Google Research publishes GlucoFM for continuous-glucose modelling

GlucoFM is a lightweight, self-supervised, dual-stream foundation model that separates slow glycaemic state from short-lived events. Google reports training on 109,066 hours from 477 participant/session records and evaluation across four cohorts and seven clinical tasks.

Why it matters: The design shows how domain structure and sparse labels can be handled without scaling to a general-purpose LLM.

Business/technical/strategic impact: The accompanying paper is an arXiv preprint and results are author-reported, not independently reproduced or clinically approved. PR-AUC and forecasting gains must not be treated as diagnostic performance in practice.