Personas Shift Clinical Action Thresholds in Large Language Models.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Natural Language Processing

Background:

• Clinical deployment of large language models (LLMs) is moving from assessing feasibility to addressing liability.
• Current guidance often frames LLM behavior as a control problem.
• The impact of system prompts on clinical judgment requires further investigation.

Purpose of the Study:

• To determine if decision-style system prompts can alter clinical action thresholds in LLMs.
• To assess the consistency of these shifts across different clinical settings and LLM architectures.
• To evaluate the influence of physician personas on LLM decision-making.

Main Methods:

• Defined nine physician personas based on ethical orientations and cognitive styles.
• Evaluated 20 open-weight LLMs using 2,500 simulated ED vignettes and 2,500 MIMIC-IV-Note discharge summaries.
• Models responded to five binary decision items per text under baseline and persona prompts, generating over 5 million decisions.

Main Results:

• Persona prompts shifted affirmative decision rates from 36.9% to 46.4%, a 9.5-percentage-point change with fixed clinical evidence.
• Prompting effects were most pronounced for autonomy and treatment decisions.
• Prompt-induced shifts were consistent across different text corpora, but susceptibility varied by LLM, with no clear benefit from medical fine-tuning or larger model size.

Conclusions:

• Decision-style system prompts demonstrably alter clinical action thresholds in LLMs when presented with consistent clinical facts.
• Prompting functions as a policy-setting mechanism, not merely a communication tool.
• System prompts should be considered a primary deployment configuration for clinical LLMs.