Coronary artery disease (CAD) continues to represent a major global health burden, underscoring the importance of noninvasive and scalable diagnostic strategies. This systematic review published in the NPJ Digital Medicine evaluated the performance of heart sound analysis techniques in detecting CAD, defined as ≥50% coronary artery stenosis. A structured search across four databases yielded 1082 records, of which 40 studies comprising 13,814 participants fulfilled the inclusion criteria. 

Studies not meeting predefined eligibility requirements, including absence of standardized CAD definitions or relevant diagnostic metrics, were excluded.
Of the included studies, 21 employed conventional signal processing techniques, while 19 utilized machine learning (ML)-based methodologies. Most signal processing studies demonstrated modest diagnostic performance, with nearly all larger studies (>50 participants, n=15) reporting accuracy below 75%. Additionally, a lack of validation using independent datasets was common, limiting confidence in reproducibility. 

In contrast, ML-based approaches showed comparatively higher diagnostic metrics, with 15 of 19 studies reporting accuracy, sensitivity, and specificity exceeding 80%. A similar number of ML studies incorporated external dataset validation, supporting improved generalizability. Furthermore, analyses suggested that models leveraging the complete heart sound waveform achieved higher sensitivity than those restricted to diastolic segments, indicating the importance of preserving full signal information.

Overall, diagnostic performance varied based on methodological approach. Further large-scale, multicenter studies are required to confirm clinical applicability and ensure robustness across diverse populations.