A two-phase machine learning framework for coronary heart disease risk prediction: Integrating population ecology with individual pathophysiology

Objective: Coronary heart disease (CHD) shows significant regional variations that cannot be fully explained by individual risk factors alone, suggesting the need for an analytical framework that integrates macro-level environmental influences with individual-level pathophysiology. This study developed and validated a novel two-phase, interpretable machine learning (ML) framework for identifying and validating multi-level predictors of CHD risk. Methods: In Phase 1 (Population-Level Discovery), we employed a consensus-driven Recursive Feature Elimination (RFE) pipeline across 12 regression algorithms (e.g., XGBoost, Random Forest, ElasticNet) on integrated ecological data from the Global Burden of Disease study (socioeconomic/environmental variables) and CDC WONDER (age-adjusted CHD mortality rates across US states, 2018-2021). Features were selected based on triangulation of the elbow method, one-standard-error rule, and stability selection. In Phase 2 (Individual-Level Validation), the prioritized features were mapped to individual-level data from the Health and Retirement Study (HRS, 2018-2020 waves). We trained and optimized 10 classification algorithms (e.g., ExtraTrees, XGBoost, SVM) to predict binary CHD status, evaluating performance via discrimination (AUROC), calibration (calibration curves), and clinical utility (Decision Curve Analysis). Model interpretability was achieved using SHapley Additive exPlanations (SHAP). Results: =0.874, RMSE=2.910). SHAP analysis identified ambient ozone pollution, population-level physical inactivity, and dietary risks as key state-level predictors. In Phase 2, the ExtraTrees classifier achieved the highest accuracy (0.897) for individual-level prediction. SHAP analysis revealed hypertension, diabetes, and smoking as the strongest individual-level risk factors. While predictors diverged between levels (environmental vs. clinical), factors like smoking and physical activity were important at both, acting as bridging mechanisms. Both models showed excellent calibration and positive net benefit on Decision Curve Analysis. Conclusions: This study presents a robust methodological framework that successfully links population-level ecological determinants with individual-level clinical risk. The findings underscore that a comprehensive understanding of CHD etiology requires simultaneous consideration of environmental context and individual biology. The interpretable ML pipeline not only provides accurate predictions but also generates actionable insights for targeted public health interventions and personalized risk stratification, offering a roadmap for two-phase analysis of complex chronic diseases.