Asking the Right Questions: Benchmarking Large Language Models in the Development of Clinical Consultation Templates

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Natural Language Processing

Background:

• Electronic consultations (eConsults) facilitate efficient physician communication.
• Structured templates are crucial for clear and concise clinical information exchange.
• Large language models (LLMs) show potential for automating template generation.

Purpose of the Study:

• To evaluate the capability of state-of-the-art LLMs in generating structured clinical consultation templates.
• To assess the clinical coherence, conciseness, and prioritization of LLM-generated templates.
• To identify limitations of current LLMs in producing clinically relevant eConsult schemas.

Main Methods:

• Utilized 145 expert-crafted eConsult templates from Stanford's eConsult team.
• Assessed frontier LLMs including o3, GPT-4o, Kimi K2, Claude 4 Sonnet, Llama 3 70B, and Gemini 2.5 Pro.
• Employed a multi-agent pipeline with prompt optimization, semantic autograding, and prioritization analysis.

Main Results:

• Models achieved high comprehensiveness (up to 92.2%) but generated overly long templates.
• LLMs failed to prioritize clinically important questions effectively under length constraints.
• Performance varied by medical specialty, with notable degradation in psychiatry and pain medicine.

Conclusions:

• LLMs show promise for improving structured clinical information exchange.
• Current LLMs require enhanced evaluation methods focusing on clinical salience and prioritization.
• Addressing length constraints and specialty-specific nuances is critical for real-world LLM application in eConsults.