AI-driven personalized metabolic models of postprandial glucose to mixed meals

Background/Objectives Diet monitoring for chronic disease management is often hindered by burdensome self-reporting. Automated techniques are possible using continuous glucose monitors (CGMs), which capture post-prandial glucose responses (PPGRs), but inter-individual variability limits predictive utility of models. We hypothesized that machine learning models that jointly represent meals and individual health parameters would improve prediction of PPGRs from meal composition (direct) and inference of meal composition from PPGRs (inverse). Subjects/Methods : Data came from the CGMacros dataset. Forty-five individuals with varied HbA1c, demographics, labs, and gut microbiome profiles were included and their PPGRs captured by CGMs. We developed JointCGMacros, a deep neural network framework that encodes PPGRs and participant health features. Using triplet loss, we aligned similar meal-health pairs while separating dissimilar ones in the embedding space. From this, we estimated the two-hour incremental area under the curve (iAUC) of PPGRs (direct) and carb-to-calorie ratio of meals (inverse). Models were trained and validated with 10-fold cross-validation, with performance measured using Pearson correlation and normalized root mean square error (NRMSE). Interventions/methods used and duration of administration CGMacros provided participants 10 breakfast shakes with varied macronutrient compositions, 10 Chipotle lunches with varied compositions, and free-choice dinners. Participants fasted before breakfasts and avoided eating for three hours after breakfast and lunch. Results: On the direct task, we achieved an NRMSE=0.22 and Pearson correlation r=0.59, outperforming baseline predictions (NRMSE=0.23, r =0.52), though not significantly at p<0.01. On the inverse task, we achieved an NRMSE=0.28 and r = 0.47, a statistically significant improvements over baseline predictions (NRMSE=0.62,p<0.01; r=0.17,p<0.001). Conclusions While PPGRs alone are not predictive of meal composition without accounting for individual health parameters, JointCGMacros integrates these factors, enabling more accurate automated dietary monitoring and estimation of meal macronutrient effects.