Development of a Precision Nutrition Algorithm for Predicting Glycemic Response in Type 2 Diabetes: Utilizing Continuous Glucose Monitoring and Gut Microbiome Data

Abstract

Background and Aims: While carbohydrates are viewed as key post-prandial glycemic response (PPGR) predictors, their reliability has been debated. Machine-learning techniques have improved PPGR predictions in healthy individuals; however, data for type 2 diabetes (T2D) patients remains sparse. Our goal is to assess carbohydrate-based PPGR predictions and develop a precision nutrition algorithm using continuous glucose monitoring (CGM) and gut microbiome data in T2D patients. Methods: We analyzed mixed-meal data from forty-nine T2D patients on antidiabetic medications, covering roughly a thousand meals. We developed a multimodal predictive deep neural network (DNN)learning model incorporating convolutional neural network and Long Short-Term Memory layers (DNN) that fuses static personal data, meal composition, time-series CGM prior to meals, and microbiome data (Fig. A). We evaluated its performance against the existing gradient boosting regressor (GBR)-based model. Pearson correlation using leave-one-person-out cross-validation was reported. Results: Prediction based on carbohydrate alone yielded correlation of 0.349 for single predictor. Incorporating additional meal composition (macronutrients and calorie) improved correlations to 0.358 for both GBR and DNN. With the inclusion of meal context (dining time and preceding meals), performance increased further to 0.488 and 0.507, respectively. GBR's best performance was 0.637 using meal composition, meal context, CGM, and microbiome. For DNN, the highest score was 0.702, incorporating meal composition, meal context, CGM, clinical parameters, microbiome, and medications (Fig. B,C). Conclusions: While carbohydrates provide limited predictive accuracy for PPGR in T2D patients, the proposed DNN model including CGM and microbiome can significantly enhance prediction rates, paving the pathway for the real-world medical application of PPGR predictor in T2D.