Development and Validation of an Interpretable Machine Learning Model for Predicting 1‐Year Cardiac Death After Percutaneous Coronary Intervention in Patients With Acute Myocardial Infarction: A Multicenter Study

Background Risk prediction of cardiac death following percutaneous coronary intervention remains suboptimal in acute myocardial infarction. This study aimed to develop and externally validate an interpretable machine learning model using only routine laboratory and demographic variables to predict 1‐year cardiac death in this population. Methods We retrospectively enrolled 19 284 patients with acute myocardial infarction who underwent percutaneous coronary intervention across 82 hospitals in Tianjin, China between January 2010 and March 2024. The cohort was randomly split into training (70%, n=13 499) and internal validation (30%, n=5785) sets. An external cohort of 2048 patients from an independent center was used for validation. A Light Gradient‐Boosting Machine model based solely on routinely available laboratory and demographic variables, with no imaging inputs, was developed and compared with GRACE (Global Registry of Acute Coronary Events) scores. Shapley Additive Explanations were used to assess model interpretability. Results In the original data set, 1984 patients experienced 1‐year cardiac death. The model achieved strong discrimination in internal validation (area under the curve 0.921, precision‐recall area under the curve 0.711, sensitivity 79.7%). In the external validation cohort, LightGBM achieved a precision‐recall area under the curve of 0.162 and significantly outperformed the GRACE score in discrimination (area under the curve, 0.811 versus 0.728; P =0.001). Complementary assessments of calibration and decision‐curve analysis supported the overall findings. Conclusions This interpretable machine learning model based exclusively on routine laboratory and demographic variables outperformed the GRACE score in predicting 1‐year cardiac death after percutaneous coronary intervention in patients with acute myocardial infarction. Its strong discrimination and external validity support its potential for real‐world risk stratification and individualized management.