Multi-Horizon Glucose Prediction Across Populations With Deep Domain Generalization

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Continuous glucose monitoring (CGM) is vital for diabetes self-management, especially for type 1 diabetes (T1D).
• Accurate glucose prediction is essential for maintaining therapeutic blood glucose ranges but is challenged by glycemic variability.
• Developing generalizable prediction algorithms across diverse patient populations remains a significant hurdle.

Purpose of the Study:

• To introduce GPFormer, a Transformer-based zero-shot learning method leveraging meta-learning for domain generalization in multi-horizon glucose prediction.
• To evaluate GPFormer's performance on diverse datasets reflecting real-world CGM usage scenarios.

Main Methods:

• Developed GPFormer, a Transformer-based meta-learning model for zero-shot glucose prediction.
• Trained and validated on the REPLACE-BG dataset (226 T1D participants).
• Evaluated on three external datasets (OhioT1DM, two proprietary) with diverse populations (T1D, T2D, non-diabetic, inpatient/outpatient).

Main Results:

• GPFormer demonstrated superior performance compared to baseline machine learning methods across all evaluated datasets.
• Achieved the lowest root mean square error (RMSE) for glucose prediction up to a two-hour horizon.
• Showcased consistent effectiveness and generalizability across varied populations and clinical settings.

Conclusions:

• GPFormer's meta-learning approach effectively addresses domain generalization challenges in glucose prediction.
• The model shows significant potential for enhancing glucose management in diverse, real-world clinical settings.
• Highlights the promise of advanced AI for improving diabetes care through accurate, real-time glucose forecasting.