Classification of segmented heart sounds with Artificial Neural Networks

Abstract

Nowadays heart diseases are the first cause of human deaths. For this reason, many studies have been carried out to reduce early diagnosis and death of heart diseases. These studies are mostly about developing computer-aided diagnosis systems by utilizing the developing technology. Some computer aided systems are clinical decision support systems developed to more easily detect heart diseases from heart sounds. These systems are used in the automatic analysis of heart sounds based on the classification of heart sounds in general. Much of the work done to diagnose heart diseases is to increase the success of classification. Segmentation of heart sound signals is also one of the frequently used methods to increase classification performance. In this study, S1-S2 sounds were segmented using the resampled energy method and the contribution to segmentation performance of the segment was examined. In practice PASCAL Btraining data set which is widely used for heart diseases application is used. The PASCAL Btraining data set contains three different heart sounds such as normal, murmur, and extrasystole. Artificial Neural Networks (ANN) were used to classify these sounds. For the comparison of the obtained results, two classifications were made for the segmented and the non-segmented sounds. As a result of the classification studies, the average all accuracy of classification 84% was achieved in the non-segmented ANN study, and the average all accuracy of classification 88.6% was obtained in the segmented S1-S2 sounds ANN study. Thus, segmentation of heart sounds increased the accuracy of classification by about 4.6%.