GraphDx: AI multi-agent framework cuts diagnosis costs by 20-54%

GraphDx introduces a cost-aware knowledge-enhanced multi-agent framework for sequential diagnosis. It addresses a key limitation in current LLM-based diagnosis: models encode vast medical knowledge but struggle to reason systematically under cost constraints, often ordering excessive tests. The framework features an automated pipeline that leverages LLMs to construct Medical Diagnosis Knowledge Graphs (MDKGs) with quantized typicality, action-centric topology, and dual-objective attributes for diagnostic relevance and cost sensitivity. Three collaborative agents—Perception, Reasoning, and Decision—handle language understanding, evidence scoring, and cost-aware planning on the MDKG.

Experiments across three LLM backbones (DeepSeek-V3, Kimi-k2, Llama-3.3) on MedQA and MIMIC-IV datasets show dramatic improvements: diagnostic success rates jump from 50-68% to 79-93%, while test costs drop by 20-54%. The reasoning agent's deterministic scoring on structured knowledge graphs makes the system interpretable and robust. GraphDx demonstrates that combining knowledge graphs with multi-agent orchestration can make AI-assisted diagnosis both more accurate and more economical, potentially reducing unnecessary medical spending while improving patient outcomes.