Classification of heart sounds using an artificial neural network

Abstract

A novel method is presented for the classification of heart sounds (HSs). Wavelet transform is applied to a window of two periods of HSs. Two analyses are realized for the signals in the window: segmentation of the first and second HSs, and extraction of the features.

After the segmentation, feature vectors are formed by using the wavelet detail coefficients at the sixth decomposition level. The best feature elements are analyzed by using dynamic programming. Grow and learn (GAL) network and linear vector quantization (LVQ) network are used for the classification of seven different HSs.

It is observed that HSs of patients are successfully classified by the GAL network compared to the LVQ network.

Introduction

Auscultation is a technique in which a stethoscope is used to listen to the sounds of a body. The structural defects of the heart are often reflected in the sounds the heart produces. Physicians use the stethoscope as a device to listen to a patient’s heart and make a diagnosis accordingly. They are particularly interested in abnormal sounds, which may suggest the presence of a cardiac pathology and also provide diagnostic information. For instance, a very important type of abnormal sound is the “murmur”, which is a sound caused by the turbulent flow of blood in the cardiovascular system. The timing and pitch of a murmur are of significant importance in the diagnosis of a heart condition, for example, murmurs during diastole are signs of malfunctioning of heart valves but murmurs during systole may correspond to either a pathological or healthy heart, depending on the acoustic characteristics of the murmurs.

Time–frequency/scale methods have been applied to characterize heart sounds (Debjais et al., 1997; Bently, 1996). In previous publications, the authors have discussed the characterization of heart murmurs using time–frequency methods over a number of cardiac cycles (Leung et al., 1998; Yoshida et al., 1997) and showed that features hence obtained were suitable for classification (Leung et al., 1999).

In this study, wavelet transform is proposed to analyze heart sounds (HSs) in time and frequency domains simultaneously. Each class of HSs contains characteristic and distinctive information that exists in time and frequency domains. Feature vectors are formed by using wavelet transform. Classification performance highly decreases if the right feature space is not constituted. Artificial neural networks (ANNs) are used as classifiers to increase the classification performance. The most prominent advantages of using an ANN as a classifier are: (i) Weights representing the solution are found by iteratively training, (ii) ANN has a simple structure for physical implementation, (iii) ANN can easily map complex class distributions, and (iv) generalization property of the ANN produces appropriate results for the input vectors that are not present in the training set.

In the literature, it is observed that multi-layer perceptron (MLP) (Lippmann, 1987) is widely used in the recognition of patterns with neural networks (Miller et al., 1992) and is also used to classify heart sounds (Barschdorff et al., 1995; Liang and Hartimo, 1998). One major problem encountered in MLP is its back-propagation algorithm (an iterative scheme) which takes too long time during learning. The second problem is the structure of the network, i.e., the number of hidden units and their interconnections, is defined by the programmer and the learning rule can modify only the connection weights. There is no rule which allows one to determine the necessary structure from a given application or training set. Lastly, the MLP may be caught by local minima, which decreases network performance.

In this study, an incremental and competitive learning network is proposed to handle the problems mentioned above and to increase the classification performance of HSs. In the literature, linear vector quantization (LVQ) and adaptive rezonance theory (ART) can be seen as the most basic schemes of the competitive learning network. A major advantage of the LVQ network is its fast learning speed. The major disadvantages are; it is not an incremental network, and the network generates feature vectors throughout the inside of a class homogeneously rather than concentrating them on the boundaries between classes. This causes the generation of an excessive number of feature vectors. ART2 (Carpenter and Grossberg, 1987) is a neural network that self-organizes stable recognition code patterns in real-time in response to arbitrary sequences of input patterns. The classification performance of the ART2 may decrease when the vigilance value is not carefully chosen. There is no direct way for choosing appropriate vigilance value, and a trial-and error process is usually time-consuming. The problem may be solved by using ‘slow learning mode’, but the learning speed of the ATR2 network slows down considerably. With ‘fast learning’ mode, the learning and clustering speed of ART2 models may improve; however, the non-centroid computation sometimes causes problems on the clustering results. The learning of a new pattern in ART2 tends to overwrite the previously stored information (Chin-Der and Stelios, 1997).

It is observed that incremental networks are widely used in the literature (Berlich et al., 1996; Burzevski and Mohan, 1996; Bruske and Sommer, 1995; Martinetz et al., 1993; Martinetz and Schulten, 1994; Fritzke, 1994, Fritzke, 1995). A number of approaches, advanced from SOM, have been proposed to achieve the objectives of retaining both the topology preserving and clustering properties. Fritzke (Fritzke, 1995) proposed a growing cell structure (GCS) for self-organizing clustering and topology preserving. The GCS approach starts the self-organizing with a k-dimensional simplex that is distributed over the input manifold. The GCS conditionally adds new nodes and removes old nodes based on a heuristic criterion that takes the relative winning frequency of a node or accumulated error into account, which is called a ‘resource’ of the output nodes. The resources of the winner and its neighbors determine the location of the added nodes. By computer simulation, Fritzke showed that output maps could be formed that resemble the topological structure of the input data in many different cases. To its simplicity, the competitive Hebbian rule has been used for topology learning in the growing neural gas (GCS) (Fritzke, 1994) and dynamic cell structure (Bruske and Sommer, 1995). However, these algorithms add and delete nodes based on the ‘resource’ used in GCS. This has created some complexity in its implementation.

In this study, grow and learn (GAL) is proposed as an incremental and competitive learning network to increase the classification performances of heart sounds. In a previous study (Ölmez et al., 1998) it is observed that GAL has fast training and classification, implementation simplicity, and satisfactory performance. Hence in order to carry out HS classification in real-time, we preferred to use GAL as an incremental neural network.