Optimizing hypoglycaemia prediction in type 1 diabetes with Ensemble Machine Learning modeling.

Area of Science:

• Endocrinology and Metabolism
• Biomedical Data Science
• Machine Learning in Healthcare

Background:

• Type 1 diabetes (T1D) affects millions globally, necessitating intensive management including insulin therapy and continuous glucose monitoring (CGM).
• Accurate short-term prediction of subcutaneous glucose levels is crucial for preventing dangerous glucose fluctuations and minimizing long-term complications in T1D patients.

Purpose of the Study:

• To develop an ensemble univariate predictive model for subcutaneous glucose concentration in T1D.
• To specifically improve prediction accuracy in the hypoglycemic glucose range.
• To reduce erroneous predictions (EP) evaluated by continuous glucose error grid analysis (CG-EGA).

Main Methods:

• An ensemble model was constructed using basis functions selected for minimizing EP in the hypoglycemic region.
• XGBoost and Support Vector Regression (SVR) were identified as top-performing basis models for the ensemble.
• The dataset included 29 individuals with T1D, monitored for 2-4 weeks in the GlucoseML study.

Main Results:

• The ensemble model demonstrated a significant reduction in EP within the hypoglycemic region, achieving 19% error at a 30-minute prediction horizon.
• This represents a substantial improvement compared to individual basis models like XGBoost and SVR.
• Overall prediction errors across hypoglycemic, euglycemic, and hyperglycemic ranges remained comparable to the best individual basis models.

Conclusions:

• Incorporating basis function performance in the hypoglycemic region during ensemble construction enhances joint performance in this critical area.
• The developed ensemble model offers potential for more precise insulin management in T1D.
• This approach may lead to a reduced risk of short-term hypoglycemic events.