Predicting the When: Multimodal AI for Time-to-Recurrence Analysis After Atrial Fibrillation Ablation
Journal:
medRxiv
Published Date:
May 15, 2026
Abstract
Background: Catheter ablation is the most effective rhythm control strategy for atrial fibrillation (AF); however, recurrence remains common. Current post-ablation management follows largely population-level protocols, constrained by the absence of tools that can anticipate not merely whether, but when, an individual patient will experience recurrence. The emergence of multimodal artificial intelligence (AI) presents a new opportunity to address this unmet clinical need. Objective: To develop a predictive model for time-to-AF-recurrence post-ablation using pre-procedural bi-atrial imaging, clinical covariates, and procedural characteristics, within a novel multimodal AI and survival analysis framework. Methods: We analyzed a retrospective cohort of 437 AF patients who underwent catheter ablation with follow-up censored at 36 months. MARTA-AF (Multimodal AI Recurrence and Time-to-event Analysis post-Ablation in AF) was trained on pre-procedural bi-atrial images, and covariates/procedural characteristics, and integrated into a survival model to generate time-varying recurrence probability estimates. Model interpretability was achieved by quantifying contribution of covariates/procedural characteristics to predicted survival probabilities. Results: MARTA-AF successfully predicted time-varying recurrence risk up to three years post-ablation. Patients were effectively stratified into low- and high-risk groups, with statistically significant discrimination sustained over the follow-up period. The model demonstrated consistent performance across clinically relevant subgroups, including sex, age, and AF type. Incorporation of right atrial shape features improved time-to-AF-recurrence prediction. Interpretability analyses identified key recurrence predictors. Conclusions: MARTA-AF delivers individualized, time-varying AF recurrence risk forecasts and enables stratification into clinically meaningful risk groups. This framework has the potential to transform post- ablation management into a proactive paradigm and to support informed clinical decision-making prior to ablation.
