Evaluating Deep Q-Learning Algorithms for Controlling Blood Glucose in In Silico Type 1 Diabetes.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Endocrinology

Background:

• Type 1 diabetes requires constant insulin management to maintain blood glucose levels.
• Reinforcement learning (RL) presents a potential solution for automated glucose control.
• Deep Q-learning (DQL) algorithms are being explored for personalized diabetes care.

Purpose of the Study:

• To evaluate DQL algorithms for automated insulin dosing in type 1 diabetes.
• To assess the personalization capabilities of model-free RL approaches.
• To compare DQL performance against standard treatments in a simulated environment.

Main Methods:

• Simulated type 1 diabetes patient using the Hovorka model.
• Incorporation of meal variations and carbohydrate counting errors in simulations.
• Testing and comparison of several DQL algorithm extensions.

Main Results:

• DQL algorithms demonstrated effective blood glucose regulation.
• Some DQL extensions outperformed standard baseline treatments.
• Model-free approaches showed feasibility without prior patient-specific data.

Conclusions:

• Deep Q-learning offers a promising avenue for automated and personalized insulin therapy.
• RL-based systems can potentially improve blood glucose control in type 1 diabetes.
• Further research into DQL for diabetes management is warranted.