(P41-24-26) Personalized Meal Optimization Through Uncertainty-Aware Time-Resolved Glucose Trajectory Prediction Using Latent Metabolic Phenotyping in Type 2 Diabetes

Objectives: Postprandial glucose responses vary widely between individuals with type 2 diabetes, making effective dietary management difficult. Recent machine learning approaches have improved the prediction of glucose responses to meals, but few frameworks translate these predictions into optimized meal recommendations. This study aimed to develop a predictive modeling framework capable of forecasting individualized postprandial glucose trajectories and enabling personalized optimization of meal composition to improve glycemic control.

Methods: Continuous glucose monitoring data from 1,712 meals in individuals with type 2 diabetes were analyzed using a recurrent neural network variational autoencoder (RNN-VAE) model. This model integrated 31 meal and individual-specific features and learned 16-dimensional latent metabolic state embeddings that captured unobserved physiological variability. The RNN-VAE forecasted glucose trajectories up to 300 minutes post-meal and estimated prediction uncertainty. A within-subject cross-validation scheme was employed to evaluate predictive performance. Meal optimization was carried out using differential evolution, balancing glycemic minimization with nutritional constraints.

Results: The latent metabolic state model achieved robust time-resolved glucose prediction accuracy (RMSE: 24.38 mg/dL, R²: 0.71). Personalized meal optimization showed clinically meaningful reductions in predicted peak glucose levels, averaging 41.6 ± 22.0 mg/dL, while enhancing nutritional diversity. Optimization strategies included carbohydrate reductions coupled with strategic increases in protein, fat, and fiber.

Conclusions: By combining latent metabolic phenotype inference with extended temporal glucose modeling and uncertainty quantification, this framework provides accurate, interpretable forecasts of individual postprandial glucose responses. The integrated meal optimization approach effectively translates predictions into nutritionally balanced, personalized dietary recommendations. Together, these advances establish a foundation for precise, proactive diabetes management through data-driven nutrition.

Funding Sources: The Swedish Foundation for International Cooperation in Research and Higher Education (STINT).