Classification of Abnormal Heart Sounds with Machine Learning

Abstract

This paper discusses the utilization of machine learning techniques in separating normal and abnormal heart sounds of public heart sound data sets. The focus of the study is to examine the information value of commonly used heart sound features that characterize heart sound signal of interest. We adopted two information measures, entropy and Gini index. They were utilized to rank the relevance of extracted heart sound features that would assist accurate classification. We examined 26 different heart sound features and ranked them accordingly. Using these two information measures, we have shown that with a minimum of six highest information valued heart sound features, satisfactory classification accuracy can be achieved. Our results were obtained by employing the ensemble decision-tree algorithm in supervised classification with both 80-20 and 90-10 splits cross-validation.