Abstract

Cardiovascular disease (CVD) is a major global cause of morbidity and mortality. This paper proposes a calibration-aware hybrid ensemble framework. It combines traditional machine learning (ML) and deep learning (DL) techniques to predict CVD from tabular clinical datasets of 70,000 records, evaluated with leakage-safe nested cross-validation and calibrated base probabilities, and reports both discrimination and calibration metrics. Six machine learning models, including LR, RF, ET, XGB, LGBM, and KNN, along with three deep learning models, namely MLP, TabNet, and FT-Transformer, are evaluated. The top four ML models and the three DL models are used to construct an ensemble via stacking and soft voting, with both weighted and unweighted variants. The proposed approach integrates hybrid ML/DL stacking, selective probability calibration, leakage-safe nested cross-validation, and feature engineering. The performance is evaluated through nine metrics (ROC-AUC, PR-AUC, accuracy, recall, precision, F1-score, Brier score, log loss, and ECE). The results show that soft voting of XGB and FT-Transformer achieved the highest discrimination performance with an ROC-AUC of 80.16% and the best calibration metrics, including a Brier Score of 0.1803, a Log Loss of 0.505, and ECE of 0.004. While the models demonstrated strong internal performance, further validation on external datasets is needed to establish their potential utility for clinical decision support.

Statistics