Detection of cardiac abnormality from PCG signal using LMS based least square SVM classifier
Introduction
Auscultation, the noninvasive cardiac testing, is used as a primary detection tool for diagnosis of heart valve disorders since invention of stethoscope in 1816 by Lannec (Hanna & Silverman, 2002). Heart sounds provide valuable diagnostic and prognostic information concerning the heart valves and hemodynamics of heart. During the last few decades, valvular heart diseases remain one of the major health concerns. Hence, early detection of heart valve diseases and accurate diagnosis of related conditions comprise a significant medical research area. In Bender (1992, chap. 13), it is reported that few heart valve diseases are best detected only by means of auscultation process. Auscultation is the most common and cost-effective technique, continues to provide an important source of clinical information related to heart valves and also cannot be totally replaced by alternative technical methods like echocardiography (Tavel, 1996). Moreover, echocardiography is not required for all patients with systolic murmurs (Shub, 2003, Tavel, 1996). Chizner (2008) has shown in his practical clinical overview of a variety of cardiac disease states and conditions that the stethoscope often enables many well-trained and experienced cardiac auscultators to make a rapid and accurate cardiac diagnosis. Physicians use the stethoscope as a device to listen the function of heart valves and make a diagnosis accordingly. However, in studying the physical characteristics of heart sounds and human hearing, it is seen that the human ear is poorly suited for cardiac auscultation (Mangione & Nieman, 1997). Therefore, relative and qualitative nature of heart sounds is limited by human perception and varies with personal aptitude and training. Phonocardiogram (PCG), a visual display of the recorded heart sounds (Durand and Pibarot, 1995, Lukkarinen et al., 1997) provides a trace of acoustic energy produced by the mechanical activity of various cardiac components and processes (Durand & Pibarot, 1995). Consequently, any abnormality of heart valves is reflected in the corresponding sounds of PCG signal. A cardiac cycle of a normal heart is comprised of two major sounds namely first heart sound, S1 followed by second heart sound, S2. These two distinct normal heart sounds are often described as lub (or lup) and dub (or dup), and occur in sequence with each heart beat. In case of abnormal heart sounds there could be several other sounds in the PCG signal besides primary heart sounds. Murmurs are abnormal heart sounds and refer to different pathological conditions (Cromwell, Weibell, & Pfeiffer, 2002, chap. 6) as per location, shape, duration and other associated features. Murmurs are generally high-frequency, noise like sounds that are produced as a result of turbulent blood flow. Different features of PCG signals like intensity, frequency content, split information, time relations etc. are helpful in detecting heart valve diseases, if any and the state of the heart function (Rangayyan, 2002, chap. 1).
Analysis of phonocardiogram signal can also be carried out by considering the heart sound cycle as a whole instead of separating the major components. Most of the heart sound classification techniques reported so far follow the same line of thought (Ari and Saha, 2008, Ari and Saha, 2009, Cathers, 1995, Gupta et al., 2007, Ölmez and Dokur, 2003, Reed et al., 2004). Time–frequency analysis techniques like wavelet transform have been widely used (Ari and Saha, 2009, Cathers, 1995, Gupta et al., 2007, Ölmez and Dokur, 2003, Reed et al., 2004) for extracting feature vectors from heart sound signal because of its ability to characterize time–frequency information which is important in this context (Debbal and Bereksi-Reguig, 2007, El-Asir et al., 1996, Lee et al., 1999, Wood and Barry, 1995). Ian Cather have presented artificial neural network (ANN) as a discriminative model for classification of five different heart sounds taken from 48 recordings of nine different subjects using wavelet based feature extraction technique (Cathers, 1995). Ölmez et al. have given a classification technique that utilizes Daubechies-2 wavelet detail coefficients at the second decomposition level for classification of seven different heart sounds collected from 28 subjects using ANN (Ölmez & Dokur, 2003). Reed et al. have described a computer-aided diagnosis mechanism for five different pathological cases using a seven level wavelet decomposition, based on a Coifman fourth order wavelet kernel (Reed et al., 2004) and ANN classifier model. Gupta et al. (2007) have used the same method as proposed by Ölmez and Dokur (2003) to classify three different heart sounds (normal, systolic murmur and diastolic murmur) recorded from 41 volunteers. In a work (Ari, Sensharma, & Saha, 2008), we make a binary decision on heart sound whether pathological or not in a Digital Signal Processor based system. Choi (2008) proposed a technique for detection of valvular heart sounds as normal or pathological using wavelet packet decomposition and support vector machine with fifth order polynomial kernel function.
In the present work, we propose a technique to improve the performance of Least Square Support Vector Machine (LSSVM) using Least Mean Square (LMS) algorithm (Haykin, 2002) for diagnosis of the valvular heart sounds as normal or pathological pattern classes. Note that, for the implementation of automatic classification of valvular heart diseases, it is necessary to detect primary heart sounds properly that serves as a reference marker to extract features from the corresponding cycle for use in classifier model. In one of authors’ previous work (Ari & Saha, 2007), such a method is described which does not need auxiliary input like electrocardiogram signal. We use the segmentation technique of Ari and Saha (2007) here as preprocessing block. In order to perform the decision making for the extracted features from wavelet based technique, the least square SVM technique is employed here as a classifier. Due to the equality constraints in the formulation, least square SVM solves a set of linear equations in the dual space instead of solving a quadratic programming problem as for the standard SVM. This simplifies the computation and enhances the speed considerably (Hanbay, 2009, Suykens and Vandewalle, 1999). The performance of the SVM mostly depends on kernel function and adjustable weight vectors. However, no such method exists which allows one to decide a good kernel function in a data-dependent way (Amari & Wu, 1999). The RBF kernel is chosen empirically in this application. The proposed technique relies on the basic idea that in order to improve the performance of the least square SVM, the pattern separability or margin between the clusters needs to be increased. Our aim is to update the adjustable weight vectors at the training phase such that all the data points fall outside the region of separation and to enlarge the width of the separable region. To implement this idea, Least Mean Square (LMS) algorithm is used to modify the Lagrange multipliers of Lagrangian function, which in turn modifies the adjustable weight vectors. Here, error is represented by the minimum distance of data points from margin of the region of separation for the data points that falls inside the region of separation or makes a misclassification. This error is minimized using modification of adjustable weight vectors. Therefore, as the number of iterations of the LMS algorithm increases, weight vector performs a random walk (Brownian motion) (Haykin, 2002) about the solution of optimum hyperplane having maximal margin (Kecman, 2001, chap. 2) that minimizes the error. Experimental results show that the proposed method identifies the valvular heart sounds with higher recognition accuracy than classical least square SVM for normal heart sound and five common pathological cases on a database of 64 different phonocardiogram recordings. A comparison with standard SVM as used in Choi (2008) is also presented here.
The rest of the paper is organized as follows. Section 2 describes the theoretical background behind support vector machine. Section 3 discusses the database that is used in the work and the methodology for the identification of the valvular heart sounds is described in Section 4. Section 5 illustrates a comprehensive evaluation of the proposed method for identification of valvular heart sounds and its performance is compared with standard SVM and least square SVM. Section 6 presents concluding remarks.
Section snippets
Support Vector Machine (SVM)
Support Vector Machines (SVM) have been successfully applied to various pattern recognition problem such as image recognition, text classification and bioinformatics since introduced by Vapnik (1995). The main idea behind this method is to map the data into higher dimensional input space where the different classes become linearly separable and to construct an optimal separating hyperplane in this space. This requires solving a quadratic programming problem and is done by using Kernel functions
Database
The heart sound recordings for normal and abnormal heart conditions were obtained from various resources, maximum being provided by Maulana Azad Medical Institute, Delhi, India. In total 512 cycles from 64 recordings of 64 volunteers are collected for five different pathological problems and normal heart conditions. These pathological problems include aortic insufficiency, aortic stenosis, atrial septal defect, mitral regurgitation and mitral stenosis. Heart sound signals were recorded by the
Methodology
The three successive stages followed in characterization of phonocardiogram signals are: preprocessing or segmentation, feature extraction and classification. This is decision making process, is shown as a flowchart in Fig. 1.
Experimental results
In this section, the performances of the proposed techniques are evaluated for the PCG datasets having 64 different recordings for normal heart sound and five different pathological problems as described in Section 3. The database is divided in two sections each of which consists of 256 cycles from 32 volunteers. One set is used for training the network and the other set is used for testing.
The experiment is evaluated using normal and five different pathological heart sounds. Sixteen different
Conclusions
In this paper, we presented a new method to improve the performance of the least square SVM classifier based on Least Mean Square (LMS) algorithm for recognition of heart sounds. The decision support system shows ‘diseased’ and ‘not-diseased’ decisions, which may be useful for mass screening camps or rural health units as a preliminary investigation tool. The idea is to enlarge the separating boundary surface so that the separability between the clusters is increased. In the proposed technique,
References (34)
- et al.
Improving support vector machine classifiers by modifying kernel functions
Neural Networks
(1999) - et al.
In search of an optimization technique for artificial neural network to classify abnormal heart sounds
Applied Soft Computing
(2009) Neural network assisted cardiac auscultation
Artificial Intelligence in Medicine
(1995)Cardiac auscultation: Rediscovering the lost art
Current Problems in Cardiology
(2008)Detection of valvular heart disorders using wavelet packet decomposition and support vector machine
Expert Systems with Application
(2008)- et al.
A new training method for support vector machines: Clustering k-NN support vector machines
Expert Systems with Application
(2008) - et al.
Time–frequency analysis of the first and the second heartbeat sounds
Applied Mathematics and Computation
(2007) - et al.
Neural Network classification of homo-morphic segmented heart sounds
Applied Soft Computing
(2007) An expert system based on least square support vector machines for diagnosis of the valvular heart disease
Expert Systems with Application
(2009)- et al.
A history of cardiac auscultation and some of its contributors
The American Journal of Cardiology
(2002)
Classification of heart sounds using an artificial neural network
Pattern Recognition Letters
Heart sound analysis for symptom detection and computer-aided diagnosis
Simulation Modelling Practice and Theory
On a robust algorithm for heart sound segmentation
Journal of Mechanics in Medicine and Biology
Classification of heart sounds using empirical mode decomposition based features
International Journal of Medical Engineering and Informatics
A DSP implementation of heart valve disorder detection system from phonocardiogram signal
Journal of Medical Engineering & Technology
Yale university school of medicine heart book
Cited by (142)
Phonocardiogram signal classification for the detection of heart valve diseases using robust conglomerated models
2023, Expert Systems with ApplicationsMulti-classification neural network model for detection of abnormal heartbeat audio signals
2022, Biomedical Engineering AdvancesEffect of Hilbert-Huang transform on classification of PCG signals using machine learning
2022, Journal of King Saud University - Computer and Information SciencesDeep attention-based neural networks for explainable heart sound classification
2022, Machine Learning with ApplicationsA customized framework for coronary artery disease detection using phonocardiogram signals
2022, Biomedical Signal Processing and Control