The MedPerturb Dataset: What Non-Content Perturbations Reveal About Human and Clinical LLM Decision Making

2025-06-20

AI Summary

This study investigates the robustness of medical large language models (LLMs) in clinical decision-making, specifically examining how non-semantic perturbations—such as patient gender, linguistic style, and output format—affect human-AI decision consistency. Method: We introduce MedPerturb, a novel dataset derived from real-world clinical cases, and employ multidimensional controllable text perturbation generation, cross-model evaluation (four LLMs), multi-expert annotation (three clinicians per case), and causal sensitivity analysis to quantify human-AI discrepancies across 800 instances. Contribution/Results: We find that LLMs exhibit significantly higher sensitivity to gender and stylistic perturbations, whereas humans are more susceptible to LLM-typical output formats (e.g., summaries, multi-turn dialogues), revealing a fundamental divergence in robustness mechanisms between humans and LLMs. These findings shift clinical AI evaluation from static correctness toward dynamic scenario adaptability and establish a new assessment paradigm grounded in real-world clinical variability.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Learning Human Values and Preferences

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Clinical robustness is critical to the safe deployment of medical Large Language Models (LLMs), but key questions remain about how LLMs and humans may differ in response to the real-world variability typified by clinical settings. To address this, we introduce MedPerturb, a dataset designed to systematically evaluate medical LLMs under controlled perturbations of clinical input. MedPerturb consists of clinical vignettes spanning a range of pathologies, each transformed along three axes: (1) gender modifications (e.g., gender-swapping or gender-removal); (2) style variation (e.g., uncertain phrasing or colloquial tone); and (3) format changes (e.g., LLM-generated multi-turn conversations or summaries). With MedPerturb, we release a dataset of 800 clinical contexts grounded in realistic input variability, outputs from four LLMs, and three human expert reads per clinical context. We use MedPerturb in two case studies to reveal how shifts in gender identity cues, language style, or format reflect diverging treatment selections between humans and LLMs. We find that LLMs are more sensitive to gender and style perturbations while human annotators are more sensitive to LLM-generated format perturbations such as clinical summaries. Our results highlight the need for evaluation frameworks that go beyond static benchmarks to assess the similarity between human clinician and LLM decisions under the variability characteristic of clinical settings.

Problem

• Evaluates medical LLM sensitivity to non-content clinical input perturbations
• Compares human and LLM decision differences under clinical variability
• Assesses impact of gender, style, format changes on treatment selections

Innovation

• Dataset MedPerturb evaluates clinical LLM robustness
• Perturbs gender, style, format in clinical inputs
• Compares human and LLM decision-making under variability