AI model finds hidden risk signals in CGM data

A generative artificial intelligence (AI) model trained on more than 10 million continuous glucose monitoring (CGM) measurements may help clinicians extract far more prognostic value from glucose traces than traditional metrics such as HbA1c, according to a study published in Nature.

The model, called GluFormer, was developed using self-supervised learning on CGM data from 10,812 adults, most of whom did not have diabetes. Rather than being trained to predict a single predefined outcome, the model learned generalizable “representations” of glucose dynamics using autoregressive prediction — an approach borrowed from large language models and other foundation models in AI, reported Guy Lutsker, of Weizmann Institute of Science, Rehovot, Israel, and colleagues.

Researchers then tested whether these learned representations transferred to new populations. Across 19 independent external cohorts totaling 6,044 participants from five countries and using eight different CGM devices, GluFormer consistently outperformed baseline blood glucose, HbA1c, and conventional CGM-derived measures in forecasting glycemic parameters.

The clinical signal was particularly notable in individuals with prediabetes. GluFormer more accurately identified those likely to experience clinically meaningful rises in HbA1c over a 2-year period compared with baseline HbA1c alone and commonly used CGM metrics.

In a separate cohort of 580 adults who underwent short-term CGM and were followed for a median of 11 years, the model stratified long-term risk of incident diabetes and cardiovascular mortality more effectively than HbA1c. Two-thirds of new diabetes cases and 69% of cardiovascular deaths occurred in the highest-risk quartile identified by GluFormer, while rates were minimal in the lowest-risk group.

The investigators also evaluated the model in clinical trial settings, where baseline CGM-derived representations improved outcome prediction. In addition, a multimodal extension integrating dietary data was able to generate plausible glucose trajectories and estimate individualized glycemic responses to food.

Together, the findings suggest that CGM data – already widely used in diabetes care – may hold underexploited prognostic information when analyzed using foundation-model approaches. By learning complex temporal patterns rather than relying on summary statistics, GluFormer offers a framework that could support precision risk stratification in prediabetes, diabetes, and cardiometabolic disease.

Two of the study authors reported that they are employed by Pheno.AI, developer of GluFormer. Another reported being a paid consultant for Pheno.AI.

AACE Endocrine AI is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (AACE^®). The ideas and opinions expressed in AACE Endocrinology do not necessarily reflect those of Conexiant or AACE. For more information, see Policies.