Deep reinforcement learning for Type 1 Diabetes: Dual PPO controller for personalized insulin management.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Effective blood glucose management in Type 1 Diabetes Mellitus (T1DM) is crucial for preventing long-term complications.
• Current insulin delivery methods often demand substantial patient engagement, hindering full automation.
• Reinforcement Learning (RL) presents a viable strategy for advancing automated insulin administration systems.

Purpose of the Study:

• To introduce and evaluate a Dual Proximal Policy Optimization (Dual PPO) controller for personalized insulin delivery within a hybrid closed-loop system.
• To optimize patient-specific insulin delivery parameters using a grid search on pre-trained models.
• To implement a safe-control mechanism to prevent hypoglycemia during automated insulin administration.

Main Methods:

• A Dual PPO controller was developed for personalized insulin delivery in a hybrid closed-loop system.
• Patient-specific insulin delivery bounds were optimized via grid search on pre-trained models.
• The system was validated using the UVA/Padova simulator on 10 in silico adult patients with randomized meal scenarios over five days.

Main Results:

• The Dual PPO controller significantly enhanced Time in Range (TIR), achieving 69.30% ±1.61, compared to a single PPO model (61.69% ±1.54).
• The system demonstrated effective reduction of severe hyperglycemia with a low incidence of severe hypoglycemia.
• The Dual PPO system minimized patient interaction, negating the need for carbohydrate estimation, unlike traditional Basal-Bolus (BBC) and PIDC controllers.

Conclusions:

• The Dual PPO controller represents a significant advancement in personalized insulin delivery for T1DM, leading to improved glycemic control.
• This approach reduces the daily burden on patients managing Type 1 Diabetes Mellitus.
• The study highlights the potential of Dual PPO in precision medicine for diabetes management, paving the way for future clinical applications.