Google Research Unveils GlucoFM for Glucose Tracking

Google Research has developed GlucoFM, a lightweight, self-supervised foundation model designed to analyze continuous glucose monitoring (CGM) data. Unlike existing models like CGMformer, GluFormer, and CGM-JEPA that process glucose through a single stream, GlucoFM uses a dual-stream architecture. This design separates slow baseline glycemic trends from short-term deviations caused by meals, physical activity, or sensor anomalies. The model was pre-trained on 109,066 hours of unlabeled CGM data from the Wear-CGM dataset and four other published sources, representing 477 participant records.

In evaluations across four clinical cohorts—CGMacros, Stanford, Hall, and ShanghaiT2DM—GlucoFM was tested on seven metabolic prediction tasks, including diabetes risk and insulin resistance. The model achieved an average precision-recall area under the curve (PR-AUC) of 58.8, representing a 4.1-point absolute increase over the strongest CGM-specific baseline's score of 54.7. It also outperformed the best-performing GluFormer variant by an average of 5.8 percentage points. For postprandial glycemic response forecasting across Dexcom and Libre devices, GlucoFM achieved the lowest mean absolute error of 21.88 mg/dL, compared to 22.90 mg/dL for the top baseline.

The model demonstrated robust generalization in cross-dataset transfer tests, leading in 11 of 12 evaluations by margins of 0.5 to 8.6 PR-AUC points, with absolute scores reaching up to 90.0 percent. GlucoFM also excelled in few-shot learning scenarios, maintaining its performance edge even when limited to one labeled participant per class or just 1 percent of observations. Furthermore, aggregating data over seven days improved subject-level predictions, yielding gains of 9.6 points for Stanford beta-cell dysfunction and 14.0 points for Hall diabetes prediction.

For clinical practitioners and medical AI developers, GlucoFM addresses the persistent challenge of sparse and expensive labeled clinical data. By extracting rich, transferable representations from unlabeled wearable data, the model allows researchers to build highly accurate metabolic phenotyping tools with minimal supervision. Its ability to generalize across different cohorts and sensor brands means diagnostic tools can be deployed more reliably in diverse real-world patient populations.