Universal glucose models for predicting subcutaneous glucose concentration in humans.

Area of Science:

• Biomedical Engineering
• Data Science in Healthcare
• Endocrinology

Background:

• Accurate glucose prediction is crucial for proactive diabetes management.
• Current predictive models often lack universality across different individuals and devices.
• Continuous glucose monitoring (CGM) devices generate valuable data for model development.

Purpose of the Study:

• To test the hypothesis of a universal, data-driven glucose model applicable across diverse diabetic subjects.
• To evaluate the model's performance using data from different CGM devices and diabetes types.
• To establish the feasibility of universal glucose models for clinical application.

Main Methods:

• Development of data-driven autoregressive models (order 30) using filtered subcutaneous glucose concentration data.
• Short-term (30-min ahead) glucose concentration predictions.
• Evaluation using root-mean-squared error (RMSE) and Clarke error grid analysis (EGA), comparing same-subject, cross-subject, and cross-study predictions.

Main Results:

• Cross-subject and cross-study prediction errors (RMSE) were small and comparable to same-subject predictions.
• Over 99.0% of predicted glucose concentrations fell within the clinically acceptable Zone A of the Clarke error grid.
• Model predictive performance was unaffected by diabetes type, subject age, CGM device, or interindividual variability.

Conclusions:

• A universal glucose model is feasible for predicting subcutaneous glucose concentrations.
• These universal models support the clinical integration of predictive algorithms with CGM devices.
• The findings pave the way for proactive therapy in diabetic patients through universal glucose modeling.