Abstract 4370598: Machine Learning Models to Predict High Atrial Fibrillation Burden Post-Catheter Ablation in Patients with Persistent AF: Insights from the DECAAF II Trial

Background: High atrial fibrillation (AF) burden is associated with increased risk of stroke and heart failure. While catheter ablation reduces AF burden in most patients, a minority remain at risk for high AF burden after the procedure. Objective: In this study, we aimed to utilize machine learning to predict high AF burden post-ablation in patients with persistent AF. Methods: This study analyzed six hundred and eighty-five with persistent AF (mean age: 62.0 ± 9.1; women: 20%) who underwent catheter ablation in the DECAAF II trial and were followed for a total of 540 days. Four machine-learning models—Elastic Net, Decision Tree, Random Forest, and XGBoost—were developed to predict each of AF recurrence and AF burden ≥10% using 200 pre-ablation variables, including clinical, MRI, and laboratory data. The models were trained and validated using stratified 5-fold cross-validation. SHapley Additive exPlanations (SHAP) were derived to explain the most impactful features collected from each patient. Results: The XGBoost models outperformed all other models in predicting AF recurrence (30 variables; cross-validated AUC = 0.64 ± 0.04) and AF burden ≥ 10% (27 variables; cross-validated AUC of 0.66 ± 0.03) (Figure 1A). SHAP analysis revealed the top predictors of high AF burden on a patient-specific level, including left atrial volume index (importance: 0.1), age (0.02), left atrial appendage enhancement (0.01), Utah stage <3 (0.01), and left pulmonary vein enhancement percentage (0.01). Fatigue at rest (0.01) and frequency of AF episodes (0.001) based on patient-filled questionnaires (University of Toronto Atrial Fibrillation Severity Scale- AFSS) did contribute to the prediction (Figure 1B). Conclusion: XGBoost models, augmented by SHAP explainability, were the most reliable and explainable models for predicting both recurrence and high post-ablation AF burden (≥10%). These models can lead to a more granular risk stratification and facilitate future patient-specific management.