GlucoFM: Foundation model for continuous glucose monitoring
Google says GlucoFM beat prior CGM models across metabolic prediction, transfer, and low-label tests.
The model separates slower glucose baselines from short-term deviations in continuous monitor data. Google evaluated it on four cohorts, seven clinical prediction tasks, and post-meal glucose forecasting. It reported higher average PR-AUC than the strongest same-data CGM baseline, plus the lowest mean error in the postprandial task. The note also says the current pre-training population is modest, with larger and more diverse training planned. Google Research's note
The model separates slower glucose baselines from short-term deviations in continuous monitor data. Google evaluated it on four cohorts, seven clinical prediction tasks, and post-meal glucose forecasting. It reported higher average PR-AUC than the strongest same-data CGM baseline, plus the lowest mean error in the postprandial task. The note also says the current pre-training population is modest, with larger and more diverse training planned. Google Research's note
score 5