KGR-MAS: Knowledge Graph-Enhanced Multi-Agent LLM Framework for Evidence-Based Medical Diagnosis and Clinical Decision Support.

Abstract

Large Language Models (LLMs) demonstrate significant potential in healthcare applications, yet they exhibit reasoning bias, information bottlenecks, and limited adaptability in complex clinical settings. To address these limitations, we propose KGR-MAS, which introduces an integrated control loop that couples (1) KG-derived diagnostic paths for executable subtask decomposition and dynamic agent allocation, while using these KG reasoning paths to initialize shared short-term memory; (2) a language-level LSTM-inspired gating mechanism with Transformer-XL short-term memory to regulate cross-agent information flow and maintain global coherence; and (3) a KG-guided RAG-STM process that triggers evidence retrieval at diagnostic checkpoints and writes the retrieved evidence back into STM for subsequent reasoning. We constructed a large-scale medical knowledge graph achieving the best embedding performance (Hits@1:32.8%, MRR:0.498). Experimental results on MedQA and MedMCQA show that KGR-MAS achieves 87.2% accuracy on MedQA($\uparrow$14.4%), and 77.4% average accuracy across four open-domain datasets ($\uparrow$29.0%), outperforming larger models like GPT-o1(72.0%). KGR-MAS provides interpretable, accurate, and clinically grounded diagnostic reasoning, highlighting its potential as a next-generation AI assistant for medical decision-making.