Safe and Personalized Meal Bolus Calculator for Type-1 Diabetes Using Bayesian Optimization.

Area of Science:

• Biomedical Engineering
• Endocrinology
• Artificial Intelligence in Medicine

Background:

• Postprandial glycemic regulation remains a significant challenge in type 1 diabetes management.
• Current meal bolus calculators often require patient-specific parameters or historical data, limiting their accessibility.
• Mealtime insulin dosing is critical for managing blood glucose spikes after eating.

Purpose of the Study:

• To develop a model-free, safe, and personalized bolus calculator algorithm for type 1 diabetes.
• To address the limitations of existing bolus calculators by eliminating the need for patient-specific parameters or historical data.
• To ensure high probability of satisfying the hypoglycemia constraint during insulin dosing.

Main Methods:

• Utilized safe contextual Bayesian optimization for a model-free bolus calculator.
• Developed an algorithm that learns optimal insulin doses using only continuous glucose monitoring (CGM) data.
• Conducted in silico experiments using the UVA/Padova T1DM simulator and Hovorka T1D model.

Main Results:

• The proposed algorithm quickly learned optimal bolus insulin doses for announced meals.
• Demonstrated effective glycemic regulation and patient safety in simulated type 1 diabetes scenarios.
• Successfully ensured satisfaction of the hypoglycemia constraint with high probability.

Conclusions:

• The model-free safe contextual Bayesian optimization algorithm offers a promising approach for personalized type 1 diabetes management.
• This novel algorithm enhances postprandial glycemic control by leveraging CGM data without prior patient-specific information.
• The approach prioritizes patient safety by minimizing the risk of hypoglycemia.