Abstract 17861: Separating Patients With Long-Term Success versus Acute Response From Atrial Fibrillation Ablation Using Explainable Machine Learning

Introduction: It is difficult to predict outcomes from atrial fibrillation (AF) ablation. In particular, it is unclear what physiology explains patients with long-term AF elimination yet no acute response, or vice versa.

Hypothesis: We hypothesize that different combinations of electrophysiological and clinical features predict long-term outcome or acute response to AF ablation. We tested this hypothesis using machine learning of a large Registry.

Methods: We studied N=561 patients at pulmonary vein isolation (65±10 years, 28% female) with rich clinical data (44 elements, ECGs), and 64 pole global electrograms (EGMs). We propensity matched the dataset 70/30 for training/testing (fig A). We extracted (i) 4 EGM features: dominant frequency (DF), similarity over time, rate (CL), and voltage; (ii) 6 ECG features; (iii) 44 clinical features. We compared 6 ML models including logistic regression to predict 1 year and acute outcomes in the test set, then used explainability (e.g. Shapley scores) to explain the top models and contributing features for each outcome.

Results: Of 561 patients, N=390 (69.5%) had 1 year success and N=278 (49.6%) had AF termination by ablation. The best ML model predicted acute response (Random Forest), provided AUROC = 0.83 (fig B) with highest predictors of AF type, RR interval, CL, DF, and voltage. Conversely, the best model for long-term success provided AUC of 0.65 (accuracy of LT vs acute: p<0.001 McNemar test), but top features were PR interval, QRS axis, AF duration, and age (shapley Rsq = 0.11, fig C). Predictors for long term outcome poorly predicted acute outcome and vice versa (fig C).

Conclusions: Clinical features for long-term and acute response to AF ablation were quite different, applying several classification methods to a well-characterized registry. Confirmation in other registries may reveal novel phenotypes to guide clinical decision making or procedural planning, and shed mechanistic insights.