Abstract
This study presents a reliability-centered and decision-aware Human–AI clinical decision-support framework for cardiovascular risk prediction using structured clinical data. Unlike conventional machine learning approaches that prioritize discrimination metrics alone, the proposed framework formulates clinical prediction as a multi-dimensional reliability optimization problem, jointly modeling discrimination, probabilistic calibration, subgroup consistency, and robustness under dataset shift. Using a benchmark dataset of 918 patients with independent external validation on the UCI Cleveland cohort (n = 303), multiple machine learning models including Logistic Regression, Random Forest, XGBoost, and CatBoost were evaluated under a unified, leakage-safe protocol. While all models achieved strong internal discrimination (AUC ≥ 0.92), statistical testing revealed no significant differences (p > 0.05), highlighting the limitations of accuracy-centric model selection. External validation demonstrated substantial variability in generalization, with Random Forest achieving the strongest performance (AUC = 0.988), indicating superior robustness under distributional shift. To address limitations of single-metric evaluation, a composite reliability score is introduced to aggregate discrimination, calibration, fairness, and robustness into a unified evaluation framework. Calibration analysis shows that raw model probabilities outperform post-hoc calibration methods (Brier = 0.111, ECE = 0.048), emphasizing the dataset-dependent nature of probabilistic reliability. Subgroup analysis further reveals heterogeneity in calibration performance across patient populations, underscoring the importance of fairness-aware evaluation. Beyond predictive performance, the framework integrates decision-aware modeling through threshold-based risk stratification and Decision Curve Analysis (DCA), enabling optimization with respect to clinical net benefit rather than accuracy alone. The proposed system is further operationalized through a deployment-oriented interface, demonstrating how reliability-aware machine learning can be translated into an interactive clinical decision-support tool with interpretable outputs and actionable recommendations. Collectively, this work advances clinical machine learning from an accuracy-centric paradigm toward a reliability- and utility-driven framework, providing a principled foundation for developing robust, interpretable, and clinically deployable AI systems.