Exploring interpretability in deep learning prediction of successful ablation therapy for atrial fibrillation

Background: Radiofrequency catheter ablation (RFCA) therapy is the first-line treatment for atrial fibrillation (AF), the most common type of cardiac arrhythmia globally. However, the procedure currently has low success rates in dealing with persistent AF, with a reoccurrence rate of ∼50% post-ablation. Therefore, deep learning (DL) has increasingly been applied to improve RFCA treatment for AF. However, for a clinician to trust the prediction of a DL model, its decision process needs to be interpretable and have biomedical relevance. Aim: This study explores interpretability in DL prediction of successful RFCA therapy for AF and evaluates if pro-arrhythmogenic regions in the left atrium (LA) were used in its decision process. Methods: AF and its termination by RFCA have been simulated in MRI-derived 2D LA tissue models with segmented fibrotic regions (n = 187). Three ablation strategies were applied for each LA model: pulmonary vein isolation (PVI), fibrosis-based ablation (FIBRO) and a rotor-based ablation (ROTOR). The DL model was trained to predict the success of each RFCA strategy for each LA model. Three feature attribution (FA) map methods were then used to investigate interpretability of the DL model: GradCAM, Occlusions and LIME. Results: The developed DL model had an AUC (area under the receiver operating characteristic curve) of 0.78 ± 0.04 for predicting the success of the PVI strategy, 0.92 ± 0.02 for FIBRO and 0.77 ± 0.02 for ROTOR. GradCAM had the highest percentage of informative regions in the FA maps (62% for FIBRO and 71% for ROTOR) that coincided with the successful RFCA lesions known from the 2D LA simulations, but unseen by the DL model. Moreover, GradCAM had the smallest coincidence of informative regions of the FA maps with non-arrhythmogenic regions (25% for FIBRO and 27% for ROTOR). Conclusion: The most informative regions of the FA maps coincided with pro-arrhythmogenic regions, suggesting that the DL model leveraged structural features of MRI images to identify such regions and make its prediction. In the future, this technique could provide a clinician with a trustworthy decision support tool.