Google Research has introduced GlucoFM, a self-supervised foundation model for continuous glucose monitor data. The model processes slow-moving glucose baselines separately from short-term deviations that may reflect meals, exercise or sensor artifacts, rather than forcing both patterns through one representation stream.
Continuous glucose monitors measure interstitial glucose every few minutes, producing detailed traces but relatively few expensive clinical labels. Google says GlucoFM learns from unlabeled daily data and then transfers those representations to tasks including diabetes-risk assessment, insulin resistance, beta-cell dysfunction and post-meal glucose response. The dual-stream design is intended to preserve both a person’s broader metabolic pattern and brief events.
The research reports stronger performance across several prediction tasks, but it is not a consumer diagnosis tool and does not replace clinical testing. Results from a research benchmark may also vary across devices and populations. The immediate value is a reusable model that researchers can adapt with smaller labeled datasets, potentially making continuous-monitor data more useful in metabolic studies.