Giudici, Fabiola

Objective To examine the quality of information provided by artificial intelligence platforms ChatGPT-4 and Claude 2 surrounding the management of vestibular schwannomas. Study design Cross-sectional. Setting Skull base surgeons were involved from different centers and countries. Intervention Thirty-six questions regarding vestibular schwannoma management were tested. Artificial intelligence responses were subsequently evaluated by 19 lateral skull base surgeons using the Quality Assessment of Medical Artificial Intelligence (QAMAI) questionnaire, assessing “Accuracy,” “Clarity,” “Relevance,” “Completeness,” “Sources,” and “Usefulness.” Main Outcome Measure The scores of the answers from both chatbots were collected and analyzed using the Student t test. Analysis of responses grouped by stakeholders was performed with McNemar test. Stuart-Maxwell test was used to compare reading level among chatbots. Intraclass correlation coefficient was calculated. Results ChatGPT-4 demonstrated significantly improved quality over Claude 2 in 14 of 36 (38.9%) questions, whereas higher-quality scores for Claude 2 were only observed in 2 (5.6%) answers. Chatbots exhibited variation across the dimensions of “Accuracy,” “Clarity,” “Completeness,” “Relevance,” and “Usefulness,” with ChatGPT-4 demonstrating a statistically significant superior performance. However, no statistically significant difference was found in the assessment of “Sources.” Additionally, ChatGPT-4 provided information at a significant lower reading grade level. Conclusions Artificial intelligence platforms failed to consistently provide accurate information surrounding the management of vestibular schwannoma, although ChatGPT-4 achieved significantly higher scores in most analyzed parameters. These findings demonstrate the potential for significant misinformation for patients seeking information through these platforms.

AbstractObjectivesAlthough the reliability of frozen sections for the intraoperative assessment of complete tumour excision has been established, the best location for collection and the impact of the type of sampling are still debated. We retrospectively investigated the reliability of frozen sections when collected from the surgical bed as tissue strips representative of the whole superficial margin and as a bowl of tissue underlying the resection site for deep margin, and the possibility of relying on frozen section negativity to consider resections complete.Materials and MethodsFrozen section reliability was calculated by comparing histology before and after formalin embedding and then categorised by sampling type, in 182 patients undergoing transoral resection of oral cancer.ResultsComparing frozen and permanent histology, sensitivity, specificity and accuracy were 69%, 98% and 96%, respectively; categorisation by sampling type failed to produce statistically significant differences. Based on frozen section negativity after formalin embedding, complete resections were obtained in 91.7% of patients with multiple‐strip and bowl frozen sections.ConclusionFrozen sections collected as tissue strips and bowl are as reliable as point sampling in the intraoperative guidance of surgical resections. They effectively provide for margin enlargement, thereby increasing the surgeon's confidence that negative margins are clear.