Deep models for phonocardiography (PCG) classification

Phonocardiography or PCG plays a vital role in the initial diagnostic screenings of subjects for evaluating the presence of cardio-vascular anomalies. Since it is low-cost and less cumbersome to perform, it is found significant application in remote health diagnostics systems. It is also used to complement the ECG based cardiac diagnosis for detecting cardio-vascular abnormalities. One of the key aspect of PCG is the accurate identification of the heart sounds. In this work we propose to classify the heart sounds by performing various deep learning techniques such as RNN, LSTM and GRU. We used the widely known Peter Bentley heart sound dataset. Our experimental results show Long Short Term Memory (LSTM) provides better accuracy in the identification of heart sounds without the need for any pre-processing of the data.