Google Research Unveils GlucoFM, a Tiny Model Predicting Diabetes Risk From CGM Data

Google Research Unveils GlucoFM, a Tiny Model Predicting Diabetes Risk From CGM Data
Illustration

Google Research detailed GlucoFM, a 0.72-million-parameter model that reads continuous glucose monitor traces to predict diabetes risk, insulin resistance and beta-cell dysfunction. Pretrained on 109,066 hours of unlabelled data from 477 people, it beat far larger baselines, scoring 58.8 against 54.7 task-averaged PR-AUC across 14 evaluations. It is a research prototype: no weights have shipped and no regulator has cleared it.

2 sources

  1. 01arXiv — GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoringarxiv.org
  2. 02MarkTechPost — Google Research Introduces GlucoFM: A 0.72M-Parameter Dual-Stream Foundation Model for Continuous Glucose Monitoringmarktechpost.com