From Plate to Glucose: ML Prediction of Postprandial Responses to Keto vs. Trditional Banquets in T2D

Abstract

This study focuses on the development and rigorous evaluation of machine learning (ML) models to predict the peak postprandial glycemic response (PPGR_peak_0_120) in individuals at metabolic risk, analyzing keto versus traditional banquets. The primary objective was to integrate a comprehensive set of variables (clinical phenotype, nutritional composition, and microbiota markers) to improve individual prediction accuracy. Penalized regression models (ElasticNet, Lasso) and a nonlinear model (GradientBoosting) were applied. Validation was performed using the strict Leave-One-Group-Out (LOGO) strategy to assess the real generalization capacity to new participants. The ElasticNet model turned out to be the best performer, reaching an RMSE of 32.13 mg/dL. Although the negative R2 obtained in the LOGO validation underlines the challenge of high inter-individual variability, the relative RMSE (∼18% of the SD) is clinically acceptable. The models converged in identifying robust risk factors: BMI_kg/m2 and HbA1c_pct were consistently the most powerful predictors of peak glycemic magnitude. The findings highlight the importance of nutritional trait engineering (diet × macronutrient interactions and caloric densities), demonstrating that nutrient proportions and context are crucial. The study concludes that, while accurate individual prediction is complex, the models are valuable tools for identifying risk hierarchies and informing personalized dietary interventions in the management of Type 2 Diabetes.