Learning control-ready forecasters for Blood Glucose Management

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Type 1 diabetes (T1D) management requires diligent blood glucose (BG) monitoring and insulin dosing.
• Current machine learning (ML) models for automated BG control struggle with accuracy due to entangled insulin and carbohydrate data.
• Basal bolus (BB) insulin strategies often lead to inaccurate ML model training, hindering effective BG management.

Purpose of the Study:

• To develop a novel algorithm for training BG forecasters that disentangles insulin and carbohydrate impacts.
• To improve the accuracy and reliability of ML-based BG management systems for T1D patients.
• To enhance automated insulin delivery systems through more precise BG prediction.

Main Methods:

• Developed a new algorithm to disentangle insulin and carbohydrate effects on BG levels.
• Utilized correction bolus values and the monotonic effect of insulin on BG for training.
• Evaluated the algorithm using an FDA-approved simulator on 10 individuals over 30 days.
• Assessed performance using proxy metrics on three real-world datasets.

Main Results:

• The novel algorithm achieved an average time in range of 81.1% in simulations, significantly outperforming standard approaches (53.6%).
• The method demonstrated reliable maintenance of healthy BG levels in simulated T1D individuals.
• Proxy metrics indicated potential for improved control in real-world T1D datasets.

Conclusions:

• The proposed algorithm effectively disentangles insulin and carbohydrate effects, leading to superior BG control in simulated T1D.
• This advancement holds promise for developing more accurate and effective ML-based automated insulin delivery systems.
• The findings pave the way for significant progress in ML-driven BG management for type 1 diabetes.