Evaluation of Five Large Language Models for Parental Education in Pediatric Anesthesia: Reliability and Readability...

Although large language models (LLMs) show potential for patient education, their accuracy, usability, and comprehensibility lack validation in high-risk pediatric anesthesia. Rigorous evaluation is therefore essential prior to widespread clinical use in perioperative parental anesthesia education.

This study aims to evaluate the accuracy, reliability, and readability of responses generated by 5 LLMs to parental inquiries regarding pediatric anesthesia, and to assess their suitability for clinical use in perioperative caregiver education.

Two expert anesthesiologists identified 33 parental questions on pediatric anesthesia by screening authoritative resources and Google Trends. On December 14, 2025, these questions were submitted to 5 LLMs (DeepSeek-V3.2, ChatGPT-5, Gemini 2.5 Flash, Copilot, and Perplexity) via official web interfaces with default settings and zero-shot prompting, with each query in a separate conversation. Responses were standardized for blinded assessment. Two pediatric anesthesiologists with ≥10 years of clinical experience independently evaluated accuracy and reliability using the 4-point Likert accuracy scale, DISCERN, Ensuring Quality Information for Patients (EQIP), Journal of the American Medical Association (JAMA) benchmark, and Global Quality Score (GQS). After text preprocessing, readability was evaluated using 6 algorithms (Automated Readability Index [ARI], Flesch Reading Ease Score [FRES], Gunning Fog Index [GFI], Flesch-Kincaid Grade Level [FKGL], Coleman-Liau Index [CL], and the Simple Measure of Gobbledygook [SMOG]) via an online calculator. Interrater reliability was analyzed using the intraclass correlation coefficient (ICC); differences across models were assessed with the Kruskal-Wallis H test; and deviations from the sixth-grade benchmark were evaluated using 1-sample Wilcoxon signed-rank tests (P<.05 considered significant).

All 5 LLMs demonstrated high clinical accuracy (>90%; P=.12), with Gemini reaching 100%. Nevertheless, safety risks and content hallucinations were still observed. Excluding Gemini and Copilot, the remaining 3 models (ChatGPT, DeepSeek, and Perplexity) each produced unsafe content in 3.03% (n=1) of the 33 queries. Hallucinations were detected in all models except Gemini, with DeepSeek and Perplexity showing the highest hallucination rate (3/33, 9.09%). Furthermore, Perplexity showed superior reliability on DISCERN (median 41; P<.05), yet no model achieved a "good" rating. Gemini achieved the highest EQIP (median 66.67%; P<.05) despite lower GQS (median 3). Transparency was universally poor (JAMA median ≤1), with DeepSeek and ChatGPT showing a "floor effect." ChatGPT had superior readability, but all models exceeded the recommended 6-grade complexity level.

In this study, 5 LLMs generally provided clinically accurate information when responding to parental questions about pediatric anesthesia. However, limitations were also identified, including hallucinated content, safety-related deficiencies, limited source transparency, and readability levels exceeding recommended standards. Therefore, LLM-generated information should be interpreted with caution and should not replace clinician guidance.