Abstract 16565: Machine Learning Model for Predicting Mortality in Acute Myocardial Infarction Patients Managed With or Without Beta-Blocker

Abstract

Introduction: Optimal personalized medical therapy is crucial for reducing the mortality risk of acute myocardial infarction (AMI). But, assessing the potential individual benefits of beta-blockers (BB) remains challenging despite the consensus of current guidelines. Hypothesis: The hypothesis is that machine learning (ML) models can predict the mortality benefit of BB for individual AMI patients, thus helping to identify subgroups with the highest clinical benefit. Methods: A total 13,104 AMI patients registered in the Korea AMI registry (KAMIR)-National Institutes of Health (NIH) were included in the study. After exclusion of in-hospital death (N = 505), datasets were randomly allocated to the training (N = 8467) and test datasets (N = 4132) by 2:1 ratio. Binary GLM Logistic Regression, Subspace Discriminant, Kernel Naive Bayes, Boosted Trees and KNN were used to train the ML model. Patients were stratified into quartiles (Q1, Q2, Q3, Q4) based on the estimated risk.The low-risk group comprised Q1 and Q2, while the high-risk group consisted of Q3 and Q4. Results: Binary GLM Logistic Regression outperformed other models in the 5-fold cross validation test (area under the curve = 0.85, sensitivity = 98.6%). Errors of estimated mortality were within 0.3% for all quartile groups. High-risk group identified by the ML model demonstrated significant reduction in mortality with the usage of BB, whereas no such benefit was observed in the low-risk group. Additionally, subgroups of chronic kidney disease (CKD), anemia, and old age (>65) displayed notable reduction in mortality when managed with BB, in the high-risk group. Conclusions: With the successful performance of ML model predicting mortality, we identified subpopulations within AMI patients who are highly likely to benefit from the usage of BB. Clinical decision-making could be potentially optimized for individual AMI patients by application of the ML model, through personalized risk scoring and identification.