Core Insights - The article discusses the development of a generative foundation model for continuous glucose monitoring (CGM) data called GluFormer, which has significant predictive capabilities for both short-term glucose parameters and long-term disease risk stratification, particularly for diabetes and cardiovascular mortality [4][6]. Group 1: Model Development - The GluFormer model was trained using over 10 million glucose measurements from 10,812 adults, primarily non-diabetic, and employs self-supervised learning [5]. - The model's representations can be transferred across 19 external cohorts, covering five countries and various CGM devices, demonstrating continuous improvement in predicting glucose parameters compared to baseline glucose and HbA1c levels [5]. Group 2: Risk Stratification - In individuals with prediabetes, GluFormer effectively stratified risk for those likely to experience clinically significant HbA1c increases within two years, outperforming baseline HbA1c and common CGM metrics [6]. - In a cohort of 580 adults with a median follow-up of 11 years, GluFormer identified 66% of new diabetes cases and 69% of cardiovascular mortality cases in the highest risk quartile, compared to only 7% and 0% in the lowest risk quartile [6]. Group 3: Multimodal Integration - The research team also developed a multimodal extension of GluFormer that integrates dietary data, allowing for the generation of reasonable glucose trajectories and predictions of individual glucose responses to food [7]. - Overall, GluFormer provides a scalable framework for encoding glucose patterns, enhancing both short-term glucose predictions and long-term disease risk stratification, thus offering a powerful tool for precision medicine and metabolic health management [7].