A foundation model for continuous glucose monitoring data

Area of Science:

• Artificial Intelligence in Healthcare
• Metabolic Health Research
• Biomedical Data Science

Background:

• Continuous glucose monitoring (CGM) offers rich temporal glucose data but its predictive potential for long-term health is underutilized.
• Existing methods for analyzing CGM data often lack generalizability across diverse populations and devices.
• Predicting future glycemic trajectories and associated health risks from CGM data remains a significant challenge.

Purpose of the Study:

• To develop and validate GluFormer, a generative foundation model for CGM data, to unlock its full potential for glucose homeostasis and long-term outcome prediction.
• To assess the generalizability and predictive performance of GluFormer's learned representations across diverse cohorts, CGM devices, and pathophysiological states.
• To evaluate GluFormer's ability to stratify risk and predict clinical outcomes, including HbA1c changes, incident diabetes, and cardiovascular mortality.

Main Methods:

• Trained GluFormer using self-supervised learning on over 10 million CGM measurements from a large cohort of adults.
• Validated the model's representations across 19 external cohorts, encompassing diverse demographics, countries, CGM devices, and health conditions (prediabetes, diabetes types, obesity).
• Employed autoregressive prediction and multimodal extensions integrating dietary data for enhanced analysis.

Main Results:

• GluFormer's representations demonstrated consistent improvements over baseline metrics for forecasting glycemic parameters across diverse cohorts.
• The model effectively stratified individuals with prediabetes based on future HbA1c increase risk, outperforming traditional measures.
• In a long-term follow-up cohort, GluFormer identified individuals at elevated risk for diabetes and cardiovascular mortality more accurately than HbA1c.

Conclusions:

• GluFormer provides a generalizable framework for encoding complex glycaemic patterns from CGM data.
• The model shows significant potential for improving risk prediction and informing precision medicine strategies in metabolic health.
• Integrating multimodal data, such as diet, further enhances GluFormer's ability to predict individual glycemic responses.