Heart Sound Analysis with Machine Learning Using Audio Features for Detecting Heart Diseases

Abstract

Society has been impacted by the influence of artificial intelligence (AI) across various aspects of daily life over the last few years. AI can positively impact healthcare by making it cheaper, quicker, more effective, and more accessible. AI has impacted the detection and treatment of cardiovascular diseases (CVD). The analysis of heart sound recordings using AI has been studied during the last decade in the study of noninvasive diagnosis of CVD. This study aims to construct a machine learning model that requires the least computational resources and computation time for the classification of heart sounds using a novel set of time-domain, frequency-domain, and statistical-domain features extracted from heart sound recordings. A public dataset of heart sound recordings comprising five classes including one normal category and four different categories of valvular diseases was used in this study. Two combinations of data were used in the experiments. The first combination dataset consisted of normal and abnormal heart sounds. The second combination consisted of heart sound recordings belonging to one normal category and four different valvular diseases of heart sounds. The model’s performance can be deemed excellent, with the first combination giving an accuracy of 99.89% and the second combination giving an overall accuracy of 99.26%.