Leveraging AI/Machine Learning as an End-to-End Differentiator

Rajneesh Patil, Vice President, Digital Innovation, IQVIA

Raja Shankar, Vice President, Machine Learning, Research and Development Solutions, IQVIA

Research and development industry stakeholders are recognizing the value of integrating artificial intelligence (AI)/machine learning(ML)-driven solutions and related expertise into clinical trial strategies, especially when companies have access to trusted data sources. Clinical trial sponsors are already seeing how AI/ML helps sift through and aggregate information to extract meaningful patterns for analysis that guides strategic decisions.

But as healthcare-grade AI methodologies and models are constantly fine-tuned and validated to meet the demands of clinical R&D, there is value in reflecting on what we’re learning as use cases grow about how and where these tech-enabled solutions are helping to address long-established roadblocks to trial success.

Well-designed AI/ML in the hands of experienced AI/ML scientists and engineers, clinical experts, medical reviewers, etc., can generate meaningful insights to guide better trial design, planning and management for better trial execution across all processes in planning, conduct, and close-out.

Ahead, we discuss several noteworthy examples.

Refining trial protocol generation and analysis

According to a September/October 2024 Tufts Center for the Study of Drug Development IMPACT report, participation burden has increased 54% in late-stage (Phase II and III) non-oncology protocols since 2011. This coincides with the increase in trial design complexities, such as more procedures during site visits and strict eligibility criteria. From 2005 to 2020, the number of required study procedures increased by 139%, and the number of endpoints increased by 214% in Phase III trials.

As a consequence of increasing trial complexity, we see an increased and intentional focus on assessing patient burden and incorporating insights earlier in the protocol and trial design process to improve engagement and to help reduce potential protocol amendments, which remains one of the biggest pain points in drug development time, cost, and efficiency.

Though sponsors have typically preferred to develop the first drafts of protocols, now, we can use large language models to help with this scientific process, grounding the initial version with data-driven evidence. Though this kind of analysis has previously been manual, sponsors and contract research organizations (CROs) can rely on generative AI (GenAI) models to generate a first draft for analysis. GenAI models allow sponsors and CROs to mine an extensive breadth of historical data about similar therapeutic indications and protocols to advance design development and to iterate on design options that are most likely to be successfully executed.

By layering advanced analytics of other key datasets (e.g., scientific literature) or other tech-driven enablers, sponsors can gain insights to optimize their protocol designs, such as:

• What are the patient and clinical site burdens caused by the protocol compared to other burdens?
• How feasible are the trial’s patient recruitment targets and project milestones?
• What are some of the potential regulatory challenges and issues based on country and site targets?

For example, study teams can calculate and quantify patient burden using qualitative data from patient surveys, patient advocacy groups, etc., to gauge motivation to participate by race, ethnicity, and other socioeconomic characteristics, and how this is affected by different trial design elements. Assessing burden scores per race and ethnicity within the protocol assessment offers more reliable predictive outcomes on patient satisfaction and recruitment success to improve diversity and inclusion by design, reach better patient subpopulations, and include participant groups that are traditionally underserved.

Trial planning and feasibility to mitigate operationalization risk

Historically, it has been difficult to gauge trial feasibility outside of reviewing similar trial programs that had been completed. For newer classes of treatments and other unique therapies, there may not be previous study programs to reference. With novel AI approaches, we are able to draw analogies from a larger body of examples, where each example may not be similar in all respects, but across many examples, we can capture the impact of different relevant variables on trial feasibility.

AI-driven methodologies can be used to generate insights to mitigate potential risk prior to trial operationalization, thereby reducing delays in downstream activities.

For site identification activities during trial planning, it is critical to pinpoint sites with the potential for strong performance based on the protocol at hand. AI models aid in site identification by using real-world data assets to gauge patient density near the site, patient proximity to the site, disease prevalence in the area, and previous site performance, among other variables. This helps sponsors and CRO teams to identify and select sites with a higher probability of enrolling target patient populations, thereby reducing non-enrolling sites, which are cost burdens for sponsors. AI models are improving first-patient-in rates and recruitment rates at country and site levels.

Trial management through improved protocol adherence

As noted, protocol complexity is increasing, and it tops the list of site challenges. Being equipped to review, understand, and adhere to these complicated study requirements can be difficult for sites already dealing with staff and resource shortages.

Dedicating resources to enhancing protocol design can only be beneficial if site teams working directly with patients are able to effectively carry out trial processes efficiently. So, how do sponsors and CROs help ensure sites have the necessary tools to navigate trial protocols, especially as scientific advances will continue to create more complex trial designs?

Following their recent exploration of GenAI assistants for protocol review, site monitors and principal investigators (PIs) have reported that GenAI-driven protocol Q&A chatbot assistants are helping site staff work through steps within protocols. This style of interface has to be developed and designed with the end-user in mind, accounting for what types of answers and guidance they may need and how to provide responses that are simple and consistent.

Unlike most generic GPT-based models for Q&A chatbots, scientifically-oriented chatbot interfaces can be trained to provide responses to protocol inquiries that support key operational decisions responsibly. This is done by specifying the datasets accessed within the model to ensure historical protocol data and other relevant information is curated to avoid hallucinations or responses that include false or misleading information.

By training models with specific scientific datasets, this type of solution also allows:

• Consistency in responses, regardless of how the protocol-related question is posed by different users. Though individuals will not ask questions the exact same way or with the same words, variability will not offset the consistency in scientific response for the document at hand.
• Chain-of-thought prompting is another methodical approach to continue user guidance after the initial interaction is complete. Users will often have follow-up questions, and helping continue that line of thinking is crucial for a true decision-support model.

Designing AI-driven solutions for protocol support requires having a deep understanding of what questions clinical research associates and PIs may need addressed from end-to-end and incorporating a question library within the execution model to provide real-time, relevant support.

The evolving journey of AI/ML use in clinical R&D

We have only explored a few limited ways AI/ML methodologies can impact clinical trials from design to planning and execution. As we build upon our knowledge, the industry will continue to expand upon the ways diverse AI/ML methodologies and tools can work together to dive deeper into the longstanding unknowns that can impact a trial’s success. With layers of data and information patterns to examine, we know there will be more questions to answer. This requires collaborating as an industry, from trial sponsors to CROs, regulators, and tech enablers, to ensure healthcare-grade AI integrated into clinical trials is designed with agreed-upon guardrails.

When well-designed and deployed within appropriate implementation frameworks, the levels of insight and automation that AI/ML and GenAI can offer will only help drive more informed decisions for an accelerated and efficient drug development process while serving clinical trial participants, trial sponsors, and sites more effectively.

Rajneesh Patil is Vice President, Digital Innovation, IQVIA, and Raja Shankar is Vice President, Machine Learning, Research and Development Solutions, IQVIA