Bolus Insulin calculation without meal information. A reinforcement learning approach

Area of Science:

• Artificial Intelligence in Medicine
• Diabetes Technology
• Computational Endocrinology

Background:

• Current insulin bolus calculations for type 1 diabetes rely on patient-specific parameters like carbohydrate-to-insulin ratio (CR), correction factor (CF), and carbohydrate (CHO) estimation.
• Inaccurate CHO estimation leads to errors in insulin dosing, increasing the management burden for patients.

Purpose of the Study:

• To develop and evaluate a Q-learning-based reinforcement learning (RL) algorithm for optimizing bolus insulin doses.
• To eliminate the need for CR, CF, and CHO content in insulin bolus calculations, thereby reducing estimation errors and patient burden.

Main Methods:

• A Q-learning-based reinforcement learning algorithm was developed.
• In-silico trials were conducted on a virtual cohort of 68 type 1 diabetes patients.
• Performance was compared against the standard bolus calculator (SBC) with and without CHO misestimation under open-loop basal insulin therapy.

Main Results:

• The RL algorithm achieved 73.4% time in the target glucose range (70-180 mg/dL), compared to 72.37% for SBC (without CHO misestimation).
• RL demonstrated superior performance over SBC when CHO content was misestimated.
• RL's performance was comparable to SBC under ideal conditions, despite not requiring CHO content information.

Conclusions:

• Reinforcement learning offers a promising approach to optimize insulin bolus dosing in type 1 diabetes, independent of carbohydrate counting.
• This algorithm can potentially alleviate patient management burden and improve glycemic control.
• The developed RL algorithm is suitable for integration into future artificial pancreas and automated insulin delivery systems.