Machine-learning based calculator for personalized risk following ST-elevation myocardial infarction

Abstract

Abstract Background Despite recent advancements in the management of patients with ST-elevation myocardial infarction (STEMI), a high risk for adverse events remains during the first year following the event. There are several risk scores which estimate the risk of major adverse cardiac events (MACE) following STEMI. However, the most established risk scores are not contemporary, their diagnostic accuracy ranges across populations, and they are far from being personalized. Purpose We aimed to develop a machine learning (ML)-based calculator to provide patients and physicians with a personalized prognostic tool based on modern-world data to determine outcomes for patients following STEMI. Our goal was to predict mortality within 1-year post-STEMI and contribute personalized recommendations to reduce the risk of events. Additionally, we aimed to have the calculator provide a personalized risk assessment for each independent variable. Methods We included patients from two large tertiary care centers in Israel and Italy. 2888 patients who suffered STEMI between the years of 2000 – 2020 were included in our training model. Our calculator leveraged CatBoost for supervised machine learning, utilizing gradient boosting on decision trees. The model was subsequently tested on 321 patients from the Israeli center and 132 patients from the Italian center. Results The calculator was able to accurately predict 1-year mortality (96.5% accuracy, ROC 0.97 in the Israeli center and 93.2% accuracy, ROC 0.89 in the Italian center). In the full cohort, left ventricular ejection fraction (LVEF) post-STEMI demonstrated the strongest contribution towards predicted outcome with a mean Shapley value (SHAP) of 0.978, with lower EFs correlating to higher likelihood of one-year mortality (figures 1 and 2). Other strong contributing factors were identified, including glomerular filtration rate (GFR), age, and Killip classification, with mean SHAPs of 0.580, 0.571, and 0.563, respectively. Importantly, the algorithm was designed to also output a personalized assessment of risk for each patient based on their individual variables. Conclusion ML-derived STEMI calculators can determine which risk factors and variables provide highest predictive value in determining outcomes in a personalized manner. They can be leveraged for high-accuracy prognostication and provide guidance for physicians and patients after STEMI.