LLM Sensitivity Evaluation Framework for Clinical Diagnosis

2025-04-18

International Conference on Computational Linguistics

AI Summary

This work addresses the insufficient sensitivity of large language models (LLMs) to critical clinical information—such as symptom duration and vital sign trends—in diagnostic reasoning, systematically characterizing their diagnostic fragility for the first time. We propose a clinical diagnosis–oriented sensitivity evaluation framework, incorporating multi-dimensional, medical-semantic-preserving perturbation strategies: adversarial perturbations, entity substitutions, and temporal deformations. To quantify robustness, we introduce a diagnostic pathway consistency metric and validate results via expert annotation. Experiments across GPT-3.5, GPT-4, Gemini, Claude-3, and LLaMA2-7b reveal an average 27.6% drop in diagnostic accuracy under minor perturbations to key clinical cues, exposing significant vulnerabilities in clinical reasoning. We publicly release the DiagnosisQA benchmark dataset and evaluation code, establishing a new standard and open toolkit for advancing the reliability and clinical trustworthiness of LLMs in medicine.

Technology Category

• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large language models (LLMs) have demonstrated impressive performance across various domains. However, for clinical diagnosis, higher expectations are required for LLM's reliability and sensitivity: thinking like physicians and remaining sensitive to key medical information that affects diagnostic reasoning, as subtle variations can lead to different diagnosis results. Yet, existing works focus mainly on investigating the sensitivity of LLMs to irrelevant context and overlook the importance of key information. In this paper, we investigate the sensitivity of LLMs, i.e. GPT-3.5, GPT-4, Gemini, Claude3 and LLaMA2-7b, to key medical information by introducing different perturbation strategies. The evaluation results highlight the limitations of current LLMs in remaining sensitive to key medical information for diagnostic decision-making. The evolution of LLMs must focus on improving their reliability, enhancing their ability to be sensitive to key information, and effectively utilizing this information. These improvements will enhance human trust in LLMs and facilitate their practical application in real-world scenarios. Our code and dataset are available at https://github.com/chenwei23333/DiagnosisQA.

Problem

• Evaluating LLM sensitivity to key medical information for diagnosis
• Assessing LLM reliability in clinical diagnostic decision-making
• Improving LLM ability to detect and utilize critical medical data

Innovation

• Evaluates LLM sensitivity to key medical information
• Introduces perturbation strategies for testing
• Highlights need for improved reliability and sensitivity