Cardiovascular diseases remain the leading cause of global mortality, highlighting the need for reliable and scalable tools for early heart-attack risk screening. However, many existing machine learning models for cardiovascular risk prediction are limited by severe class imbalance, lack of uncertainty awareness, and poor interpretability, which restrict their reliability in real-world clinical screening. To address these challenges, this paper proposes CRuSE-Heart, an explainable risk-controlled selective ensemble framework for heart-attack screening using large-scale population health data. The proposed method combines LightGBM and XGBoost through a weighted ensemble. Further, it incorporates a conformal prediction-based selective mechanism that abstains from low-confidence predictions to control error on accepted cases. In addition, SHAP and LIME are integrated to provide global and local explanations, thereby improving model transparency and clinical interpretability. Experiments conducted on the BRFSS Personal Key Indicators of Heart Disease dataset, comprising approximately 319,795 records, show that the proposed selective model achieves 97.28% accuracy, 95.01% sensitivity, and 97.41% specificity, outperforming several baseline methods. The selective prediction setting also provides a practical coverage–risk trade-off, improving reliability in high-stakes screening scenarios. These results demonstrate that CRuSE-Heart offers an effective, interpretable, and safety-aware framework for population-scale heart-attack risk assessment.
Asad et al. (2026) studied this question.