Evaluating the Performance and Fragility of Large Language Models on the Self-Assessment for Neurological Surgeons

Area of Science:

• Artificial Intelligence in Medicine
• Neurosurgery Education
• Natural Language Processing

Background:

• Neurosurgery residents use the Congress of Neurological Surgeons Self-Assessment for Neurological Surgeons for board exam preparation.
• Large language models (LLMs) are increasingly evaluated for their neurosurgical knowledge.
• Clinical text is susceptible to distractions from generative AI and ambient dictation.

Purpose of the Study:

• To assess the performance of state-of-the-art LLMs on neurosurgery board-like questions.
• To evaluate the robustness of LLMs to in-text distractor statements.

Main Methods:

• Evaluated 28 LLMs using 2904 neurosurgery board examination questions.
• Introduced a distraction framework with irrelevant statements containing polysemous words.
• Assessed the degradation of model performance due to distractors.

Main Results:

• Six of 28 LLMs achieved board-passing scores.
• Distractions reduced accuracy by up to 20.4%, causing one model to fail.
• Open-source models showed greater performance decline than proprietary variants.

Conclusions:

• Current LLMs can answer neurosurgery board-like questions but are vulnerable to distractions.
• Developing mitigation strategies is crucial for safe clinical deployment of LLMs.
• LLM resilience against in-text distractions needs enhancement.