AI in clinical trial design can help research teams examine feasibility assumptions, identify suitable patient populations, and use historical data to inform study planning. In this episode of QCast, co-hosts Jullia and Tom discuss where AI and machine learning can contribute to these decisions, from assessing eligibility criteria to exploring treatment effects in different patient subgroups. They also consider how AI can support data review during trial conduct and where conventional statistical methods may be sufficient.
A central challenge is determining whether an AI-generated prediction is relevant to the study being planned. Historical datasets may help identify recruitment barriers or potential protocol amendments, but their usefulness depends on how closely they represent the intended patient population. The discussion also addresses the need for pre-specified rules in adaptive trials, the risks of relying on exploratory findings, and how sponsors can evaluate AI tools against practical study outcomes.
AI can analyse historical trial data to identify restrictive eligibility criteria, recruitment challenges, and potential patient burden. These findings can help teams review design assumptions before finalising a protocol and consider whether adjustments are scientifically appropriate.
Predictive modelling can help identify potential participants using electronic health records, biomarkers, and genomic information. Retrospective analysis can also reveal differences in treatment response between patient subgroups, although these findings require clinical interpretation and further investigation.
The reliability of AI-generated insights depends on data quality, model validation, and the intended use of the technology. Sponsors should assess whether AI supports measurable improvements in study planning or delivery while maintaining appropriate human oversight.
Jullia
Welcome to QCast, the show where biometric expertise meets data-driven dialogue. I’m Jullia.
Tom
I’m Tom, and in each episode, we dive into the methodologies, case studies, regulatory shifts, and industry trends shaping modern drug development.
Jullia
Whether you’re in biotech, pharma or life sciences, we’re here to bring you practical insights straight from a leading biometrics CRO. Let’s get started.
Tom
Nowadays, AI is being discussed across almost every stage of clinical development. But when we narrow that down to clinical trial design, where can it actually make a difference?
Jullia
A lot of the value is in decisions made before the first patient is enrolled. You're deciding who should enter the study, which endpoints to measure, and whether the protocol is realistic for patients and sites.
AI can help teams examine those decisions using historical trial data, healthcare records, and other information.
Tom
And presumably some of those assumptions are quite difficult to test using conventional feasibility methods?
Jullia
They can be, particularly when you're working with complex patient populations.
Take an oncology trial targeting a specific genetic mutation, for example. You might have strong scientific reasons for your eligibility criteria, but you still need enough patients who meet them.
Machine learning can analyse patterns across large datasets to estimate how many patients might qualify. There, you can then examine which criteria are limiting the available population.
Tom
Can you give us an example of how that might influence the protocol itself?
Jullia
Imagine a proposed study where patients need a particular biomarker and a specific treatment history.
The analysis suggests that one additional exclusion criterion could substantially reduce the eligible population. The team can investigate whether that criterion is essential for safety or scientific validity.
If it isn't, there may be a reasonable case for changing it.
Tom
There's also the question of whether patients can realistically manage the study. Can AI help with that side of feasibility?
Jullia
Yes. See, protocol feasibility goes beyond estimating patient numbers.
Historical trial data can help teams examine visit schedules, assessment requirements, site workload, and possible amendment risks.
For example, a proposed schedule might require frequent hospital visits from patients with significant disease burden. Analysing previous trial experience could help the team assess whether a different schedule would be more workable, provided the required data can still be collected.
Tom
But those predictions are only going to be as useful as the data behind them, aren't they?
Jullia
Exactly. If historical data don't represent the intended population or the sites being considered, the predictions may be misleading.
You also need to know what the model was trained to predict. A tool that performs well at estimating recruitment in one setting might not perform equally well in a different therapeutic area.
Tom
We've talked about models analysing structured data. What about information sitting in clinical notes or previous protocols?
Jullia
That's where natural language processing, or NLP, becomes useful.
NLP allows systems to extract and organise information from text, including patient records, protocols, and scientific literature.
Novartis, for example, used NLP in its data42 project to analyse more than 20 years of clinical trial protocols and patient data. The aim was to identify relationships and generate hypotheses that would be difficult to find through manual review alone.
Tom
And could previous trial results tell us something about which patients might respond differently to a treatment?
Jullia
Yes, and that's an interesting application of retrospective analysis.
Researchers can examine historical results to identify heterogeneous treatment effects. That simply means looking for differences in treatment response between patient subgroups.
Some companies even apply advanced analytical methods to previous trial data to investigate those differences.
The findings might suggest a subgroup worth studying further, or help researchers reconsider an endpoint for a future trial.
Tom
I imagine there's a danger of taking an interesting pattern and treating it as established evidence.
Jullia
There is. Finding a pattern in historical data doesn't establish that the same effect will appear in a new study.
Researchers still need to examine whether the finding makes clinical and biological sense, how reliable the analysis is, and whether it could reflect chance or bias.
It may help shape a hypothesis, but the subsequent trial needs to be designed to test that hypothesis appropriately.
Tom
Adaptive designs are another area where AI often comes up. Does that mean a model can decide to change the trial while it's running?
Jullia
No, that would be an oversimplification.
Adaptive designs allow certain modifications based on interim data. Depending on the design, those might involve dose allocation or patient groups.
AI may support analysis or planning, but adaptations need to follow pre-specified rules and appropriate statistical oversight.
Tom
Let's move to recruitment. We've already discussed estimating patient availability, but identifying actual patients seems like a different challenge.
Jullia
It is. AI-supported patient matching can examine electronic health records, laboratory results, genomic information, and pathology reports to identify patients who may meet a study's eligibility criteria.
For example, an oncology study may require patients with an EGFR mutation or a KRAS G12C alteration. AI can help locate potentially relevant records.
Tom
There's a useful FDA example involving anakinra and COVID-19, isn't there?
Jullia
Yes. The FDA used machine learning during its review of an emergency use authorisation for anakinra.
The supporting SAVEMORE trial had enrolled patients with elevated levels of a biomarker called suPAR. Because the relevant test wasn't commercially available in the United States, the FDA used machine learning to develop a scoring approach based on other patient characteristics.
That helped identify patients likely to have elevated suPAR levels. It's an example of predictive modelling supporting a specific regulatory decision, with validation and recognised limitations.
Tom
One of those limitations is that a model might miss someone who could benefit. How do teams manage that risk in recruitment?
Jullia
They need to look beyond the overall accuracy of the model.
An AI system might perform well on the population used to develop it but be less reliable for patients who are underrepresented in those data.
Clinicians must still confirm eligibility, and teams need to consider whether the matching process is systematically overlooking relevant groups.
And identification is only part of recruitment. Patients still need appropriate information, consent processes, and a study they can realistically attend.
Tom
Once the trial is underway, does AI have a different role?
Jullia
Potentially. During trial conduct, AI can support data review and operational monitoring.
For example, it might help prioritise data queries, examine digital health technology data, or identify patterns associated with missed visits and protocol deviations.
A system could flag an unusual pattern in incoming patient-reported outcome data for review. But it shouldn't automatically determine the clinical significance of that pattern.
Tom
And the closer an AI output gets to a safety or clinical decision, the more carefully it needs to be controlled?
Jullia
Yes. A useful starting point is the model's context of use.
That means being clear about the specific task the model supports and what could happen if its output is wrong.
A tool used to explore recruitment assumptions presents different risks from one involved in safety-related data review.
Validation, documentation, access to the underlying evidence, and human oversight should reflect those differences.
Tom
What about models changing over time? If they perform well initially, can sponsors assume they'll keep performing the same way?
Jullia
They shouldn't. Performance can change when the underlying data or patient populations change. That's sometimes called model drift.
Sponsors also need to consider incomplete records, differences between data systems, privacy requirements, and whether the model's outputs can be examined and explained.
These are ongoing governance questions. They don't disappear once a model has passed its initial validation.
Tom
That brings us to the commercial decision. With so many possible applications, how should a sponsor decide whether AI is worth introducing into a study?
Jullia
I'd start with a specific problem that the study team is trying to solve.
If it's protocol feasibility, perhaps the question is whether AI reduces the time needed to assess eligibility assumptions or helps identify amendment risks earlier.
For recruitment, you might examine screen-failure rates, recruitment performance, or whether the matching process represents the intended patient population.
Tom
And there must be situations where the additional technology simply isn't necessary?
Jullia
Absolutely. If conventional statistical analysis or feasibility review can answer the question adequately, adding AI may create unnecessary work.
There's also a cost to preparing data, validating models, documenting decisions, and maintaining oversight.
For sponsors, the sensible approach is to assess whether the potential benefit justifies that additional work, particularly when the data are complex or difficult to review manually.
Tom
Before we finish, what would you want a clinical development team to remember when considering AI for its next trial?
Jullia
I'd keep three things in mind. First, start with the design decision rather than the technology. Second, examine the data and the model's limitations before relying on the results. And finally, keep scientific judgement central.
This gives sponsors a reasonable basis for deciding where AI genuinely adds value, and where existing methods are sufficient.
With that, we’ve come to the end of today’s episode on AI in clinical trial design. If you found this discussion useful, don’t forget to subscribe to QCast so you never miss an episode and share it with a colleague. And if you’d like to learn more about how Quanticate supports data-driven solutions in clinical trials, head to our website or get in touch.
Tom
Thanks for tuning in, and we’ll see you in the next episode.
QCast by Quanticate is the podcast for biotech, pharma, and life science leaders looking to deepen their understanding of biometrics and modern drug development. Join co-hosts Tom and Jullia as they explore methodologies, case studies, regulatory shifts, and industry trends shaping the future of clinical research. Where biometric expertise meets data-driven dialogue, QCast delivers practical insights and thought leadership to inform your next breakthrough.
Subscribe to QCast on Apple Podcasts or Spotify to never miss an episode.