Investigators enrolled 5,829 patients with STEMI who underwent emergency percutaneous coronary intervention at a single emergency critical care center between January 1, 2015, and July 1, 2025. The cohort's mean age was 58.1±12.3 years, and 4,875 patients (83.63%) were male. The study team defined in-hospital major adverse cardiovascular events (MACE) to include cardiogenic shock, severe heart failure, severe arrhythmia, acute stent thrombosis or recurrent acute myocardial infarction, non-fatal stroke, and cardiogenic death. A total of 524 patients experienced MACE and 5,305 did not. For model development, the investigators split the dataset into a training set (4,080 patients) and a test set (1,749 patients) at a 7:3 ratio for internal validation and performance evaluation.
Data elements collected encompassed demographics, lifestyle factors, prior medical history, admission vital signs, laboratory tests performed within 24 hours of admission, echocardiographic parameters including left ventricular ejection fraction (LVEF), Killip classification, and coronary intervention details. The authors reported using univariate analysis followed by Lasso regression to select predictive variables. Fourteen features remained for model construction: B-type natriuretic peptide, myoglobin, LVEF, high-density lipoprotein cholesterol, high-sensitivity troponin I, triglycerides, triglyceride-glucose index, low-density lipoprotein cholesterol, glycated hemoglobin, creatine kinase-MB, smoking history, heart rate, Killip class, and age.
The study team trained five classification algorithms: support vector machine (SVM), logistic regression, random forest, naive Bayes, and XGBoost. Model discrimination was evaluated using the area under the receiver-operating characteristic curve (AUC-ROC) and the area under the precision-recall curve (AUPRC). The investigators also used calibration curves to assess agreement between predicted probabilities and observed outcomes, and decision curve analysis to quantify clinical net benefit across decision thresholds. The abstract reports that these metrics guided selection of the optimal model, although specific hyperparameter settings or training procedures were not provided in the abstract.
Among the five candidate algorithms, the XGBoost model produced the highest reported discrimination and precision-recall performance in the test set. The authors reported an AUC-ROC of 0.992 and an AUPRC of 0.947 for XGBoost. They also noted that the model exhibited favorable probability calibration and provided clinical net benefit on decision curve analysis. The abstract presents these results as evidence that the model can distinguish patients at high risk of in-hospital MACE after emergency PCI and that its probability estimates aligned well with observed event rates in the internal test cohort.
To increase interpretability, the investigators applied SHapley Additive exPlanations (SHAP) to evaluate each feature's contribution to model predictions. SHAP analysis ranked the top five contributors as B-type natriuretic peptide, creatine kinase-MB, myoglobin, LVEF, and triglyceride-glucose index. These variables, according to the authors, carried the largest influence on the predicted probability of in-hospital MACE in the trained XGBoost model.
The study performed internal validation using a held-out test set drawn from the same institution and time span; investigators reported discrimination, precision-recall, calibration, and decision-curve results for that internal cohort. The abstract did not report external validation across other centers or populations. The authors did not provide details in the abstract about model hyperparameters, class-imbalance handling, cross-validation strategy, or thresholds used for clinical decision rules. Implementation specifics such as prospective testing, integration into clinical workflows, or software packaging were not reported in the abstract. These omissions limit assessment of generalizability and operational readiness.
Limitations reported implicitly in the abstract include that validation was internal only; the authors presented their model as a reference for early in-hospital risk stratification and clinical decision-making but did not document external testing or deployment data within the abstract. Future steps that remain unreported in this abstract include external validation in other centers, prospective impact assessment on clinical decision-making, and transparent reporting of modeling choices (for example, hyperparameters and handling of missing data).