Key result
XGBoost outperforms other models predicting in-hospital MACCE after STEMI PCI with ~0.85 AUROC.
Why the study?
Despite timely reperfusion, 7-9% of patients undergoing primary PCI for STEMI experience in-hospital MACCE, motivating the development and temporal validation of predictive machine-learning models.
Do machine-learning models improve the prediction of in-hospital MACCE in STEMI patients undergoing primary PCI compared to logistic regression?
Do machine-learning models improve the prediction of in-hospital MACCE in STEMI patients undergoing primary PCI compared to logistic regression?
An explainable gradient-boosted machine learning model (XGBoost) provided superior prediction of in-hospital MACCE after primary PCI for STEMI compared with traditional logistic regression.
Answers are grounded in this paper and cite their sources.
May support ML risk stratification after STEMI PCI; hypothesis-generating and should not yet change practice.
Background Despite timely reperfusion, 7-9% of patients undergoing primary percutaneous coronary intervention (PCI) for ST-elevation myocardial infarction (STEMI) experience major adverse cardiac and cerebrovascular events (MACCE) within the index hospitalization. Purpose To develop and temporally validate machine-learning (ML) models that predict in-hospital MACCE after primary PCI in STEMI admissions using a nationally representative dataset. Methods We analyzed the Nationwide Inpatient Sample (2016-2022). After survey weighting, 822 980 STEMI admissions treated with PCI were included. Predictors comprised demographics, socioeconomic factors, hospital characteristics, and 29 Elixhauser comorbidities. MACCE was defined as a composite of in-hospital mortality, ischemic stroke, coronary artery bypass grafting, cardiogenic shock, or cardiac arrest. Five models—Extreme Gradient Boosting (XGBoost), LightGBM, random forest, multilayer perceptron, and logistic regression—were trained on 2016-2021 data and temporally validated on 2022 admissions. Performance metrics included discrimination (area under the receiver operating characteristic curve [AUROC], 95% confidence intervals [CI] via 1,000 bootstraps), calibration (Brier score), pairwise model comparison (bootstrap test, n=1,000), and interpretability analysis (Shapley Additive Explanations values). Sex- and race-stratified AUROCs evaluated fairness. Sex- and race-stratified AUROCs evaluated fairness. Results In-hospital MACCE occurred in 8.8% of admissions. In the 2022 test set, XGBoost achieved the highest discrimination (AUROC 0.846, 95% CI 0.839–0.853; Brier 0.062), outperforming LightGBM (0.842, 95% CI 0.835–0.850), logistic regression (0.840, 95% CI 0.833–0.848; p < 0.001 vs XGBoost), random forest (0.838, 95% CI 0.830–0.846), and multilayer perceptron (0.821, 95% CI 0.812–0.829). Calibration was preserved across models (all Brier ≤ 0.066). AUROC drift was minimal across sex (men 0.847 vs women 0.839) and race (Black 0.848, White 0.847, Hispanic 0.837; largest gap = 0.011) [Figure 1]. SHAP interpretation showed that patients at lowest risk for MACCE typically presented without cardiac arrhythmias, electrolyte abnormalities, or neurological comorbidities. Conclusion An explainable gradient-boosted model trained on routinely captured admission data provided superior prediction of in-hospital MACCE after primary PCI for STEMI compared with logistic regression and other ML approaches, while maintaining equitable performance across sex and racial groups. These findings need to be validated using prospective data to confirm generalizability and clinical utility.
No takes yet. Share an insight, caveat, or question.
Singh et al. (2026) studied this question. The XGBoost model achieved the highest discrimination for predicting in-hospital MACCE after PCI for STEMI, with an AUROC of 0.846 compared to other models.
Synapse has enriched 10 related papers. Consider them for comparative context: