XGBoost and LightGBM machine learning models accurately predicted post-ablation recurrence for paroxysmal (AUC 0.831) and persistent (AUC 0.917) atrial fibrillation.
Cohort (n=297)
No
Do machine learning models incorporating right atrial structural parameters improve the prediction of atrial fibrillation recurrence in patients undergoing radiofrequency catheter ablation?
Machine learning models incorporating right atrial structural parameters from cardiac CT can accurately predict AF recurrence after catheter ablation, highlighting the right atrial appendage short diameter and right atrial volume as key predictors for paroxysmal and persistent AF, respectively.
Background The high recurrence rate following catheter ablation for atrial fibrillation (AF) remains a significant clinical challenge. Existing prediction models are predominantly limited to left atrial parameters and often fail to distinguish between AF types, resulting in suboptimal predictive accuracy and clinical utility. This study aimed to develop distinct machine learning (ML) models for predicting recurrence in patients with paroxysmal AF (PaAF) and persistent AF (PeAF), with a specific focus on evaluating the predictive value of right atrial structural parameters. Furthermore, explainable artificial intelligence (XAI) techniques were employed to decipher the decision-making mechanisms of the models. Methods We retrospectively enrolled 297 patients who underwent radiofrequency catheter ablation (RFCA) for AF (230 in the PaAF group; 67 in the PeAF group). A total of 37 clinical and cardiac computed tomography (CT) imaging features were collected. Following feature selection, eight ML algorithms were trained and evaluated. The SHapley Additive exPlanations (SHAP) framework was used to provide model interpretability, and a clinical decision support tool was developed. Results The XGBoost and LightGBM models demonstrated superior predictive performance for PaAF Area Under the Receiver Operating Characteristic Curve (AUC): 0.831 and PeAF (AUC: 0.917), respectively. SHAP analysis identified the right atrial appendage (RAA) short diameter as the most important predictor for PaAF recurrence, whereas right atrial (RA) volume was the top contributor to predicting PeAF recurrence. Feature dependence plots further revealed complex nonlinear relationships and interaction effects. Conclusion Machine learning models exhibited excellent performance in predicting post-ablation AF recurrence, with right atrial structural parameters emerging as key predictors. The explainable framework and clinical decision support tool developed in this study provide a new paradigm for precise prognosis assessment and personalized management following AF ablation.
Han et al. (Wed,) conducted a cohort in Atrial fibrillation (n=297). Machine learning prediction models (XGBoost and LightGBM) was evaluated on Prediction of atrial fibrillation recurrence (AUC). XGBoost and LightGBM machine learning models accurately predicted post-ablation recurrence for paroxysmal (AUC 0.831) and persistent (AUC 0.917) atrial fibrillation.