Atrial fibrillation recurrence after atrial flutter ablation in subjects initially diagnosed as atrial fibrillation: A machine learning approach using XGBoost

Abstract Background Some patients initially diagnosed with atrial fibrillation (AF) experience recurrence only as atrial flutter (AFL) after starting antiarrhythmic drugs. Under the current Korean health insurance policy, these patients are eligible for AFL ablation but not AF ablation. However, a portion of these patients later experience AF recurrence. The rate of AF recurrence after AFL ablation in patients initially diagnosed with AF remains unclear. Purpose This study aims to assess the incidence of AF recurrence following AFL ablation and to evaluate the predictive performance of an XGBoost machine learning algorithm using clinical variables. Methods Patients who underwent AFL ablation between September 2003 and October 2024 were included in the study. Exclusion criteria were: no initial diagnosis of AF, failed cavotricuspid isthmus ablation, concurrent pulmonary vein antrum isolation, or a previous history of AFL or AF ablation. The primary outcome was AF recurrence. Kaplan-Meier analysis was used to estimate the incidence probability of AF recurrence over time. Additionally, an XGBoost algorithm was developed to predict AF recurrence using clinical variables. Results Among the 49 patients (mean age 64.7 ± 12.0 years, 85.7% male), 27 (55.1%) experienced AF recurrence during a mean follow-up period of 4.8 ± 4.4 years. Kaplan-Meier analysis showed that the cumulative incidence probability of AF reached approximately 50% by year 2 post-ablation (Figure 1). Patients with AF recurrence had significantly lower CHA2DS2-VASc scores (1.7 ± 1.4 vs. 2.6 ± 1.4, p = 0.039) and lower CHARGE-AF scores (11.8 ± 4.1 vs. 12.2 ± 1.1, p = 0.033) with no difference in ARIC-AF score (10.9 ± 4.2 vs 12.5 ± 4.0, p = 0.163) compared to those without recurrence. The XGBoost algorithm demonstrated strong predictive performance for AF recurrence with an area under the receiver operating characteristic curve (AUROC) of 0.898, an F1 score of 0.857, accuracy of 0.857, precision of 0.857, and recall of 0.857 (Figure 2). The model was trained using clinical variables including age, weight, systolic blood pressure, AF duration, smoking status, and anti-hypertensive medication use. Conclusion AF recurrence after AFL ablation among patients with initial diagnosis as AF is common. The XGBoost algorithm demonstrated high accuracy and discrimination power in predicting AF recurrence using readily available clinical variables. This machine learning approach may help identify patients who could benefit from prophylactic pulmonary vein isolation when undergoing AFL ablation. Further validation in a bigger population is needed to confirm these findings.