Variation in Large Language Model Recommendations in Challenging Inpatient Management Scenarios.

Area of Science:

• Clinical decision support systems
• Artificial intelligence in medicine
• Medical informatics

Background:

• Large language models (LLMs) are increasingly integrated into clinical workflows.
• Their performance in judgment-dependent bedside decisions remains largely uncharacterized.
• Understanding LLM behavior is crucial for safe and effective clinical implementation.

Purpose of the Study:

• To evaluate the variability of recommendations from commercially available LLMs.
• To assess consistency within and across different LLMs for common inpatient management scenarios.
• To identify potential limitations of LLMs in nuanced clinical decision-making.

Main Methods:

• A cross-sectional simulation study was conducted using four clinical vignettes.
• Six LLMs, including general-purpose and domain-specific models, were queried.
• Each LLM was prompted five times per vignette to assess internal consistency and inter-model agreement.

Main Results:

• LLM recommendations diverged significantly across all tested scenarios.
• Inter-model agreement was low, with substantial variation in management decisions (e.g., transfusion, anticoagulation, discharge).
• Internal consistency varied, with some models altering recommendations across repeated queries.

Conclusions:

• Widely used LLMs exhibit considerable variability in recommendations for complex inpatient management decisions.
• Clinicians should treat LLM outputs as supplementary information, not definitive guidance.
• Further research is needed to develop methods for surfacing model uncertainty and ensuring safe AI integration into clinical practice.