MetaPlate: Counterfactual-Guided RAG-LLM Tool for Personalized Food Recommendation and Hyperglycemia Prevention

MetaPlate is an AI-powered dietary decision-support system that combines counterfactual explanations, continuous glucose monitoring data, and large language models to generate personalized meal recommendations for preventing postprandial hyperglycemia. The system demonstrated improved clinical plausibility and actionability through expert validation with registered dietitians, showcasing how domain-specific constraints enhance LLM reliability in healthcare applications.

MetaPlate addresses a significant gap in personalized nutrition by moving beyond static dietary guidelines to generate context-aware, real-time meal recommendations. The system integrates multimodal physiological data with machine learning glucose prediction and counterfactual optimization to suggest meal adjustments that keep blood glucose levels within healthy ranges. This represents a meaningful advancement in translating predictive health models into actionable guidance that users can actually follow.

The research builds on growing interest in AI-driven personalization within healthcare, particularly in metabolic disease prevention. Continuous glucose monitoring has become more accessible, and machine learning models have improved at predicting individual glycemic responses. However, previous systems largely stopped at prediction without providing practical dietary interventions. MetaPlate closes this gap by combining three layers: data integration for personalized context, ML-based glucose prediction, and an LLM interface that searches constrained food databases to generate realistic, interpretable recommendations.

The expert validation methodology is particularly noteworthy. Rather than assuming LLM outputs are suitable for clinical use, the researchers conducted structured assessments with registered dietitians before and after prompt refinement. This revealed significant improvements in meal realism and clinical appropriateness, demonstrating that thoughtful prompt engineering and domain constraints can substantially enhance LLM utility in specialized fields. The findings underscore a critical lesson: healthcare AI systems require domain expertise integration, not just raw model capability.

Looking forward, this work hints at broader potential for AI-assisted dietary intervention in metabolic disorder management. Success at scale would depend on integration with existing healthcare workflows, validation in diverse populations, and regulatory clarity around clinical decision-support tools.

• →MetaPlate combines counterfactual explanations with LLMs to generate personalized, actionable meal recommendations for glucose management.
• →Expert validation with dietitians showed substantial improvements in clinical plausibility after prompt refinement, highlighting the importance of domain knowledge in healthcare AI.
• →The system integrates multimodal data including CGM readings and wearable signals to model individual pre-meal context for personalization.
• →Constrained search of the USDA food database enhances both interpretability and practical feasibility of LLM-generated recommendations.
• →Results demonstrate that structured constraints and expert-in-the-loop refinement can significantly improve LLM reliability in clinical applications.