Truace
Subscribe

Large Language Model-Assisted Review of Clinical Trial Informed Consent for Data Reuse: Retrieval-Augmented Generation Approach

Background: Regulatory frameworks such as the Belmont Report, the Common Rule, and the Declaration of Helsinki require informed consent to ensure participants understand a study's purpose and can make voluntary decisions about their involvement. Regulations, including the General Data Protection Regulation (Regulation [EU] 2016/679), further emphasize that consent must be freely given and revocable without disadvantage. Although informed consent forms (ICFs) are intended to be clear and accessible, they have bec…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1539· cite & verify
Large Language Model-Assisted Review of Clinical Trial Informed Consent for Data Reuse: Retrieval-Augmented Generation Approach
"Clinical Trial Volunteer Signs Informed Consent Form (45116548901)" by NIAID is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

In brief

Researchers tested whether GPT-4.1 combined with retrieval-augmented generation could review how clinical trial informed consent forms describe reuse of participant data. They first processed 438 forms from different trials, countries, languages, and versions, comparing the model's outputs with human expert ground truth.

The work matters because consent forms have grown lengthy and complex, and manual review is hard to keep consistent. The authors conclude the workflow can support review and improve consistency, though the abstract frames this as support for the process, and accuracy below 100% means errors remain in a high-stakes setting.

Main points

  1. The initial evaluation covered 438 consent forms spanning multiple countries, languages, and form versions, and reached 81.6% accuracy.
  2. A separate validation set of 488 forms began with a baseline accuracy of 84.0% on 89 documents.
  3. Three optimization rounds on 14, 281, and 104 documents produced accuracy of 82.3%, 90.7%, and 92.1%.
  4. Evaluation also assessed the quality of supporting evidence the model provided, plus the time needed to extract information per form.

The gain

A retrieval-augmented GPT-4.1 workflow extracted data reuse information from clinical trial consent forms across languages, reaching 92.1% accuracy on a holdout set after iterative prompt optimization.

The rundown

The study was motivated by regulatory frameworks such as the Belmont Report, the Common Rule, the Declaration of Helsinki, and the General Data Protection Regulation, which require that consent be informed, voluntary, and revocable. Consent forms have become increasingly long and complex despite being meant to be accessible.

Retrieval-augmented generation was used because large language models are prone to hallucinations, which limit reliability in high-stakes settings. Expertly curated prompts guided extraction of data reuse information, and outputs were compared with expert-derived ground truth.

In the validation phase, the final round was run on an independent holdout cohort. The authors conclude the approach supports consent form review and interprets complex consent language, contributing to more consistent evaluations of data reuse.

Sources

  1. Peer-reviewedJMIR Medical Informatics2026-10-15

The debate