DxDirector: an agentic large language model driving the full-process clinical diagnosis

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Medical Diagnostics

Background:

• Real-world clinical diagnosis involves complex reasoning from ambiguous patient complaints.
• Current large language models (LLMs) assist with specific queries but cannot autonomously manage the entire diagnostic workflow.
• Physician workload is a significant challenge in healthcare systems.

Purpose of the Study:

• To introduce DxDirector-7B, an agentic LLM capable of autonomously navigating the complete clinical diagnostic process.
• To evaluate DxDirector-7B's diagnostic accuracy and efficiency compared to existing LLMs and human physicians.
• To assess the potential of AI to lead clinical reasoning and reduce physician involvement.

Main Methods:

• Development of DxDirector-7B, an agentic LLM with "slow thinking" capabilities for strategic diagnostic planning.
• Autonomous determination of diagnostic strategies with targeted requests for physician intervention.
• Evaluation on rare diseases and complex, real-world clinical cases.

Main Results:

• DxDirector-7B demonstrated superior diagnostic accuracy compared to state-of-the-art medical and general-purpose LLMs.
• The model significantly reduced the need for physician involvement in the diagnostic process.
• A robust safety and accountability framework was maintained for high-risk conditions.

Conclusions:

• DxDirector-7B represents a paradigm shift in AI-driven clinical reasoning, effectively leading the diagnostic workflow.
• This agentic LLM offers a scalable solution for enhancing diagnostic efficiency and accessibility.
• AI can significantly alleviate physician workload while maintaining diagnostic quality and safety.