Med-KAG: Preliminary Results of a Medical Knowledge-Augmented Generation Approach

We propose in this paper Med-KAG, a novel Artificial Intelligence (AI) assistant architecture for clinical decision support. Med-KAG extends the Retrieval-Augmented Generation (RAG) paradigm by integrating a medical Knowledge Graph built from the Unified Medical Language System (UMLS) Metathesaurus. This approach grounds the model's responses in verified biomedical relationships between diseases, symptoms, and treatments, reducing hallucinations and improving transparency. Our preliminary evaluation on MedQA-US compares a baseline large language model (Qwen3-235B-A22B) with its knowledge-enhanced variant, showing comparable accuracy while identifying the retriever as the primary source of error. These results highlight both the potential and current limitations of combining structured medical knowledge with generative AI to achieve more reliable and explainable clinical assistance.