TLDR: A study evaluated how large language models (LLMs) generate structured clinical consultation templates for physician-to-physician communication. It found that while LLMs can create comprehensive templates, they struggle with prioritizing the most important information, often producing excessively long outputs. Performance varied by specialty, with narrative-heavy fields like psychiatry showing significant degradation. The research highlights the need for LLMs to better curate information under real-world clinical constraints.
Large Language Models (LLMs) are rapidly changing various fields, and medicine is no exception. A recent study delves into how well these advanced AI models can create structured clinical consultation templates, which are crucial for effective communication between generalist doctors and specialists.
The Challenge of Clinical Communication
High-quality communication is vital in medicine to ensure accurate, timely, and efficient patient care. Breakdowns in communication can lead to medical errors and patient harm. Structured consultation templates, like those used by Stanford’s eConsult team, aim to streamline interactions by guiding primary care physicians on what information to include when seeking specialist advice. However, maintaining and updating these expert-crafted templates is resource-intensive and they often struggle to adapt to the wide variety of clinical situations.
This research explores whether LLMs can generate these templates, adapting to individual consults while maintaining clinical accuracy and relevance. It’s a unique challenge for AI, as it requires the models to anticipate what another clinician needs to know, rather than just recalling facts or summarizing text.
How the Study Was Conducted
The researchers used a dataset of 145 expert-generated templates from 20 medical specialties at Stanford. They developed a multi-agent pipeline to evaluate the LLMs. This involved:
-
Prompt Optimization: Using a framework called DSPy, model-specific prompts were refined to help LLMs generate high-quality templates that adhered to a specific structure.
-
Semantic Autograding: An AI-powered grader (OpenAI’s o3 model) was used to automatically assess if the LLM-generated items were clinically equivalent to the expert-created ones, even if the wording was different. This autograder was validated against human expert reviews, showing high concordance.
-
Template Generation: After optimization, various frontier LLMs, including OpenAI’s o3 and GPT-4o, Google’s Gemini 2.5 Pro, Anthropic’s Claude 4 Sonnet, and open-source models Kimi K2 and Llama 3 70B, were tasked with generating templates based only on the specialty and condition name.
-
Prioritization Analysis: A secondary AI agent ranked the generated template items by clinical importance. This allowed the researchers to assess how well models prioritized information when faced with length constraints, simulating real-world scenarios where conciseness is key.
Key Findings
The study revealed several important insights into LLMs’ capabilities in this domain:
-
Comprehensiveness vs. Conciseness: The top-performing model, o3, achieved high comprehensiveness (92.2% coverage of expert items) when unconstrained. However, it struggled with conciseness, generating 2.58 times more items than the original templates. When forced to prioritize to the original template length, o3’s coverage dropped significantly to 60.4%.
-
Prioritization Challenges: Kimi K2, another LLM, showed lower overall comprehensiveness (78.2%) but demonstrated stronger prioritization, retaining 69.5% coverage when constrained to the original template length. This highlights a core limitation: LLMs can list relevant content but struggle to curate it effectively under real-world constraints.
-
Specialty-Specific Performance: Performance varied greatly across medical specialties. Structured fields like nephrology (97.4% coverage) and infectious disease (93.9%) saw high performance, while narrative-driven fields such as psychiatry (61.5%) and pain medicine (68.2%) showed significant degradation, especially when length was constrained.
-
Novel Template Generation: For newly generated templates on novel topics, the AI-generated content was highly clinically appropriate (96.2%) and complete (human experts added only 1.2 items on average). However, these templates were often not as concise as they could be, with human experts deeming only 73.1% of items as necessary.
Also Read:
- Advancing Medical AI: A Deep Dive into Reasoning Capabilities of Large Language Models
- AI Framework for Smarter Pre-Consultation in Healthcare
Implications for Healthcare AI
This research underscores that while LLMs are powerful tools for generating medical content, their integration into physician-to-physician communication requires further refinement. The ability to prioritize clinically salient information within the time and attention constraints of real-world practice is crucial. Current benchmarks often don’t capture this nuanced need for curation and conciseness.
The study suggests that future solutions might involve combining LLMs’ broad knowledge with external structures defined by human expertise, perhaps through length-aware prompting or integrating data from electronic health records. The ultimate goal is to move towards “intelligence-augmented clinical dialogue” where LLMs enhance communication efficiency, diagnostic precision, and decision-making accuracy.
For more details, you can refer to the full research paper: Asking the Right Questions: Benchmarking Large Language Models in the Development of Clinical Consultation Templates.
