MetaPlate uses counterfactual RAG-LLM to personalize meals and prevent blood sugar spikes

MetaPlate, developed by Asiful Arefeen, Carol Johnston, and Hassan Ghasemzadeh at Arizona State University, addresses a critical gap in dietary guidance for preventing postprandial hyperglycemia—a key risk factor for metabolic disorders. Unlike static dietary advice or purely predictive machine learning models, MetaPlate integrates multimodal data (continuous glucose monitoring readings, wearable physiological signals, and user-provided meal logs from 25 individuals) to model pre-meal context. A machine learning model predicts glucose response, and a counterfactual explanation (CF) module adjusts macronutrient amounts to keep glucose within a target range (≤140 mg/dL). An LLM-based retrieval-augmented generation (RAG) layer then queries the USDA food database to produce human-readable, context-aware meal recommendations.

Expert evaluation with registered dietitians compared outputs before and after prompt refinement. Results showed significant improvements in meal realism, portion suitability, and overall recommendation likelihood—shifting from clinically implausible suggestions to actionable, contextually appropriate advice. The work emphasizes the importance of domain knowledge and structured constraints in LLM-driven health systems. MetaPlate demonstrates potential as a real-time personalized dietary decision-support tool that could help millions manage blood sugar through practical, data-driven meal adjustments.