Leakage-Safe Precision-Aware Dual-Branch FT-Transformer for Population-Scale Heart Disease Risk Prediction.

Heart disease remains one of the leading causes of mortality worldwide, creating a strong need for reliable population-scale risk prediction models for large-scale screening and preventive monitoring. However, existing machine learning and deep learning approaches often struggle under severe class imbalance, data leakage risks, and unstable precision-recall trade-offs, limiting reliability in population-scale health-monitoring settings. To address these challenges, this study proposes a precision-aware Dual-Branch FT-Transformer framework for cardiovascular risk prediction using the BRFSS-2024 dataset. The proposed architecture separates recall-oriented detection and precision-oriented verification through two specialized prediction heads and integrates them using a lightweight gating mechanism trained strictly within training folds to prevent information leakage and enable controlled error arbitration. Under a strict leakage-safe 5-fold cross-validation protocol, the proposed model achieves an F1-score of 0.43, recall of 0.59, and AUPRC of 0.38 at a fixed threshold of 0.50 while reducing false negatives by more than 50% compared to LightGBM without excessive false positives. Although some baseline models achieve higher AUROC values, the proposed framework demonstrates more balanced and clinically meaningful precision-recall behaviour at operational screening thresholds. Additional evaluation on an independent NHANES cohort under the same leakage-safe re-training protocol further suggests robustness across heterogeneous population-health settings. Overall, the proposed dual-objective learning framework provides a practical and robust approach for imbalanced tabular prediction in population-scale cardiovascular risk assessment.