Advancing ST-elevated myocardial infarction mortality risk prediction in Asian populations through explainable and calibrated machine learning.

Background: Traditional risk scores for ST-segment elevation myocardial infarction (STEMI), such as the Thrombolysis in Myocardial Infarction (TIMI) score, were developed predominantly in Western populations and may exhibit suboptimal performance in Asian patients due to differing clinical and genetic profiles. Accurate in-hospital mortality prediction is essential for optimizing clinical management in this high-risk group. However, key aspects such as model explainability (e.g., SHapley Additive exPlanations (SHAP) analysis) and probability calibration are often neglected, limiting the clinical utility and trustworthiness of predictive models. Objectives: This study aimed to develop and validate explainable, well-calibrated machine learning models for predicting in-hospital mortality among Asian STEMI patients, benchmarking their performance against the TIMI risk score. Interpretability was enhanced using SHAP for both global and local explanations, and model calibration was systematically addressed. Methods: We conducted a retrospective cohort study using data from 49,574 Asian STEMI patients in the Malaysian National Cardiovascular Disease registry (2006-2021). A temporal split was applied to simulate prospective deployment: data from 2006-2018 were used for model training, 2019 data for calibration, and 2020-2021 data as an independent test set. Multiple ML algorithms including logistic regression (LR), support vector machine, random forest, gradient boosting machine, and XGBoost were developed and compared. Stacked ensemble models were constructed using combinations of these base learners. Performance was evaluated on the independent test set using area under the receiver operating characteristic curve (AUC-ROC), accuracy, recall, specificity, Brier score (for calibration), and Net Reclassification Index (NRI), with benchmarking against the TIMI score. Isotonic regression was applied for probability calibration, and SHAP was used for interpretability. Results: The calibrated LR model achieved the best overall performance (AUC = 0.8884, 95% CI (confidence interval): 0.8756-0.9011; accuracy = 0.8538; recall = 0.7746; specificity = 0.8617; Brier score = 0.0598; NRI = 0.5828 vs. TIMI). SHAP analysis confirmed that model predictions were aligned with established clinical reasoning, enhancing interpretability. Probability calibration further improved model reliability, as evidenced by a reduced Brier score. Conclusions: Calibrated LR, supported by SHAP-based explainability and robust probability calibration, significantly outperformed the TIMI score for in-hospital mortality prediction in Asian STEMI patients. This approach improves predictive accuracy, reliability, and interpretability, supporting more personalized and clinically trustworthy risk stratification. These findings highlight strong potential for real-world clinical integration and improved patient outcomes in diverse Asian populations.