Estimated reading time: 7 minutes

Clinical trials are costly lengthy processes but could AI usher in a new age of drug testing that gets treatments to patients faster? We explore some of the emerging innovations in this space, and the challenges that need careful consideration when entering this new era of health technology

How AI can help run clinical trials

AI can already draft clinical development plans in a matter of hours potentially saving clinicians weeks’ worth of work. Crucially, AI is taking on two of the biggest challenges involved in running clinical trials. That’s recruiting candidates, which is particularly challenging in rare diseases and retaining them. 

Some researchers already use AI to identify candidates for clinical trials by going through electronic health records for example. Our Future Health, which aims to be the UK’s largest ever health research programme is currently collecting data from up to 5 million people with their consent¹. It runs a clinical research recruitment service where participants who match a defined set of characteristics can be invited to take part in clinical trials.

Dr Raymond A Huml, MS, DVM, RAC, who’s worked in the biopharmaceutical industry for over 25 years, says AI can help make unstructured data like clinician notes more clinically useful by spotting trends, for example.

He also says: “Some companies are using AI to listen more systematically to rare disease patient and caregiver voices.”

When it comes to retention, AI can help facilitate decentralised clinical trials through remote patient monitoring. Decentralised trials can make it easier for patients with rare diseases to take part in clinical trials. The NHS Health Research Authority has outlined key considerations for using decentralised trial methods. These include verifying technology used to collect data from participants².

Challenges of using AI in clinical trials

A 2024 review looking at the benefits and risks of AI in healthcare highlighted that “acquiring enough data to train precise algorithms is a continual effort that necessitates a shift in attitude towards data sharing that promotes technical advancement.³” 

But perhaps one of the biggest concerns is safely and ethically handling people’s health data.

A team including Professor Sofia Villar, Professor Thomas Jaki and Dr Pavel Mozgunov from the Medical Research Council’s (MRC) Biostatistics Unit (BSU) at the University of Cambridge⁴ gave us their view on AI use. They say the primary barrier to adoption is a “trust and transparency crisis” as regulators and statisticians are deeply uncomfortable with the use of a “black box” AI model that can hallucinate, or fail to explain how they processed data, or drafted a clinical document. (An AI hallucination occurs when an AI model confidently generates false or misleading information.)

Patients may also be wary about how their private health information will be used, while data breaches, accuracy and safety are concerning for all parties, from clinicians to regulators. 

The research leads at the MRC Biostatistics Unit go on to explain: “There is a growing ethical concern over professional deskilling, as outsourcing routine data wrangling to machines risks eroding the core foundational expertise that human reviewers need to catch critical AI errors.” Such errors may arise from AI hallucinating or replicating the biases inherent in some data sets for instance.

Marko Balabanovic, chief technology officer at Our Future Health, says: “The cautions raised by UK bodies such as the Alan Turing Institute, Health Data Research UK, the UK Research Integrity Office and the National Institute for Health and Care Excellence (NICE) point to a broader set of risks around AI in health research.” 

These include results that can’t be reproduced when a model is used in a different setting or population and strong performance on a benchmark that could be mistaken for real benefit to patients.

But he highlights that this doesn’t mean AI should be avoided. Instead, “its outputs need careful validation before anyone relies on them.” 

While Ray says most users still need stronger habits around fact-checking, source verification and appropriate data use.

Potential solutions to using AI in health

To build trust in AI use it needs to be consistently governed and fully standardised. Marko explains the set of standards researchers at Our Future Health must adhere to when using AI.

“If an approved researcher wants to use AI methods within our Trusted Research Environment (TRE), they must clearly explain what data and methods they will use, why the approach is scientifically justified, how they will address risks such as bias or poor performance and what public benefit the work is intended to deliver,” he says.

He goes on to explain that Our Future Health is the most ethnically diverse large-scale UK research programme with about 24% of its cohort from at least one group traditionally under-represented in health research. This makes it more representative of the general population and therefore helps reduce bias.

However, the pace of change makes it challenging to put guidelines in place. The European Medicines Agency’s (EMA) ‘Regulatory Science to 2025’ strategy⁵ aims to build a more adaptive regulatory system to encourage the use of innovations like AI. It has put together a series of guidelines for using AI⁶. These include upholding continuous learning and knowing whom to consult when concerns arise.

The future of AI in clinical trials

Ray goes on to explain that as AI tools are validated for specific uses, they could support many additional functions like the use of digital biomarkers, automated safety-signal detection, predictive enrolment modelling, protocol feasibility assessment, clinical-trial matching and patient-facing education.

The research leads at the MRC Biostatistics Unit say AI’s true future potential lies in making complex adaptive trial designs practically viable at scale. It can do this in several ways including:

• Early stopping via biomarkers: by analysing real-time streaming data from wearables, AI can identify powerful digital biomarkers that trigger early stopping rules to quickly cut poor-performing treatments.
• Viable response-adaptive randomisation: real-time AI analysis removes the high cost and complexity of response-adaptive randomisation, making it operationally feasible to dynamically shift patients to superior treatment arms.
• The use of existing historical borrowing methods through digital twins: digital twins will expand existing methods of dynamic borrowing, using deep learning on historical data to minimise placebo groups.  

Marko’s view of the future is that “it will be possible for researchers working within approved studies to create large-scale AI models based on Our Future Health and other health data such as health records, genetic data, questionnaire responses and data from wearable tech like smart watches.

“With participants’ consent, these could help generate synthetic research participants that reflect patterns in the real population⁷. Virtual studies could then be conducted at high speed to inform the design of real-world studies.”

With this in mind, a European survey on the use of synthetic data in healthcare is underway⁸. The survey is being carried out by Synthia, a consortium that includes data experts, clinical researchers, legal experts, regulatory and policy experts.

In the future, the research leads at the MRC Biostatistics Unit expect that trust is likely to be less of an issue as they expect standardised regulatory frameworks to lead to a rise of auditable and explainable AI systems built specifically for medicine.

They say the next major hurdle will be creating new evaluation methodologies, like paired clinician-AI trials and long-term real-world monitoring, to prove these diagnostic tools translate into a genuine, measurable clinical benefit for patients.

We might even enter an era where medical avatars are used to help communicate to patients, Ray says, while Marko expects AI assistants to become co-scientists.

So, are we entering a brave new world of AI-led clinical trials? That remains to be seen but nonetheless, AI shows great promise as a potential transformative tool that could help cut costs and time when running clinical trials.

References
[1] https://ourfuturehealth.org.uk/our-research-mission/
[2] https://www.hra.nhs.uk/planning-and-improving-research/policies-standards-legislation/clinical-trials-investigational-medicinal-products-ctimps/decentralised-trial-methods-position-statement/
[3] https://pmc.ncbi.nlm.nih.gov/articles/PMC11612599/
[4] https://www.mrc-bsu.cam.ac.uk/
[5] https://www.ema.europa.eu/en/about-us/how-we-work/regulatory-science-strategy
[6] https://www.ema.europa.eu/en/about-us/how-we-work/data-regulation-big-data-other-sources/artificial-intelligence
[7] https://pmc.ncbi.nlm.nih.gov/articles/PMC7218288/
[8] https://www.ihi-synthia.eu/news/what-do-you-think-about-synthetic-data-in-healthcare-europe-we-want-to-hear-from-you

innovation spotlights the future of RARE through a technological lens. To access more innovation articles click below.