Once confined to post-market surveillance, real-world evidence is now being used to emulate trials, guide approvals and even run studies in real time, forcing a rethink of what constitutes a clinical trial. It’s the end of the day at Providence Health’s network of hospitals and clinics. Clinicians pack their bags and tidy up their consultation notes. As the halls grow quieter, a different kind of worker begins its shift — an artificial intelligence (AI) system that turns routine care into structured data — and, increasingly, into something that resembles a clinical trial. Developed with Microsoft, this biomedical language model scans, curates and cleans hundreds of thousands of patient notes written that day, translating clinical text into structured records. “If you had to employ a traditional trial coordinator to analyze these data, that could cost a million hours and tens of millions of dollars,” says Hoifung Poon, managing director of health futures at Microsoft Research.
Once confined to post-market surveillance, real-world evidence is now being used to emulate trials, guide approvals and even run studies in real time, forcing a rethink of what constitutes a clinical trial.
It’s the end of the day at Providence Health’s network of hospitals and clinics. Clinicians pack their bags and tidy up their consultation notes. As the halls grow quieter, a different kind of worker begins its shift — an artificial intelligence (AI) system that turns routine care into structured data — and, increasingly, into something that resembles a clinical trial.
Developed with Microsoft, this biomedical language model scans, curates and cleans hundreds of thousands of patient notes written that day, translating clinical text into structured records. “If you had to employ a traditional trial coordinator to analyze these data, that could cost a million hours and tens of millions of dollars,” says Hoifung Poon, managing director of health futures at Microsoft Research. “With generative AI, it takes seconds and costs just cents per patient.”
Feeding these insights — along with imaging reports, medication and lab orders, and sequencing results — into a framework called TRIALSCOPE, Poon and colleagues analyzed data from over a million patients with cancer. Their 2025 findings 1 suggest that well-curated real‑world data — extracted from messy and incomplete digital records — can reliably replicate the outcomes seen in randomized controlled trials (RCTs).
“What we show is that this kind of virtual trial can be just as accurate as a real trial compared to the gold standard,” says Poon.
The days when real-world evidence could be brushed off with a simple “correlation is not causation” are fading, says Maya Petersen, professor and co‑director of the Center for Targeted Machine Learning and Causal Inference at the University of California, Berkeley. Thanks to advances in statistical methods, machine learning and generative AI, investigators can now use real‑world evidence from electronic health records (EHRs), claims data, registries, digital health devices and more, to complement RCTs and generate robust insights.
Over the past few years, regulators have cautiously opened the door to these approaches, stressing that rigor must not be compromised for the sake of innovation and hype (Fig. 1 ). However, a recent push from the US Food and Drug Administration (FDA) to radically overhaul its approach to clinical trials — notably reporting trial data in ‘real time’ — hints that the wall between trial and real-world data could be weakening.
Drug regulators such as the US Food and Drug Administration (FDA) or the European Medicines Agency (EMA) are publishing guidance and investing in infrastructure to support the use of real-world evidence in clinical trials. © European Union, 2020 (left) and Michael Ventura / Alamy (right).
Will real-world evidence eclipse randomized controlled trials? “Absolutely not,” says Petersen (Box 1 ). But will real-world evidence change how and when we do RCTs? “Absolutely.”
When it comes to RCTs versus real-world evidence, there’s dogma on either side, Arlett notes. “People who have worked mainly in clinical trials think real‑world evidence is a dirty word, and to some extent vice versa,” he says.
Arlett urges investigators to let the research question dictate the approach 10 . “Patient voice should come first,” he says, followed by a systematic look at existing data, before deciding whether a trial, a real-world study, or a hybrid approach is most appropriate.
Petersen agrees. “There are all sorts of things we do in giant trials now that we could do much more effectively and efficiently if we brought in real-world data.”
With real-world evidence, Arlett suggests proceeding with caution. “I think it is important to make sure you’ve got a good pharmacoepidemiologist on the case,” he advises.
Petersen also shudders at the idea of investigators parachuting into the field without the appropriate knowledge. Part of her work is aimed at education 11 and democratizing access. Her center convenes meetings with regulatory bodies, academics and industry professionals to discuss how to formalize these methods, and is also working on an open‑source toolkit for investigators .
“When you tackle these harder problems, you do need advanced methods. You really do,” she says, adding: “You may understand what it’s doing from a sort of textbook statistics perspective, but you may not understand all of the ways it’s falling short in a real-world setting.”
Real-world evidence has long supported post-approval surveillance — such as monitoring drug safety after market release. In recent years, however, its growing role in pre-authorization and regulatory meetings marks a notable shift, says Peter Arlett, head of data analytics and methods task force at the European Medicines Agency (EMA).
The EMA is seeing “more and more interest in supplementing randomized trials with real-world evidence”, says Arlett.
The momentum has been building in recent years, says Harpreet Singh, a former FDA oncology division director who is now chief medical officer at the US-based clinical research organization Precision for Medicine.