Abstract
Oral anticoagulation (OAC) following catheter ablation (CA) of nonvalvular atrial fibrillation (NVAF) is essential for the prevention of thrombosis events. Inappropriate application of OACs does not benefit stroke prevention but may be associated with a higher risk of bleeding. Therefore, this study aims to develop clinical data-driven machine learning (ML) methods to predict the risk of thrombosis and bleeding to establish more precise anticoagulation strategies for patients with NVAF.Patients with NVAF who underwent CA therapy were enrolled from Southwest Hospital from 2015 to 2023. This study compared eight ML algorithms to evaluate the predictive power for both thrombosis and bleeding. Model interpretations were recognized by feature importance and SHapley Additive exPlanations methods. With potential essential risk factors, simplified ML models were proposed to improve the feasibility of the tool.A total of 1,055 participants were recruited, including 105 patients with thrombosis and 252 patients with bleeding. The models based on XGBoost achieved the best performance with accuracies of 0.740 and 0.781 for thrombosis and bleeding, respectively. Age, BNP, and the duration of heparin are closely related to the high risk of thrombosis, whereas the anticoagulation strategy, BNP, and lipids play a crucial role in the occurrence of bleeding. The optimized models enrolling crucial risk factors, RF-T for thrombosis and Xw-B for bleeding, achieved the best recalls of 0.774 and 0.780, respectively.The optimized models will have a great application potential in predicting thrombosis and bleeding among patients with NVAF and will form the basis for future score scales.