A NOVEL DEEP LEARNING NEURAL NETWORK SYSTEM FOR IMBALANCED HEART SOUNDS CLASSIFICATION

Abstract

This study proposed a novel TFNNS method, which aimed to solve the imbalanced phonocardiogram (PCG) signals’ classification. TFFNS consisted of three submodules: HeartNet, 2D-Maps transformation, and TF-Mask augmentation. HeartNet, deep neural networks (CNNs), was designed to recognize the categories of PCG signals. In particular, on the basis of short-time Fourier transform and Mel filtering, 2D-Maps transformation was used to convert one-dimensional PCG into two-dimensional Savitzky-MFSC feature maps that were fed into HeartNet; TF-Mask augmentation was designed to augment the training datasets by randomly shielded Savitzky-MFSC maps in the domains of time and frequency. We trained our model on the PASCAL heart sounds’ datasets to classify three categories of heart sounds including normal, murmur, and extrasystole. We also evaluated and compared the model with the baselines on the consistent evaluation protocols. The experimental results showed that the proposed TFFNS method significantly promoted the performance of the PCG signals’ classification and exceeded the baselines by giving the mean precision of 94%, heart problem specificity of 99%, and discriminant power of 1.317.