source: Google Research: GlucoFM: Foundation model for continuous glucose monitoring
level: research
google research built glucofm, a lightweight self-supervised foundation model for continuous glucose monitoring data. it uses a dual-stream design that separates slower glycemic trends from short-term deviations, while preserving time-of-day and missingness. the model was pre-trained on 109,066 hours of unlabeled cgm data from five sources. glucofm learns reusable representations through latent predictive tasks rather than reconstructing raw glucose readings, which can be noisy.
across 14 cohort-task evaluations, glucofm improved average pr-auc by 4.1 points over the strongest cgm-specific baseline, from 54.7 to 58.8. it led all diabetes-risk and beta-cell-dysfunction tasks and three of four insulin-resistance tasks. for postprandial glucose forecasting, glucofm achieved the lowest mean absolute error of 21.88 mg/dl versus 22.90 for the best baseline. in few-shot settings, it outperformed competitors even with one labeled participant per class or 1% of observations.
the dual-stream design outperformed single-stream variants, showing that separating slow and fast dynamics helps. multi-day averaging improved predictions, with gains up to 14 pr-auc points. cross-dataset transfer was strong, leading in 11 of 12 evaluations. limitations include a modest pre-training population and independent 24-hour window processing. future work aims to train on larger, more diverse cohorts and model multi-day trends natively.
why it matters: glucofm shows that separating slow and fast glucose signals improves prediction accuracy, which could help clinicians assess diabetes risk and insulin resistance with limited labeled data.
source: Google Research: GlucoFM: Foundation model for continuous glucose monitoring