Predicting Atrial Fibrillation in Patients with Ischemic Heart Disease Based on Multilevel Categorization

The aim of the study was to develop new prognostic models of postoperative atrial fibrillation (PoAF) in patients with ischemic heart disease (IHD) after coronary artery bypass grafting (CABG) based on preoperative predictors and to assess the effectiveness of their multilevel categorization to improve the quality of the prognosis and its clinical interpretation. A single-center retrospective cohort study was conducted, analyzing the data of 1305 medical histories of patients with IHD who underwent elective isolated CABG. Two groups were identified, the first group included 280 (21.5%) patients with PoAF, and the second group included 1025 (78.5%) patients without rhythm disturbances. Prognostic models of PoAF were developed using multifactorial logistic regression (MLR), random forest (RF), and stochastic gradient boosting (SGB) methods. The predictors were dichotomized using optimal cutoff points grid search methods, centroid calculation, and Shapley additive explanations (SHAP). For multilevel categorization, it was proposed to combine the threshold values identified during dichotomization and rank them based on cutoff thresholds using MLR weight coefficients (multi-metric categorization method). As a result of the multi-stage selection, 9 PoAF predictors were identified, validated and categorized. Prognostic models were developed with continuous, dichotomous, and multilevel categorical variables. The best SGB model with continuous predictors had an AUC of 0.795. Models with predictors identified by the multi-metric categorization method showed better performance than the models with continuous variables (AUC—0.802).