Development and explainability of a machine-learning model for predicting in-hospital mortality after percutaneous coronary intervention for unprotected left main coronary artery disease

Background Accurate prediction of mortality after percutaneous coronary intervention (PCI) for unprotected left main coronary artery (LMCA) disease is essential for risk stratification. We compared machine-learning (ML) models with the EuroSCORE II and the SYNTAX score for predicting in-hospital mortality and used explainable artificial intelligence to identify the most important predictors. Methods In this subanalysis of the Gulf-LM registry, 1,205 patients who underwent PCI for unprotected LMCA disease between 2015 and 2019 were analyzed; 45 (3.7%) died in hospital. Twenty-four baseline pre-procedural variables were used, with the EuroSCORE II and SYNTAX score retained continuously. Fourteen ML algorithms and the traditional scores were evaluated by cross-validation; performance was summarized from pooled out-of-fold predictions with bootstrap 95% confidence intervals. Discrimination was assessed by the AUROC and AUPRC, and calibration by the Brier score, calibration slope, and calibration-in-the-large; the model was interpreted using SHapley Additive exPlanations (SHAP). Results The random forest achieved the best performance, with an AUROC of 0.904 (bootstrap 95% CI, 0.859–0.941), an AUPRC of 0.335, and a Brier score of 0.030, outperforming the EuroSCORE II (AUROC 0.82) and the SYNTAX score (AUROC 0.56). In patient-level paired comparisons, the random forest improved discrimination over the EuroSCORE II (ΔAUROC +0.083) but not significantly over a parsimonious four-predictor model, and a nested cross-validation gave a selection-corrected AUROC of 0.885 (0.841–0.924), the internally estimated discrimination of the AUROC-based selection procedure. Calibration estimates were compatible with ideal calibration but imprecise (calibration-in-the-large 0.02; slope 1.20, 95% CI 0.93–1.57; mean predicted mortality 3.7% versus observed 3.7%), and the model stratified patients into groups with observed in-hospital mortality of 0.3%, 4.0%, and 15.1%. SHAP analysis, corroborated by permutation importance, identified the EuroSCORE II, STEMI presentation, cardiogenic shock, and tachyarrhythmia as the strongest predictors of mortality. Conclusion An explainable random-forest model discriminated better than the EuroSCORE II and the SYNTAX score for predicting in-hospital mortality after PCI for unprotected LMCA disease, with descriptive risk stratification including a low observed mortality rate in the lowest-risk group. This is a model development study with internal validation only, and external validation is needed.