Google Research Unveils GlucoFM for Continuous Glucose Monitoring

Google Research and UNSW Sydney have introduced GlucoFM, a novel self-supervised foundation model designed for continuous glucose monitoring (CGM). This model distinguishes itself by employing a dual-stream architecture, a departure from prior CGM models such as CGMformer, GluFormer, and CGM-JEPA, which treated glucose traces as a single, entangled sequence. GlucoFM instead decomposes the glucose trace into two distinct streams: one representing the slow physiological "state" and another capturing transient "event" fluctuations. This approach maintains the observation mask intact and utilizes two Joint Embedding Predictive Architecture (JEPA)-style latent objectives for pretraining.

The resulting encoder, comprising 0.72 million trainable parameters, achieved a task-averaged Precision-Recall Area Under the Curve (PR-AUC) of 58.8 across 14 cohort-task evaluations. This performance surpasses the strongest CGM-specific baseline, which attained a PR-AUC of 54.7 when retrained on the same dataset. The pretraining process involved 109,066 hours of unlabeled CGM data from 477 subjects, executed on a single NVIDIA H100 GPU. While GlucoFM is presented as a research prototype and is not yet cleared or approved by any regulatory authority for clinical or consumer use, its underlying methodology is intended to be deployable. The research team explicitly states that the model is not intended to diagnose, treat, cure, or prevent any disease.

Each evaluation conducted was retrospective, and the largest pretraining cohort remains non-public. As of August 26, 2026, no model checkpoint has been released. However, the paper commits to making the code and reproducibility scripts publicly available, enabling other research teams to replicate the findings. The core innovation lies in the model's architecture and pretraining strategy, which can be implemented by any team possessing a CGM dataset. With 0.72 million trainable parameters and 120 training epochs on a single NVIDIA H100, the model can be reproduced. Furthermore, 24-hour window inference can be performed on a CPU container or directly on a device, indicating potential for efficient deployment.

The rationale behind GlucoFM's dual-stream approach addresses a fundamental limitation in existing CGM models. By treating CGM data as a single signal, previous models failed to adequately capture the dual nature of glucose readings: the slow, underlying regulatory baseline and the rapid, transient deviations caused by factors like meals, physical activity, stress, or sensor artifacts. The expense and cohort-specific nature of clinical labels also pose a significant barrier to effective supervised training in traditional models. GlucoFM's self-supervised, dual-stream architecture offers a more nuanced and potentially more accurate method for analyzing CGM data, paving the way for future advancements in diabetes management technology.