AI and Clinical Trials

From Scientific to Plain Language

People often underestimate the efforts and skillset needed to develop high-quality and balanced clinical results summaries in plain language. It’s a multidisciplinary task requiring language, communication, and visual design skills, as well as expertise within:

• Clinical trial research methodology
• Regulatory requirements
• Biomedical statistics
• Health literacy principles

Notably, the summation of clinical research results is prone to risks, including:

• Unintentional bias when selecting protocol endpoints for a short summary
• Inaccurate presentation of trial results, which threatens scientific integrity
• Undesirable changes in style or language tone, e.g., introduction of promotional language or unsubstantiated performance claims
• Retaining language style, terminology, and accuracy during translation of the summary
• Failure to consider health literacy and numeracy principles to ensure readability of summaries to non-scientific audiences

Given the risks of summarizing and publicly disclosing clinical research results, many trial sponsors apprehensively deploy AI in developing their PLSs. Defining a use case for AI deployment and the right balance between AI and human involvement may be challenging for sponsors lacking necessary multidisciplinary expertise.

What are the Potential Benefits of AI in Clinical Trials?

Under a proper use case, AI can significantly reduce the burden of summarizing and drafting plain language summaries. The involvement of AI and clinical trials doesn’t threaten to replace human resources. According to the EU AI Act, deployers of AI systems must assign human oversight to individuals with the required competence, training, and authority. Humans must define context and validate AI output based on:

• Unique knowledge within clinical research
• Cultural nuances
• PLS tone

Additionally, overreliance on AI and clinical trials is risky, given the potential damage or distrust an inaccurate summary could cause.

Considering the PLS communication’s sensitivity and linguistic challenges, Lionbridge recommends using AI to automate the initial PLS draft stage. Then, AI can be used to optimize during a summary’s other stages:

• Revision
• Refinement
• Translation

Sponsors can generate a first-draft PLS with automation using prompt engineering excellence. A Large Language Model (LLM) can be instructed to extract and summarize pre-defined information from the Clinical Trial Report and other relevant reference documents, such as the Clinical Trial Protocol, the Informed Consent Form or Tables, Listings, and Graphs. A bad prompt will generate a low-quality response, thus increasing the PLS writing burden. To avoid this burden, engineering experts should design prompt engineering to prevent:

• LLM hallucination
• Context window limitations
• Presentation of erroneous information

Avoid Bias, Omission, or Distortion During PLS Development

Bias is a known risk within research summaries, and trial sponsors must mitigate it when developing plain language summaries. Notably, it’s essential in the selection of endpoints to include in the summary. A phase 3 or 4 protocol may incorporate primary, secondary, and tertiary endpoints. The sponsor may determine that some are patient-relevant and should be included in the PLS. The Good Lay Summary Practice (GLSP) guidance recommends sponsors pre-define which endpoints to include according to an established documented framework for endpoint selection. This will help avoid bias risk in endpoint selection for the PLS. Trial sponsors can build such pre-selected endpoints into a PLS prompt template, thus ensuring the LLM presents and extracts correct results.

Bias, omission, or misrepresentation may also arise in the PLS development process when the LLM is instructed to summarize common adverse events or in- and exclusion criteria for patient clinical trial recruitment. Without advanced prompting and human oversight, the LLM may omit adverse events of interest to patients or even misinterpret safety data due to lack of context. Or it may miss key eligibility criteria and summarize less relevant criteria in the trial context. The trial team and medical writers should ensure the PLS:

• Addresses patient interests
• Presents favorable and unfavorable results
• Aligns with scientific nuances

Prepare, Refine, and Scale Across Clinical Development Programs

Skilled prompt engineers can prepare and refine prompt templates, automate prompt recycling, and execute post-processing prompts for an optimal workflow. Putting in the work upfront to develop and test a master prompt template enables sponsors to improve the automated PLS draft. Depending on the trial sponsor’s methodology, a PLS may need several prompt iterations. A prompt template can also be reused and configured from one study to another in a clinical development program. As AI evolves and the corpus improves across natural languages, the entire process, from drafting a PLS in source language to translating it into target languages, may be further streamlined.

An LLM is trained on an extensive corpus of datasets, which means fine-tuning and prompt engineering excellence is necessary to generate the desired PLS output. Sponsors should remember that people generate prompts, and may introduce bias that LLMs can amplify or reproduce. Such bias derived from AI and clinical trials (intentional or not) can lead to skewed or unbalanced result summaries.

Get in touch

Responsible AI usage is transparent AI usage. According to the EU AI ACT, using AI to generate or manipulate text for public access must be disclosed by the entity deploying the AI and clinical trials. Where and how trial sponsors apply LLMs in PLS processes should be clear.

Lionbridge has developed a TRUST framework to build confidence in our use of AI and ensure AI trust across the research community, customers, and the public. For more information on AI and clinical trials, get more insights in our eBook: AI and Language Strategy in Life Sciences. Let’s get in touch.