Machine Learning and End-to-End Deep Learning for the Detection of Chronic Heart Failure From Heart Sounds

Abstract

Chronic Heart Failure (CHF) is a major cardiovascular disorder affecting millions of individuals worldwide and remains a leading cause of hospitalization and mortality. Early and accurate detection of CHF is essential for improving patient outcomes and reducing healthcare burdens. Heart sound analysis provides a noninvasive, cost-effective, and accessible approach for automated CHF screening. This study presents an intelligent CHF detection framework that integrates traditional Machine Learning (ML) and End-to-End Deep Learning (DL) techniques using heart sound recordings. The ML approach employs expert-engineered acoustic features with Random Forest classification, while the DL approach learns discriminative spectro-temporal representations directly from raw heart sound signals. Experimental evaluation was conducted on heart sound recordings collected from multiple publicly available datasets and CHF-specific data. The segmented ML model achieved an accuracy of 82.72%, sensitivity of 68.42%, and specificity of 87.10%. The end-to-end DL model significantly improved performance, attaining an accuracy of 95.12%, sensitivity of 100.00%, and specificity of 93.55%. Furthermore, the recording feature aggregate model based on Random Forest delivered the highest performance, achieving an accuracy of 96.34%, sensitivity of 100.00%, and specificity of 94.83%. The results demonstrate that combining advanced feature learning with robust classification techniques enables reliable CHF detection and stage differentiation, highlighting the potential of automated heart sound analysis for clinical decision support and remote health monitoring applications.