Missing-data-aware machine learning prediction of in-hospital major adverse cardiovascular events after primary percutaneous coronary intervention for ST-segment elevation myocardial infarction.

Machine learning (ML) offers opportunities to improve prognostication after ST-segment elevation myocardial infarction (STEMI), but real-world registries frequently contain incomplete data, and inappropriate handling of missingness may degrade performance. We retrospectively evaluated 659 consecutive STEMI patients undergoing PCI during the index admission. The primary outcome was in-hospital major adverse cardiovascular events (MACE). Eighty clinical, electrocardiographic, laboratory, echocardiographic, and angiographic variables were analyzed, with up to 40% missingness in some predictors. A hybrid feature-selection approach incorporating CatBoost feature importance, SHAP values, variance filtering, and correlation screening identified the most informative predictors. Models were trained using stratified 5-fold cross-validation, comparing native missing-value handling in CatBoost with transformer-based imputation (TabImpute) and a pretrained tabular model (TabPFN). MACE occurred in 282 patients (42.8%). CatBoost trained directly on incompletely observed data achieved the best performance using approximately 8-10 predictors (AUC 0.73), with balanced accuracy 0.69, precision 0.68, and recall 0.53. Transformer-based imputation did not improve discrimination. SHAP analysis indicated that impaired pre-PCI TIMI flow, reduced LVEF, elevated troponin, greater ST-segment deviation, inflammation, and renal dysfunction were the strongest contributors. A class-weighted CatBoost model for in-hospital mortality achieved excellent accuracy (AUC = 0.93). These findings support the use of missing-data-aware ML methods for outcome prediction in STEMI registries and warrant external validation.