• Ethan D L Brown, Apratim Maity, Max Ward, Daniel Toscano, Griffin R Baum, Mark A Mittler, Sheng-Fu Larry Lo, and Randy S D'Amico.
• Department of Neurological Surgery, Donald and Barbara Zucker School of Medicine at Hofstra/Northwell, Lake Success, New York, USA. Electronic address: ebrown35@northwell.edu.
• World Neurosurg. 2024 Aug 28.

ObjectiveChatGPT has been increasingly investigated for its ability to provide clinical decision support in the management of neurosurgical pathologies. However, concerns exist regarding the validity of its responses. To assess the reliability of ChatGPT, we compared its responses against the 2023 Congress of Neurological Surgeons (CNS) guidelines for patients with Chiari I Malformation (CIM).MethodsChatGPT-3.5 and ChatGPT-4 were prompted with revised questions from the 2023 CNS guidelines for patients with CIM. ChatGPT provided responses were compared to CNS guideline recommendations using cosine similarity scores and reviewer assessments of 1) contradiction with guidelines, 2) recommendations not contained in guidelines, and 3) failure to include guideline recommendations. Scoping review was conducted to investigate reviewer-identified discrepancies between CNS recommendations and GPT-4 responses.ResultsA majority of ChatGPT responses were coherent with CNS recommendations. However, moderate contradiction was observed between responses and guidelines (15.3% ChatGPT-3.5 responses, 38.5% ChatGPT-4 responses). Additionally, a tendency toward over-recommendation (30.8% ChatGPT-3.5 responses, 46.1% ChatGPT-4 responses) rather than under-recommendation (15.4% ChatGPT-3.5 responses, 7.7% ChatGPT-4 responses) was observed. Cosine similarity scores revealed moderate similarity between CNS and ChatGPT recommendations (0.553 ChatGPT-3.5, 0.549 ChatGPT-4). Scoping review revealed 19 studies relevant to CNS-ChatGPT substantive contradictions, with mixed support for recommendations contradicting official guidelines.ConclusionsModerate incoherence was observed between ChatGPT responses and CNS guidelines on the diagnosis and management of CIM. The recency of the CNS guidelines and mixed support for contradictory ChatGPT responses highlights a need for further refinement of large language models prior to their application as clinical decision support tools.Copyright © 2024 Elsevier Inc. All rights reserved.

5 keywords...

hide…