A Comparison of Interpretable Machine Learning Models to Predict In-Hospital Mortality After Myocardial Infarction: Analyzing Two Years Data from a High-Volume Interventional Center

Abstract

The most common cause of death among patients with cardiovascular diseases is myocardial infarction (MI). Identifying predictors for in-hospital mortality is an essential step toward MI prevention and consequent reduction in mortality. We aimed to develop machine learning (ML) methods for predicting in-hospital mortality in MI patients and apply novel techniques for models’ interpretability to detect the predictive importance of the variables. Random forest (RF) and extreme gradient boosting (XGB) are applied to a dataset of 2035 MI patients who underwent percutaneous coronary intervention. When comparing the models’ AUC (RF—0.9712 vs. XGB—0.9666) and accuracy (RF—97% versus XGB—98%), both techniques achieved similar performance. However, the RF model obtained a higher sensitivity (86%) than the XGBoost classifier (80%). Hypertension, cardiogenic shock, ejection fraction were identified as some of the main contributors to the outcome. Our paper contributes to the global effort of reducing mortality in patients with myocardial infarction by proposing two interpretable ML models that accurately predict in-hospital mortality in MI patients. These results are essential steps in improving current preventive strategies. However, future studies on larger datasets enriched with both categorical and continuous variables, and models validated on external data from other centers are needed to accurately assume generalizability in clinical practice.