Heart Sounds Segmentation Using Neural Networks

Abstract

The segmentation of heart sounds plays a major role in the automatic classification and analysis of a Phonocardiogram (PCG). Segmentation is the processes by which a heart cycle of PCG is segmented into its four sounds (S1, Systole, S2 and Diastole). In this project we aim to build a framework for classification and segmentation of PCG sounds. Objective was to employ neural network based approaches for sequence modeling and prediction such as RNNs and LSTMs for this purpose. The proposed algorithm was long short term memory network that consist forget gate input gate and output gate. We used the long short term memory algorithm to find the emission probability matrix this network was combined with the hidden Markov model for the segmentation of the heart sound and we get 80 % of F1 score. The second experiment was to use the LSTM network solely for the whole seg mentation processes, for that we have used the 792 audios of different lengths from 5 sec to 90 sec. By LSTM we get the 80% average F1 score for S1 sound and 82% average score for S2 sound. The proposed method is useful for the practical applications because it has been trained on the both noisy and clean data. That shows that if the noise is added while recording the heart sounds the algorithm can give good segmentation results as compare to the previous segmentation methods