Control of Blood Glucose for Type-1 Diabetes by Using Reinforcement Learning with Feedforward Algorithm

Area of Science:

• Biomedical Engineering
• Control Systems
• Artificial Intelligence in Medicine

Background:

• Type-1 diabetes is characterized by insulin deficiency, leading to elevated blood glucose.
• The complex, nonlinear nature of glucose kinetics poses significant control challenges.
• Difficulty in measuring key physiological variables complicates diabetes management.

Purpose of the Study:

• To develop an automated system for calculating insulin doses in type-1 diabetes.
• To maintain stable blood glucose levels by directly compensating for external factors like food intake.
• To utilize reinforcement learning and feedforward control for improved glucose regulation.

Main Methods:

• A reinforcement learning algorithm with a feedforward controller was designed for basal and bolus insulin dose calculation.
• Simulations on a glucose-kinetic model were used to assess controller performance.
• A Kalman filter was integrated to estimate unmeasurable state variables.

Main Results:

• The proposed controller demonstrated superior performance in regulating blood glucose fluctuations compared to traditional PID controllers.
• The system effectively improved glucose responses and prevented hypoglycemia.
• Simulations under uncertainty highlighted the impact of inaccurate meal information on control system performance.

Conclusions:

• The developed controller is effective in mitigating post-meal glucose spikes.
• It successfully counteracts external events like meals and maintains healthy glucose levels despite uncertainties.
• This approach offers a promising tool for enhanced type-1 diabetes management.