Predicting long‐term freedom from atrial fibrillation after catheter ablation by a machine learning algorithm: Validation of the CAAP‐AF score

Background Preprocedural clinical predictors of the successful maintenance of sinus rhythm may contribute to optimal treatment strategies for atrial fibrillation (AF). The CAAP‐AF score, a novel simple tool scored as 0‐13 points (including six independent variables) has been proposed to predict long‐term freedom from AF after catheter ablation. To clarify its reproducibility, we examined the CAAP‐AF score's predictive performance and then created subgroups to best predict AF recurrence by using a machine learning algorithm. Methods We studied 583 consecutive patients who underwent initial AF catheter ablation at our institute (median CAAP‐AF score, 5; age, 66 ± 10 years old; female, 28.3%; coronary artery disease, 10.8%; left atrial diameter, 39.9 ± 6.6 mm; number of antiarrhythmic drugs failed, 0.4 ± 0.6; nonparoxysmal AF, 45.3%). All were systematically followed up with an endpoint of atrial tachyarrhythmia recurrence after the last ablation procedure. Results During the 1.8 ± 1.2‐year follow‐up, 157 patients had atrial tachyarrhythmia recurrence. Repeated procedures were performed (n = 115). Arrhythmia recurrence after the last session occurred in 69 patients. We created Kaplan‐Meier curves for freedom from AF after final AF ablation for ranges of CAAP‐AF scores; these confirmed the original study results. The machine learning using Classification and Regression Trees divided the patients into three categories by the risk score: low (score ≤5), intermediate (score 6‐8), and high (score ≥9). Conclusions The CAAP‐AF score was useful to stratify the atrial tachyarrhythmia recurrence risk in AF patients undergoing catheter ablation into three categories. The score should be considered when deciding whether to perform AF ablation in clinical practice.
We validated the CAAP‐AF score's effectiveness in an all‐comer setting to predict long‐term freedom from AF postcatheter ablation. We further created subgroups by using a machine learning method to best risk‐stratify the patients: low‐risk (the CAAP‐AF score ≤5), intermediate‐risk (score 6–8), and high‐risk (score ≥9).