Abstract
Heart sounds contain useful information that can help in early diagnose of heart disease. Therefore, the analysis of such signals has been an active research point for many groups. In this work, we present a new processing and classification system for heart sounds. We introduce a new technique to convert the time series representation of heart sound signal into time-frequency heat map representation based on fractional Fourier transform based mel-frequency spectral coefficients. Such representation is then classified using a stacked sparse autoencoder deep neural network. The proposed system is experimentally verified on the heart sounds database of the PhysioNet/Computing in Cardiology Challenge 2016. The proposed system achieves an accuracy of 0.9550 with 0.8930 sensitivity and specificity 0.9700. The average between sensitivity and specificity (score) is 0.9315. The details of the methodology and its implementation are presented and discussed.