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

Figures

• 

  Fig. 1. Flow diagram of patient selection for the final study cohort. This figure illustrates the selection process for patients with ST-segment elevation myocardial infarction (STEMI) admitted to Shahid Madani Heart Center between April 10, 2019 and September 25, 2021 (N = 897). Patients were included if they met ESC/AHA diagnostic criteria for STEMI, were aged 18 years or older, and underwent primary percutaneous coronary intervention (PCI). Exclusion criteria included not undergoing PCI (N = 37), receiving thrombolysis only (N = 52), or having missing key baseline or outcome data (N = 149). The final study population consisted of 659 STEMI patients treated with primary PCI. Among these, 377 patients experienced no in-hospital major adverse cardiovascular events (MACE), while 282 patients developed at least one in-hospital MACE.

• 

  Fig. 2. Machine learning workflow for MACE risk prediction. This figure summarizes the workflow used to develop machine learning models for predicting in-hospital MACE in STEMI patients. Registry data containing 80 clinical, ECG, laboratory, echocardiographic, and angiographic variables were used as input, with in-hospital MACE as the primary outcome. The dataset included substantial missingness and multicollinearity. After preprocessing, a missing-data–aware modeling strategy was applied, comparing CatBoost with native missing-value handling, CatBoost with TabImpute transformer-based imputation, and the TabPFN pretrained classifier. A hybrid feature selection process combined model-based and statistical criteria to rank variables and select the top 8–10 predictors. Models were trained with stratified 5-fold cross-validation and evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. CatBoost without prior imputation showed the best overall performance (AUC = 0.73). Model explainability was achieved using SHAP-based feature importance.

• 

  Fig. 3. Effect of feature number on CatBoost model performance for MACE prediction. This figure shows the relationship between the number of selected predictors and the performance of the CatBoost model under the hybrid feature selection framework. Accuracy, precision, recall, specificity, and ROC-AUC were evaluated across models trained with increasing numbers of features. Performance stabilized after inclusion of approximately 8–10 predictors, beyond which additional features provided minimal improvement.

• 

  Fig. 4. Performance of the CatBoost model for in-hospital MACE prediction. Panels illustrate the performance of the CatBoost model trained using ten selected predictors with native handling of missing data. (a) Confusion matrices (absolute and normalized) aggregated across stratified cross-validation folds, showing observed versus predicted MACE outcomes. (b) Receiver operating characteristic (ROC) curve demonstrating overall discriminative ability (AUC = 0.73). (c) Calibration curve comparing predicted probabilities with observed event rates, indicating good agreement across risk strata.

• 

  Fig. 5. SHAP analysis of key predictors in the MACE prediction model. This figure presents the SHAP summary plot for the CatBoost model, illustrating the relative importance and directional effect of the ten most influential predictors of in-hospital MACE. Each point represents an individual patient, with color indicating the magnitude of the corresponding feature value. Positive SHAP values indicate increased predicted risk, whereas negative values indicate decreased risk. TIMI flow before PCI, LVEF, cTnI, baseline cumulative ST-segment deviation, WBC count, culprit lesion characteristics, BUN, creatinine, random blood sugar, and serum potassium were among the strongest contributors to model output.

• 

  Fig. 6. Performance of the CatBoost model for in-hospital mortality prediction. This figure summarizes the predictive performance of the CatBoost model trained to estimate in-hospital mortality using ten key predictors. (a) Confusion matrix aggregated across stratified cross-validation folds, showing observed versus predicted mortality outcomes. (b) Receiver operating characteristic (ROC) curve demonstrating excellent discriminative performance (AUC = 0.93).