Scholarlink

Background <p>Large language models (LLMs) are increasingly used by the public to seek health information, including explanations of medical treatments and clinical trials.</p> Objective <p>The objective of this study was to evaluate the completeness of LLM-generated informed consent information for clinical trials involving adolescent and young adult (AYA) patients with central nervous system (CNS) tumors, using U.S. Food and Drug Administration (FDA)–aligned consent requirements as the benchmark.</p> Methods <p>We conducted a structured audit of five publicly available LLMs, ChatGPT 5.2, Claude Sonnet 4.5, Claude Haiku 4.5, Gemini (Fast), and Perplexity. Six standardized prompts were developed based on FDA informed consent requirements (21 CFR §&#xa0;50.25), corresponding to key consent domains. Prompts were applied to five CNS tumor diagnoses prevalent in the AYA population. Responses were scored using an FDA-aligned checklist, with items coded as present or absent/unclear and summed to produce a consent completeness score.</p> Results <p>The sample included 25 LLM-generated informed consent responses evenly distributed across five cancer diagnoses. The median consent completeness score was 33 (interquartile range: 31–34), with scores ranging from 22 to 35, indicating variability in disclosure completeness. Completeness scores overlapped across cancer diagnoses, suggesting that variability was driven by response-level heterogeneity rather than clinical context. Domain-level completeness was generally high but showed modest variability across consent domains. Median scores were highest and most consistent for eligibility and procedures domains (median 6, IQR 6–6), while greater variability was observed in domains related to study purpose and practical details (median 5, IQR 4–5 and 5–6, respectively), contributing to overall heterogeneity in completeness.</p> Conclusions <p>Publicly available LLMs variably and inconsistently include essential informed consent elements when responding to standardized clinical trial prompts. These findings suggest areas for further consideration regarding how general-purpose LLM-based tools are used in consent-related health communication.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating the completeness of large language model-generated clinical trial informed consent information for adolescent and young adult patients with central nervous system tumors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月26日

查看全文
  • Corey H. Basch,
  • Erin T. Jacques,
  • Erela Datuowei,
  • Griselda Chapa

摘要

Background

Large language models (LLMs) are increasingly used by the public to seek health information, including explanations of medical treatments and clinical trials.

Objective

The objective of this study was to evaluate the completeness of LLM-generated informed consent information for clinical trials involving adolescent and young adult (AYA) patients with central nervous system (CNS) tumors, using U.S. Food and Drug Administration (FDA)–aligned consent requirements as the benchmark.

Methods

We conducted a structured audit of five publicly available LLMs, ChatGPT 5.2, Claude Sonnet 4.5, Claude Haiku 4.5, Gemini (Fast), and Perplexity. Six standardized prompts were developed based on FDA informed consent requirements (21 CFR § 50.25), corresponding to key consent domains. Prompts were applied to five CNS tumor diagnoses prevalent in the AYA population. Responses were scored using an FDA-aligned checklist, with items coded as present or absent/unclear and summed to produce a consent completeness score.

Results

The sample included 25 LLM-generated informed consent responses evenly distributed across five cancer diagnoses. The median consent completeness score was 33 (interquartile range: 31–34), with scores ranging from 22 to 35, indicating variability in disclosure completeness. Completeness scores overlapped across cancer diagnoses, suggesting that variability was driven by response-level heterogeneity rather than clinical context. Domain-level completeness was generally high but showed modest variability across consent domains. Median scores were highest and most consistent for eligibility and procedures domains (median 6, IQR 6–6), while greater variability was observed in domains related to study purpose and practical details (median 5, IQR 4–5 and 5–6, respectively), contributing to overall heterogeneity in completeness.

Conclusions

Publicly available LLMs variably and inconsistently include essential informed consent elements when responding to standardized clinical trial prompts. These findings suggest areas for further consideration regarding how general-purpose LLM-based tools are used in consent-related health communication.