Screening system for heart valve disease

Abstract

The first method applied by physician and cardiologists to detect heart disease is heart sound auscultation. However the skill of auscultation takes many years to acquire. Acknowledging the importance of heart sound auscultation, this research is conducted to develop a screening system that can classify normal and abnormal heart sound as well as categorizing the abnormal heart sound signal into four common categories of heart valve disease. The diseases are Aortic Regurgitation (AR), Aortic Stenosis (AS), Mitral Regurgitation (MR) and Mitral Stenosis (MS). The screening system is able to perform automated process of segmentation, feature extraction and classification of the heart sound signals. The segmentation process that is based on the time properties of the heart sound is introduced in this study to produce samples for feature extraction. The sample is converted to frequency domain and power spectrum of the signal is calculated. Power spectrum of the signal is used to get the heart sound features using cross-correlation method. The proposed method is a robust method where samples duration, cycle sequence and amplitude of the heart sound and murmur are not significantly affecting the power spectrum itself. The extracted frequency features are classified using standard Multi-Layer Perceptron (MLP) network and hierarchical Multi-Layer Perceptron network. Classification accuracy obtained from hierarchical MLP network is 100%, better than standard MLP network with accuracy of 85.71%. This is due to the complexity in classification of 5 types of heart sound signals has been reduced to two parts by using Hierarchical MLP network. A complete system that includes the process of segmentation, feature extraction and segmentation of the heart sound signal is developed in PC based platform and implemented in an embedded system. The embedded system is consists of electronic stethoscope, multimedia board (VC21PC1) and a single board computer (VCMX212) as the core. Efficiency of both PC based and embedded system is investigated in this study. A total of 646 samples from 39 subjects have been used in this study. The results show that both PC based and embedded system has produced 96.3%, 92.59% and 94.44% of screening specificity, sensitivity and accuracy for normal and abnormal classification, respectively. This showed that the proposed method is good and reliable. However, for specific classification on the other 4 type of abnormal heart sound signal, the PC based system has produced 94.44% accuracy while the embedded system only produced 87.04% accuracy. The reason is that a few approximations were applied in calculating the features and output of the MLP network. Comparison is also made with other existing systems and it is found that the proposed system has produced a comparable screening accuracy (94.44%) for normal and abnormal classification and generally better accuracy for heart valve diseases classification (87.04%). Direct comparison cannot be made because the data and method used by the other researchers are totally different.