.88 ± 0.36 98.48 ± 0.28 99.51 ± 0.30 99.30 ± 0.35 99.40 ± 0.24

PNN 32 features 89.83 ± 0.43 87.26 ± 0.53 88.49 ± 0.40 99.79 ± 0.18 99.95 ± 0.04 99.87 ± 0.08 99.74 ± 0.17 99.94 ± 0.04 99.84 ± 0.09 99.56 ± 0.44 99.57 ± 0.07 99.56 ± 0.22PCA 88.64 ± 0.56 85.82 ± 0.45 87.16 ± 0.39 99.91 ± 0.05 99.53 ± 0.06 99.72 ± 0.03 99.80 ± 0.07 99.25 ± 0.13 99.52 ± 0.05 99.81 ± 0.07 98.33 ± 0.17 99.06 ± 0.08LDA 83.32 ± 0.47 72.48 ± 0.30 76.83 ± 0.22 98.80 ± 0.14 99.69 ± 0.10 99.24 ± 0.06 98.24 ± 0.14 98.85 ± 0.12 98.54 ± 0.06 99.60 ± 0.07 99.30 ± 0.11 99.45 ± 0.06

KNN 32 features 93.03 ± 1.70 86.60 ± 1.32 89.50 ± 0.81 99.96 ± 0.03 99.93 ± 0.05 99.95 ± 0.03 99.34 ± 0.87 99.08 ± 0.67 99.21 ± 0.74 99.95 ± 0.05 99.41 ± 0.20 99.68 ± 0.10PCA 92.35 ± 1.77 86.19 ± 1.09 88.98 ± 0.91 99.97 ± 0.03 99.63 ± 0.33 99.80 ± 0.16 99.05 ± 0.95 98.72 ± 0.51 98.88 ± 0.70 99.94 ± 0.05 98.58 ± 0.94 99.24 ± 0.49LDA 85.95 ± 1.57 80.70 ± 1.00 83.07 ± 0.67 99.68 ± 0.43 99.50 ± 0.35 99.59 ± 0.34 98.14 ± 0.23 98.32 ± 0.43 98.22 ± 0.26 99.78 ± 0.26 98.80 ± 0.83 99.28 ± 0.41

CART 32 features 83.16 ± 0.33 82.73 ± 0.55 82.92 ± 0.25 99.33 ± 0.14 99.29 ± 0.08 99.31 ± 0.10 99.71 ± 0.08 99.70 ± 0.09 99.71 ± 0.06 99.38 ± 0.17 99.04 ± 0.12 99.21 ± 0.13PCA 79.14 ± 0.56 78.35 ± 0.32 78.71 ± 0.38 99.09 ± 0.24 98.85 ± 0.08 98.97 ± 0.11 98.85 ± 0.06 98.66 ± 0.25 98.75 ± 0.09 95.57 ± 0.33 95.21 ± 0.38 95.38 ± 0.35LDA 69.30 ± 0.63 67.71 ± 0.30 68.44 ± 0.40 99.08 ± 0.17 99.11 ± 0.20 99.09 ± 0.03 97.84 ± 0.27 98.01 ± 0.20 97.92 ± 0.13 99.19 ± 0.06 99.13 ± 0.08 99.16 ± 0.01

Page 8: A hybrid expert system approach for telemonitoring of vocal fold pathology

M.

Hariharan

et al.

/ A

pplied Soft

Computing

13 (2013)

4148–4161

4155

Table 5Classification accuracies for MEEI database [entropy features, 10-fold cross validation and conventional validation].

Classifiers Entropy features Raw features Weighted features

KMC FCM SBC

SE SP ACC SE SP ACC SE SP ACC SE SP ACC

10-fold cross validationLS-SVM 32 features 94.31 ± 0.23 92.81 ± 0.13 93.55 ± 0.13 99.77 ± 0.05 99.77 ± 0.05 99.53 ± 0.05 99.78 ± 0.06 99.64 ± 0.05 99.71 ± 0.02 99.99 ± 0.03 99.78 ± 0.10 99.88 ± 0.04

PCA 91.68 ± 0.15 91.24 ± 0.17 91.46 ± 0.11 100 ± 0.00 100 ± 0.00 100 ± 0.00 99.45 ± 0.05 98.73 ± 0.08 99.09 ± 0.06 99.85 ± 0.00 98.22 ± 0.12 99.03 ± 0.06LDA 90.15 ± 0.04 87.46 ± 0.06 88.76 ± 0.04 98.28 ± 0.06 96.51 ± 0.00 97.38 ± 0.03 99.98 ± 0.03 100 ± 0.00 99.99 ± 0.02 99.02 ± 0.03 97.60 ± 0.05 98.30 ± 0.03

PNN 32 features 94.06 ± 2.79 91.58 ± 0.39 92.74 ± 1.45 98.92 ± 3.28 100.00 ± 0.00 99.40 ± 1.84 98.95 ± 3.19 100.00 ± 0.00 99.42 ± 1.78 98.91 ± 2.87 99.00 ± 0.16 98.91 ± 1.55PCA 93.33 ± 0.76 89.16 ± 0.64 91.15 ± 0.68 99.86 ± 0.25 100.00 ± 0.00 99.93 ± 0.13 99.89 ± 0.20 100.00 ± 0.00 99.95 ± 0.10 99.76 ± 0.17 97.53 ± 1.02 98.62 ± 0.48LDA 91.05 ± 0.62 86.65 ± 0.49 88.73 ± 0.07 99.91 ± 0.11 100.00 ± 0.00 99.95 ± 0.05 99.88 ± 0.14 100.00 ± 0.00 99.94 ± 0.07 99.03 ± 0.15 97.45 ± 0.36 98.23 ± 0.14

KNN 32 features 95.43 ± 0.52 90.22 ± 0.70 92.67 ± 0.50 99.77 ± 0.11 97.07 ± 0.56 98.38 ± 0.29 100.00 ± 0.00 99.99 ± 0.03 99.99 ± 0.02 99.99 ± 0.02 98.63 ± 0.25 99.30 ± 0.12PCA 93.91 ± 0.81 89.03 ± 0.44 91.33 ± 0.52 99.58 ± 0.11 96.54 ± 0.63 98.02 ± 0.32 100.00 ± 0.00 100.00 ± 0.00 100.00 ± 0.00 99.71 ± 0.13 98.19 ± 0.17 98.94 ± 0.10LDA 87.50 ± 2.64 86.06 ± 1.77 86.76 ± 2.19 97.50 ± 0.75 96.22 ± 0.12 96.85 ± 0.39 100.00 ± 0.00 99.97 ± 0.04 99.99 ± 0.02 98.21 ± 0.43 97.90 ± 0.26 98.06 ± 0.33

CART 32 features 87.27 ± 0.60 86.37 ± 0.82 86.81 ± 0.65 99.84 ± 0.03 99.77 ± 0.06 99.80 ± 0.04 99.94 ± 0.06 99.84 ± 0.12 99.89 ± 0.06 99.73 ± 0.06 99.85 ± 0.00 99.79 ± 0.03PCA 86.91 ± 0.78 85.63 ± 0.74 86.26 ± 0.72 99.87 ± 0.03 99.91 ± 0.12 99.89 ± 0.07 99.68 ± 0.04 99.83 ± 0.06 99.75 ± 0.03 96.51 ± 0.24 96.74 ± 0.25 96.63 ± 0.18LDA 84.38 ± 0.30 83.48 ± 0.36 83.92 ± 0.32 100.00 ± 0.00 100.00 ± 0.00 100.00 ± 0.00 100.00 ± 0.00 100.00 ± 0.00 100.00 ± 0.00 97.85 ± 0.18 97.43 ± 0.23 97.64 ± 0.14

Conventional validationLS-SVM 32 features 94.48 ± 1.47 92.18 ± 1.24 93.29 ± 1.03 99.74 ± 0.29 99.45 ± 0.38 99.59 ± 0.22 99.69 ± 0.36 99.35 ± 0.64 99.52 ± 0.33 99.95 ± 0.10 99.76 ± 0.30 99.86 ± 0.15

PCA 91.87 ± 0.88 90.62 ± 1.16 91.22 ± 0.66 99.49 ± 0.26 98.85 ± 0.66 99.17 ± 0.32 99.30 ± 0.52 98.52 ± 0.64 98.90 ± 0.32 99.83 ± 0.12 98.03 ± 0.60 98.91 ± 0.29LDA 89.73 ± 1.08 87.32 ± 0.64 88.48 ± 0.61 98.18 ± 0.68 96.12 ± 0.65 97.13 ± 0.48 99.95 ± 0.10 100 ± 0.00 99.98 ± 0.05 98.78 ± 0.46 97.54 ± 0.54 98.15 ± 0.36

PNN 32 features 93.47 ± 3.30 91.13 ± 0.44 92.20 ± 1.75 99.93 ± 0.11 100.00 ± 0.01 99.96 ± 0.06 98.04 ± 3.04 97.06 ± 0.37 97.49 ± 1.58 98.79 ± 3.22 98.87 ± 0.12 98.78 ± 1.77PCA 92.92 ± 0.84 88.92 ± 0.61 90.81 ± 0.67 99.88 ± 0.23 100.00 ± 0.00 99.94 ± 0.12 98.67 ± 0.70 96.67 ± 1.07 97.64 ± 0.88 99.74 ± 0.17 97.55 ± 1.01 98.61 ± 0.49LDA 91.17 ± 0.61 86.78 ± 0.66 88.84 ± 0.31 99.91 ± 0.11 100.00 ± 0.01 99.95 ± 0.06 97.07 ± 0.47 95.14 ± 0.42 96.08 ± 0.13 98.97 ± 0.17 97.49 ± 0.19 98.21 ± 0.07

KNN 32 features 91.95 ± 3.42 88.13 ± 1.99 89.92 ± 2.61 100.00 ± 0.00 100.00 ± 0.01 100.00 ± 0.01 100.00 ± 0.00 99.97 ± 0.03 99.98 ± 0.01 99.91 ± 0.07 98.48 ± 0.35 99.19 ± 0.17PCA 90.57 ± 3.42 87.24 ± 1.80 88.80 ± 2.50 100.00 ± 0.00 99.99 ± 0.01 100.00 ± 0.01 100.00 ± 0.00 99.98 ± 0.03 99.99 ± 0.01 99.81 ± 0.16 98.23 ± 0.42 99.01 ± 0.24LDA 87.68 ± 2.27 86.02 ± 1.75 86.81 ± 1.98 100.00 ± 0.00 99.99 ± 0.02 99.99 ± 0.01 100.00 ± 0.00 99.98 ± 0.03 99.99 ± 0.02 99.28 ± 0.82 98.08 ± 0.41 98.67 ± 0.5432 features 87.07 ± 0.45 85.95 ± 0.55 86.48 ± 0.21 99.64 ± 0.10 99.66 ± 0.11 99.65 ± 0.07 99.64 ± 0.07 99.62 ± 0.09 99.63 ± 0.07 99.67 ± 0.13 99.62 ± 0.16 99.64 ± 0.09

CART PCA 84.23 ± 0.52 82.89 ± 0.48 83.50 ± 0.27 99.50 ± 0.11 99.36 ± 0.12 99.43 ± 0.10 99.35 ± 0.13 99.14 ± 0.13 99.24 ± 0.08 94.16 ± 0.43 93.71 ± 0.42 93.92 ± 0.31LDA 85.21 ± 0.46 84.27 ± 0.24 84.71 ± 0.23 99.75 ± 0.08 99.64 ± 0.05 99.69 ± 0.04 99.72 ± 0.11 99.47 ± 0.10 99.59 ± 0.06 98.41 ± 0.12 97.96 ± 0.10 98.18 ± 0.10

Page 9: A hybrid expert system approach for telemonitoring of vocal fold pathology

4156

M.

Hariharan

et al.

/ A

pplied Soft

Computing

13 (2013)

4148–4161

Table 6Classification accuracies for MAPACI database [entropy features, 10-fold cross validation and conventional validation].

Classifiers Entropy features Raw features Weighted features

KMC FCM SBC

SE SP ACC SE SP ACC SE SP ACC SE SP ACC

10-fold cross validationLS-SVM 32 features 95.38 ± 0.33 92.93 ± 0.19 94.12 ± 0.24 99.36 ± 0.08 97.94 ± 0.09 98.64 ± 0.07 98.42 ± 0.11 96.16 ± 0.12 97.27 ± 0.10 100 ± 0.00 99.74 ± 0.00 99.87 ± 0.00

PCA 90.37 ± 0.14 84.36 ± 0.18 87.12 ± 0.13 99.85 ± 0.03 99.92 ± 0.03 99.89 ± 0.02 99.88 ± 0.02 99.76 ± 0.03 99.82 ± 0.02 99.87 ± 0.02 100 ± 0.00 99.93 ± 0.01LDA 82.98 ± 0.10 73.67 ± 0.08 77.56 ± 0.08 98.98 ± 0.04 99.17 ± 0.02 99.08 ± 0.01 98.60 ± 0.04 98.83 ± 0.04 98.71 ± 0.03 99.94 ± 0.00 100 ± 0.00 99.97 ± 0.00

PNN 32 features 93.47 ± 2.02 89.61 ± 0.63 91.42 ± 0.75 99.57 ± 1.32 99.98 ± 0.03 99.77 ± 0.68 99.53 ± 1.41 99.97 ± 0.03 99.74 ± 0.73 99.57 ± 1.34 100.00 ± 0.00 99.78 ± 0.70PCA 92.29 ± 2.81 83.03 ± 3.09 86.90 ± 1.43 99.60 ± 0.10 99.65 ± 0.11 99.63 ± 0.08 99.67 ± 0.08 99.40 ± 0.13 99.53 ± 0.05 99.14 ± 0.71 99.96 ± 0.05 99.54 ± 0.34LDA 86.53 ± 0.42 72.45 ± 0.37 77.80 ± 0.16 99.05 ± 0.05 99.16 ± 0.02 99.10 ± 0.02 98.74 ± 0.12 98.72 ± 0.08 98.73 ± 0.04 99.94 ± 0.00 99.97 ± 0.03 99.96 ± 0.02

KNN 32 features 94.35 ± 1.08 89.99 ± 0.53 92.05 ± 0.65 100.00 ± 0.00 99.94 ± 0.03 99.97 ± 0.02 99.99 ± 0.02 99.90 ± 0.05 99.95 ± 0.03 100.00 ± 0.00 100.00 ± 0.00 100.00 ± 0.00PCA 91.05 ± 1.87 85.32 ± 0.52 87.94 ± 0.67 99.61 ± 0.06 99.68 ± 0.09 99.64 ± 0.05 99.59 ± 0.04 99.44 ± 0.13 99.52 ± 0.07 99.56 ± 0.15 99.93 ± 0.05 99.74 ± 0.08LDA 75.27 ± 4.29 71.29 ± 1.07 73.03 ± 2.47 98.92 ± 0.24 98.85 ± 0.32 98.88 ± 0.27 98.61 ± 0.35 98.45 ± 0.24 98.53 ± 0.29 99.94 ± 0.00 99.94 ± 0.00 99.94 ± 0.00

CART 32 features 84.51 ± 0.55 83.20 ± 0.49 83.84 ± 0.37 99.58 ± 0.15 99.60 ± 0.12 99.59 ± 0.09 99.20 ± 0.09 99.31 ± 0.21 99.26 ± 0.12 99.33 ± 0.11 99.26 ± 0.19 99.29 ± 0.13PCA 81.18 ± 1.03 80.43 ± 0.77 80.79 ± 0.67 99.68 ± 0.12 99.25 ± 0.13 99.46 ± 0.11 99.60 ± 0.16 99.41 ± 0.11 99.50 ± 0.09 99.22 ± 0.15 99.20 ± 0.10 99.21 ± 0.07LDA 71.01 ± 0.47 69.51 ± 0.41 70.23 ± 0.43 98.56 ± 0.12 98.43 ± 0.09 98.49 ± 0.08 98.36 ± 0.17 98.37 ± 0.12 98.37 ± 0.08 99.94 ± 0.00 99.94 ± 0.00 99.94 ± 0.00

Conventional validationLS-SVM 32 features 94.74 ± 0.47 92.55 ± 1.01 93.61 ± 0.50 100 ± 0.00 99.68 ± 0.21 99.84 ± 0.10 98.15 ± 0.76 96.33 ± 0.56 97.22 ± 0.44 100 ± 0.00 99.64 ± 0.33 99.82 ± 0.17

PCA 90.06 ± 1.02 84.20 ± 1.18 86.88 ± 0.71 99.87 ± 0.15 99.91 ± 0.11 99.89 ± 0.10 99.77 ± 0.16 99.77 ± 0.21 99.76 ± 0.14 99.74 ± 0.13 99.98 ± 0.07 99.86 ± 0.09LDA 82.16 ± 1.18 73.58 ± 1.21 77.19 ± 0.93 98.98 ± 0.34 99.36 ± 0.32 99.17 ± 0.19 98.80 ± 0.47 98.64 ± 0.56 98.72 ± 0.31 99.96 ± 0.09 100 ± 0.00 99.98 ± 0.05

PNN 32 features 90.93 ± 1.62 89.10 ± 0.45 89.96 ± 0.80 98.95 ± 1.54 99.97 ± 0.03 99.44 ± 0.81 98.89 ± 1.58 99.95 ± 0.03 99.40 ± 0.83 98.87 ± 1.64 100.00 ± 0.00 99.42 ± 0.86PCA 87.58 ± 0.48 84.86 ± 0.36 86.15 ± 0.33 99.60 ± 0.12 99.64 ± 0.07 99.62 ± 0.06 99.58 ± 0.14 99.51 ± 0.10 99.55 ± 0.08 99.69 ± 0.08 99.85 ± 0.05 99.77 ± 0.05LDA 85.61 ± 0.35 73.20 ± 0.33 78.08 ± 0.25 98.98 ± 0.11 99.10 ± 0.12 99.04 ± 0.08 98.80 ± 0.13 98.69 ± 0.15 98.75 ± 0.09 99.93 ± 0.03 99.94 ± 0.04 99.93 ± 0.03

KNN 32 features 93.83 ± 0.97 89.38 ± 0.70 91.47 ± 0.61 99.49 ± 0.49 99.48 ± 0.49 99.48 ± 0.48 99.98 ± 0.03 99.88 ± 0.05 99.93 ± 0.03 99.80 ± 0.26 99.96 ± 0.04 99.88 ± 0.14PCA 87.58 ± 0.48 84.86 ± 0.36 86.15 ± 0.33 99.60 ± 0.12 99.64 ± 0.07 99.62 ± 0.06 99.58 ± 0.14 99.51 ± 0.10 99.55 ± 0.08 99.69 ± 0.08 99.85 ± 0.05 99.77 ± 0.05LDA 82.96 ± 8.46 78.03 ± 6.81 80.23 ± 7.49 99.24 ± 0.40 99.24 ± 0.44 99.24 ± 0.41 99.77 ± 0.22 99.60 ± 0.32 99.68 ± 0.26 99.70 ± 0.27 99.94 ± 0.04 99.82 ± 0.14

CART 32 features 83.25 ± 0.61 82.49 ± 0.32 82.83 ± 0.29 99.48 ± 0.09 99.44 ± 0.08 99.46 ± 0.03 99.48 ± 0.09 99.44 ± 0.08 99.46 ± 0.03 99.23 ± 0.10 99.33 ± 0.12 99.28 ± 0.07PCA 80.79 ± 0.51 79.66 ± 0.38 80.18 ± 0.24 99.40 ± 0.26 99.09 ± 0.25 99.24 ± 0.10 99.41 ± 0.15 99.16 ± 0.12 99.28 ± 0.13 99.50 ± 0.18 99.18 ± 0.10 99.34 ± 0.12LDA 71.48 ± 0.47 69.84 ± 0.44 70.60 ± 0.41 98.67 ± 0.24 98.42 ± 0.14 98.54 ± 0.11 98.72 ± 0.17 98.53 ± 0.30 98.62 ± 0.18 98.40 ± 0.06 98.38 ± 0.21 98.39 ± 0.11

Page 10: A hybrid expert system approach for telemonitoring of vocal fold pathology

M. Hariharan et al. / Applied Soft Computing 13 (2013) 4148–4161 4157

Fig. 6. The class distribution of raw and weighted LDA entropy feature: (a) raw LDA feature, (b) KMC, (c) FCM and (d) SBC.

Table 7Computational time (s) for classification using LS-SVM, PNN, kNN and CART [MEEI database].

Classifiers Different experiments MEEI database

Energy features Entropy features

10-fold Cross validation Conventional validation 10-fold cross validation Conventional validation

Raw KMC FCM SBC Raw KMC FCM SBC Raw KMC FCM SBC Raw KMC FCM SBC

LS-SVM 32 features 1.34 1.15 1.05 1.13 0.66 0.60 0.68 0.59 1.37 1.15 1.13 1.10 0.70 0.62 0.59 0.59PCA 1.31 1.13 1.10 1.06 0.60 0.64 0.59 0.57 1.35 1.05 1.11 1.04 0.61 0.60 0.58 0.57LDA 1.31 1.15 1.05 1.01 0.58 0.59 0.58 0.56 1.27 1.28 1.11 1.10 0.58 0.63 0.61 0.59

PNN 32 features 2.22 2.37 2.28 2.21 9.49 9.20 10.18 10.10 2.54 2.13 2.21 2.22 9.32 7.34 9.51 9.38PCA 1.87 1.64 1.56 1.53 8.09 7.83 8.37 8.48 1.77 1.55 1.62 1.54 7.91 7.84 7.92 7.92LDA 1.57 1.42 1.30 1.29 7.42 7.28 8.23 7.42 1.47 1.37 1.38 1.27 7.31 7.25 7.48 7.47

kNN 32 features 0.61 0.60 0.60 0.60 6.43 6.56 7.82 6.55 0.62 0.61 0.61 0.61 6.29 6.24 6.51 6.57PCA 0.58 0.53 0.54 0.56 6.38 6.41 7.11 6.60 0.61 0.57 0.55 0.54 6.17 6.12 6.41 6.47LDA 0.41 0.40 0.40 0.39 6.20 6.23 6.70 5.96 0.42 0.40 0.38 0.38 5.89 6.00 6.24 6.28

.15

.94

.83

mf

Y

wdl

W

B

CART 32 features 2.30 0.67 0.75 0.83 7.35 6PCA 1.24 0.49 0.56 0.85 6.49 5LDA 0.74 0.42 0.38 0.46 6.09 5

aximize the ratio of the between and within class scatters of theeature set as shown in the following equation [38,45,46].

opt = arg maxy

∣∣YTSbY∣∣∣∣YTSwY∣∣ = [y1, y2, · · ·, yP] (3)

here {yi|1 ≤ i ≤ P} are the LDA subspace base vectors, P is theimension of the subspaces Sb and Sw are represented by the fol-

owing equations:

ithin class scatter Sw =c∑

i=1

∑xk ∈ Xi

(xk − �i)(xk − �i)T (4)

etween class scatter Sb =c∑

i=1

Ni(�i − �)(�i − �)T (5)

6.23 7.19 2.31 0.63 0.64 0.85 7.39 6.04 6.11 6.216.02 6.20 1.12 0.44 0.42 0.70 6.41 5.90 5.89 6.125.85 5.85 0.74 0.36 0.36 0.42 6.08 5.86 5.81 5.88

where c is the number of classes. x ∈ RN is a data sample. Xiis the set of samples with class label i. �i is the mean for the allthe samples with the class label i. Ni is the number of samples inthe class i. Using the transformation matrix Y, the between-classscattering was maximized whereas the within-class scattering wasminimized and hence LDA seeks to reduce dimensionality whilepreserving as much of the class discriminative power as possible.LDA feature reduction method was used to map the thirty twoenergy/entropy feature space into a one-dimensional feature space(1 feature) based on the criterion given in the Eq. (3).

3.4. Classifiers

In this work, four different classifiers were used to study theefficiency of the proposed features and their brief descriptions weregiven in this section.

Page 11: A hybrid expert system approach for telemonitoring of vocal fold pathology

4158 M. Hariharan et al. / Applied Soft Computing 13 (2013) 4148–4161

Table 8Computational time(s) for classification using LS-SVM, PNN, kNN and CART [MAPACI speech pathology database].

Classifiers Different experiments MAPACI speech pathology database

Energy features Entropy features

10-fold cross validation Conventional validation 10-fold cross validation Conventional validation

Raw KMC FCM SBC Raw KMC FCM SBC Raw KMC FCM SBC Raw KMC FCM SBC

LS-SVM 32 features 1.30 1.34 1.31 1.40 0.77 0.76 0.78 0.76 1.45 1.53 1.38 1.48 0.95 0.76 0.75 0.73PCA 1.47 1.52 1.40 1.34 0.90 0.74 0.77 0.76 1.33 1.52 1.39 1.33 0.92 0.74 0.74 0.76LDA 1.26 1.32 1.27 1.28 0.87 0.72 0.72 0.73 1.48 1.47 1.30 1.25 0.88 0.72 0.72 0.74

PNN 32 features 2.47 2.63 2.48 2.37 10.17 10.14 10.10 10.14 2.50 2.51 2.46 2.51 10.22 10.03 10.11 10.06PCA 1.68 1.81 1.68 1.69 8.48 8.23 8.28 8.49 1.77 1.76 1.68 1.67 8.48 8.27 8.46 8.46LDA 1.43 1.49 1.36 1.43 7.75 7.61 7.93 7.77 1.44 1.38 1.35 1.36 7.83 7.65 7.82 7.83

kNN 32 features 0.69 0.67 0.68 0.68 7.15 7.02 7.04 7.66 0.68 0.69 0.67 0.69 7.04 6.70 7.03 7.54PCA 0.70 0.61 0.63 0.66 7.08 6.91 6.79 7.41 0.69 0.62 0.68 0.59 6.63 6.55 7.20 7.48LDA 0.40 0.41 0.40 0.41 6.81 6.69 6.41 7.17 0.40 0.39 0.42 0.40 6.26 6.26 7.52 7.17

CART 32 features 2.91 0.93 0.90 1.13 8.56 6.60 6.80 7.02 2.90 0.95 1.11 0.81 8.53 6.77 6.77 6.85.37

.23

3

ur[Irsnb[ekawrLp

3

nifnaktTa

3

Bsstvoacfiw

PCA 1.49 0.61 0.62 0.85 7.21 6LDA 1.08 0.39 0.43 0.42 6.82 6

.4.1. LS-SVMSupport vector machine based classifier has been selected and

sed in this work since it always gives higher classification accu-acy for the classification of normal and pathological voice samples5,7,16] and also for the different practical applications [24,47–49].n this work, LS-SVM was used to gauge the effectiveness of theaw and weighted features in distinguishing two classes of voiceamples such as normal or pathological ones. Three kinds of ker-el functions such as linear kernel, multilayer kernel and radialasis function (RBF) kernel were normally used by researchers5,7,16,24,47–49]. RBF kernel function was used since it gives anxcellent generalization and low computational cost. In the RBFernel, there were two regularization parameters namely � (gam)nd �2 (sig2, the squared bandwidth of the RBF kernel). Their valuesere optimally found using empirical study to obtain a better accu-

acy. In this paper, classification algorithm was implemented usingS-SVMLab toolbox [50–52] to perform classification of normal andathological voice samples.

.4.2. kNNkNN is a instance based and non-parametric classifier. This tech-

ique does not use any assumptions on the data distribution andt classifies the objects based on closest training examples in theeature space [30]. An object is classified by a majority vote of itseighbor, with the object being assigned to the class most commonmongst its k nearest neighbors, where k is a positive integer. In the-nearest neighbor algorithm, the classification of a new test fea-ure vector is determined by the class of its k-nearest neighbors.he appropriate value for k was fixed during simulation between 1nd 10.

.4.3. Probabilistic neural networkSpecht has proposed the probabilistic neural net based on

ayesian classification and classical estimators for probability den-ity function. PNN was used in many applications, since its highpeed of learning, simple learning rule and requires single passraining, and new training patterns can be incorporated into a pre-iously trained classifier quite easily [11,12,31,38]. PNN comprisesf four layers such as input layer, pattern layer, summation layernd output layer. Neurons of all the four layers are fully inter-

onnected and the pattern neurons are activated by exponentialunction. The pattern neuron computes distances from the testnput vector to the training input vectors and produces a vector

hose elements indicate how close the test input is to a training

6.58 7.02 1.42 0.56 0.56 0.54 7.18 6.42 6.40 6.416.33 6.39 1.04 0.41 0.44 0.36 6.84 6.31 6.24 6.34

input. The summation unit sums these contributions for each classof inputs and produces a net output which is a vector of prob-abilities. From the maximum of these probabilities, output unitsproduce a 1 for that class and a 0 for the other classes using com-pete transfer function. The performance of the PNN classifier highlydepends upon the smoothing parameter or spread factor (�). Basedon the experimental investigations, the � value was varied between0.01 and 0.1 in steps of 0.01.

3.4.4. CARTBreiman et al. have developed CART model and it is tree-building

technique and found to be effective in producing better decisionrules from the set of observations described in terms of featuresand class labels [33]. The observations are successfully separatedinto two subsets based on the features significantly with respectto the class labels [32–34]. CART produces a tree-structured modelusing recursive binary partitioning and it is constructed by eithersplitting or not splitting each on the tree into two daughter nodesusing training set. To gauge the goodness of a potential split, theGini impurity measure was used in this work. When this impuritymeasure is at maximum, a node is equally divided among all theclasses, whereas the node is all one class for an impurity measureof zero [32–34]. CART is a non-parametric procedure and a naturalfit for prediction of two classes of voice signals with the featureset chosen for this study. All the algorithms were developed underMATLAB environment using a LAPTOP of Intel Core i7-2670 QM(2.2 GHz) with 4 GB RAM.

4. Results and discussions

Raw and weighted features were computed using featureextraction and feature weighting methods. Three different exper-iments were conducted such as classification using raw andweighted original 32 energy/entropy features, PCA energy/entropyfeatures (10 principal components) and LDA energy/entropy fea-ture (1 feature). In order to gauge the effectiveness of the proposedfeatures, we performed both conventional validation and 10-foldcross validation. In 10-fold cross validation, the feature set wasdivided randomly into 10 sets and training and testing wererepeated for 10 times. Overall accuracy was obtained from the aver-

age of the 10 iterations. In 10-fold cross validation, the standarddeviation of the classification results was less than the conventionalvalidation method and this can be seen from the Tables 3 and 4.In the conventional validation method, 70% of data were used for
Page 12: A hybrid expert system approach for telemonitoring of vocal fold pathology

ft Com

tg

mwuwspsaw

S

S

O

o

4

Pvosrtpb

4

tfa9fi8vt9C9fiotsbt

taf

4

t

M. Hariharan et al. / Applied So

raining and remaining 30% of data were used for testing. Table 2ives the details of training and testing set.

In order to gauge the classifier performance, three performanceeasures namely sensitivity, specificity, and the overall accuracyere considered. These measures were calculated from the meas-res namely true positive (TP, the classifier classified as pathologyhen pathological samples are present), true negative (TN, the clas-

ifier classified as normal when normal samples are present), falseositive (FP, the classifier classified as pathological when normalamples are present), and false negative (FN, the classifier classifieds normal when pathological samples are present). These measuresere calculated using the Eqs. (6)–(8).

ensitivity(SE) = TPTP + FN

(6)

pecificity(SP) = TNTN + FP

(7)

verall accuracy(ACC) = TP + TNTP + TN + FP + FN

(8)

The following section gives a detailed description of the resultsbtained using raw and weighted features.

.1. Class distribution of the raw and weighted features

Fig. 5(a)–(d) depicts the class distribution of raw and weightedCA entropy features for the compressed voice samples of MEEIoice disorder database. Fig. 6(a)–(d) exhibits the class distributionf raw and weighted LDA entropy feature for the compressed voiceamples of MEEI voice disorder database. The class distribution ofaw features of both databases has many overlap among the fea-ures. The given figures have demonstrated that the discriminativeower of the raw PCA and LDA entropy features has been increasedy using clustering based feature weighting methods.

.2. Classification results for MEEI database

Tables 3 and 5 shows the results of energy and entropy fea-ures for both 10-fold cross validation and conventional validationor MEEI database. From the results, it can be seen that theverage classification accuracy of 32 raw energy features was2.24 ± 0.24% for LS-SVM, 89.54 ± 1.44% for PNN, 89.82 ± 0.28%or kNN, and 86.97 ± 0.20% for CART using 10-fold cross val-dation and 91.51 ± 1.97% for LS-SVM, 88.94 ± 1.37% for PNN,9.35 ± 0.36% for kNN, and 86.48 ± 0.21% for CART using con-entional validation. Similarly for the 32 raw entropy features,he average classification accuracy was 93.55 ± 0.13% for LS-SVM,2.74 ± 1.45% for PNN, 92.67 ± 0.50% for kNN, 86.81 ± 0.65%, and forART using 10-fold cross validation and 93.29 ± 1.03% for LS-SVM,2.20 ± 1.75% for PNN, 89.92 ± 2.61% for kNN, and 86.48 ± 0.21%or CART using conventional validation. After employing cluster-ng based feature weighting methods, the discriminative abilityf the raw energy/entropy features were increased which inurn improves the average classification accuracy to 99%. Sen-itivity and specificity results were also improved to 99% foroth 10-fold cross validation and conventional validation respec-ively.

For the MEEI voice disorder database, FCM clustering based fea-ure weighting method gives a maximum average classificationccuracy of 99% under three different experiments with four dif-erent classifiers.

.3. Classification results for MAPACI speech pathology database

Tables 4 and 6 shows the results of energy and entropy fea-ures for both 10-fold cross validation and conventional validation

puting 13 (2013) 4148–4161 4159

for MAPACI speech pathology database. From the results, it canbe inferred that the average classification accuracy of 32 rawenergy features was 92.38 ± 0.18% for LS-SVM, 90.30 ± 0.67% forPNN, 90.32 ± 0.62% for kNN, and 83.06 ± 0.62% for CART using 10-fold cross validation and 91.78 ± 0.55% for LS-SVM, 88.49 ± 0.40%for PNN, 89.50 ± 0.81% for kNN, and 82.92 ± 0.25% for CART usingconventional validation. Similarly for the 32 raw entropy features,the average classification accuracy was 94.12 ± 0.24% for LS-SVM,91.42 ± 0.75% for PNN, 92.05 ± 0.65% for kNN, and 83.84 ± 0.37%forCART using 10-foldcross validation and 93.61 ± 0.50% for LS-SVM,89.96 ± 0.80% for PNN, 91.47 ± 0.61% for kNN, and 82.83 ± 0.29%for CART using conventional validation. Clustering based featureweighting methods, not only helps to improve the discrimina-tive ability of the raw energy/entropy features but they also usedto increase the average classification accuracy to 99%. In thisstudy, PCA and LDA were also applied on the weighted energyor entropy features to obtain small set of uncorrelated features.We obtained 10 features from PCA and 1 feature from LDA. Fromthe results, it can be inferred that there was no significant per-formance loss (classification accuracy) after applying PCA and LDAon the features for LS-SVM, kNN, and PNN whereas the classifi-cation accuracy was decreased to 93% for CART. For the MAPACIspeech pathology database, FCM clustering based feature weight-ing method gives a maximum average classification accuracy of97% under three different experiments with four different classi-fiers.

In Ref. [26], authors have obtained the maximum average clas-sification accuracy of 89% for the voice samples compressed at abit rate of 8 kbps for MEEI database during frame-based analysis.After compression, the harmonic structure of voice samples wereseverely affected, which results in low quality features and henceits quality could be improved by means of the proposed method.

Tables 7 and 8 reports the average computational time in s dur-ing classification process with energy and entropy features. Directcomparison of our work with the previous works in literature can-not be performed, since most of the works in the literature usedhigh quality voice samples and they have not also reported aboutthe computational time during classification process. Though thevoice samples are of low quality, our suggested clustering basedfeature weighting and LS-SVM provides best classification accuracywith less computational time.

5. Conclusions

In the area of automatic detection of vocal fold pathology, theperformance of classification algorithm degrades smoothly dueto the irrelevant and noisy features which were extracted fromthe voice samples. The proposed study shows the effectivenessof the clustering based feature weighting/preprocessing methodsfor improving the discriminative ability of the raw features, whichimproves the performance of the classifiers. Among the three clus-tering based feature weighting methods, we obtained a maximumaverage classification accuracy of 99% for the compressed voicesamples of MEEI voice disorder database and above 97% for thecompressed voice samples of MAPACI speech pathology databaseusing FCM clustering feature weighting under all the experiments.Among four different classifiers, LS-SVM outperformed kNN, PNNand CART in terms of classification accuracy and time taken forclassification under all the experiments. From the results, it canbe concluded that the proposed feature weighting methods/datapreprocessing methods could be applied prior to classification pro-

cess to improve the robustness of the features. The proposed hybridexpert system approach could be used in remote diagnosis of vocalfold pathology from the voice samples which were transmitted overnarrow-band communications channels.
Page 13: A hybrid expert system approach for telemonitoring of vocal fold pathology

4 ft Com

A

Mt

R

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

160 M. Hariharan et al. / Applied So

cknowledgment

This work is done in part with data transferred from the databaseAPACI: http://www.mapaci.com. The authors would like to thank

he anonymous reviewers for their valuable comments.

eferences

[1] J.F. Battery, Statistics on Voice, Speech, and Language, 2012, Available:http://www.nidcd.nih.gov/health/statistics/pages/vsl.aspx (accessed10.08.12).

[2] J. Godino-Llorente, P. Gomez-Vilda, Automatic detection of voice impairmentsby means of short-term cepstral parameters and neural network based detec-tors, IEEE Transactions on Biomedical Engineering 51 (2004) 380–384.

[3] J. Godino-Llorente, P. Gomez-Vilda, M. Blanco-Velasco, Dimensionality reduc-tion of a pathological voice quality assessment system based on gaussianmixture models and short-term cepstral parameters, IEEE Transactions onBiomedical Engineering 53 (2006) 1943–1953.

[4] J.B. Alonso, J. de Leon, I. Alonso, M.A. Ferrer, Automatic detection of pathologiesin the voice by HOS based parameters, European Association for Speech signaland Image Processing 4 (2001) 275–284.

[5] R. Behroozmand, F. Almasganj, Optimal selection of wavelet-packet-based fea-tures using genetic algorithm in pathological assessment of patients’ speechsignal with unilateral vocal fold paralysis, Computers in Biology and Medicine(Elsevier) 37 (2007) 474–485.

[6] C. Crovato, A. Schuck, The use of wavelet packet transform and artificial neuralnetworks in analysis and classification of dysphonic voices, IEEE Transactionson Biomedical Engineering 54 (2007) 1898–1900.

[7] E. Fonseca, et al., Wavelet time-frequency analysis and least squares supportvector machines for the identification of voice disorders, Computers in Biologyand Medicine (Elsevier) 37 (2007) 571–578.

[8] H. Kasuya, Y. Endo, S. Saliu, Novel acoustic measurements of jitter and shimmercharacteristics from pathological voice, in: Proceedings from EUROSPEECH’93:The 3rd European Conference on Speech Communication and Technology,Berlin, Germany, 1993, pp. 1973–1976.

[9] H. Kasuya, S. Ogawa, K. Mashima, S. Ebihara, Normalized noise energy as anacoustic measure to evaluate pathologic voice, The Journal of the AcousticalSociety of America 80 (1986) 1329–1334.

10] K. Umapathy, S. Krishnan, V. Parsa, D. Jamieson, Discrimination of pathologi-cal voices using a time-frequency approach, IEEE Transactions on BiomedicalEngineering vol. 52 (2005) 421–430.

11] M. Hariharan, M. Paulraj, S. Yaacob, Detection of vocal fold paralysis andoedema using time-domain features and probabilistic neural network, Inter-national Journal of Biomedical Engineering and Technology 6 (2011) 46–57.

12] M. Hariharan, Y. Sazali, Time-domain features and probabilistic neural networkfor the detection of vocal fold pathology, Malaysian Journal of Computer Science23 (2010) 60–67.

13] M. Paulraj, Y. Sazali, S. Sivanandam, M. Hariharan, Improved back propagationneural network for the diagnosis of pathological voices, International Associa-tion for Modelling and Simulation Technique in Enterprise (AMSE Journal) 51(2008) 33–46.

14] M. Paulraj, S. Yaacob, M. Hariharan, Diagnosis of voice disorders using melscaled wpt and functional link neural network, International Journal of Biomed-ical Soft Computing and Human Sciences, Special Issue: Biosensors: DataAcquisition, Processing and Control 14 (2009) 57–62.

15] B. Boyanov, T. Ivanov, S. Hadjitodorov, G. Chollet, Robust hybrid pitch detector,IET Electronics Letters 29 (1993) 1924–1926.

16] J. Godino-Llorente, et al., Support vector machines applied to the detectionof voice disorders, Lecture Notes in Computer Science (Springer) 3817 (2005)219–230.

17] J. Jiang, Y. Zhang, Nonlinear dynamic analysis of speech from pathological sub-jects, Electronics Letters 38 (2002) 294–295.

18] M.A. Little, et al., Exploiting nonlinear recurrence and fractal scaling propertiesfor voice disorder detection, BioMedical Engineering OnLine 6 (2007) 1–19.

19] J. Gonzalez, T. Cervera, M.J. Llau, Acoustic analysis of pathological voices com-pressed with MPEG system, Journal of Voice 17 (2003) 126–139.

20] N. Sáenz-Lechón, et al., Effects of audio compression in automatic detectionof voice pathologies, IEEE Transactions on Biomedical Engineering 55 (2008)2831–2835.

21] R. Moran, R.B. Reilly, P. Lacy, Telephone based voice pathology assessment usingautomated speech analysis and VoiceXML, in: Proceedings of IEE ISSC’04: TheIrish Signals and Systems Conference, Belfast, Ireland, 2004, pp. 413–418.

22] R.J. Moran, R.B. Reilly, P. de Chazal, P.D. Lacy, Telephony-based voice pathologyassessment using automated speech analysis, IEEE Transactions on BiomedicalEngineering 53 (2006) 468–477.

23] Audacity® software for recording and editing sounds. Available:http://audacity.sourceforge.net/ (accessed 10.05.12).

24] P. Konar, P. Chattopadhyay, Bearing fault detection of induction motor using

wavelet and support vector machines (SVMs), Applied Soft Computing 11(2011) 4203–4211.

25] G. Ranganathan, R. Rangarajan, V. Bindhu, Estimation of heart rate signals formental stress assessment using neuro fuzzy technique, Applied Soft Computing12 (2012) 1978–1984.

puting 13 (2013) 4148–4161

26] A. Subasi, Classification of EMG signals using combined features and soft com-puting techniques, Applied Soft Computing 12 (2012) 2188–2198.

27] Kay Elemetrics Inc. Voice disorders database, version 1.03 [CD-ROM] [Online].Available: http://www.kaypentax.com/Product%20Info/CSL%20Options/4337/4337.htm

28] P. MAPACI. Voice Disorder Database [Online]. Available: http://www.mapaci.com/index-ingles.php

29] A. Aaccardo, F. Fabbro, E. Mumolo, Analysis of Normal and Pathological VoicesVia Short Time Fractal Dimension, in: Proceedings from IEEE EMBS’92: AnnualInternational Conference of the IEEE Engineering in Medicine and Biology Soci-ety, Paris, France, 1992, pp. 1270–1271.

30] T. Ritchings, M. McGillion, C. Moore, Objective assessment of pathologi-cal voice quality usingmulti-layer perceptrons, in: Proceedings from IEEEBMES/EMBS’99: The 1st Joint BMES/EMBS Conference in Engineering inMedicine and Biology, Atlanta, GA, USA, 1999, p. p. 925.

31] D.F. Specht, Probabilistic neural networks, Neural networks 3 (1990) 109–118.32] I. Kurt, M. Ture, A.T. Kurum, Comparing performances of logistic regression,

classification and regression tree, and neural networks for predicting coronaryartery disease, Expert Systems with Applications 34 (2008) 366–374.

33] L. Breiman, J. Friedman, R. Olshen, C.J. Stone, Classification and Regression Trees(1984).

34] M.A. Razi, K. Athappilly, A comparative predictive analysis of neural networks(NNs), nonlinear regression and classification and regression tree (CART) mod-els, Expert Systems with Applications 29 (2005) 65–74.

35] V. Parsa, D.G. Jamieson, Identification of pathological voices using glottalnoise measures, Journal of Speech, Language, and Hearing Research 43 (2000)469–485.

36] R.A. Prosek, A.A. Montgomery, B.E. Walden, D.B. Hawkins, An evaluation ofresidue features as correlates of voice disorders, Journal of CommunicationDisorders 20 (1987) 105–117.

37] C. Manfredi, Adaptive noise energy estimation in pathological speech signals,IEEE Transactions on Biomedical Engineering 47 (2000) 1538–1543.

38] M. Hariharan, et al., A comparative study of wavelet families for classi-fication of wrist motions, Computers & Electrical Engineering 38 (2012)1798–1807.

39] F. Latifoglu, K. Polat, S. Kara, S. Günes , Medical diagnosis of atherosclerosisfrom carotid artery doppler signals using principal component analysis (PCA),kNN based weighting pre-processing and artificial immune recognition system(AIRS), Journal of Biomedical Informatics 41 (2008) 15–23.

40] K. Polat, S. Günes , Principles component analysis, fuzzy weighting pre-processing and artificial immune recognition system based diagnostic systemfor diagnosis of lung cancer, Expert Systems with Applications 34 (2008)214–221.

41] S. Günes , K. Polat, S . Yosunkaya, Efficient sleep stage recognition system basedon EEG signal using k-means clustering based feature weighting, Expert Sys-tems with Applications 37 (2010) 7922–7928.

42] K. Polat, Application of attribute weighting method based on clustering centersto discrimination of linearly non-separable medical datasets, Journal of MedicalSystems 36 (4) (2012) 2657–2673.

43] A. Quinquis, Digital Signal Processing using MATLAB, John Wiley &Sons, NJ,USA, 2008.

44] L.I. Smith, A Tutorial on Principal Components Analysis, vol. 51, Cornell Univer-sity, USA, 2002, pp. 1–26.

45] P.N. Belhumeur, J.P. Hespanha, D.J. Kriegman, Eigenfaces vs. fisherfaces: recog-nition using class specific linear projection, IEEE Transactions on PatternAnalysis and Machine Intelligence 19 (1997) 711–720.

46] F. Tang, H. Tao, Fast linear discriminant analysis using binary bases, PatternRecognition Letters 28 (2007) 2209–2218.

47] N. Das, et al., A statistical–topological feature combination for recognition ofhandwritten numerals, Applied Soft Computing 12 (2012) 2486–2495.

48] S. Ekici, Support vector machines for classification and locating faults on trans-mission lines, Applied Soft Computing 12 (2012) 1650–1658.

49] T. Gruber, B. Meixner, J. Prosser, B. Sick, Handedness tests for preschool chil-dren: a novel approach based on graphics tablets and support vector machines,Applied Soft Computing 12 (2012) 1390–1398.

50] K. Pelckmans, et al., LS-SVMlab toolbox user’s guide, Pattern Recognition Letters24 (2003) 659–675.

51] J.A.K. Suykens, et al., Least squares support vector machines, World ScientificCo. Pte. Ltd, Singapore, 2002.

52] LS-SVMlab Toolbox. Available: http://www.esat.kuleuven.be/sista/lssvmlab/

M. Hariharan received Ph.D. degree in MechatronicEngineering from Universiti Malaysia Perlis (UniMAP),Malaysia. He is currently working as a Senior Lecturer inthe School of Mechatronic Engineering, UniMAP, Malaysia.He has published more than 50 papers in referred jour-nals and conferences. He has been a reviewer in variousjournals including Computers and Electrical Engineering,Computer Methods and Programs in Biomedicine, Artifi-cial Intelligence in Biomedicine, International Journal of

Phoniatrics, Speech Therapy and Communication Pathol-ogy, and Medical Engineering & Physics. His researchinterests include speech signal processing, biomedical sig-nal and image processing, and artificial intelligence. He is

a Member of IEEE.

Page 14: A hybrid expert system approach for telemonitoring of vocal fold pathology

ft Com

RUoi

robotics. He has published more than 70 papers in Journalsand 200 papers in Conference Proceedings. He receivedhis professional qualification as Charted Engineer from theEngineering Council, UK in 2005 and also a member of IET,UK since 2003.

M. Hariharan et al. / Applied So

Kemal Polat received Ph.D degrees in Electrical and Elec-tronics Engineering from Selcuk University in 2008. He iscurrently working as an Assistant Professor in Departmentof Electrical and Electronic Engineering, Faculty of Engi-neering and Architecture, Abant Izzet Baysal Universitysince September, 2011. His current research interests arebiomedical signal classification, sleep staging, biometrics,signal processing, data preprocessing, and classificationapplications. He has published 50 papers in Journals and18 papers in Conference Proceedings. He is the member ofeditorial board of Journal of Neural Computing and Appli-cations (SCI expanded).

. Sindhu is a Postgraduate student in the School of Microelectronic Engineering,niversiti Malaysia Perlis, Malaysia. She has completed her Bachelor of Technol-gy in Information Technology from Anna University in 2009. Her research interestncludes machine learning algorithms and hybrid optimization algorithms.

puting 13 (2013) 4148–4161 4161

Sazali Yaacob received his PhD in Control Engineer-ing from University of Sheffield, UK. He has successfullysupervised 8 PhD candidates and more than 20 MSc gra-duates through research mode. Currently, he has 10 PhDand 8 MSc candidates. His research interests are Con-trol, Modelling and Signal Processing with applications inthe fields of satellite, bio-medical, applied mechanics and