MetaboNet-Bench is a standardized benchmark for blood-glucose forecasting, built on the MetaboNet dataset and developed in collaboration with Stanford. Using MetaboNet's fixed train, validation, and test splits, it defines consistent forecasting tasks and evaluation metrics across modalities — CGM, insulin, carbohydrates, and activity — so that competing forecasting models can finally be compared on equal footing.
A key finding: adding insulin and carbohydrate context improves forecasts, and the benefit grows with model scale — evidence that richer, multi-modal context is central to the next generation of glucose prediction.
Read the paper: https://arxiv.org/abs/2606.18640
Evaluate your own model with the submission toolkit on GitHub, or explore the underlying dataset at metabo-net.org.