Development and external validation of an interpretable machine learning model for in-hospital major adverse cardiovascular events after PCI in acute coronary syndrome

Abstract Background Electronic medical record (EMR) data provide an opportunity to support early risk stratification in acute coronary syndrome (ACS), but models with large predictor sets may be difficult to implement. We aimed to develop and externally validate a parsimonious, interpretable machine learning model for in-hospital major adverse cardiovascular events (MACE) in patients with ACS undergoing percutaneous coronary intervention (PCI). Methods: This two-center retrospective cohort study included 1,726 patients: 1,218 in the derivation cohort and 508 in an independent external validation cohort. The primary outcome was in-hospital MACE, defined as a composite of death, cerebral hemorrhage, cerebral infarction, cardiogenic shock, cardiac rupture, recurrent myocardial infarction, stent thrombosis, or Bleeding Academic Research Consortium bleeding events. A total of 125 candidate EMR variables were evaluated. LASSO and Boruta-based feature selection, alternative preprocessing strategies including ADASYN, seven machine learning algorithms, and grid-search cross-validation were compared. Discrimination was assessed using AUROC and AUPRC, with calibration and decision-curve analyses used to assess agreement and potential clinical utility. Results: In-hospital MACE occurred in 138 of 1,218 patients (11.3%) in the derivation cohort and 15 of 508 patients (3.0%) in the external cohort. The final random forest model retained eight predictors: Killip class, inorganic phosphorus, extracorporeal membrane oxygenation, left ventricular ejection fraction, absolute neutrophil count, total protein, alanine aminotransferase, and potassium. The model achieved an AUROC of 0.82 (95% CI 0.73–0.90) in internal validation and 0.90 (95% CI 0.85–0.95) in external validation. At the reported operating point, sensitivity was 0.488 and specificity was 0.991. Conclusions: An eight-variable random forest model derived from routinely collected EMR data showed good discrimination and preserved ranking performance in an independent hospital cohort. The model is best considered an early in-hospital risk-stratification tool; prospective validation with a prespecified prediction time is required before clinical implementation.