Clinicians are increasingly exploring the integration of artificial intelligence into specialized medical fields. A recent study evaluated LLMs in Neurosurgical Diagnosis to determine their current clinical utility. Researchers compared five prominent models using 148 clinical vignettes from a standardized neurosurgery review textbook. Specifically, they analyzed OpenAI's ChatGPT-3.5 and ChatGPT-4, Google’s Gemini, Microsoft Copilot, and the specialty-specific AtlasGPT.

Performance of LLMs in Neurosurgical Diagnosis

The results highlighted a significant performance gap between the various artificial intelligence systems. ChatGPT-4 emerged as the most accurate model, achieving a 74% success rate in providing correct diagnoses. AtlasGPT followed at 63%, while ChatGPT-3.5 and Microsoft Copilot scored 53% and 48%, respectively. Gemini demonstrated the lowest accuracy at 36%. Furthermore, statistical analysis confirmed that ChatGPT-4 significantly outperformed its counterparts with a p-value of 0.005.

However, the researchers identified specific hurdles in the diagnostic process. Most errors occurred when models failed to attribute critical imaging data to the final diagnosis. Despite this, the models generally exhibited logical, stepwise reasoning. Consequently, adding image processing capabilities appears vital for enhancing accuracy. Practitioners must remain cautious, as these tools require detailed prompting and critical oversight. Moreover, while LLMs offer potential for common conditions, they are not yet a replacement for clinical expertise.

How accurate is ChatGPT-4 in neurosurgery?

ChatGPT-4 achieved a 74% accuracy rate in diagnosing neurosurgical vignettes. This makes it the top-performing general model in recent comparative studies.

What are the main limitations of AI in neurosurgery?

The primary limitation is the difficulty models face when integrating imaging data into their diagnostic reasoning. This often leads to errors despite otherwise logical text-based analysis.

Disclaimer: This content is for informational and educational purposes only. It does not constitute medical advice or a professional relationship. AI tools should be used as assistants, not as primary diagnostic tools. Refer to the latest local and national guidelines for clinical practice.

References

Warrier A et al. Comparative diagnostic capability of large language models in neurosurgery. J Neurosurg. 2026 Feb 06. doi: 10.3171/2025.9.JNS25846. PMID: 41650460.

McNulty AM et al. Performance evaluation of ChatGPT-4.0 and Gemini on image-based neurosurgery board practice questions: A comparative analysis. J Clin Neurosci. 2025 Apr;134:111097. doi: 10.1016/j.jocn.2025.111097.

Su H et al. Accuracy and quality of ChatGPT-4o and Google Gemini performance on image-based neurosurgery board questions. Neuroradiology. 2025 Mar 25. doi: 10.1007/s10143-025-03472-7.