부산대학교 도서관

Introduction:Left ventricular ejection fraction (LVEF) of ≤ 40% in patients with coronary artery disease (CAD) is associated with an increased risk of sudden cardiac death. Accurately predicting LVEF at clinical follow-up can identify patients needing an implantable cardioverter-defibrillator device.Hypothesis:To develop a predictive algorithm using ML methods for predicting LVEF on follow-up echocardiogram for individuals with CAD.Methods:We identified the 10 most significant predictors of follow-up LVEF using a variety of feature selection algorithms from 109 variables based on clinical, demographic, ECG, and echocardiography data from 2116 hospitalized patients diagnosed with CAD. Various predictive ML algorithms were constructed to predict follow-up LVEF at 60 to 730 days from admission through regression, binary classification (LVEF ≤ 40% and LVEF > 40%), and multi-classification analysis. AUC scores, accuracy, precision, F Score, and recall were calculated for each model to evaluate outcome performance.Results:N = 421 patients had LVEF ≤ 40% on presentation and 535 had LVEF ≤ 40% on follow-up. We observed that binary classification tasks outperformed regression and multi-classification tasks (Graph 1). In a binary classification setting, the top-performing models were simple neural networks, random forests, or logistic regression. (Table 1).Discussion:We developed various ML algorithms in multiple settings to optimize predicting follow-up LVEF with patients with CAD. With an enhanced ability to predict reduced LVEF, we aspire to improve the timing and efficiency of appropriate ICD implantation in patients based on current guidelines.