A Deep Convolutional Neural Network Classification of Heart Sounds using Fractional Fourier Transform

A computer-aided auscultation system can help in the initial diagnosis of heart diseases. In this work, we propose a binary classification system that uses fractional Fourier transform based Mel-frequency spectral coefficients (FrFT-MFSC) and a 1D deep convolutional neural network. FrFt-MFSC is used to convert the phonocardiogram (PCG) into heat maps using four fractional orders (0.9, 0.95, 1.0, 1.10). We verify the performance of our proposed system using a publicly available data set that was provided by 2016 Physionet/Computing in Cardiology Challenge. Ten-fold cross-validation and holdout test methods are used to evaluate the performance of the system. Classifier performance for various features using different fractional orders is also studied. The 10-fold cross-validation provides a good performance and balanced specificity and sensitivity of 0.97 and 0.95 respectively despite using imbalance data set. The proposed system performance is superior to all the current state-of-the art binary human PCG classification systems.