GluFormer: Learning Generalizable Representations from Continuous Glucose Monitoring Data

Guy Lutsker, Gal Sapir, Smadar Shilo, Jordi Merino, Anastasia Godneva, Jerry R Greenfield, Dorit Samocha-Bonet, Raja Dhir, Francisco Gude, Shie Mannor, Eli Meirom, Gal Chechik, Hagai Rossman, Eran Segal

Track: Full Paper Track

Keywords: Continuous Glucose Monitoring (CGM), Transformers, Self-Supervised Learning, Representation Learning, Autoregressive Modeling, Representation Learning, Generative Model, Metabolic Health, Diabetes, Multimodal, Foundation Model

TL;DR: GluFormer a self-supervised Transformer for continuous glucose monitoring data. Trained on > 10 million CGM measurements from 10K healthy adults, it learns embeddings that outperform standard CGM metrics in predicting short & long-term outcomes.

Abstract: Continuous glucose monitoring (CGM) enables near-continuous measurement of glucose trends, offering detailed insight into metabolic health. However, existing CGM-based metrics (e.g., time in range, glucose management indicator) only partially capture the complexities of glycemic variability. In this work, we present \textit{GluFormer}, a generative foundation model employing self-supervised representation learning on over 10 million CGM measurements from 10,812 participants without a known diabetes diagnosis. By predicting future tokens in an autoregressive fashion, GluFormer learns latent representations that generalize across 19 additional cohorts ($n=6{,}044$) with differing devices, ethnicities, and clinical contexts (from prediabetes and gestational diabetes to type 1/2 diabetes). GluFormer outperforms standard CGM metrics in forecasting clinical measures (e.g., A1c, visceral adipose tissue, and liver function) and in risk stratification for longer-term outcomes such as incidence of diabetes and cardiovascular mortality. Beyond single-number CGM summaries, the model generates realistic glucose curves that align with real-world data, and its performance further improves when including discrete dietary tokens in a multimodal framework. These findings suggest that large-scale self-supervised learning on continuous physiological signals can improve our ability to identify and manage metabolic risks, as well as simulate personalized glycemic trajectories.

Attendance: Guy Lutsker

Submission Number: 3