Managing Blood Glucose Concentration in Type 1 Diabetes with Deep Learning-Based Methodologies

This thesis presents groundbreaking research on advanced models and automated systems designed to revolutionize Blood Glucose (BG) management for individuals with type 1 diabetes (T1D). Integrating cutting-edge deep learning algorithms, phys- iological modeling, and RL techniques, the primary objective is to improve diabetes care, optimize glucose control, and enhance the overall quality of life for T1D patients. The journey commences with the development of state-of-the-art BG predictive mod- els, proposing a powerful deep learning-based model to achieve remarkable accuracy in glucose predictions. Additionally, the impact of behavioral factors, such as physical activity and stress, on BG fluctuations, is explored, further enhancing model accuracy and generalizability. Furthermore, innovative BG testing platforms are introduced, including a data- driven model predictive control algorithm integrated with a BG predictor, showcasing superior glucose control compared to traditional linear control models, and deriving BG dynamic models in response to physical activity, providing invaluable insights for the design of automated closed-loop systems aimed at improved glucose control in daily life. The final frontier lies in RL-based automated insulin delivery systems, with two pivotal papers presenting closed-loop insulin administration frameworks dynamically adjusting insulin dosages based on real-time glucose readings and meal intakes, show- casing significant reductions in glucose variability and improvements in time spent within the target glucose range. In conclusion, this comprehensive thesis showcases innovative models and systems to elevate BG management in T1Ds, offering hope for a brighter future for those living with this condition as advancements continue to reshape diabetes care through personalized and automated approaches.