Cardiovascular diseases detection from phonocardiograms using deep learning

Abstract

Cardiovascular diseases are a significant public health concern, responsible for a high number of global deaths. Manual diagnosis of CVDs using heart sound signals requires extensive clinical expertise. In recent years, researchers have explored signal processing and machine learning techniques to automate the early detection of cardiovascular diseases from Phonocardiograms. However, the majority of these approaches depend on traditional features and classifiers, which may experience difficulties capturing the complexity of heart sounds. This study aims to develop a deep learning model capable of accurately classifying heart sounds as normal or abnormal. Making use of the publicly available PhysioNet 2016 dataset to train a hybrid CNN-LSTM model, a comprehensive comparison between different sound segmentation (windowed segments and heart cycle segments) and feature extraction techniques (Mel Spectrograms and Mel Frequency Cepstrum Coefficients) are conducted. The goal of this comparative study is to identify the optimal combination of segmentation and feature extraction methods to effectively represent heart sounds for efficient training of the adopted deep neural network architecture. We achieved an overall final score of 93% and an accuracy of 92% using the heart cycle segments and spectrogram features setting. Performance comparisons with the existing literature indicate the efficiency of this approach. This research aims to contribute to the advancement of automated CVD detection from Phonocardiograms, potentially aiding in early diagnosis and intervention.