GlucoFM: Foundation model for continuous glucose monitoring (research.google)

Google Research has introduced GlucoFM, a pioneering self-supervised foundation model designed for continuous glucose monitoring (CGM) that adeptly separates slow glucose trends from short-term fluctuations. This dual-stream model innovatively handles diverse metabolic prediction tasks, such as diabetes risk assessment and insulin resistance, while significantly enhancing prediction accuracy compared to existing models. GlucoFM achieved an average PR-AUC increase of 5.8 percentage points across multiple evaluation tasks and cohorts, underscoring its efficacy even in scenarios with limited labeled clinical data. What sets GlucoFM apart is its ability to process CGM data in a way that accounts for the unique multi-scale structure of glucose dynamics. By employing latent predictive pre-training and handling gaps and sensor artifacts effectively, GlucoFM provides robust representations that excel in cross-cohort transfer scenarios. The model's efficiency in few-shot settings indicates its potential for broad clinical applications, particularly as the need for high-quality, labeled data grows. Moving forward, researchers aim to scale the training to larger, more heterogeneous populations and enhance the model's capability to capture long-term glucose patterns—an advancement that could transform metabolic health monitoring and intervention strategies.