Machine learning prediction of post-CABG atrial fibrillation using clinical and pharmacogenomic biomarkers

Background Postoperative atrial fibrillation (POAF) is a frequent complication following coronary artery bypass grafting (CABG), significantly impacting patient prognosis and healthcare costs. This study aimed to develop an integrated predictive model for POAF risk stratification to optimize clinical management. Methods We retrospectively analyzed 2,528 patients undergoing 21-gene pharmacogenetic testing for cardiovascular therapy. After stringent data curation, 576 CABG patients were enrolled and randomly allocated into training and test sets. Eight machine learning algorithms were trained using clinical variables and genetic variants. An independent validation set was performed on 61 patients from a subsequent 1,075-patient cohort of 21-gene pharmacogenetic testing. Results Eight machine learning algorithms were trained, tested, and validated, with the Gaussian Naive Bayes (GNB) model demonstrating robust performance (Accuracy: 0.81 in test set and 0.79 in independent validation set). SHapley Additive exPlanations analysis identified four key predictors: multivessel CABG (CABGVx ≥ 3), history of heart failure (HFHx), rs5219 ( KCNJ11 ), and prolonged bypass duration (CABGTime). To facilitate clinical translation, we developed an accessible web-based tool ( https://www.xingyeyard.site/cabg/ ) for real-time POAF risk stratification. Conclusion This GNB-based classifier synergistically integrates Pharmacogenomic and clinical predictors to predict POAF risk following CABG. The combination of rigorous validation and user-centered design positions this model as a valuable clinical decision-support tool for optimizing personalized perioperative care.