Basal Glucose Control in Type 1 Diabetes Using Deep Reinforcement Learning: An In Silico Validation

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Endocrinology

Background:

• Type 1 diabetes (T1D) management requires precise insulin delivery to maintain blood glucose within target ranges.
• Current artificial pancreas and continuous glucose monitoring systems face challenges due to complex glucose dynamics and technological limitations.

Purpose of the Study:

• To develop and evaluate a novel deep reinforcement learning (DRL) model for closed-loop glucose control in T1D.
• To investigate both single-hormone (insulin) and dual-hormone (insulin and glucagon) delivery strategies using DRL.

Main Methods:

• Developed DRL strategies using double Q-learning with dilated recurrent neural networks.
• Utilized the FDA-accepted UVA/Padova Type 1 diabetes simulator for in silico testing.
• Employed generalized training for a population model, followed by subject-specific personalization.

Main Results:

• Both single and dual-hormone DRL strategies significantly improved time in the target glucose range (70-180 mg/dL) compared to standard basal-bolus therapy.
• In adults, time in range increased from 77.6% to 80.9% (single-hormone) and 85.6% (dual-hormone).
• In adolescents, time in range improved from 55.5% to [Formula: see text] (single-hormone) and 78.8% (dual-hormone), with significant hypoglycemia reduction across all scenarios.

Conclusions:

• Deep reinforcement learning offers a viable and effective approach for advanced closed-loop glucose control in T1D.
• DRL-based dual-hormone delivery shows particular promise for optimizing glycemic control and minimizing hypoglycemia.
• Personalized DRL models can enhance the efficacy of artificial pancreas systems for individuals with T1D.