New Benchmark for Multimodal Glucose Forecasting in Type 1 Diabetes

A new benchmark, MetaboNet-Bench, has been developed to standardize the evaluation of glucose forecasting algorithms for Type 1 Diabetes patients. This open-source framework addresses the current challenge of inconsistent model performance comparisons, which hinders innovation in the field. Unlike many existing algorithms that rely solely on Continuous Glucose Monitoring (CGM) data, MetaboNet-Bench emphasizes multimodal signals. It integrates data from glucose levels, insulin dosing, and carbohydrate intake, providing a more comprehensive view for forecasting. The utility of the benchmark was demonstrated by evaluating several published glucose forecasting models and a custom multimodal time-series model. The results indicate that the advantage of incorporating additional data modalities depends on the model's complexity, and using more clinical metrics helps identify areas for future research and improvement.