Model Overview
Google Research developed GlucoFM, a foundation model designed for continuous glucose monitoring. The model utilizes a dual-stream architecture to process CGM data, distinguishing between slower glycemic trends and short-term deviations. This approach aims to improve the accuracy of metabolic predictions, including diabetes risk assessment and insulin resistance estimation.
Training and Architecture
The model was pre-trained on 109,066 hours of unlabeled CGM data from Wear-CGM and four published datasets. GlucoFM’s encoder separates a lower-frequency state component representing slow trends from a residual event component capturing short-term deviations. The model employs latent-prediction objectives, including contextual prediction and temporal dynamics, to learn daily glucose patterns and their evolution over time. CGM-aware augmentations simulate real-world data variations, such as sensor artifacts and missing measurements.
Evaluation Results
Evaluations across four cohorts and seven clinical prediction tasks showed GlucoFM achieving a PR-AUC 5.8 percentage points higher than GluFormer, with the model leading in diabetes risk and beta-cell dysfunction evaluations. Specifically, the model demonstrated the lowest mean absolute error (MAE) when forecasting postprandial glycemic response, achieving the best cross-dataset transfer performance and strong few-shot adaptation capabilities.
Evaluation Methodology
To assess GlucoFM’s capabilities, researchers used subject-disjoint window-level linear probing. This involved freezing the model’s encoder, training a linear classifier on individual 24-hour representations, and ensuring no participant appeared in both the training and test folds. This method evaluated whether single-day representations are informative for unseen participants and retained day-to-day variability.
Source: https://research.google/blog/glucofm-foundation-model-for-continuous-glucose-monitoring/