Integrative ensemble and meta-learning frameworks for high-precision cardiovascular risk prediction

Abstract

Cardiovascular diseases remain the leading cause of global mortality, yet conventional risk models often fail to capture complex, non-linear dependencies in patient data, limiting their clinical applicability. We designed a comprehensive ensemble and meta-learning framework integrating multiple boosting algorithms (AdaBoost, Gradient Boosting, XGBoost, LightGBM, etc.) and bagging approaches (Random Forest, Bagged Logistic Regression) trained independently and as base learners within a Super Learner framework, with Logistic Regression, HistBoost, etc. as meta-learners to capture non-linear feature relationships. Two custom Blending strategies were applied for practical implementation. Models were trained and validated on harmonized data from five heterogeneous cohorts, with hyperparameter optimization and K-fold cross-validation ensuring robust performance. Ensemble approaches achieved strong predictive accuracy, with meta-learning consistently outperforming base learners. The Comprehensive Blending model achieved the highest AUC (0.972) and average precision (96.9%), exceeding LightGBM (AUC: 0.96). Super Learners using Logistic Regression as a meta-learner provided balanced, generalizable predictions (AUC up to 0.97; F1-score up to 96%). Carefully tuned ensemble and meta-learning frameworks achieved state-of-the-art cardiovascular risk prediction, where RF Boosting excelled in classification, Super Learners provided balance, and Blending models offered the highest AUC, supporting early detection and precision cardiovascular care.