GlucoFM

GlucoFM is a self-supervised foundation model for continuous glucose monitoring (CGM) data, introduced by researchers at Google Research and the University of New South Wales in May 2026. CGM sensors produce dense, multi-day glucose time series that are increasingly central to diabetes management and metabolic health research, yet most analyses still reduce these rich signals to a handful of hand-crafted summary statistics (time-in-range, mean glucose, glucose variability). GlucoFM instead learns general-purpose representations directly from raw CGM traces, with the goal of transferring across cohorts, devices, and clinical tasks without task-specific labels.

The model's central design choice is a dual-stream decomposition of glucose dynamics. Rather than treating a CGM recording as a single undifferentiated sequence, GlucoFM separates the signal into a slow-moving physiological state stream (an individual's baseline glycemic regime) and a transient event stream (the short-term excursions driven by meals, activity, and other perturbations). This separation reflects a physiological intuition: long-term metabolic status and acute glucose responses carry distinct information, and modeling them jointly but explicitly can produce representations that generalize better than a monolithic sequence model.

GlucoFM sits alongside a small but growing family of CGM foundation models—including GluFormer and predictive self-supervised approaches such as CGM-JEPA—that aim to bring the pretrain-then-transfer paradigm to wearable biosignals. Its distinguishing contribution is the explicit state/event factorization paired with complementary self-supervised objectives.

#Key Features

• Dual-stream decomposition: Glucose dynamics are factored into a slow physiological state stream and a transient event stream, so baseline glycemic regime and acute excursions are represented separately rather than entangled in one sequence.
• Self-supervised pretraining: The model is trained on unlabeled CGM recordings using complementary objectives covering temporal dynamics and cross-stream alignment, removing the need for large labeled clinical datasets.
• Cross-cohort transfer: Learned representations are evaluated for transfer across subjects and cohorts, targeting the generalization gap that limits CGM models trained on a single population or device.
• Wearable-native modeling: GlucoFM is designed specifically for the dense, irregular, multi-day structure of CGM signals, treating glucose as a continuous physiological biosignal rather than a tabular feature set.

#Technical Details

GlucoFM is a transformer-based dual-stream architecture pretrained on 109,066 hours of unlabeled CGM recordings drawn from 477 subjects. Pretraining uses two complementary self-supervised objectives: one modeling the temporal dynamics of the glucose signal and one aligning the slow physiological-state and transient-event streams. Downstream evaluation spans glucose forecasting at multiple horizons, glycemic-control assessment, and clinical risk stratification, with cross-cohort and cross-device generalization examined using public benchmarks including the PhysioNet BIG IDEAs Lab Glycemic Variability and Wearable Device dataset. As an arXiv preprint (submitted 29 May 2026), these results have not yet undergone peer review. The authors state that they will release code and reproducibility scripts, but at the time of writing no public code repository or pretrained weights are available; potential users should treat the model as not-yet-released.

#Applications

GlucoFM is aimed at researchers and clinicians working with CGM data in diabetes and metabolic-health settings. Pretrained CGM representations can serve as a shared backbone for downstream tasks such as forecasting future glucose, characterizing glycemic control, and stratifying individuals by metabolic risk, potentially reducing the labeled data needed to build each new predictor. Because the model is designed to transfer across cohorts and devices, it is particularly relevant for studies that pool heterogeneous CGM sources or seek to deploy a single model across diverse populations and sensor hardware.

#Impact

GlucoFM contributes to an emerging shift from hand-crafted CGM summary metrics toward learned, transferable representations of glucose biosignals, extending the foundation-model paradigm into wearable metabolic monitoring. Its dual-stream state/event decomposition offers a physiologically motivated alternative to single-sequence modeling that may inform future biosignal foundation models. As an unrefereed preprint without a current public code or weights release, its real-world adoption and independent validation remain to be established, and the relatively modest pretraining cohort (477 subjects) leaves open questions about scaling to larger and more diverse populations.