Reinforcement and federated learning for privacy-preserving glucose regulation: A review

Type 1 diabetes mellitus (T1DM) requires continuous monitoring of blood glucose levels and careful insulin dosing to maintain safe glycemic control. Reinforcement learning (RL) has emerged as a promising approach for automated glucose regulation in T1DM by enabling adaptive insulin dosing through sequential decision-making. However, real-world deployment remains constrained by safety requirements, inter-patient variability, limited clinical data, and strict privacy regulations that hinder centralized model development. This review aims to synthesize recent advances in RL-based automated insulin delivery (AID) and examine how federated learning (FL) can enable privacy-preserving personalization across distributed clinical and device data. We organize the literature into key themes, including: (i) online RL controllers developed primarily in simulation environments, (ii) offline RL methods that learn dosing policies from logged glucose-insulin trajectories (e.g., batch-constrained and conservative value-learning approaches), (iii) safety-aware strategies aimed at minimizing hypoglycemia risk and improving robustness under practical disturbances, and (iv) hybrid architectures integrating glucose prediction and control. We further discuss FL designs for personalization in heterogeneous, non-identically distributed patient data and highlight their relevance to diabetes management. Across the studies reviewed, most reinforcement learning approaches for glucose control are evaluated in simulation environments, while only a limited number use real-world clinical or continuous glucose monitoring (CGM) datasets. Federated learning studies are emerging for privacy-preserving personalization, and only a small subset of works currently explore hybrid RL–FL or federated reinforcement learning frameworks. Finally, we identify open challenges related to sim-to-real generalization, clinical validation, interpretability, scalability, and privacy risks in distributed learning, and outline future directions for robust, clinically viable next-generation diabetes management systems.