Reinforcement Learning Adjustment of Conventional Insulin Therapy for People with Type 1 Diabetes

During the last decades, the use of Automated Insulin Delivery systems (AIDs) for people with Type I Diabetes (TID) has shown great promise and improvement for glycemic control. However, most people with TID still rely on Multiple Daily Injections (MDI) therapy to regulate their glycemia, es-pecially in developing countries. In this work, a Reinforcement Learning (RL) agent is trained to provide suggestions to adapt the long-lasting insulin dose used in MDI therapies. The proposed RL agent aims to rapidly improve time-in-range while minimizing the risk of hypoglycemia, compared to standard clinical guidelines. The strategy is evaluated in silico using the UVAlPadova simulator and contrasted with Standard of Care (SoC) dosing recommendations. Results show that, even under carbohydrates (CHO) estimation errors, the RL agent is able to adapt the long-acting insulin dosage in less than 3 weeks, while the SoC approach only achieves comparable glucose outcomes after 4 weeks across all virtual patients. These findings suggest that RL-based decision support systems have great potential to empower MDI users to optimize their insulin therapy between clinical visits, leading to improved long-term glycemic control.