Abstract:
Non-Communicable Diseases (NCDs) such as diabetes, hypertension, and cardiovascular
conditions remain among the most pressing global health challenges. These are often
exacerbated by poor dietary habits and the complexities of comorbidity management. Existing
digital meal planning solutions typically offer generic advice, lacking the capacity to reason
across individual health profiles or provide clinically grounded dietary interventions for users
with multiple overlapping conditions.
To address this gap, this research introduces MealFit, an AI-powered web application that
delivers personalized dietary recommendations tailored to both NCD and non-NCD users. The
system integrates a multi-agent architecture with a Retrieval-Augmented Generation (RAG)
pipeline and a domain-specific Bio-Medical LLaMA model, enabling the interpretation of
structured health data (e.g., BMI, age, blood pressure) and unstructured user input (e.g.,
ingredient preferences, disease mentions). Agents collaborate to perform risk assessment,
disease detection, and meal plan generation, grounded in clinical guidelines from sources such
as the WHO and ADA.
The system was developed using FastAPI, LangChain, and ChromaDB, and tested using real
world health profiles. Manual evaluation across diverse test cases demonstrated a 80% clinical
accuracy rate, while domain expert reviews confirmed the relevance, personalization, and
interpretability of recommendations. Benchmarking against general-purpose models like GPT
4, Gemini, and DeepSeek highlighted the system’s superior performance in medical alignment
and dietary precision. The platform’s modular design supports future enhancements including
ingredient substitution, real-time biomarker integration, and multilingual expansion
positioning MealFit as a scalable and clinically meaningful tool in AI-driven personalized
nutrition.
