An Insulin Bolus Advisor for Type 1 Diabetes Using Deep Reinforcement Learning

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Type 1 diabetes (T1D) management requires precise exogenous insulin delivery to maintain blood glucose (BG) within a therapeutic range.
• Optimizing insulin doses is challenging due to complex glucose dynamics, variability, and uncertainty, leading to risks of hyperglycemia and hypoglycemia.

Purpose of the Study:

• To develop and evaluate a novel insulin bolus advisor utilizing deep reinforcement learning (DRL) for optimizing mealtime insulin delivery in individuals with T1D.
• To enhance blood glucose control and minimize glycemic excursions through intelligent insulin dosing.

Main Methods:

• A deep reinforcement learning (DRL) approach, specifically an actor-critic model based on deep deterministic policy gradient, was employed to compute mealtime insulin doses.
• A two-step learning framework involved initial population model training followed by subject-specific personalization.
• Prioritized memory replay was utilized to accelerate training.
• The algorithm was validated using the FDA-accepted UVA/Padova T1D simulator for in silico trials on adult and adolescent subjects.

Main Results:

• The DRL insulin bolus advisor significantly increased the average percentage of time in the target blood glucose range (70-180 mg/dL) compared to a standard bolus calculator.
• Improvements were observed in both adult (74.1% to 80.9%) and adolescent (54.9% to 61.6%) cohorts.
• The DRL system also demonstrated a reduction in hypoglycemic events.

Conclusions:

• The proposed DRL-based insulin bolus advisor shows significant potential for improving mealtime insulin delivery in people with T1D.
• The algorithm is a feasible candidate for future clinical validation and may enhance glycemic control and patient outcomes.