# Multilevel predictors categorization for post-CABG atrial fibrillation prediction

## TL;DR

This study develops predictive models for post-CABG atrial fibrillation using multilevel categorization to improve clinical interpretability.

## Contribution

The novel multimetric categorization method enhances model explainability without sacrificing predictive accuracy.

## Key findings

- The best XGB model with continuous predictors achieved an AUC of 0.76.
- Multimetric categorization models showed comparable performance (AUC = 0.758) with better interpretability.
- Nine PoAF predictors were identified and validated using multistage selection.

## Abstract

Postoperative atrial fibrillation (PoAF) is a common complication after coronary artery bypass grafting (CABG). Despite its association with increased risk of ischemic stroke, bleeding, acute renal failure and mortality there is still no ideal predictive tool with proper clinical interpretability. A retrospective single-center cohort study enrolled 1305 electronic medical records of patients with elective isolated CABG. PoAF was identified in 280 (21.5%) patients. Prognostic models with continuous variables were developed utilizing multivariate logistic regression (MLR), random forest and eXtreme gradient boosting methods. Predictors were dichotomized via grid search for optimal cut-off points, centroid calculation, and Shapley additive explanation (SHAP). For multilevel categorization, we proposed to use threshold values combination identified during dichotomization, as well as ranking cut-off thresholds by MLR weighting coefficients (multimetric categorization method). Based on multistage selection, nine PoAF predictors were identified and validated. After categorization, prognostic models with continuous and multilevel categorical variables were developed. The best XGB model employing continuous predictors demonstrated an AUC = 0.76. Models in which predictors were derived utilizing the multimetric categorization approach showed comparable predictive performance (AUC = 0.758). The main advantage of models with multilevel predictors categorization was their superior explainability and clinical interpretability in predicting POAF. Multilevel predictors categorization represents a promising tool for improving the explainability of POAF predictive development estimates. Using the developed prognostic models, it was demonstrated that the categorization procedures proposed by the authors ensure both high predictive accuracy and transparency of the generated clinical conclusions.