Cardiac disorder classification by heart sound signals using murmur ...

1 downloads 0 Views 596KB Size Report
Jan 27, 2012 - Cardiac disorder classification by heart sound signals using murmur likelihood and hidden Markov model state likelihood. C. Kwak O.-W. Kwon.
www.ietdl.org Published in IET Signal Processing Received on 2nd May 2011 Revised on 27th January 2012 doi: 10.1049/iet-spr.2011.0170

ISSN 1751-9675

Cardiac disorder classification by heart sound signals using murmur likelihood and hidden Markov model state likelihood C. Kwak O.-W. Kwon Department of Control and Robot Engineering, Chungbuk National University, Cheongju, South Korea E-mail: [email protected]

Abstract: This study proposes a new algorithm for cardiac disorder classification by heart sound signals. The algorithm consists of three steps: segmentation, likelihood computation and classification. In the segmentation step, the authors convert heart sound signals into mel-frequency cepstral coefficient features and then partition input signals into S1/S2 intervals by using a hidden Markov model (HMM). In the likelihood computation step, using only a period of heart sound signals, the authors compute the HMM ‘state’ likelihood and murmur likelihood. The ‘state’ likelihood is computed for each state of HMM-based cardiac disorder models, and the murmur likelihood is obtained by probabilistically modelling the energies of band-pass filtered signals for the heart pulse and murmur classes. In the classification step, the authors decided the final cardiac disorder by combining the state likelihood and the murmur likelihood by using a support vector machine. In computer experiments, the authors show that the proposed algorithm greatly improve classification accuracy by effectively reducing the classification errors for the cardiac disorder categories where the temporal murmur position plays an important role in detecting disorders.

1

Introduction

Cardiac disorders are critical and must be detected as soon as possible. Physicians should pay special attention to diagnose heart disorders by listening to heart sounds using a stethoscope. Nowadays two diagnosis methods are commonly used in hospital: electrocardiogram (ECG) and echocardiography. However, the ECG method takes long diagnosis time and high cost; the echocardiography method does not produce good images for overweight or lungdisease patients [1]. Cardiac disorder classification systems by heart sound signals of stethoscopes are to reduce diagnosis time and cost. On this background, this work aims to develop a simple cardiac disorder detector by using an electronic stethoscope so that it could help physicians discern the necessity of close diagnosis in a short time. The previous approaches for automatic classification of heart sound signals were based on artificial neural network (ANN) [2–6] or hidden Markov model (HMM) [7, 8]. Wavelet coefficients and multilayer perceptron-backpropagation (MLP-BP) neural networks were widely used as a feature vector and a pattern classifier, respectively [2–6]. ANNbased classification systems require an automatic segmentation algorithm in order to extract a feature vector from a single period of heart sound signals since ANNs cannot deal with variable-length input patterns. However the automatic segmentation algorithm often causes segmentation errors, which result in lower classification accuracy. On the contrary, the HMM-based classification method can operate on multiple periods of heart sound signals, but it has low 326 & The Institution of Engineering and Technology 2012

discriminability compared with the ANN-based method. In [7], the authors analysed the inter-subject variation in mean and variance of the envelopes and used the durationdependent HMM. A few researchers proposed improved algorithms to combine HMM and support vector machine (SVM) [9, 10], where the HMM ‘model’ likelihood was used as the input of the succeeding SVM which made the final decision on cardiac disorders. Various kinds of features have been proposed for cardiac disorder classification: wavelet coefficients [2 – 6] and temporal features [7, 8]. In perceiving cardiac disorders, murmur and click sound signals are known to be critical [11]. Accordingly, the cardiac disorder classification systems using spectral or cepstral features [4 – 6] have classification errors because murmur and click sound signals were smoothed. On the other hand, the cardiac disorder classification systems using temporal features [7, 8] often make classification errors for the cardiac disorders with similar murmur positions but different spectral characteristics. In this work, we propose an HMM-based heart sound segmentation algorithm to correctly spotting a single period of heart sound signals. Then we propose to add murmur likelihood as a temporal feature, which can probabilistically model the temporal position of murmur and click sounds. Finally, we use an SVM-based classifier to combine the HMM ‘state’ likelihood and the murmur likelihood reflecting spectral and temporal characteristics, respectively. The remainder of the paper is organised as follows: Section 2 describes spectral and temporal characteristics of heart IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326 –334 doi: 10.1049/iet-spr.2011.0170

www.ietdl.org sound signals useful for diagnosis. Section 3 describes the baseline system with envelope-based segmentation and cepstral features. Section 4 describes the proposed HMMbased segmentation algorithm and combining the HMM state likelihood and murmur likelihood by using SVM. Section 5 gives the experimental results and discussion, and finally Section 6 draws conclusion.

2

Heart sound signals

The heart has four chambers. The upper two are the right and left atrias. The lower two are the right and left ventricles. Blood is pumped through the chambers aided by four heart valves. As shown in Fig. 1, a period of heart sound signals make up the sequence of the first heart pulse (S1), systole, the second heart pulse (S2) and diastole intervals [11]. Fig. 1 shows a normal heart sound and an abnormal heart sound with aortic regurgitation. Normal heart sounds are produced by closure of the valves of the heart. Flow through the valves will affect the sound that the valves make. Thus, in situations of increased flow (e.g. exercise), the intensity of the heart sounds increases. In situations of low flow (e.g. shock), the intensity of the heart sounds decreases. The S1 pulse is normally the first heart sound heard. The S1 sound is best heard in the mitral area, and corresponds to closure of the mitral and tricuspid valves. A normal S1 sound is low pitched and of longer duration than S2. The S2 sound is normally heard second. The S2 sound is best heard over the aortic area, and corresponds to closure of the pulmonic and aortic valves. A normal S2 sound is higher pitched and of shorter duration than a normal S1 sound. The heart murmur and click sound is a cue to detect heart disorders [11]. The abnormal heart sounds have the murmur

or click sound in the systole and diastole intervals, but the normal heart sounds do not. Systolic murmurs occur between S1 and S2, and therefore are associated with mechanical systolic and ventricular ejection [11]. This type of murmur is caused by aortic stenosis (AS), or by mitral regurgitation (MR) or by ventricular septal defect (VSD). Diastolic murmurs occur after S2 and are therefore associated with ventricular relaxation and filling. This type of murmur is caused by aortic regurgitation (AR) or by mitral stenosis (MS). Mitral valve prolapse (MVP) is a disease where the click sound exists in the systole section. The murmur of AR is high pitched, often faint, puffing and blowing quality. The murmur of AS is low pitched, rough and rasping. The S1 of MR is absent, soft or buried in the systolic murmur, and the S2 of MR is widely split. The murmur of MS is low pitched and rumbling murmur is heard throughout diastole. The MVP is a disease where the click sound exists in the systole interval. The VSD is difficult to distinguish from MR because they are almost similar.

3

Baseline algorithm

The structure of the baseline algorithm is shown in Fig. 2. From continuous heart sound signals, we extract a single period of heart sound signals using an envelope-based segmentation algorithm. We obtain the cepstral characteristics from a single period of heart sound signals, then we use an HMM to classify the cardiac disorders. 3.1

Envelope-based segmentation

The heart sound segmentation algorithm [12] is based on the envelope calculated using the normalised average Shannon energy. The peaks of each part whose levels exceed the

Fig. 1 Period of heart sound signals a Normal heart sound b Abnormal heart sound with aortic regurgitation IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326– 334 doi: 10.1049/iet-spr.2011.0170

327

& The Institution of Engineering and Technology 2012

www.ietdl.org

Fig. 4 HMM structure of a single period of heart sound signals a Period of heart sound signals is segmented into 6 states b Single period of heart sound is modelled by a 6-state left-to-right HMM

Fig. 2 Baseline heart sound classification algorithm

diastolic murmur. Therefore a single period of heart sound is modelled by a six-state left-to-right HMM for each cardiac disorder. To train the acoustic model of HMM, we performed five iterations of the forward – backward algorithm [13] starting from initially uniform segmentation. The final cardiac disorder is determined as the one with the maximum log-likelihood.

4 Fig. 3 Example of segmentation results using envelope-based segmentation

Proposed algorithm

The structure of proposed algorithm is shown in Fig. 5. To extract a single period of heart sound signals, we used

threshold are picked up and assumed temporarily to be the first or the second heart sound. The extra peaks are rejected and the weak peaks are picked up by an iterative process. The starting points of S1 and S2 are identified and then a single period of heart sound signals is obtained. Fig. 3 shows an example of segmentation results with the cardiac disorder of MR using envelope-based segmentation. In case that the S2 sound is buried in murmur, the envelope-based segmentation algorithm cannot extract the starting points of S2. 3.2

Feature extraction and classification

From the continuous heart sound signals, we obtain a single period of heart sound signals by using the automatic S1/S2 segmentation algorithm. The heart sound signals have most of energy below 1 kHz. Hence, we convert the sampling rate down to 2 kHz for feature extraction. Parameters such as the frame/shift size, the order of mel-frequency cepstral coefficients (MFCCs), the number of states of HMM and the number of Gaussian distributions for each state are determined through computer experiments. We extract MFCCs with order M and log energy for each frame. The window size and the shift size were determined from experimental results. We add delta (D) and acceleration (DD) components [13] to consider the dynamic property of heart sound signals. Thus, a baseline feature vector comprises 3(M + 1) components. As shown in Fig. 4, a period of abnormal heart sound signals can be segmented into six intervals: S1, early systolic murmur and late systolic murmur; S2, early diastolic murmur and late 328 & The Institution of Engineering and Technology 2012

Fig. 5 Proposed heart sound classification algorithm IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326 –334 doi: 10.1049/iet-spr.2011.0170

www.ietdl.org an HMM-based segmentation algorithm. From a single period of heart sound signals, we obtain the MFCC features and compute the HMM state likelihood using the Viterbi algorithm. To exploit the temporal position characteristics of murmur signals, we compute the murmur likelihood for two sub-bands. Then SVM classifies the cardiac disorders by combining the HMM state likelihood and the murmur likelihood. 4.1

HMM-based segmentation

Heart sound signals are sampled at 8 kHz and digitised to 16-bit pulse-coded modulation data. The starting points of the S1/S2 pulses should be correctly detected because the position of murmur or click sounds is important in cardiac disorder classification. From continuous heart sound signals, we apply the Wiener filter to reduce the background noise. We use 2 three-state left-to-right HMMs, which ‘model S1 or S2 intervals’ and ‘murmur or silence intervals’. The same baseline features were used for S1/S2 segmentation as the previously described HMM-based classifier. The parameters of the two HMMs were learned by using a manually labelled database. Then, the starting points of S1 and S2 pulses of new input signals are determined by the Viterbi algorithm [13]. We partition a period of signals into S1-systole or S2-diastole intervals. Then we correct the ill-positioned starting points of S1/S2 pulses and complement the missed starting points by utilising the average duration of the intervals, which is obtained by averaging the duration of the intervals within a period of correct segments only. Here a correct segment means the one having two consecutive S1/ S2 intervals within a time period, which is reliably estimated by finding a cepstral peak within the possible time range [14]. Fig. 6 shows an example of automatic segmentation results, where murmur sound signals exist in the systole interval. Whereas the conventional envelopebased segmentation algorithm has missed a S2 starting point between 1150 and 1300 ms (see Fig. 3), the proposed HMM-based segmentation algorithm correctly adds the starting points of S2 and also corrects the starting points of S1 between 900 and 1050 ms. 4.2

HMM state likelihood

The cardiac disorder is decided on the position of murmur and click sound. A single period of heart sound signals is modelled by a six-state HMM, the state likelihood reflects the information of murmur positions. Therefore the state likelihood can be used as a cue to classify the cardiac disorder. For each input signal, the Viterbi segmentation algorithm is performed on the six-state HMM to obtain the state

segmentation result. Then the HMM state likelihood for each model Li is obtained as follows T 

LLHMM (i, j) =

t=1

i = 1, . . . , I, j = 1, . . . , J

IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326– 334 doi: 10.1049/iet-spr.2011.0170

(1)

where sit is the state index at time t according to the segmentation result obtained from the HMM-based classifier with the ith cardiac disorder model, and j denotes an index to HMM states. Accordingly, we obtain the 54HMM state likelihood values in total. 4.3

Murmur likelihood

The spectral or cepstral features of heart sound signals often smoothed murmur and click sound signals, the temporal position of murmurs becomes hard to identify. To utilise the fact that the position of murmur sounds depends on the category of cardiac disorders, we compute the murmur probability for every frame, and then obtain the murmur likelihood corresponding to each state of HMM. Fig. 7 shows the abnormal heart sound signals of time domain and spectral domain. The input signals are filtered by two band-pass filters since the heart pulses (S1 or S2) of heart sound signals have most of energy between the 20 and 200 Hz, and the murmurs between the 200 and 700 Hz [11]. We obtain the mean absolute value of the sub-band signals representing energy for each frame as follows x=

N 1 |s(n)| N n=1

(2)

The s(n) is the nth sample value of sub-band signals and N is the frame size. A gamma probability distribution function (pdf) is a better choice than a Gaussian pdf because the sample values are non-negative. We model the mean absolute value to have a gamma probability distribution representing the heart pulse class (v1) or the murmur class (v2) as follows p(x|vi ) = xkvi −1

e−x/uvi k uvvi i (kvi

− 1)!

,

i = 1, 2

(3)

where p(x|v1) is the gamma probability distribution of the ith class with the integer shape parameter k and the scale parameter u. Given a mean absolute value x, we compute the a posteriori probability of murmur p(v2|x) as follows p(v2 |x) =

Fig. 6 Example of segmentation results using HMM-based segmentation

log p(xt |sit = j, Li ),

p(x|v2 ) p(x|v1 ) + p(x|v2 )

(4)

Fig. 8 shows an example of murmur probability computation. Figs. 8a – c show the filtered heart sound signals, the gamma probability distributions and a posteriori probabilities of murmur, for band 1 of 20– 200 Hz, respectively. Figs. 8d – f show the same information for band 2 of 200 –700 Hz as the left-sided figures, respectively. The murmur likelihood is obtained by accumulating frame-level log probabilities belonging to each state of 329

& The Institution of Engineering and Technology 2012

www.ietdl.org

Fig. 7 Characteristics of heart sound signals: the heart pulses (S1/S2) of heart sound signals have most of energy between the 20 and 200 Hz, and the murmurs between the 200 and 700 Hz a Time domain b Spectral domain

Fig. 8 Example of murmur probability computation a b c d e f

Output signal of band-pass filter of band 1 Gamma probability distributions of band 1 where the dotted and solid lines represent the heart pulse class and the murmur class, respectively Murmur probabilities (dotted line) overlapped on the mean absolute values (solid line) of band 1 Output signal of band-pass filter of band 2 Gamma probability distributions of band 2 where the dotted and solid lines represent the heart pulse class and the murmur class, respectively Murmur probabilities (dotted line) overlapped on the mean absolute values (solid line) of band 2

330 & The Institution of Engineering and Technology 2012

IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326 –334 doi: 10.1049/iet-spr.2011.0170

www.ietdl.org experiments to compensate the small sample size problem. All data have been down sampled to 2 kHz.

HMM as follows LLMPk (i, j) =

T 

log p(v2 |xt , sit = j|Li ),

k = 1, . . . , K

5.2

Cardiac disorder classification

t=1

(5) where k is the band index. Consequently, we obtain 108 murmur likelihood values in total. 4.4 Combination of HMM state likelihood and murmur likelihood In the previous algorithms [9, 10], the HMM ‘model’ likelihood for the cardiac disorder categories was often used as input of an SVM-based classifier. However, in this work, instead of HMM ‘model’ likelihood, the HMM ‘state’ likelihood together with the murmur likelihood is input to the SVM-based classifier to decide the final cardiac disorder category. SVM-based classifiers [15, 16] have shown successful results compared with other classifiers based on maximum likelihood, decision tree and neural network. An SVM aims to fit an optimal separating hyperplane (OSH) between classes by focusing on the training samples that lie at the edge of the class distributions, which are called the support vectors. The OSH is oriented so that it is placed at the maximum distance between the sets of support vectors. It is because of this orientation that SVM is expected to generalise more accurately on unseen cases than the classifiers that aim to minimise the training error such as neural networks. Thus, with SVM classification, only some of the training samples that lie at the edge of the class distributions in feature space (support vectors) are needed in the establishment of the decision surface unlike statistical classifiers such as the widely used maximum-likelihood classifiers, in which all training samples are used to characterise the classes. Conventional maximum-likelihood classification, therefore, requires much larger training sample size than SVM to derive an accurate classification. Moreover, SVM is able to identify the most useful training data for the provision of support vectors before classification [17]. The SVM in this work uses the kernel function based on radial-basis function (RBF) networks. The SVM-based classifier was expanded to multi-class by using the oneagainst-all algorithm [18]. The trade-off weight value (C ) was set to 500 and the kernel width (s) was set to 1 for computer experiments. The input vector of the SVM-based classifier is the HMM state likelihood and murmur likelihood. The output nodes of SVM-based classifier represent nine cardiac disorder categories.

5 5.1

Experimental results Database

We used the database which includes two sets of data. The first set consists of continuous abnormal heart sound signals extracted from a collection of heart sounds and murmurs [11]. The second set includes only normal heart sound signals added for reliable evaluation. Cardiac disorders were grouped into nine categories (I ) including the normal category. The number of samples was 80 for the normal category and 80 for abnormal categories as shown in Table 1. We used leave-one-out cross validation in all IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326– 334 doi: 10.1049/iet-spr.2011.0170

Table 2 shows the classification accuracy of the baseline algorithm (Fig. 2) with envelope-based segmentation along with different MFCC orders (M ). The best accuracy was obtained with the feature dimension of 39, similar to speech signal processing. Table 3 shows the classification accuracy of the baseline algorithm with different window sizes and shift sizes. The results showed that six-state HMMs with a 39-dimensional feature vector performed best with the window size of 75 ms and the shift size of 25 ms, which implies that the window size and shift size in cardiac disorder classification should be longer than those (usually 25 and 10 ms) in speech signal processing. The reason is because the duration of murmur and click sound is longer than 25 ms. Table 4 shows the classification accuracy with the varying number of HMM states. As indicated in the table, we set the number of states to six in all the next experiments. Table 5 shows the classification accuracy with the varying number of Gaussian distributions for each state. For this experiment, the number of mixtures in HMMs was set to three, which was optimised in the preliminary experiments. In summary, the baseline algorithm yielded the best accuracy of 80.6% with the configuration of 39-dimensional feature vector, 75 ms window size, 25 ms shift size, six HMM states and three Gaussian mixtures. Table 1

Cardiac disorder categories and the number of heart sound data Cardiac disorder category normal, N abnormal

Number of heart sound data 80 6 9 12 9 12 5 14 13 160

AR AS AR + AS MR MS MR + MS MVP VSD

total

Table 2

Classification accuracy of the baseline algorithm with different MFCC orders MFCC order 8 12 16 20

Feature dimension

Accuracy, %

27 39 51 63

78.6 80.6 78.5 78.3

Table 3

Classification accuracy with different combinations of the window size and the shift size Window size, ms 25 75 150

Shift size, ms

Accuracy, %

10 25 50

77.5 80.6 78.8

331

& The Institution of Engineering and Technology 2012

www.ietdl.org Table 4

Classification accuracy with the varying number of HMM states Number of HMM states, J

Accuracy, %

3 4 5 6 7 8

79.4 80.0 80.4 80.6 80.3 80.0

Table 5

Classification accuracy with the varying number of Gaussian distributions for each state Number of Gaussian distributions

Accuracy, %

1 2 3 4 5 6 7 8

78.7 79.4 80.6 80.4 80.0 79.5 79.2 78.8

Table 6 shows the classification accuracy of the proposed method. First, we used the HMM-based segmentation algorithm instead of the envelope-based segmentation algorithm to reduce the segmentation errors (‘A1’), and used an HMM-based classifier. Then we used an SVM to discriminatively weight 54 state likelihood values (‘A2’). Next, all the 54 state likelihood values and 108 murmur likelihood values were used as the input of SVM (‘proposed’ algorithm is shown in Fig. 5). The structure of the proposed algorithms with different configurations is shown in Fig. 9. Even though heart pulses (S1/S2) are buried in murmur, the A1 algorithm can extract the correct starting points. Therefore the algorithm achieved the classification accuracy of 81.9%. The A2 algorithm achieved the classification accuracy of 83.8%. The improvement was largely because of reduction of the classification errors for cardiac disorders with different murmur position. In particular, for AR and MS cardiac disorder categories, the A2 algorithm yielded relative error rate reduction of 34.0 and 32.0% compared to the baseline algorithm. Table 6

Classification accuracy (%) of the proposed algorithm compared with different algorithms Cardiac disorder category

N AR AS AR + AS MR MS MR + MS MVP VSD average

Algorithm Baseline

A1

A2

Proposed

96 50 56 67 56 75 80 71 62 80.6

96 50 56 67 67 75 80 71 69 81.9

98 67 56 58 56 83 80 79 77 83.8

99 67 67 67 67 83 80 71 77 85.6

332

& The Institution of Engineering and Technology 2012

Fig. 9 Block diagrams of the reference systems A1 and A2 used for performance comparison a HMM-based S1/S2 segmentation and HMM-based classifier (A1) b HMM-based S1/S2 segmentation, HMM state likelihood and SVM-based classifier (A2)

The proposed algorithm yielded the best accuracy of 85.6%, which is relative error rate reduction of 25.8% compared to the baseline algorithm. We then checked the performance of the proposed algorithm by substituting the SVM-based classifier with other neural network-based classifiers: MLP- or RBF-based classifiers. For the other classifiers, we used the same number of input nodes and output nodes as for the SVM-based classifier. The MLP-based classifier had two hidden layers, which are organised as 100 neurons of the first hidden layer and 20 neurons of the second hidden layer. The MLP network [19] with the sigmoid activation function was trained for up to 300 epochs by using the scaled conjugate gradient algorithm. We used the learning rate of 0.1, the momentum constant of 0.5 and the error goal of 0.0001. For the RBF network [19], we used the error goal of 0.0001, the spread of 1.0 and the maximum number of neurons of 300. The neurons in the hidden layer were added until the RBF network meets the specified error goal. Table 7 shows the classification accuracy with different pattern classifiers, where the best accuracy was obtained with the SVM-based classifier. This fact implies that SVM generalises more accurately on unseen data than the MLP- or RBF-based classifiers. Table 8 shows the confusion matrix of cardiac disorder classification with the proposed algorithm. For AR, AS and MS cardiac disorder categories where the position and the duration of murmurs are important cues for classification,

Table 7

Classification accuracy (%) of the proposed algorithm with different pattern classifiers Pattern classifier MLP RBF SVM

Accuracy, % 83.1 84.4 85.6

IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326 –334 doi: 10.1049/iet-spr.2011.0170

www.ietdl.org Table 8

Average confusion matrix of cardiac disorder classification with the proposed algorithm

Cardiac disorder category Input a b c d e f g h i

N AR AS AR + AS MR MS MR + MS MVP VSD

Output

a

b

c

d

e

f

g

h

i

79 0 1 0 0 0 0 1 0

0 4 0 0 0 1 0 0 0

0 1 6 1 1 0 0 1 0

0 1 0 8 0 0 0 0 0

0 0 1 0 6 1 0 2 2

0 0 0 1 0 10 1 0 0

0 0 0 1 1 0 4 0 1

1 0 0 0 0 0 0 10 0

0 0 1 1 1 0 0 0 10

Table 9

Detection accuracy (%) of cardiac disorder with the proposed algorithm Input

Result

normal abnormal

Normal

Abnormal

98.7 2.5

1.3 97.5

the proposed algorithm yielded relative error rate reduction of 34.0, 25.0 and 32.0% compared to the baseline algorithm, respectively. In [2], cardiac disorder categories were classified by using four types of ANN based on wavelet decomposition. The authors reported to achieve the classification accuracy of 81.3, 87.2, 91.6 and 96.4% with 100% features by using MLP, BP algorithm, Elman neural network and RBF, respectively. Our results cannot be directly compared with [2] because the previous work used a database with four subjects for each category with seven cardiac disorder categories, which is smaller than our database shown in Table 1. More comparative research is needed in case when a common database will be available in public. 5.3

Cardiac disorder detection

Next, we performed cardiac disorder detection experiment to check whether the proposed algorithm can be applied for initial diagnosis. We mapped nine cardiac disorder categories of the classification results to the two categories: normal and abnormal categories. Table 9 shows the detection accuracy of cardiac disorders with the proposed algorithm. The proposed algorithm yielded the false alarm rate of 1.3% and the false rejection rate of 2.5%. These results are significantly better than those of the previous work [5] where ANN was used to classify the normal category and the abnormal category with mitral valve regurgitation and achieved the false alarm rate of 2.0% and the false rejection rate of 7.0%.

6

Conclusion

We proposed a new cardiac disorder classification method, which includes two major contributions: HMM-based segmentation and combination of the HMM state likelihood and murmur likelihood by an SVM. The conventional IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326– 334 doi: 10.1049/iet-spr.2011.0170

envelope-based segmentation method often missed the starting points of heart pulses when the heart pulse is buried in murmur. However, our proposed HMM-based segmentation method reduced such segmentation errors by adding the missing starting points of heart pulses (S1/S2) and correcting the ill-positioned starting points. As the duration of murmur and click sounds is mostly longer than 25 ms, we used the window size of 75 ms and the shift size of 25 ms. Whereas the envelope-based segmentation method using the segmentation yielded the classification accuracy of 80.6%, the proposed HMM-based segmentation method achieved 81.9%. In order to reflect the fact that murmur position is important in identifying the cardiac disorders, we extracted the HMM state likelihood and murmur likelihood. The HMM state likelihood was shown to effectively discriminate cardiac disorders with different murmur positions. By using SVM to classify the cardiac disorders based on the HMM state likelihood, we improved the classification accuracy to 83.8%. To reflect the position and spectral characteristics of murmurs, we computed the murmur likelihood within the corresponding HMM states. By adding the murmur likelihood to the input vector of SVM, we obtained our best classification accuracy of 85.6%, which means classification error reduction of 25.8%. Computer experiments showed that the murmur likelihood effectively reduces the classification errors for the AR, AS and MS cardiac disorder categories. The proposed method can be applied to develop an electronic stethoscope with the cardiac disorder detection capability, which can assist physicians to diagnose cardiac disorders especially in a remote health-care monitoring situation. Further study is required to measure the classification and detection accuracy by using a larger and richer clinical database.

7

Acknowledgments

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MEST) (No. 2011-0004063).

8

References

1 Iacobellis, G., Willens, H.J., Barbaro, G., Sharma, A.M.: ‘Threshold values of high-risk echocardiographic epicardial fat thickness’, Obesity (Silver Spring), 2008, 160, pp. 887– 892 2 Mokhlessi, O., Rad, H.M., Mehrshad, N.: ‘Utilization of 4 types of artificial neural network on the diagnosis of valve-physiological heart 333

& The Institution of Engineering and Technology 2012

www.ietdl.org 3

4

5 6 7

8 9 10

disease from heart sounds’. Proc. 2010 17th Iranian Conf. om Biomedical Engineering, November 2010, pp. 1– 4 Tian-hua, C., Su-xia, X., Pei-yuan, G., Zheng, Y.: ‘The research of non-invasive method of coronary heart disease based on neural network and heart sound signal’. Proc. Int. Conf. on Information Engineering and Computer Science, Wuhan, China, December 2009, pp. 1–4 Azra’ai, R.A., bin Taib, M.N., Tahir, N.M.: ‘Artificial neural network for identification of heart problem’. Proc. Second Int. Conf. on Signal Processing and Communication Systems, Gold Coast, Australia, December 2008, pp. 1 –6 Sinha, R.K., Aggarwal, Y., Das, B.N.: ‘Backpropagation artificial neural network classifier to detect changes in heart sound due to mitral valve regurgitation’, J. Med. Syst., 2007, 31, pp. 205 –209 Gupta, C.N., Palaniappan, R., Swaminathan, S., Krishnan, S.M.: ‘Neural network classification of homomorphic segmented heart sounds’, Appl. Soft Comput., 2006, 7, pp. 286–297 Schmidt, S.E., Toft, E., Holst-Hansen, C., Graff, C., Struijk, J.J.: ‘Segmentation of heart sound recordings from an electronic stethoscope by a duration dependent hidden-Markov model’. Proc. Computers in Cardiology, Bologna, Italy, September 2008, pp. 345–348 Obayya, M., Abou-Chadi, F.: ‘Classifying some cardiac abnormalities using heart rate variability signals’. Proc. 25th National Radio Science Conf., March 2008, pp. 1 –8 Donghui, C., Zhijing, L.: ‘New text categorization method based on HMM and SVM’. Proc. 2010 Second Int. Conf. on Computer Engineering and Technology, Chengdu, China, April 2010, pp. 383–386 Cai, J.D., Yan, R.W.: ‘Application of FCM-HMM-SVM based mixed method for fault diagnosis of power electronic circuit’. Proc. Int.

334

& The Institution of Engineering and Technology 2012

11 12 13 14 15 16 17 18

19

Conf. on Mechanic Automation and Control Engineering, Wuhan, China, June 2010, pp. 3982–3985 Mason, D.: ‘Listening to the heart: a comprehensive collection of heart sounds and murmurs’ (F. A. Davis Company, Philadelphia, 2000) Liang, H., Lukkarinen, S., Hartimo, I.: ‘Heart sound segmentation algorithm based on heart sound envelogram’. Proc. Computers in Cardiology, Lund, Sweden, September 1997, pp. 105–108 Young, S., Evermann, G., Gales, M., et al.: ‘Hidden Markov model toolkit v3.4’ (Cambridge University, 2006) Ahmadi, S., Spanias, A.S.: ‘Cepstrum-based pitch detection using a new statistical V/UV classification algorithm’, IEEE Trans. Speech Audio Process., 1999, 7, (3), pp. 333– 338. Huang, C., Davis, L.S., Townshend, J.R.G.: ‘An assessment of support vector machines for land cover classification’, Int. J. Remote Sens., 2002, 3, (4), pp. 725 –749 Bazi, Y., Melgani, F.: ‘Toward an optimal SVM classification system for hyperspectral remote sensing images’, IEEE Trans. Geosci. Remote Sens., 2006, 44, (11), pp. 3374– 3385 Mathur, A., Foody, G.M.: ‘Multiclass and binary SVM classification: Implications for training and classification users’, IEEE Geosci. Remote Sens. Lett., 2008, 5, (3), pp. 241– 245 Tatsumi, K., Tai, M., Tanino, T.: ‘Multiobjective multiclass support vector machine based on the one-against-all method’. Proc. 2010 Int. Joint Conf. on Neural Networks, Barcelona, Spain, July 2010, pp. 1– 7 Chen, S., Cowan, C.F.N., Grant, P.M.: ‘Orthogonal least squares learning algorithm for radial basis function networks’, IEEE Trans. Neural Netw., 1991, 2, (2), pp. 302–309

IET Signal Process., 2012, Vol. 6, Iss. 4, pp. 326 –334 doi: 10.1049/iet-spr.2011.0170