Application of Deep Reinforcement Learning to Event-Triggered Control for Networked Artificial Pancreas Systems

Researchers Junya Ikemoto, Satoshi Maruyama, and Kazumune Hashimoto from the Systems and Control group on arXiv have proposed a novel deep reinforcement learning (DRL)-based event-triggered controller for networked artificial pancreas (AP) systems. Traditional DRL-based AP controllers rely on periodic control updates, which is inefficient for networked control systems (NCSs) that require reduced communication frequency for energy-efficient operation. The team tackled the complexity of jointly learning insulin dosing and update timing by introducing a rule-based criterion based on blood glucose changes, avoiding explicit learning of update timing. This allows decision-making at irregular intervals, naturally formulated as a semi-Markov decision process (SMDP), extending standard DRL algorithms.

Numerical experiments demonstrate that the proposed method significantly improves communication efficiency while maintaining control performance, a critical trade-off in medical devices. The work addresses a key challenge in NCSs: balancing energy savings with reliable control. By reducing communication frequency, the approach could extend battery life in wearable AP systems, making them more practical for long-term use. The paper is available on arXiv under ID 2604.26126 and is pending DOI registration via DataCite. This research has implications for other networked medical devices, such as pacemakers or insulin pumps, where energy efficiency is paramount.