In-Silico Deep Reinforcement Learning for Effective Cardiac Ablation Strategy

We propose an in-silico deep reinforcement learning scheme that combines a computer simulation model of cardiac tissue and deep reinforcement learning to unveil an effective ablation strategy for atrial fibrillation (AF). The deep neural network-based ablation model (DAM) was designed to input a membrane potential movie and output an ablation pattern. As a virtual environment for in-silico learning, a numerical 2D cardiac tissue model was used. After training, the trained DAM was compared with two different strategies: the random ablation strategy (RND) and the rotor ablation strategy (ROT). We demonstrated the ablation area percentage of RND, ROT, and the best learned DAM was 7.0 ± 2.8%, 12.5 ± 5.7%, 6.5 ± 2.4%, and the AF termination rate was 12.6%, 8.5%, 74.1%, respectively. Results suggest that the DAM learned to effectively terminate spiral excitations by ablating the area from near the spiral center towards the tissue boundary without any prior knowledge. The learned DAM achieved a better AF termination rate compared to the other ablation strategies. This study showed feasibility of in-silico learning for an effective ablation strategy. To the best of our knowledge, this is the first approach to optimize spatial ablation patterns by deep learning. The proposed learning method has possibility to contribute to establishing an effective ablation strategy for more complex excitations for which conventional heuristic ablation is ineffective.