Development and Validation of an Explainable Machine Learning Model for Predicting Repeat Catheter Ablation for Atrial Fibrillation: A Single-Center Retrospective Cohort Study

Shuai Shang,1,2,* Huasheng Lv,1,2,* Guoxiang Ma,3 Meng Wei,1,2 Kai Wang,3 Yanmei Lu,1,2 Baopeng Tang1,2 1Department of Cardiac Pacing and Electrophysiology, The First Affiliated Hospital of Xinjiang Medical University, Urumqi, People’s Republic of China; 2Xinjiang Key Laboratory of Cardiac Electrophysiology and Remodeling, The First Affiliated Hospital of Xinjiang Medical University, Urumqi, People’s Republic of China; 3Department of Medical Engineering and Technology, Xinjiang Medical University, Urumqi, People’s Republic of China*These authors contributed equally to this workCorrespondence: Baopeng Tang, Email tangbaopeng1111 @163.com Yanmei Lu, Email gracy @189.cnBackground: Atrial fibrillation (AF) is the most prevalent sustained cardiac arrhythmia worldwide. Catheter ablation is the first-line therapy for symptomatic/refractory AF, yet post-procedural recurrence remains extremely common, driving a high rate of repeat ablation procedures. Repeat ablation is associated with elevated medical costs, incremental procedural risks, and impaired quality of life and clinical outcomes in affected patients. Existing clinical risk scores for predicting repeat AF ablation have limited discriminative ability, poor interpretability, and suboptimal clinical utility. This study aimed to develop and validate an explainable machine learning model, using routine clinical and echocardiographic features, to predict the risk of requiring repeat catheter ablation for AF.Methods: A retrospective cohort of 1073 patients undergoing AF ablation from 2012 to 2023 was analyzed, with data split into training (70%) and testing (30%) sets. Feature selection was performed using LASSO regression and the Boruta algorithm, followed by the construction of eight machine learning models. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, F1 score, balanced accuracy, Brier score, and clinical utility via decision curve analysis. Interpretability was enhanced using Shapley Additive Explanations (SHAP).Results: Among 1073 patients undergoing AF ablation, 352 (32.8%) required a second procedure. LASSO regression combined with the Boruta algorithm identified nine predictive features: NT-proBNP, age, globulin (GLO), direct bilirubin (DBIL), left ventricular ejection fraction (LVEF), cystatin C (Cys-C), smoking history, creatine kinase (CK), and urea. Among the eight models evaluated, XGBoost demonstrated the best overall performance, achieving an AUC of 0.811 (95% CI: 0.762– 0.859) in the testing cohort, with a sensitivity of 0.748, specificity of 0.726, and Brier score of 0.1682. It also outperformed alternative models in terms of F1 score and clinical net benefit. SHAP analysis confirmed NT-proBNP and age as the most influential predictors, alongside non-linear contributions from the remaining variables.Conclusion: The XGBoost model may provide a useful and interpretable tool for predicting repeat AF ablation, providing clinical insights to guide patient management and optimize procedural outcomes.Keywords: atrial fibrillation, catheter ablation, machine learning, repeat ablation, XGBoost, SHAP, prediction model, interpretability