Evaluating Insomnia Queries from an Artificial Intelligence Chatbot for Patient Education.

STUDY OBJECTIVES: We evaluated the accuracy of ChatGPT in addressing insomnia-related queries for patient education and assessed ChatGPT's ability to provide varied responses based on differing prompting scenarios. METHODS: Four identical sets of 20 insomnia-related queries were posed to ChatGPT. Each set differed by the context in which ChatGPT was prompted: no prompt, patient-centered, physician-centered, and with references and statistics. Responses were reviewed by 2 academic sleep surgeons, 1 academic sleep medicine physician, and 2 sleep medicine fellows across 4 domains: clinical accuracy, prompt adherence, referencing, and statistical precision, using a binary grading system. Flesch-Kincaid grade-level scores were calculated to estimate the grade level of the responses, with statistical differences between prompts analyzed via analysis of variance and Tukey's test. Interrater reliability was calculated using Fleiss's kappa. RESULTS: The study revealed significant variations in the Flesch-Kincaid grade-level scores across 4 prompts: unprompted (13.2 +/- 2.2), patient-centered (8.1 +/- 1.9), physician-centered (15.4 +/- 2.8), and with references and statistics (17.3 +/- 2.3, P < .001). Despite poor Fleiss kappa scores, indicating low interrater reliability for clinical accuracy and relevance, all evaluators agreed that the majority of ChatGPT's responses were clinically accurate, with the highest variability on Form 4. The responses were also uniformly relevant to the given prompts (100% agreement). Eighty percent of the references ChatGPT cited were verified as both real and relevant, and only 25% of cited statistics were corroborated within referenced articles. CONCLUSIONS: ChatGPT can be used to generate clinically accurate responses to insomnia-related inquiries.