Explainable Clinical Risk Prediction from EHR Tabular Data using Monotonic Constraints and Calibrated Probabilities

Tabular-based clinical risk prediction models are extensively applied in medical decision support systems; however, two major challenges often reduce their reliability: predictions that contradict basic clinical logic and poorly calibrated probability outputs that weaken threshold-based decision making. This study investigates explainable binary risk prediction using the processed Cleveland subset of the UCI Heart Disease dataset as a public clinical benchmark. A lightweight and CPU-efficient pipeline is proposed by employing an XGBoost classifier integrated with monotonic constraints on clinically relevant features, followed by probability calibration through post-hoc methods, including Platt scaling, temperature scaling, and isotonic regression on a separate validation set. Model performance is assessed in terms of discrimination capability using AUROC, AUPRC, F1-score, sensitivity, and specificity, while probability reliability is evaluated using ECE and Brier score metrics. A monotonicity audit is also conducted through counterfactual feature sweeps to measure violation rates. In addition, the model is applied for risk stratification into low-, medium-, and high-risk categories with corresponding event-rate reporting. The findings demonstrate that isotonic regression improves probability reliability without degrading discrimination performance. Furthermore, the monotonicity audit reveals no observed violations for constrained features. Overall, the integration of monotonic constraints and probability calibration produces more decision-ready risk estimates for threshold-based clinical decision support while maintaining transparency through SHAP-based analysis.