AI in clinical trials: Building better trials before they start

Drug development is an inherently risky endeavour, with the average probability of approval for a drug entering Phase I studies now sitting at just 6.7%. While much of this can be attributed to lack of efficacy, factors such as poor clinical trial design and challenges with enrolling and retaining patients can lead to significant roadblocks.

Meanwhile, studies have become increasingly complex and the advancement of technology means that we are now able to collect a higher volume of data than ever, with Phase III studies now producing an average of 5.9 million data points, an increase of 283% over the past decade.

Against this backdrop, artificial intelligence (AI) is shifting how clinical development is approached. Instead of designing trials based on institutional precedent and legacy processes, sponsors are now reducing risk, accelerating timelines, and driving innovation by modelling what is most likely to work in the future.

Shift towards predictive modelling

AI-driven simulation and predictive modelling have emerged as a turning point. Virtual twins enable teams to optimise their protocol designs, test eligibility criteria, recruitment assumptions, dropout risk, and site performance in advance. The concept of a virtual twin is rooted in product life cycle management, historically applied to designing complex products like an aircraft. Similarly, a virtual twin in clinical research is a digital model of the entire trial, combining data and knowledge that enables earlier testing of design choices, such as optimising the design of the protocol.

From eligibility criteria through objectives and endpoints to the schedule of activities, clinical and operational teams can anticipate potential outcomes like the number of patients to enrol in a given country or the risk of patient dropout. These teams can account for preventable causes of delay before a single patient is enrolled by reducing complexity, anticipating increased cost, and mitigating the risk of operational failure. Virtual twins enable clinical programmes and studies to move from static planning closer to structured reality.

Recent research has shown that high-fidelity, fit-for-purpose data sets derived from historical clinical trials and real world data can be reliably used to construct external control arms that support evidence generation in oncology. In some late-stage neurological studies, simulation approaches have been used to reduce control arm sizes by up to 33%. For patients, this is transformative as, with fewer people on the control arm of the trial, more participants can take the study drug and, consequently, gain access to life-changing or even life-saving treatments.

While results vary by therapeutic area, it is clear that trial design is becoming something that can be tested and improved before patients are even enrolled.

This also changes how inefficiency is addressed. Mid-study protocol amendments have long been a major source of delays, cost, and patient burden. Predictive modelling allows teams to conduct scenario-based risk assessments of the likely impacts of these changes before implementation, reducing the need for mid-study changes and improving confidence in decision-making.

Patient and sponsor impact

As AI tools quickly evolve, the quality of impactful AI-driven outcomes is closely tied to the volume, fragmentation, and knowledge infused into clinical data transactions across a variety of systems. Trials now generate huge sets of patient and study data; however, a lot of this information has historically been difficult to connect.

AI enables these datasets to be integrated, enriched with industry standards, and analysed together, improving forecasting, feasibility planning, and study execution decisions. The benefits of this are already being felt throughout the industry with many clinical trials leaders already using or actively exploring AI and reporting early value in design and operational efficiency.

Within this context, many companies are focused on applying AI across the full clinical lifecycle using large scale, high-fidelity clinical datasets to improve outcomes for patients, sponsors, and CROs. The emphasis is not only on analysing data faster, but on increasing impact, making it actionable at the point decisions are made.

For patients, more accurate eligibility criteria reduces unnecessary screening, better forecasting improves site readiness, and digital tools reduce patient burden to make participation more accessible. Additionally, wearables and remote monitoring allow patients to contribute to trials while minimising disruption to their everyday schedules, leading to data that better reflects real life.

For sponsors and CROs, these same capabilities reduce inefficiency, improve predictability, and limit costly amendments that disrupt timelines and study operations.

More evolution to come

Looking ahead five to ten years, clinical programmes are likely to become more adaptive by using environments that model patient populations and behaviours before enrolment. These models will be refined throughout the trial as new data is collected, helping sponsors compare predicted outcomes with actual results and respond more quickly when issues arise. Trials will continue to evolve to be more flexible and data-driven throughout their lifecycle.

This shift is expected to improve the efficiency of patient recruitment, reduce late-stage failure rates, and shorten overall drug development timelines by identifying issues early.

Against this background, the regulatory environment is becoming more nuanced. Data sovereignty requirements in some regions now require clinical data to remain within national borders. Thus, organisations and sponsors are adopting more adaptable approaches to their trial design, infrastructure, and analysis.

Consequently, sponsors and CROs need to balance the speed and efficiency gained through the use of AI with the flexibility needed to meet different regulatory requirements around the world.

Throughout this transition, trust remains essential. AI systems are only as reliable as the data and knowledge they are trained on. Bias or gaps in representation can affect outcomes. Therefore, having a large repository of historical data along with the knowledge and experience to use it effectively is essential to successfully executing trials and generating accurate results.

Exploration to execution

The life sciences industry is entering a structural shift in how research is conducted. AI is moving from experimentation into execution, reshaping how trials are designed and delivered. Yet, it is worth remembering that the most important change is not the technology itself, but the shift in mindset that it enables, from reactive correction to predictive design.

As this shift accelerates, success will depend on how effectively the industry connects AI, data, knowledge, and regulatory strategy into a coherent approach to bring better treatments to patients faster.

About the author

Josh Hartman is SVP for Platform AI at Medidata. With over two decades of experience at the intersection of data and AI, Josh has developed advanced solutions that optimise clinical development decision-making and is the author of multiple patents.