Explainable Deep Reinforcement Learning for Optimizing Insulin Dosage in Type-۱ Diabetes Patients

Type-۱ Diabetes (T۱D) management requires precise insulin dosing to maintain Blood Glucose (BG) levels within a safe range and prevent long-term complications. Traditional control strategies, such as Model Predictive Control (MPC), provide effective regulation but often lack adaptability to patient-specific dynamics and daily variability. Recently, Deep Reinforcement Learning (DRL) has emerged as a promising approach for personalized insulin therapy, offering the ability learn optimal policies through interaction with simulated or real environments. However, the inherent black- box nature of DRL limits its adoption in clinical practice, where transparency and interpretability are essential for trust and safety. This study proposes an Explainable Deep Reinforcement Learning (XDRL) framework for insulin dosage optimization in T۱D patients. The framework integrates DRL with explainability techniques, including feature attribution and policy visualization, to provide clinicians with interpretable decision insights. Simulation experiments conducted on the UVA/Padova T۱D simulator demonstrate that the proposed XDRL approach achieves superior glycemic control compared to baseline methods, while offering transparent reasoning behind insulin dosing decisions. These findings highlight the potential of XDRL to bridge the gap between advanced AI-driven control strategies and clinical applicability in diabetes management.

کلیدواژه ها:

Explainable Artificial Intelligence ، Deep Reinforcement Learning ، Type-۱ Diabetes ، Insulin Dosage Optimization ، Glycemic Control ، Clinical Decision Support