level: research
large language models often order too many tests when diagnosing patients step by step. they know a lot about medicine but struggle to reason about costs. graphdx addresses this by building a medical diagnosis knowledge graph automatically. the graph includes typicality scores, action-focused connections, and attributes for both diagnostic value and cost. this structure helps the system decide which test to run next without wasting resources.
the framework uses three agents that work together. the perception agent understands language input. the decision agent generates responses. the reasoning agent does the core work: it scores evidence and plans actions while respecting cost limits. this separation keeps the reasoning deterministic and transparent. the knowledge graph gives the reasoning agent a clear map of diseases, symptoms, and tests, so it can weigh the usefulness of each possible next step against its cost.
graphdx aims to close the gap between knowing medical facts and using them wisely under constraints. by combining structured knowledge with agent-based control, it tries to avoid the common llm pitfall of excessive testing. the approach could make automated diagnostic tools more practical in settings where time and money are limited. the paper describes the design and the automated graph-building pipeline, showing how llms can help create the knowledge base that then guides more disciplined reasoning.
why it matters: it shows a way to make ai diagnostic tools more cost-effective, which is important for real-world healthcare where resources are limited.