Artificial Intelligence–Based Prediction of Major Adverse Cardiovascular Events Following Acute Myocardial Infarction: A Retrospective Cohort Study
Jeffrey Chidumebi Nebonta *
Enugu State University of Science and Technology, Enugu, Nigeria.
Chipuruime I. Ideh
Imo State University College of Medicine, Owerri, Nigeria.
Benjamin O. C. Echendu
Vinnytsia National Pirogov Medical University, Ukraine.
Chimaobi Nwevo
NYC Health + Hospital / South Brooklyn Health, 11235, Brooklyn NY, United States.
*Author to whom correspondence should be addressed.
Abstract
Background: Major adverse cardiovascular events (MACE) remain a substantial contributor to morbidity and mortality following acute myocardial infarction (AMI). Conventional risk prediction models may inadequately capture complex relationships among clinical predictors. This study evaluated the predictive performance of artificial intelligence–based machine learning algorithms for 12-month MACE following AMI.
Methods: This multicentre retrospective cohort study included adults hospitalised with AMI between January 2016 and December 2024. The primary outcome was 12-month MACE, comprising cardiovascular death, recurrent myocardial infarction, ischaemic stroke, or heart failure hospitalisation. Logistic regression, random forest, extreme gradient boosting (XGBoost), support vector machine, and artificial neural network models were developed using routinely collected clinical variables. Patients were allocated to training (80%) and held-out testing (20%) datasets. Performance was assessed using discrimination, calibration, classification metrics, and decision curve analysis.
Results: Among 20,846 patients (mean age, 64.8 years; 64.6% male), 3,684 (17.7%) experienced MACE. XGBoost achieved the highest area under the receiver operating characteristic curve (AUROC; 0.862; 95% CI, 0.846–0.878), outperforming random forest (0.841), artificial neural network (0.833), support vector machine (0.819), and logistic regression (0.784). The AUROC improvement over logistic regression was 0.078 (95% CI, 0.057–0.099; P < .001). XGBoost demonstrated 81.6% sensitivity, 77.9% specificity, and favourable calibration (slope, 0.96; Brier score, 0.112). Left ventricular ejection fraction, age, renal function, Killip classification, and previous heart failure were influential predictors. Discrimination remained broadly consistent across clinical subgroups and sensitivity analyses.
Conclusions: XGBoost demonstrated superior predictive discrimination and favourable calibration compared with conventional logistic regression for 12-month MACE following AMI. These findings suggest potential value for individualized post-infarction risk stratification, although external validation and prospective clinical evaluation are required before implementation.
Keywords: Acute myocardial infarction, artificial intelligence, machine learning, major adverse cardiovascular events, XGBoost, risk prediction