Determine generalization of macronutrient-based GluFormer to novel food items

Determine whether the multimodal GluFormer variant that represents diet via macronutrient tokens (e.g., calories, carbohydrates, proteins, lipids, water, sugars, alcohol, caffeine) can generalize to accurately predict glucose responses to previously unseen food items based solely on nutrient content rather than specific food identities.

Background

The model encodes dietary intake using macronutrient-level tokens rather than specific food identities, which the authors argue should confer generalizability across diverse dietary patterns and cultures. They explicitly state that this hypothesized generalization has not been tested.

Validating generalization to unseen foods would substantiate the nutrient-based representation approach and expand GluFormer's applicability in personalized nutrition and cross-cultural dietary contexts.