Can atrial fibrillation ablation outcomes be properly predicted with electrocardiography and artificial intelligence?

Aims: The success of ablation for atrial fibrillation (AF) varies, often leading to repeat ablation. Reliable prediction of repeat ablation remains challenging. This study aimed to investigate if AF ablation outcomes can be predicted with an electrocardiogram (ECG)-based deep learning (DL) algorithm. Methods and results: We included 865 patients undergoing AF ablation, of whom 163 (18.8%) required a repeat procedure during a minimum follow-up of 572 days. A deep neural network was trained on the raw data of the standard 12-lead ECG obtained within 3 months prior to the index ablation, using stratified nine-fold nested cross-validation. Unfortunately, the model achieved a nested cross-validation area under the receiver operating characteristic curve (AUC) of only 0.61 (95% CI: 0.57-0.64). For comparison, the same analytic approach achieved significantly higher accuracy for sex classification (AUC = 0.87, 95% CI: 0.86-0.89). A random forest model only using clinical variables (age, sex, body mass index, AF pattern) yielded a similar performance for a repeat ablation (cross-validated AUC = 0.59, 95% CI: 0.55-0.63), suggesting limited added value of ECG-based prediction. SHapley Additive exPlanations was used to pinpoint the most relevant ECG segments and highlighted contributions from P-wave, QRS-complex, and T-wave features. Conclusion: The DL model demonstrated limited ability to predict repeat AF ablation based on the standard 10-second 12-lead ECG. Ablation outcomes may be influenced more by non-ECG parameters or require larger datasets or long-term ECG monitor data, and multi-modality inputs to be accurately predicted.