Google's GlucoFM Lifts CGM Prediction Accuracy by 5.8 Points Across 14 Evaluations
Updated
Updated · Google Research · Aug 26
Google's GlucoFM Lifts CGM Prediction Accuracy by 5.8 Points Across 14 Evaluations
1 articles · Updated · Google Research · Aug 26
Summary
GlucoFM beat the best GluFormer variant by an average 5.8 PR-AUC points across 14 cohort-task evaluations, improving diabetes-risk, insulin-resistance and beta-cell-dysfunction prediction from continuous glucose monitor data.
109,066 hours of unlabeled CGM data powered the model’s self-supervised pretraining, with a dual-stream design that separates slower glucose trends from short-term deviations instead of treating readings as one signal.
21.88 mg/dL was GlucoFM’s mean MAE in two-hour post-meal glucose forecasting, better than 22.90 for the best baseline under matched Dexcom and Libre tests using 874 meal events from 34 participants.
11 of 12 cross-dataset transfer tests favored GlucoFM by 0.5 to 8.6 PR-AUC points, and its edge persisted in few-shot settings with as little as one labeled subject per class or 1% of observations.
Google said the current training population is still modest and plans to expand to larger, more diverse cohorts and native multi-day modeling beyond separate 24-hour windows.
Will Google's new self-supervised glucose model finally solve the problem of inaccurate real-world sensor data caused by sweat, temperature, and drift?
How does a new AI model unlock hidden health secrets from messy continuous glucose monitoring data without needing expensive clinical labels?