Susceptibility of Large Language Models to User-Driven Factors in Medical Queries.

Area of Science:

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

Background:

• Large language models (LLMs) are increasingly adopted in healthcare settings.
• LLM reliability is influenced by factors beyond model design, including user input.
• Understanding user-driven impacts is crucial for safe and effective LLM deployment in medicine.

Purpose of the Study:

• To evaluate how user-driven factors affect the diagnostic accuracy and reliability of LLM-generated medical responses.
• To assess the influence of misinformation framing, source authority, model personas, and data omission on LLM performance.
• To compare the susceptibility of proprietary and open-source LLMs to input biases.

Main Methods:

• Two testing methodologies were employed: perturbation (evaluating persona, source authority, tone) and ablation (omitting key clinical data).
• Evaluated proprietary LLMs (GPT-4o, Claude-3·5 Sonnet/Haiku, Gemini-1·5 Pro/Flash) and open-source LLMs (LLaMA-3 8B, LLaMA-3 Med42 8B, DeepSeek-R1 8B).
• Utilized MedQA and Medbullets datasets for diagnostic accuracy assessments.

Main Results:

• All LLMs demonstrated susceptibility to user-driven misinformation, with assertive tones having the most significant impact.
• Proprietary models were more vulnerable to authoritative misinformation and showed sharper accuracy declines with incomplete input.
• Omission of physical examination findings and laboratory results led to the most substantial decrease in diagnostic accuracy.

Conclusions:

• Structured prompts and comprehensive clinical context are essential for accurate LLM responses in healthcare.
• Users must avoid authoritative misinformation and provide complete data, especially for complex cases.
• Findings provide insights into mitigating user-driven biases for safer integration of LLMs into clinical practice.