Principal ResultsThis study evaluated the performance of the Mistral 8x22B LLM in generating key information sections for ICFs in clinical trials. The comparison between LLM-generated content and human-generated ICFs revealed that LLMs demonstrate considerable potential for improving the efficiency, readability, and actionability of ICFs, while maintaining comparable accuracy and completeness across most assessed categories.
Accuracy and CompletenessThe LLM-generated ICFs achieved comparable performance to human-generated content across most areas. Both LLM and human outputs were similar in conveying the purpose of the study. They also performed equally well in describing the duration and procedures. While the human-generated content slightly outperformed the LLM in discussing alternatives, the LLM performed better in explaining the benefits. The overall impression score favored the LLM slightly. These results suggest that while the LLM performs similarly to humans in most areas, further refinement in prompt engineering may be required to improve its performance in more complex sections, such as alternatives. With additional fine-tuning, LLMs could potentially match or exceed the quality of human-generated ICFs across all categories.
ReadabilityThe LLM outperformed human-generated ICFs in readability, as demonstrated by higher RUAKI scores. Both LLM- and human-generated ICFs exhibited good readability according to the Flesch-Kincaid grade level, but the LLM achieved a lower average grade level, reflecting superior readability and closer alignment with institutional requirements for eighth-grade reading level content. This highlights the LLM’s strength in tackling one of the primary challenges of ICF creation: producing documents that are both comprehensive and easily understood by a general audience. Given that many ICFs often exceed the recommended reading level, the LLM’s consistent ability to generate readable content is a significant advantage, ensuring accessibility without sacrificing detail.
Understandability and ActionabilityThe LLM significantly outperformed human-generated ICFs in both understandability and actionability, as reflected in the higher RUAKI scores. The LLM’s output was not only more comprehensible but also consistently included actionable next steps, a critical component that was often missing from the human-generated content. The perfect actionability score of the LLM-generated ICFs suggests that these models can enhance participant comprehension and facilitate informed decision-making. These findings demonstrate the potential of LLMs to create ICFs that are not only easier to read but also more effective in guiding participants through the consent process.
Rater ConsistencyThe ICC of 0.83 indicates a very high level of agreement among raters, reflecting the reliability of the evaluation process. The narrow CI further supports the robustness of these ratings, ensuring the consistency and validity of the results across different protocols.
Lessons LearnedLesson 1: The Importance of Precise Temperature SettingsFor all LLMs, the model temperature parameter controls the diversity of responses. A higher temperature, such as 0.8, produces more varied answers, while a lower temperature, such as 0.2, results in more focused and deterministic outputs. In our experiments, we found that setting the temperature to 0 was the most effective choice for this task. Temperatures introduce a level of randomness that can lead to hallucinations—unwanted deviations from the source material. Since our goal is to extract information directly from the research protocol and treat it as the sole source of truth, it is essential to minimize any creative output from the LLM. Although a temperature of 0.2 is already fairly focused [27,28], the need for absolute accuracy in clinical trial documents led us to set the temperature to 0, ensuring that the content remains strictly aligned with the provided data.
Lesson 2: Addressing Readability Challenges With Cross-Model Few-Shot PromptingIt was observed that Mistral struggled to generate content at a Flesch-Kincaid grade level of 8 when prompted directly with instructions such as, “The content should meet literacy standards, specifically an 8th-grade reading level or lower.” To overcome this challenge, unlike the other zero-shot prompts used in this project, additional prompts and a technique known as “few-shot training” [29] were introduced. This involved providing the model with examples of text at both below 8 and above 9 Flesch-Kincaid grade levels, helping to guide the model in producing content at the desired reading level. These examples, generated using ChatGPT 4, were incorporated (as shown in Supplementary 3 in Multimedia Appendix 1) enabling Mistral to produce content more consistently at the desired grade level 8 or below. This approach, known as cross-model few-shot prompting, involves using one model to generate examples (or “shots”) that are subsequently fed into another model to enhance its performance on a specific task. It is essential to apply this step after any content-editing prompts are used, as further content edits could inadvertently raise the reading level above the target. Format editing should be the final step in the process, as any subsequent prompts could alter the formatting.
Lesson 3: Least-to-Most Prompt EngineeringEffective prompt engineering involves breaking down tasks into manageable steps. When LLMs are given multiple instructions in a single prompt, they often struggle to follow all directions accurately. By adopting a Least-to-Most approach [30]—where each prompt contains a focused set of instructions and builds on the previous output—we achieved more consistent and reliable results. This is analogous to how clinical workflows are built iteratively in electronic health record (EHR) systems to ensure accuracy in decision-making. Much like building clinical templates or order sets, prompt engineering ensures that each phase of content generation is guided to avoid ambiguity, ensuring accuracy and relevance to the context of informed consent. For example, if we were building a medication alert in an EHR, breaking the alert logic into separate steps—from checking allergies to suggesting alternatives—ensures clarity and avoids overwhelming the user. Similarly, breaking down the prompt for generating ICF sections helps the LLM focus on retrieving the right information from the protocol. This method involves 2 stages: first, decomposing a complex problem into a series of simpler subproblems and then sequentially solving these subproblems, with each solution informed by the answers to the previous ones. By guiding the LLM to work incrementally, we not only enhanced its accuracy but also ensured that human oversight remained integral to the process, leading to optimal outcomes and greater control over the final product. For instance, one reviewer noted that the LLM-generated procedures were missing some procedure information. This issue likely stemmed from the lengthy and poorly structured procedure section in the original research protocol. To address this, we designed a workflow in which we first asked the LLM to summarize the study’s procedures and timeline. We then instructed the LLM to extract the necessary information from this summarized output. This step-by-step Least-to-Most approach allowed us to successfully extract the missing information and integrate it into the key information.
Lesson 4: Real-World Application and Integration ChallengesImplementing LLMs in clinical workflows requires more than just improving readability or accuracy—it necessitates seamless integration with existing clinical systems and processes, such as EHRs and IRB workflows. For LLMs to have a real-world impact, models need to be adaptable to diverse clinical environments and meet regulatory and ethical standards. One way to ensure this is by developing interfaces that allow researchers to fine-tune LLM outputs while ensuring compliance with clinical trial guidelines.
Lesson 5: The Need for Detailed Source Material and Human OversightIn one research protocol, a reviewer found that the human-generated ICF was more accurate in detailing risks and discomforts compared with the LLM-generated version. This discrepancy was evident in the omission of common discomforts associated with COVID-19 tests in the LLM-generated ICF, which occurred because this information was not included in the original research protocol. This underscores a key lesson: “garbage in, garbage out.” For LLMs to produce a comprehensive and accurate ICF, the original research protocol must be thorough and detailed. Furthermore, this finding highlights the importance of having a human-in-the-loop to review and refine the output from LLMs. While LLMs can significantly reduce the effort required to create an ICF—potentially saving up to 90% of the work—the final product still benefits from human oversight. For example, after generating a well-structured ICF key information section with the LLM, researchers can easily tweak the content to better suit their specific audience. One reviewer noted that the LLM-generated ICF had a more technical and clinical tone than the human-generated version. By having researchers to customize the LLM-generated content, the ICF can be tailored to the study audience, while most of the heavy lifting has been accomplished by the LLM.
Lesson 6: Ethical and Regulatory Considerations for LLM-Generated ContentWith the increasing role of LLMs in generating participant-facing documents, ethical and regulatory concerns must be addressed. Key considerations include ensuring that AI-generated ICFs do not inadvertently introduce bias or misinformation. Furthermore, as LLMs take on more responsibility in clinical settings, regulatory bodies may need to establish guidelines to govern their use. These guidelines could include stipulations on the necessity of human oversight to verify that LLM-generated content is accurate, participant-friendly, and compliant with ethical standards for informed consent.
Lesson 7: Techniques for Structured and Accurate Information ExtractionBy using effective prompt engineering strategies [31,32] and crafting prompts that were precisely focused on extracting information from the research protocol, we were able to generate a well-structured and neatly formatted output. Key components—such as the introduction, study purpose, procedures, risks and discomforts, benefits, alternatives, cost and compensation, and consent process—were clearly delineated with headers, which improved the organization of the document and made it easier to locate specific information. When crafting the initial prompt to extract key information, we used a combination of techniques:
Delimiter usage: Delimiters like ### and [] were used to clearly define boundaries between different sections of the text.
Role-playing: Assigning the LLM a specific role, such as “As a Clinical Trial Informed Consent Writer,” provided contextual guidance, resulting in improved performance by making the model’s responses more relevant and focused.
Lesson 8: Addressing Practical Benefits and Cost-SavingsBeyond improving the quality of ICFs, LLMs offer the potential to reduce operational costs and administrative burdens in clinical trials. Automating the creation of ICFs can significantly cut down the time spent on document preparation while maintaining compliance with regulatory standards. By quantifying these savings—such as estimating the reduction in hours spent on ICF creation—future studies could further demonstrate the tangible benefits of incorporating LLMs into clinical workflows.
Lesson 9: Anticipating Future AdvancesAs of this publication, GPT-4o mini [33], Mistral Large 2 [34], and Meta Llama 3.1 [35] have been released, each featuring an expanded context window of 128k tokens, making them ideal for these tasks. However, they were not available during the development phase. While 64k tokens are sufficient for handling most clinical research protocols, for more extensive content, these new models would be preferable. That said, the prompts are compatible with all of these models.
Limitations and Future DirectionsThis study’s findings must be interpreted in light of certain limitations. The LLM’s performance is closely tied to the quality of the input research protocol. The LLM’s performance is closely tied to the clarity and completeness of the source material. Ambiguities or inconsistencies in the research protocols can hinder the model’s ability to capture all relevant details to generate accurate and comprehensive ICFs. Future research should focus on improving the clarity of source materials and refining prompt-engineering strategies to optimize LLM performance, particularly in more complex sections such as study procedures.
To address these challenges with procedural details, we used a targeted prompt-engineering approach, which involved having the LLM first summarize the study’s procedures and timeline, and then extract specific details from the summarized text. This method improved accuracy, but ongoing refinement of these strategies is needed to enhance the LLM’s ability to process complex and lengthy sections more effectively.
Another limitation relates to the potential recognizability of LLM-generated text. Although evaluators were blinded to the source of the ICFs and presented with human- and LLM-generated documents in a randomized order, the distinctive textual style of LLM outputs—characterized by active voice, well-organized structure, simplified language, and consistent adherence to readability guidelines—may have inadvertently revealed their origin. This recognizability could introduce subconscious bias into the evaluation process. To address this in future studies, we plan to use text obfuscation techniques, such as paraphrasing or reformatting outputs, to minimize stylistic differences and ensure true blinding. This approach will help strengthen the validity of future comparisons.
Another important limitation of this study is the small sample size, which consisted of only 4 clinical trial protocols. While these protocols provided a useful test bed, the relatively small sample size limits the generalizability of the findings. Future studies should incorporate a broader range of clinical trials from diverse therapeutic areas, phases, and levels of complexity to fully validate the model’s performance. Furthermore, the limited sample size may have contributed to some statistically nonsignificant results, such as those related to procedural details or study alternatives. A larger sample would provide greater statistical power, enabling the detection of smaller but practically significant differences between human- and LLM-generated ICFs.
A larger sample would also better capture the diversity of challenges involved in ICF creation, such as variations in regulatory requirements, medical procedure complexity, and considerations for vulnerable populations. Future research should aim to assess the LLM’s robustness and adaptability across a wider array of clinical trial contexts.
The process of designing and refining prompts, as well as generating the AI-generated ICFs, required a moderate time investment during initial development. However, leveraging the existing prompts and lessons learned in this study would enable future users to complete the process more efficiently and at reduced cost, enhancing the scalability of this approach.
The 0% actionability score for human-generated ICFs reflects a structural issue rather than a methodological flaw. Most institutions do not include explicit actionable sections in the key information portions of their templates. Updating these templates to include actionable instructions would likely improve scores significantly. This highlights a strength of LLM-generated ICFs, which inherently include actionable elements, enhancing the clarity and use of consent forms.
While the LLM demonstrated strong performance, there were instances where it missed details related to the duration and procedural elements of the study. This likely stems from 2 primary challenges: ambiguous or inconsistent presentation of these sections in the research protocols and the verbosity of the text, which can hinder the LLM’s ability to process details efficiently. Our targeted approach of summarizing and then extracting procedural details helped address this issue, but further enhancements are needed to ensure that the LLM can consistently handle such challenges.
Looking ahead, our next goal is to automate the creation of entire ICFs directly from research protocols. This would significantly reduce the time and effort required for ICF development while maintaining consistency and quality. While this study highlights the potential for human oversight to address issues in LLM-generated content, our future research will aim to quantify the time and effort required for revisions to better assess the practical efficiency gains of integrating these models into clinical workflows. We also plan to explore the LLM’s multilingual capabilities to generate ICFs in multiple languages, broadening the recruitment base and promoting diversity and equity in clinical trials. Ensuring that non–English-speaking participants receive ICFs that are as readable and understandable as those in English is crucial for improving inclusivity and representation in research. Future work should also prioritize a thorough examination of ethical concerns, including potential biases in AI-generated content, the need for transparency in AI decision-making processes, and the legal implications of deploying LLMs in clinical trial workflows, to ensure that these tools are implemented responsibly and equitably.