Answers are grounded in this paper and cite their sources.
Background Multiple-choice board examinations are a primary objective measure of competency in medicine. Large language models (LLMs) have demonstrated rapid improvements in performance on medical board examinations in the past two years. We evaluated five leading LLMs on neurosurgical board exam questions. Methods We evaluated five LLMs (OpenAI o1, OpenEvidence, Claude 3.5 Sonnet, Gemini 2.0, and xAI Grok2) on 500 multiple-choice questions from the Self-Assessment in Neurological Surgery (SANS) American Board of Neurological Surgery (ABNS) Primary Board Examination Review. Performance was analyzed across 12 subspecialty categories and compared to established passing thresholds. Results All models exceeded the threshold for passing, with OpenAI o1 achieving the highest accuracy (87.6%), followed by OpenEvidence (84.2%), Claude 3.5 Sonnet (83.2%), Gemini 2.0 (81.0%) and xAI Grok2 (79.0%). Performance was strongest in Other General (97.4%) and Peripheral Nerve (97.1%) categories, while Neuroradiology showed the lowest accuracy (57.4%) across all models. Conclusions State of the art LLMs continue to improve, and all models demonstrated strong performance on neurosurgical board examination questions. Medical image analysis continues to be a limitation of current LLMs. The current level of LLM performance challenges the relevance of written board examinations in trainee evaluation and suggests that LLMs are ready for implementation in clinical medicine and medical education.
No takes yet. Share an insight, caveat, or question.
Andrade et al. (2025) studied this question.
Synapse has enriched 6 closely related papers on similar clinical questions. Consider them for comparative context: