A multimodal artificial intelligence approach for atrial fibrillation subtype classification and post-ablation recurrence risk stratification
O S Kwon, H J Park, D Kim, H N PakAbstract
Background
Clinical classification of atrial fibrillation (AF) subtypes plays an important role in treatment decisions and catheter ablation strategies, but a one-time binary classification may not appropriately capture the characteristics of atrial remodeling as AF progresses. We aimed to determine whether a multimodal ensemble artificial intelligence (AI) model combining structural, electrical, and functional features could predict AF subtype and provide prognostic utility for AF recurrence after catheter ablation.
Methods
A multimodal ensemble AI model was developed to classify AF subtypes using the Yonsei AF cohort (n = 4,067). The model incorporated left atrial wall thickness (LAWT), full-length and median 12-lead electrocardiograms (ECGs), ECG-derived features, transthoracic echocardiographic parameters, and clinical variables. The dataset was randomly divided into training (75%) and internal holdout test sets (25%, n = 1,017); all performance metrics were computed on the internal holdout test set. We also assessed the association between AI-clinical subtype discordance and markers of atrial remodeling.
Results
The ensemble model achieved the highest performance (AUC, 0.84; F1 score, 0.68) compared with single-modality models (AUC range, 0.73–0.78). SHapley Additive exPlanations (SHAP) analysis indicated that median ECG and LAWT contributed most significantly. AI-predicted persistent AF (PeAF) probability showed slightly higher discrimination for 3-year recurrence than clinical PeAF classification (AUC, 0.61 vs. 0.59; C-index, 0.59 vs. 0.57, respectively). In survival analysis, AI-predicted PeAF probability (per 1 standard deviation increase) remained independently associated with recurrence even after adjustment for clinical factors and AF subtype (P = 0.008). Consistently, patients with clinical paroxysmal AF predicted using AI as PeAF showed a significantly higher AF recurrence risk (log-rank, P = 0.005).
Conclusion
Our findings suggest that AI-predicted AF subtypes may better reflect atrial remodeling and post-ablation recurrence risk. Continuous AF subtype monitoring using AI may improve risk stratification and support personalized post-ablation monitoring strategies.