Lessening the shock of defibrillation with machine learning

Patients who suffer from life-threatening cardiac arrhythmias on a regular basis often receive implanted cardioverter-defibrillators, which stop these arrhythmias. While these can improve the life expectancy of patients, the high-energy, painful shocks applied by the devices cause tissue damage as well as psychological side effects, including post-traumatic stress disorder, depressive symptoms, and anxiety disorders.

In the hope of reducing these side effects, Frühwald et al. developed a machine learning agent to control the chaotic electrical excitation patterns underlying arrhythmias using lower amplitude pulses. The authors trained the agent with reinforcement learning (RL), in which desired behaviors are rewarded and the agent aims to maximize the amount of reward over time. To tailor their agent to solve their particular problem, they presented it with two-dimensional numerical simulations of the chaotic electrical excitation patterns that underlie cardiac arrhythmias.

The RL agent produced a protocol that delivers pulses with only 20 percent of the amplitude required by a single conventional pulse method and 66 percent of the amplitude used by state-of-the-art low-amplitude methods. The flexibility of the RL agent compared to previous methods has the potential to contribute to the development of patient-specific defibrillation strategies.

“This work represents a promising first step towards RL-driven defibrillation,” said author Daniel Frühwald. “We believe that RL poses a very promising alternative in the field of medical AI, in which wide-reaching decisions with imperfect knowledge have to be made. However, a lot of work has to be done before this proof-of-concept can be transferred to experimental setups.”

Next, the authors plan to apply RL methods to more realistic numerical simulations of arrhythmia, such as ventricular fibrillation, in three-dimensional heart models.

Source: “Deep reinforcement learning to control cardiac arrhythmias via pulse sequences,” by Daniel Frühwald, Christopher Odefey, and Thomas Lilienkamp, Chaos (2026). The article can be accessed at https://doi.org/10.1063/5.0346786 .