GraphDx framework improves LLM-based clinical diagnosis accuracy and reduces test costs in study

Researchers propose GraphDx, a knowledge-enhanced multi-agent framework designed to address a persistent challenge in LLM-based clinical diagnosis: balancing diagnostic accuracy with resource costs during iterative information gathering.

The framework introduces three collaborative agents: a Perception Agent for language understanding, a Reasoning Agent for deterministic evidence scoring and cost-aware planning on Medical Diagnosis Knowledge Graphs (MDKGs), and a Decision Agent for language generation.

GraphDx automates the construction of MDKGs, which incorporate quantized typicality, action-centric topology, and dual-objective attributes for both diagnostic relevance and cost sensitivity. These graphs serve as the substrate for the Reasoning Agent’s cost-aware planning.

In experiments conducted on the MedQA and MIMIC-IV benchmarks using three LLM backbones—DeepSeek-V3, Kimi-k2, and Llama-3.3—the framework improved diagnostic success rates from a baseline of 50–68% to 79–93%. Concurrently, GraphDx reduced test costs by 20–54%, demonstrating both accuracy gains and economic efficiency.

The authors argue that existing LLM approaches often resort to excessive testing due to a knowledge-reasoning gap, and that GraphDx mitigates this by embedding cost constraints directly into the diagnostic reasoning process via the MDKG and collaborative agent architecture.