Development and validation of a large language model-powered chatbot for neurosurgery : mixed methods study on enhancing perioperative patient education

Abstract

Background: Perioperative education is crucial for optimizing outcomes in neuroendovascular procedures, where inadequate understanding can heighten patient anxiety and hinder care plan adherence. Current education models, reliant on traditional consultations and printed materials, often lack scalability and personalization. Artificial intelligence (AI)–powered chatbots have demonstrated efficacy in various health care contexts; however, their role in neuroendovascular perioperative support remains underexplored. Given the complexity of neuroendovascular procedures and the need for continuous, tailored patient education, AI chatbots have the potential to offer tailored perioperative guidance to improve patient education in this specialty. Objective: We aimed to develop, validate, and assess NeuroBot, an AI-driven system that uses large language models (LLMs) with retrieval-augmented generation to deliver timely, accurate, and evidence-based responses to patient inquiries in neurosurgery, ultimately improving the effectiveness of patient education. Methods: A mixed methods approach was used, consisting of 3 phases. In the first phase, internal validation, we compared the performance of Assistants API, ChatGPT, and Qwen by evaluating their responses to 306 bilingual neuroendovascular-related questions. The accuracy, relevance, and completeness of the responses were evaluated using a Likert scale; statistical analyses included ANOVA and paired t tests. In the second phase, external validation, 10 neurosurgical experts rated the responses generated by NeuroBot using the same evaluation metrics applied in the internal validation phase. The consistency of their ratings was measured using the intraclass correlation coefficient. Finally, in the third phase, a qualitative study was conducted through interviews with 18 health care providers, which helped identify key themes related to the NeuroBot’s usability and perceived benefits. Thematic analysis was performed using NVivo and interrater reliability was confirmed through Cohen κ. Results: The Assistants API outperformed both ChatGPT and Qwen, achieving a mean accuracy score of 5.28 out of 6 (95% CI 5.21-5.35), with a statistically significant result (P<.001). External expert ratings for NeuroBot demonstrated significant improvements, with scores of 5.70 out of 6 (95% CI 5.46-5.94) for accuracy, 5.58 out of 6 (95% CI 5.45-5.94) for relevance, and 2.70 out of 3 (95% CI 2.73-2.97) for completeness. Qualitative insights highlighted NeuroBot’s potential to reduce staff workload, enhance patient education, and deliver evidence-based responses. Conclusions: NeuroBot, leveraging LLMs with the retrieval-augmented generation technique, demonstrates the potential of LLM-based chatbots in perioperative neuroendovascular care, offering scalable and continuous support. By integrating domain-specific knowledge, NeuroBot simplifies communication between professionals and patients while ensuring patients have 24-7 access to reliable, evidence-based information. Further refinement and research will enhance NeuroBot’s ability to foster patient-centered communication, optimize clinical outcomes, and advance AI-driven innovations in health care delivery.