GraphDx: A Cost-Aware Knowledge-Enhanced Multi-Agent Framework for Sequential Diagnosis

Short summary

GraphDx is a multi-agent framework that enhances LLM-based sequential medical diagnosis by combining a Medical Diagnosis Knowledge Graph with three collaborative agents for perception, reasoning, and decision-making. It improves diagnostic success rates from 50–68% to 79–93% on MedQA and MIMIC-IV while reducing test costs by 20–54% across three LLM backbones. The system addresses the key gap where LLMs encode medical knowledge but fail to reason systematically under cost constraints.

• •GraphDx pairs a Medical Diagnosis Knowledge Graph with three collaborative agents for cost-aware sequential diagnosis
• •Diagnostic accuracy improved from 50–68% to 79–93% on MedQA and MIMIC-IV while cutting test costs 20–54%
• •Tested across DeepSeek-V3, Kimi-k2, and Llama-3.3 backbones

Generated with AI, which can make mistakes.