Can artificial intelligence prediction of successful atrial fibrillation catheter ablation therapy be interpretable?

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, artificial intelligence (AI), particularly deep learning (DL), has increasingly been applied to improve and optimise treatments for AF. However, AI is limited by its black-box nature. Therefore, a factor hindering the broad clinical application of DL in AF is the lack of assurance that the DL model is using physiological relevant features when making its prediction. In order to provide this assurance, its decision process needs to be interpretable and have biomedical relevance. AIM: This study aims to explore interpretability in DL prediction of successful ablation therapy for AF and evaluate if pro-arrhythmogenic regions in the left atrium (LA) were used in its decision process. METHODS: LA models with segmented fibrotic tissue were derived from 122 late gadolinium-enhanced magnetic resonance (LGE MR) images of persistent AF patients. To increase the dataset size, an additional 199 synthetic LA tissue models were generated from the LGE MR images. Two ablation strategies were simulated: fibrosis-based ablation (FIBRO) and a rotor-based ablation (ROTOR). RFCA strategy success was determined by its ability to terminate persistent AF in 2s with less than 40% of the tissue ablated [1]. A convolutional neural network (CNN) was developed to predict the success of each RFCA strategy, and gradient weighted class activation maps (GradCAM) were used to assess if the CNN was using locations of pro-arrhythmogenic regions in its decision process on an independent test set of 50 LA tissue models. RESULTS: For predicting the success of the FIBRO strategy, the CNN model had an AUC (area under the receiver operating characteristic curve) of 0.92±0.02, recall of 0.89±0.03 and precision of 0.82±0.02. For the ROTOR strategy, the AUC was 0.77±0.02, the recall was 0.93±0.04 and the precision was 0.76±0.02. Finally, the independent test set's GradCAM saliency maps showed that 62±25% and 71±13% of ablation lesions (known from the LA model simulations, but unseen by the CNN during training) coincided with informative regions in the saliency maps for the FIBRO and ROTOR strategies, respectively (Figure 1). CONCLUSION: The most informative regions of the saliency maps coincided with the successful ablation lesions, suggesting that the DL model was able to identify ablated pro-arrhythmogenic regions by leveraging structural features of LGE MR images. In the future, this technique could provide a clinician with a decision support tool and increase confidence in the AI prediction. FUNDING ACKNOWLEDGEMENT: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): UK Medical Research Council