A Safety Aware Deep Reinforcement Learning Technique for Automated Insulin Delivery

Automated Insulin Delivery (AID) systems have shown great promise in managing diabetes by automating insulin administration. However, a significant challenge remains: preventing hypoglycemia or hyperglycemia during dynamic glucose fluctuations while minimizing the daily insulin dosage, or control effort. This study explores ways to enhance AID systems using a Reinforcement Learning (RL) algorithm called Maskable Proximal Policy Optimization (Maskable PPO), based on invalid action masking. Our findings demonstrate that this approach leads to a safety-aware framework for AIDs, providing highly realistic simulation scenarios for individual adult, adolescent, and child patients. The results show improved Time In Range (TIR) metrics (96.39%, 96.85%, and 54.43%), prevention of emergency bolus administration in adolescent and adult patients, and a reduction in Total Injected Insulin (TII) (17.32 U, 13.52 U, and 4.35 U per day).