Machine learning improves diagnosis of coronary artery disease using heart sounds

Coronary artery disease (CAD) is a leading cause of morbidity and mortality globally which requires timely diagnosis and interventions. Heart sounds have been proposed as a noninvasive method for CAD diagnosis, but consistency and quality of evidence remain variable.

This study aims to systematically evaluate the diagnostic value of signal processing-only and machine learning techniques for heart sound-based CAD diagnosis.

A comprehensive search was performed on PubMed, Web of Science, Embase, and the Cochrane Library in June 2025. Two reviewers independently screened articles and extracted data based on PRISMA guidelines. Eligible studies were assessed for their techniques, accuracy and limitations in CAD diagnosis.

Of 1082 identified records, 41 studies involving 14266 patients were included in the review. Among the 22 studies on signal processing-only, those with a sample size over 50 (n = 14) reported lower diagnostic accuracy (<75%) for CAD diagnosis than the remaining studies. Sixteen of the 19 machine learning studies demonstrated high accuracy, sensitivity and specificity, all exceeding 80%. Moreover, the use of full-cycle heart sound signals yielded a higher diagnostic sensitivity of CAD than using only the diastolic phase.

This first systematic review shows that machine learning is more accurate than signal processing-only algorithms for heart sound-based CAD diagnosis. Future studies may validate these machine learning algorithms in real-world environments in larger, multicenter studies.