Heart sound signals contain bioacoustics itiformation that reflecting the heart operation. The signal analysis can give itiformation about the heart condition whether normal or abnormal. The abnormal signals usually contain murmurs or other sounds that can give information about the type of heart disease. Therefore, this research is focusing on classification of normal and several type of heart valve disease into two types of classification. The first classification is into normal and abnormal. The second classification is into normal and types of heart valve diseases. The cases of heart valve diseases that chosen are aortic stenosis, aortic regurgitation, mitral stenosis and mitral regurgitation. There are two data sets that have been used. The first data set is
simulated data which generated by using sojtware 'Heart Sounds Made Easy'. This data is recorded using software 'Nero Wave Editor'. The second data set is combination of patients' data (self recorded) and internet data. Patients' data are recorded using electronic stethoscope. Each data set consists of 250 samples. Four techniques have been chosen to analyse heart sound signals. The jirst technique is wavelet transform. The second technique is empirical mode decomposition. The third and forth techniques are based on S-Tramform. They are also called asS-Transform I and S-Tran.iform fl. Several experiments have been carried out to extract certain ji.wtures from each technique. The features will be used as inputs to the classification part. Two class(fiers
are used to see the accuracy performance of each feature extraction method which are multilayered perceptron neMork and support vector machine. For each classification, the analysis of number of input and analysis of parameters were done to get the optimum accuracy performance. Comparisons were done to determine the best method of feature extraction for classification. The results showed that for classification of normal and abnormal, generally the feature extraction method based on empirical mode decomposition or S-Transfhrm I was the best choice. For classification of normal and cases of heart valve diseases, generally feature extraction that based on S-Transform I was the best when used with MLP that can achieved testing accuracy of 99% while for classification with SVM, either empirical mode decomposition or S-Transjorm I was the best choice. Both MLP network and SVM give a comparable accuracy results generally.
