Korean Artificial Intelligence-Based Risk Model for Predicting Mortality After Acute Myocardial Infarction

Acute myocardial infarction is associated with substantial mortality risk that persists beyond the acute phase. Many existing post-acute myocardial infarction risk models were developed before contemporary advances in treatment, potentially limiting their relevance in current practice. We aimed to develop simplified machine learning-based models to predict short- and long-term mortality after acute myocardial infarction in a contemporary therapeutic context.

The Korean Artificial Intelligence-Based Risk Model for Acute Myocardial Infarction (KARMA) was developed to predict 3-month and 3-year mortality using a boosted decision tree algorithm. Model development and internal validation were performed using the KAMIR (Korea Acute Myocardial Infarction Registry)-National Institutes of Health registry (2011-2015), with external validation in the KAMIR-V registry (2016-2020). Seven routinely available clinical variables were included.

The KARMA models demonstrated excellent discrimination for both early and late mortality. Areas under the receiver operating characteristic curves for 3-month KARMA and 3-year KARMA were 0.91 (95% CI, 0.88-0.94) and 0.85 (95% CI, 0.82-0.87), respectively, significantly outperforming the GRACE (Global Registry of Acute Coronary Events) score (3-month score: 0.79; 3-year score: 0.80) and the KAMIR score (3-month score: 0.84; 3-year score: 0.82). Kaplan-Meier curves showed clear and sustained separation across KARMA risk groups. Predictive performance was consistent across subgroups defined by age, sex, MI type, left ventricular dysfunction, and renal dysfunction. External validation in the KAMIR-V cohort confirmed robust performance (area under the receiver operating characteristic curve, 0.92 for 3-month and 0.86 for 3-year mortality).

The KARMA scores provide a reliable and readily accessible tool for predicting short- and long-term post-acute myocardial infarction mortality. Identification of high-risk individuals would enable providers to optimize management strategies that could improve outcomes.