Abstract: Predicting 2-hour postprandial blood glucose (BGAfter) is important for diabetes management, but many meal-aware models require manually entered nutritional information. This study investigates whether image-derived meal representations combined with pre-meal blood glucose (BGBefore) can provide comparable prediction performance with lower input burden. The proposed multimodal framework includes a nutritionally supervised meal-image model and a BG prediction model that combines BGBefore or 17 physiological features with nutritional facts, PCA features, VAE features, or joint embeddings. Performance was evaluated using RMSE, MAE, R-squared, and Clarke Error Grid analysis. Image-derived meal features achieved broadly comparable performance to expert-estimated nutritional facts, although differences were generally modest relative to fold-wise variability. In the BGBefore-only setting, the best image-based configuration achieved an RMSE of 43.52 mg/dL, compared with 47.20 mg/dL for the nutritional-fact-based model. With 17 physiological features, the best model achieved an RMSE of 32.10 mg/dL. Across the main configurations, 98.15%–98.89% of prediction–reference pairs were located in Clarke Error Grid Zones A+B. These findings suggest that image-derived representations may reduce manual nutritional entry at inference time. However, the small cohort and absence of temporal and behavioural modelling limit generalizability and clinical interpretation.

 

Presenter: Md Benzir Ahmed (Std No. 1016054001)

Venue: Graduate Seminar Room