ResHNet: Spectrograms Based Efficient Heart Sounds Classification Using Stacked Residual Networks

Abstract

Cardiovascular disease (CVD) is one of the major contributors of global mortality rate as it accounts for almost 31% of the worldwide deaths. As per World Health Organization (WHO), CVD continues to be the number one cause of death in the world. In some parts of the world, access to expert doctors and diagnosis are difficult. Thus an efficient and quick diagnosis of heart disease method is needed especially for low-income and middle-income countries where Magnetic Resonance Imaging (MRI) and Ultrasound becomes a constraint in terms of the resources to save human life. With the tremendous technology advancement in the medical field, deep learning has gained more attention to automate most of the initial diagnosis of diseases. This fosters continuous research in adopting deep learning methods for automatic classification of heart sounds to identify any abnormalities. In this work, we aim to investigate the efficiency of introducing residual modules in heart sounds classification using a deep neural network. This approach involves the following steps: (i) Generation of Spectrograms for every 1D audio signal using Spectrogram generator module (ii) Training of residual network based classifier for identifying normal and abnormal heart sounds based on the Spectrograms. The standard dataset given for 2016 PhysioNet/CinC Challenge has been used here for validating our residual network. This method achieved around 97% accuracy on the independent hidden test set performing best without incorporating any segmentation or additional Mel-Frequency Cepstrum Coefficients (MFCC) features of the audio signals and just the learned features from the image-based representations. Various baseline results of other deep learning based approaches have also been considered for evaluating the robustness of this framework.