A HYBRID STATISTICAL–MACHINE LEARNING FRAMEWORK FOR EARLY DISEASE PREDICTION IN CLINICAL DATA

Abstract

Coronary heart disease (CHD) remains a leading cause of global mortality, emphasizing the need for accurate and interpretable early risk prediction models. Traditional statistical models provide clinically interpretable risk estimates but may fail to capture nonlinear interactions among cardiometabolic factors. Conversely, machine learning approaches improve predictive performance yet often lack transparency and calibration reliability. This study proposes a hybrid statistical–machine learning framework integrating Logistic Regression, Cox proportional hazards modeling, Random Forest, and Gradient Boosting to enhance early CHD prediction using longitudinal clinical data. Models were evaluated using discrimination (ROC-AUC), precision–recall analysis, calibration metrics (Brier score and decile plots), threshold-optimized confusion matrices, and survival concordance index. To ensure true primary prevention capability, analyses were conducted excluding prior disease history variables. Results demonstrate moderate discrimination (ROC-AUC ≈ 0.70) using baseline predictors alone, with acceptable calibration and meaningful survival stratification. SHAP-based explainability confirmed the clinical relevance of age, blood pressure, cholesterol, diabetes, and smoking as dominant contributors. The proposed hybrid framework balances predictive performance, interpretability, calibration, and explainability, supporting clinically actionable early CHD risk stratification and translational deployment in preventive cardiology.