MedRef-KG: Multi-agent KG-grounded reasoning framework for medical specialist referrals

Abstract

Patients often describe their symptoms in broad or subjective terms (e.g., “I feel tired,” “my chest feels weird”), which can make early clinical triage difficult and delay referral to the right specialty. Many conversational AI tools still struggle in this setting because they rely heavily on complete, well-structured symptom descriptions. In this work, we introduce MedRef-KG, a framework that combines a medical knowledge graph with a multi-agent dialogue process to turn incomplete patient narratives into specialist referral recommendations. We build a Neo4j medical knowledge graph using an ensemble of five large language models, resulting in 5773 nodes and 23,752 relationships. The system follows three main steps: (1) extracting key symptoms and asking targeted follow-up questions when information is missing, (2) managing a goal-driven doctor–patient dialogue, and (3) predicting the appropriate specialty using Qwen3-30B-A3B. On our evaluation, MedRef-KG reached a weighted F1-score of 0.9051, outperforming strong baselines such as fine-tuned medical GPT-4 (＋6.01%), OpenBioLLM-Llama3-70B (＋4.41%), Me-LLaMA (＋4.51%), BioBERT (＋8.51%), and Meditron-70B (＋20.01%). Performance was particularly strong for Musculoskeletal (F1 = 0.957) and Respiratory (F1 = 0.919) referrals. Overall, the results suggest that pairing structured medical knowledge with coordinated multi-agent reasoning can improve referral quality when patient input is incomplete—especially in real-world and resource-limited triage settings.