Deep Reinforcement Learning and Edge Computing for Type-1 Diabetes Management

Type 1 Diabetes (T1D) is an autoimmune disorder characterized by the destruction of pancreatic β-cells, leading to an absolute deficiency of insulin production and a consequent inability to regulate blood glucose levels. Globally, T1D affects a smaller proportion of individuals compared to Type 2 Diabetes (T2D), yet both types significantly contribute to the global health burden. T1D requires lifelong insulin therapy, while T2D which accounts for over 90% of diabetes cases is primarily driven by lifestyle factors such as obesity and physical inactivity. Together, diabetes contributes to a rise in associated complications, including cardiovascular disease and kidney failure, with the International Diabetes Federation estimating that approximately 1 in 10 adults worldwide live with some form of diabetes, highlighting the urgent need for effective management and prevention strategies. To mitigate acute complications such as hypoglycemia and hyperglycemia, as well as prevent long-term organ damage, modern therapeutic approaches often include continuous glucose monitoring (CGM) sensors and insulin pumps to provide real-time insights and precision in insulin delivery. The rapid proliferation of machine learning (ML) and artificial intelligence (AI) techniques, particularly deep learning and deep reinforcement learning, has opened new avenues for personalized and proactive diabetes management. By leveraging these techniques, researchers aim to develop data-driven models capable of predicting shortterm and long-term fluctuations in blood glucose levels, thereby enabling early intervention and tailored insulin dosing. This capability is crucial because accurate blood glucose level (BGL) forecasting not only allows for improved glycemic control but also reduces the risk of severe complications through preemptive actions, thus transforming the clinical paradigm from a reactive to a preventive approach. Nonetheless, deploying AI solutions in T1D care is not without challenges: the need for high predictive accuracy, real-time computational efficiency, and robust data privacy protections underlines the complexity of constructing scalable and trustworthy systems. Consequently, integrating AI-driven forecasting and control mechanisms into the existing framework of CGM sensors and insulin pumps holds significant promise in improving patient outcomes, but it necessitates concerted efforts in algorithm development, hardware 2 optimization, and ethical data governance. This thesis integrates advanced predictive modeling and control frameworks to address the challenges of managing Type 1 Diabetes, combining state-of-the-art algorithms with personalized solutions. The contributions begin with a layered meta-learning approach that introduces a multi-expert architecture for predicting adverse glycemic events. By leveraging continuous glucose monitoring data and specializing in hypoglycemia, hyperglycemia, and normoglycemia detection, this method not only anticipates critical events with high precision but also enhances generalization through a meta-learner trained on limited patientspecific data. Complementing this, a Federated Online Extreme Learning Machine framework demonstrates the efficacy of decentralized learning in blood glucose level forecasting. This system achieves computational efficiency while maintaining robust data privacy, offering a scalable and secure solution for personalized diabetes management across distributed devices. Then this thesis advances the control of blood glucose levels through a novel dual Deep Reinforcement Learning (DRL) framework. This approach introduces a hybrid closed-loop control system for optimizing insulin delivery in real-time, achieving superior glycemic stability with minimal patient intervention. A safe-control mechanism and adaptive insulin caps ensure the mitigation of hypo- and hyperglycemic risks. Moreover, as a natural extension, the thesis investigates the application of Multi-Agent Reinforcement Learning (MARL), emphasizing inter-agent communication and collaboration to optimize decision-making in complex scenarios. This framework enables agents to dynamically share information and coordinate strategies, effectively capturing cooperative dynamics while enhancing the system’s capability to deliver highly personalized and adaptive medical interventions tailored to the unique physiological and pathological profiles of individual patients.