Heart Sound Classification using Convolutional Neural Network

Abstract

Cardiovascular Diseases (CDV) is a term that groups the disorders related to the heart and blood vessels. One way to diagnose the CDV is using the heart sound where an abnormal sound is heard which indicates a problem. In this work we perform a heart sound classification using Convolutional Neural Network (CNN) and log mel spetrogram where we test different models on the available datasets. The experiments included in this work are a 2D-CNN model and adaptations of ResNet-18 and VGG-11 architectures. Results were evaluated based on accuracy, precision, recall, and F1-score metrics, with the pre-trained ResNet-18 model demonstrating superior performance, achieving an accuracy of 86% on the PASCAL dataset and 70% accuracy on the Physionet datasets of 2016 and 2022.