Explainable Machine Learning to Predict Anchored Reentry Substrate Created by Persistent Atrial Fibrillation Ablation in Computational Models

BACKGROUND: Postablation arrhythmia recurrence occurs in ~40% of patients with persistent atrial fibrillation. Fibrotic remodeling exacerbates arrhythmic activity in persistent atrial fibrillation and can play a key role in reentrant arrhythmia, but emergent interaction between nonconductive ablation‐induced scar and native fibrosis (ie, residual fibrosis) is poorly understood. METHODS AND RESULTS: We conducted computational simulations in pre‐ and postablation left atrial models reconstructed from late gadolinium enhanced magnetic resonance imaging scans to test the hypothesis that ablation in patients with persistent atrial fibrillation creates new substrate conducive to recurrent arrhythmia mediated by anchored reentry. We trained a random forest machine learning classifier to accurately pinpoint specific nonconductive tissue regions (ie, areas of ablation‐delivered scar or vein/valve boundaries) with the capacity to serve as substrate for anchored reentry‐driven recurrent arrhythmia (area under the curve: 0.91±0.03). Our analysis suggests there is a distinctive nonconductive tissue pattern prone to serving as arrhythmogenic substrate in postablation models, defined by a specific size and proximity to residual fibrosis. CONCLUSIONS: Overall, this suggests persistent atrial fibrillation ablation transforms substrate that favors functional reentry (ie, rotors meandering in excitable tissue) into an arrhythmogenic milieu more conducive to anchored reentry. Our work also indicates that explainable machine learning and computational simulations can be combined to effectively probe mechanisms of recurrent arrhythmia.