Sense and Sensitivity: Benchmarking LLM Clinical Triage Recommendations with Physician Experts

2026-09-29

AI Summary

This study addresses the unclear reliability of large language models (LLMs) in clinical triage when facing textual perturbations. To investigate this, we introduce a novel, large-scale evaluation benchmark that integrates physician expert behaviors, encompassing over 6,000 clinical scenarios and 7,000 physician annotations. By combining adversarial example generation with clinical text perturbation techniques, this work systematically quantifies how semantically preserving yet irrelevant textual variations affect decision-making. Our findings reveal that LLMs are more prone than physicians to recommending unnecessary care and exhibit heightened sensitivity to gender and tone perturbations. This research underscores the necessity of deployment-oriented evaluations grounded in expert behavior, providing critical evidence for enhancing the robustness of medical AI systems.

Abstract

As large language models (LLMs) are increasingly used in clinical settings, it is critical to evaluate their reliability under realistic variation in clinical text. We study this question in clinical triage, comparing LLMs to practicing physicians under text perturbations that preserve the underlying clinical setting. We introduce a benchmark of over 6,000 clinical scenarios, 7,000 physician annotations, and 225,000 model responses. Using this benchmark, we make two key observations. First, LLMs are more likely than physicians to recommend unnecessary care at baseline, and this tendency increases under perturbed inputs. Further, we find that LLM recommendations are more sensitive to gender and tone perturbations than human recommendations. Together, these results demonstrate that LLMs can vary under clinically irrelevant textual changes, highlighting the need for deployment-oriented evaluations grounded in expert physician behavior.

Problem

• Large Language Models
• Clinical Triage
• Reliability
• Text Perturbation
• Benchmarking

Innovation

• Large Language Models
• Clinical Triage
• Benchmarking
• Text Perturbation
• Physician Experts