Deep reinforcement learning to control cardiac arrhythmias via pulse sequences

Life-threatening arrhythmias are the leading cause of sudden cardiac death. The standard clinical treatment involves delivering a high-energy defibrillation shock, which comes along with severe side effects like potential tissue damage or adverse psychological outcomes for patients with implantable cardioverter-defibrillators. We examine whether Reinforcement Learning (RL) can be used to derive low-energy pulse sequences to control the chaotic myocardial excitation dynamics during cardiac arrhythmias. We demonstrate that policies can be trained using numerical simulations and that the resulting multi-pulse protocols offer amplitude reductions of up to 34% compared to state-of-the-art multi-pulse methods. We further show that RL-trained policies increasingly focus on phase singularity dynamics throughout the pulse sequence and that established system observables can be used to explain RL-induced pulse timings. These observations indicate that RL-trained policies are more flexible than previously published protocols and may open up the way toward patient-specific defibrillation strategies.