GraphDx Improves AI Medical Diagnosis, Cuts Costs by Leveraging Knowledge Graphs.

Sequential diagnosis in medicine requires careful balancing of diagnostic accuracy with the associated costs of tests and procedures. Current Large Language Model (LLM) based systems, despite their vast knowledge, often struggle with this balance, frequently recommending more tests than necessary due to a lack of systematic, cost-aware reasoning. A new framework, GraphDx, tackles this challenge by integrating a knowledge-enhanced, multi-agent system. It first constructs detailed Medical Diagnosis Knowledge Graphs (MDKGs) using LLMs, embedding both diagnostic relevance and cost-sensitivity. These graphs feature quantized typicality and an action-centric topology. GraphDx then employs three specialized agents: a Perception Agent for language understanding, a Decision Agent for output generation, and a Reasoning Agent that deterministically scores evidence and plans cost-aware actions directly on the MDKG. This collaborative approach has shown substantial improvements in diagnostic success rates and significant reductions in test costs across various LLM backbones and medical datasets.