1935-P: Multimodal CGM Forecasting: Fine-Tuned Foundation Models and Dietary Context Improve Postprandial Prediction

Introduction and Objective: Accurate forecasting of continuous glucose monitoring (CGM) trajectories can support proactive diabetes management. We evaluated time-series foundation models (FMs) for CGM forecasting in two populations: type 1 diabetes and a combined non-diabetes/type 2 diabetes group. We assessed whether adding multimodal dietary context improves predictive performance.Methods: Using 8 CGM datasets, we compared FMs (zero-shot and fine-tuned) with baseline methods across multiple forecast horizons and input history lengths. To test dietary effects, we augmented CGM inputs in the CGMacros dataset with (1) food-image embeddings generated by a ConvNeXT model fine-tuned on Food-101 and (2) structured macronutrient features. Models were evaluated on held-out test sets and postprandial subset.Results: In zero-shot settings, FMs provided little improvement over baseline models. After task-specific fine-tuning, FMs achieved lower prediction error across horizons. Adding meal information improved postprandial forecasting vs CGM-only models. Food-image embeddings were most beneficial for visually distinctive meals, whereas macronutrient features were more robust for visually complex or ambiguous meals (Fig).Conclusion: FMs require task-specific fine-tuning to improve CGM forecasting. Dietary context improves post-meal prediction, with images and macronutrients contributing complementary value for meal-aware forecasting.

B. Zhang: None. K. Zhou: None. H. Cheng: None. H. Yang: None. J. Zhou: None. H. Zhou: None.

National Institutes of Health under grants R35 GM141798 (HZ), R01 HG006139 (HZ and JZ), and R01 DK142026 (HZ and JZ); National Science Foundation under grants DMS 2054253 (HZ and JZ) and IIS 2205441 (HZ and JZ).



Source link