GlucoAgent: Personalized Glucose Forecasting from Multimodal Personal Context
| Source: arXiv AI
Tags: GlucoAgent, glucose forecasting, CGMacros, Chronos-2, continuous glucose monitoring
GlucoAgent, an LLM agent that writes and runs code over a person's CGM, wearable and meal data, beat Chronos-2 on glucose forecasting with an average RMSE of 22.5 vs 23.5 mg/dL on CGMacros.
Details
Glucose monitors, smartwatches, meal logs and health profiles hold useful signals, but their relevance varies widely between people and over time. GlucoAgent handles this case by case at inference: for each forecast it writes and runs code over the individual's data and returns a prediction with reasoning and evidence. Generated programs range from retrieving similar past meals to fitting physiological models. On the CGMacros dataset it beat time series foundation models at every horizon, with average RMSE of 22.5 mg/dL versus 23.5 for Chronos-2. Multimodal data cut its RMSE by 1.4 mg/dL versus 0.3 for Chronos-2, and by 4.6 mg/dL in 60-minute post-meal forecasts.