SHAP-enabled explainable AI framework for clinical interpretation of valvular heart diseases via digital acoustic features.

Abstract

Valvular heart diseases (VHDs), including mitral regurgitation, aortic stenosis, mitral stenosis, and mitral valve prolapse represent a significant global health burden particularly among older adults. Digital auscultation platforms can transform traditional cardiac assessments that are primarily subjective and clinician-driven into a data-driven diagnostic tool enabling advanced signal processing, feature extraction and machine learning (ML)-based interpretation. This creates new opportunities for early, objective, and precise disease screening, diagnosis, longitudinal monitoring, and personalized clinical decision-making. In this study, we present an explainable ML framework for early detection and precise classification of VHDs using digital auscultation data. Acoustic features extracted from digital auscultation data are used to build ML models on a public dataset followed by validation using clinical hospital data. Shapley additive explanations (SHAP) helps create more understandable models for early detection by pinpointing unique acoustic characteristics of VHD, which enhances the interpretability and accuracy of ML models. The SHAP tree explainer is utilized to improve interpretability and guide feature selection by identifying unique, consistent, and overlapping features relevant to VHDs, providing physiological insights and enhancing model transparency. Among the five models assessed, XGBoost with SHAP stood out as the most reliable, delivering high interpretability and 85% accuracy on the clinical dataset, achieving condition-specific accuracies with minimal variability across different practitioners. By combining predictive performance with explainability, the proposed framework shows high promise in objective screening, early diagnosis, and informed clinical decision-making for VHDs.