MealRes-Gate: Forecasting Glucose Dynamics from CGM and Sparse Meal Logs Using Residual-Gated Multimodal Transformer

Continuous glucose monitoring (CGM) enables personalized metabolic health support, but forecasting glucose in free-living settings remains challenging because future trajectories depend on both endogenous dynamics and sparsely recorded meals. We developed MealRes-Gate, a multimodal transformer that incorporates meal information as a gated residual refinement to a strong CGM-based backbone. In 1,752 non-diabetic and pre-diabetic adults from the Framingham Heart Study, MealRes-Gate consistently outperformed recurrent, Transformer-based, and GluFormer baselines across 30-, 60-, 90-, and 120-minute horizons. Gains were largest in postprandial, high glucose, and low glucose windows, and extended to clinically relevant postprandial summaries that included peak glucose, time-to-peak, and glucose area under the curve. Ablation analysis showed that engineered CGM features provided the dominant predictive backbone, while explicit meal features contributed smaller but meaningful gains when integrated through the proposed residual-gating mechanism. These results demonstrate that sparse dietary information can improve glucose forecasting without destabilizing prediction, provided it is incorporated through a selective, residual gating mechanism.