Evaluating Time-Series Foundation Models and Multimodal Dietary Context for CGM Forecasting
Zero-shot forecasting models were not reliable winners on CGM data.
The paper tests eight public CGM datasets across Type 1 diabetes, Type 2 diabetes, and non-diabetes groups. Strong task-specific baselines, including Elastic Net and PatchTST, often matched or beat zero-shot foundation models. Lightweight fine-tuning changed that, with Chronos-Bolt cutting RMSE by 6.5% to 18.4% in the T1D cohort and 8.6% to 18.2% in the non-diabetes/T2D cohort. Adding dietary context from CGMacros improved errors further, especially after meals, where postprandial RMSE fell by about 15%. Source: ArXiv · AI/CL/LG's note.
The paper tests eight public CGM datasets across Type 1 diabetes, Type 2 diabetes, and non-diabetes groups. Strong task-specific baselines, including Elastic Net and PatchTST, often matched or beat zero-shot foundation models. Lightweight fine-tuning changed that, with Chronos-Bolt cutting RMSE by 6.5% to 18.4% in the T1D cohort and 8.6% to 18.2% in the non-diabetes/T2D cohort. Adding dietary context from CGMacros improved errors further, especially after meals, where postprandial RMSE fell by about 15%. Source: ArXiv · AI/CL/LG's note.
score 4