Forecasting glucose from CGM and sparse meal logs with a residual-gated multimodal transformer

Forecasting glucose from continuous glucose monitoring in free-living settings is challenging because trajectories depend on endogenous dynamics and sparsely recorded meals. We developed $\our$, a multimodal transformer that produces a complete CGM-based forecast and then adds a gated, meal-informed residual correction. The correction is controlled by combined self-reported and CGM-derived evidence of meal presence, scaled per forecast horizon, allowing for dietary context to improve predictions without requiring meal records. We evaluated $\our$ in 1,752 adults without diabetes from the Framingham Heart Study. Participants wore Dexcom G6 Pro sensors and completed paired ASA24 dietary recalls. $\our$ achieved the lowest mean absolute error across forecast horizons, meal-state strata, and glucose ranges compared with long short-term memory, transformer, and GluFormer model architectures. It also improved prediction of postprandial peak glucose, time-to-peak, positive incremental area under the curve, and time in the 70--140~mg/dL range. Participant-level cross-validation confirmed generalization to unseen participants.