INTRODUCTION
As one of the most critical ethical constructs, informed consent is a prerequisite for all research involving human subjects, respecting autonomy, mitigating potential harm, and upholding research integrity. 1 In the present, clinical trials often involve intricate procedures and novel therapies, which often prioritize legal compliance over participant understanding. Complex informed consent forms can hinder participant comprehension, leading individuals to agree to participate without fully grasping the implications, which undermines the trial’s ethical validity and raises concerns about how consent is currently obtained. 2–5 Therefore, to achieve truly informed consent, researchers must provide clear and accessible information tailored to participants’ understanding and provide patient information leaflets and informed consent forms (ICFs), which are essential components of the informed consent process, adequate to guarantee that participants can comprehend and make free decisions about their involvement.
To address this challenge, researchers are exploring innovative ways to improve the clarity and accessibility of informed consent materials. Large Language Models (LLMs), with their ability to process and generate human-like text, show promise in creating more effective informed consent documents. 6 This study aims to answer the question: “How can we utilize LLMs to assist in developing participant information sheets (PISs) and ICFs for clinical trials?”
LLMs can simplify medical terminology, tailor information to individual needs, and identify potential ambiguities, ultimately promoting greater participant understanding. Results from a recent study indicate that LLMs can effectively rephrase complex medical jargon into more accessible language, enhancing readability for patients and research participants, particularly those with lower health literacy. 7,8 Additionally, LLMs can identify and address potential ambiguities, ensuring clarity and reducing the risk of misunderstanding. Finally, they can automate the analysis of existing ICFs, identifying areas for improvement and ensuring compliance with ethical and regulatory standards. 9–11
Furthermore, recent research demonstrates that digital tools improved participants’/patients’ understanding of and satisfaction with the IC process. Digital tools, particularly interactive multimedia tools, may help develop more personalized IC processes tailored to an individual’s socio-cultural characteristics. 12 A systematic review 13 showed that compared with patients using paper-based consenting, patients using eConsent better understood clinical trial information, showed greater engagement with the content, and rated the consenting process as more acceptable and usable.
Even if these tools increase the participants’ comprehension and decision-making capacity, contrary to LLMs, all these strategies are time- and resource-consuming. 14 Our work is innovative because it proposes an automated solution to generate PIS and ICF, potentially saving time and resources while maintaining high quality and ethics compliance. Regulatory challenges in using eConsent would also apply to, and be even more complex, LLM-assisted consent strategies for the preparation of documents. 15 Despite efforts to harmonize the rules on data protection and clinical trials in the EU, the legal acceptance of eConsent differs significantly among the Member States. LLMs may encounter even more hurdles regarding clinical trial regulations and the one-size-fits-all requirements for PISs and ICFs. These models also raise concerns about misinformation, biases in the training data, and the potential for discrimination (like ageism) in their application. Caution is necessary due to inherent imprecision and a propensity for disseminating misinformation. 16
This study will investigate the potential of LLMs to assist in creating and analyzing PISs and ICFs within the REPO4EU project (Precision drug REPurpOsing For EUrope and the world). We will use REPO4EU clinical trials to test the feasibility and effectiveness of LLMs for generating and analyzing informed consent documents.
METHODOLOGY
Participant Information Sheets and Informed Consent Forms
Approach
Standardized tools for informed consent ensure transparency and enable autonomous participant decision-making. These templates aim to ensure clear and consistent communication with potential research participants, providing them with all the necessary information to make informed decisions about their involvement. This task aimed to develop PISs and ICFs templates specific to the REPO4EU project. To prepare information sheets and informed consent templates for one of the trials under the REPO4EU project (REPO-HYPER II), we conducted a multistep methodological approach and two pilot exercises:
Bibliographic Review: We comprehensively searched scientific literature to identify the essential elements of PISs for clinical trials. This search included peer-reviewed articles and relevant guidelines. Additionally, “gray literature” sources, such as reports and informed consent templates from regulatory bodies and other institutions, were explored.
Checklist Development: The bibliographic review informed the creation of a preliminary checklist table to capture the key information required for both documents. We then validated this initial checklist by analyzing 15 PISs and ICFs from ongoing clinical trials registered in the EU Clinical Trials database. The analysis focused on how these trials met the information needs of potential participants. The initial checklist was subsequently modified based on the analysis of existing clinical trials.
Template Design: After ensuring the text included information on all the previously checked requirements, we considered the recommendations outlined in Coleman et al. 17 to improve readability and, consequently, enhance the participants’ decision-making capacity. We used the PIS and ICF templates from the Swedish regulatory agency responsible for approving the REPO4EU trial as the main template. We modified this template using other templates in the corpus of documents used in this study and the requirements from our checklist.
We performed two pilot exercises to prepare PISs and consent templates specific to REPO4EU. These exercises include the PISs and ICFs template for the REPO4EU HYPER II trial. The templates provided in this report are provisional and require further refinement in close collaboration with the relevant partners and in accordance with any specific requirements by local/National Ethics Committees or Institutional Review Boards. However, manually creating PISs and ICFs that meet these requirements can be time-consuming 14 and prone to errors, prompting the exploration of automated solutions based on LLMs.
We amended the templates following the recommendations outlined in the study by Coleman et al., 17 summarized in Table 1 .
Recommendations for Participant Information/Informed Consent Form
| Category | Recommendation |
|---|---|
| Use leaflet format | |
| Use line spacing (1.2–1.5) | |
| Layout | If appropriate to support the main message, use simple images or illustrations |
| Use text boxes if highlight is needed | |
| Align text left | |
| Type size 12 | |
| Use a sans serif font (e.g., Arial, Verdana, Tahoma) | |
| “All Capitals” should be avoided | |
| Formatting | Avoid underlining |
| Avoid italics | |
| Use clear headings | |
| Use clear contrast between text and background | |
| Avoid long sentences | |
| Use short paragraphs | |
| Use of questions in headings | |
| Use bullet points or numbered lists instead of long sentences | |
| Language | Minimize technical language or jargon |
| Specify numbers, avoiding the use words like “multiple” | |
| Use words for numbers 0–9; for 10+ use the digit | |
| Use whole numbers (avoiding percentage) for risk or benefits | |
| Avoid subordinate clauses |
The table is adapted from the study by Coleman et al. 17 .
Large Language Models
As explained in the previous section, drafting PISs and ICFs for clinical trials is complex and time-consuming, demanding meticulous attention to detail. LLMs can potentially solve this by automating tedious and error-prone tasks. We aim to leverage LLMs to create a tool that automatically generates PISs and ICFs for various clinical trials, adhering to the templates and checklists designed and described in the previous section.
LLMs are deep learning models trained on significant quantities of data to learn and generate text that mimics the natural language and other forms of content. Regarding text, during their training stage, the models learn to predict a sentence’s following word by considering the sentence’s previous words and assigning a probability score to the overall frequency of words. Once trained, LLMs can generate text by predicting the sequence of words considering the input given as context and the knowledge they received during training. 18 A well-known approach that facilitates the application of LLMs is Langchain. 19,20 Existing work shows LLMs’ excellent performance in a plethora of tasks, 21 including assisting in patient communication and simplifying documentation tasks. 16
Because they are trained with general information from multiple sources and, therefore, lack domain-specific knowledge. 22 LLMs alone can be inconsistent and inaccurate for specific applications. A Retrieval Augmented Generation (RAG) framework 23 addresses this limitation by retrieving relevant documents to provide context for the LLMs. RAG systems consist of two major components: the retriever, which gathers relevant documents or other information based on the provided input, and the generator, which produces suitable answers according to the retrieved information. Studies have shown that retrieval augmentation effectively helps LLMs to surpass knowledge boundaries when supplementary context is required and enhances LLMs’ capacity to answer questions. 24 By incorporating questions from a checklist, the RAG framework identifies relevant documents, which the generator can use to produce a more reliable and accurate final output.
The RAG system forms the second major component of our methodology. It significantly enhances the accuracy and compliance of generated documents. This system maintains a knowledge base of regulatory documents using vector storage and embeddings, allowing for real-time verification of compliance requirements. The RAG implementation ensures that all generated content aligns with current regulatory standards while maintaining context-specific accuracy.
In generating the PISs and the ICFs with the LLM, we plan to use a prompt template to ensure all crucial elements of a clinical trial—trial purpose, procedures, and risk factors—are consistently included. The templates maintain a balance between regulatory compliance and readability, addressing one of the critical challenges identified in the current process.
Figure 1 summarizes our proposed methodology. We will use the tables with the required PIS and ICF information as checklists of questions. Each question is a query that serves as a prompt for the LLM supplemented with the context in the knowledge database. This additional knowledge enhances the accuracy of generated text. Therefore, with the prompt and retrieved documents, the LLM may generate a new PIS and ICF and a completed checklist.
Proposed methodology for creating PISs (participant information sheets) and ICFs (informed consent forms) using LLMs (Large Language Models).
To evaluate the results of our RAG-based LLM application, we need to address two main components: the retriever and the generator of the RAG framework. 25 Hence, we will assess metrics such as the following:
Contextual relevance, which evaluates if the information in the retrieval context is relevant given a specific input;
Contextual precision, which evaluates if the retriever properly ranks relevant information;
Contextual recall, which evaluates if the retrieved information complies with the expected output;
Faithfulness, which evaluates if the generator’s output complies with the information presented in the retrieval context;
Answer relevance evaluates if the generator’s output is relevant given a particular context.
Since statistical-based scorers are inaccurate because they fail to consider semantics, we plan to resort to scorers that are model-based or both statistical and model-based for better results. 26 We will also consider user feedback to corroborate our result documents in real-life clinical trials.
Our methodology also incorporates a robust memory component that maintains consistency across documents and versions. This system tracks document history, maintains audit trails, and ensures version control, which are crucial for regulatory compliance. Additionally, the memory component helps maintain consistency in terminology and explanations across different sections of the documents.
The generated documents may only sometimes meet the desired quality. As such, quality control validates multiple aspects of the generated documents. This process includes automated assessment of readability levels, regulatory compliance verification, and consistency checking. The system uses a feedback loop to continuously improve document quality based on expert review and user feedback. This iterative improvement process helps maintain high standards while adapting to specific trial requirements.