Advancing cardiovascular disease diagnosis with an interpretable and responsible AI framework

Cardiovascular disease (CVD) remains a leading global health threat, responsible for one in five deaths worldwide. Early detection is critical to mitigate morbidity and mortality, yet traditional diagnostic methods often rely on reactive clinical assessments, missing opportunities for preventive intervention. In this study, a machine learning (ML) ecosystem is developed to enhance CVD diagnosis through two key approaches: (1) an early warning system using non-clinical, self-reported features for accessible risk stratification, and (2) specialized diagnostic models integrating clinical and non-clinical data. The framework leverages advanced ML techniques, including tabular neural networks (TabNet, TabPFN) and ensemble methods (XGBoost, Random Forest), validated on multi-regional datasets. Shapley Additive Explanations (SHAP) analysis identified ECG-related features as dominant predictors of CVD risk, with ST-segment slope (+0.93) and ST depression (+0.63) exhibiting the strongest effects. Counterfactual explanations from the non-clinical model further revealed actionable preventive measures: reducing exercise-induced angina and chest pain severity, alongside increasing exercise heart rate, could shift predictions from diseased to healthy, highlighting the model's utility for lifestyle interventions. To address ethical and clinical trustworthiness, interpretability tools (SHAP, counterfactuals), fairness mitigation (FairLearn), and uncertainty quantification (Bayesian Neural Networks) are incorporated. Causal inference identified key predictors and their Average Treatment Effects (ATEs) such as exercise-induced angina (ATE: 0.36) and ST slope (ATE: 0.33), informing a hybrid ensemble model that achieved 89% accuracy while reducing dimensionality. The system aligns with FDA Good ML Practices and EU Trustworthy AI guidelines, offering a scalable solution for early detection and equitable diagnosis.