Digital twins gain traction

A drug trial usually works like a coin toss: one group gets the new treatment, one group gets the standard treatment or placebo, and regulators compare the two groups at the end. The reason for the coin toss is simple: it is still the cleanest way to make sure the groups are alike before the drug starts changing outcomes. (fda.gov) A digital twin is an artificial intelligence-built stand-in for the patient you did not enroll, using old trial records, registries, scans, or medical charts to predict what likely would have happened without the new drug. In trial design, that stand-in usually shows up as an external control, which means the comparison group sits partly or entirely outside the live randomized trial. (fda.gov) Regulators have been inching toward this for years, but the rulebook was written for classic control groups, not computer-built ones. The International Council for Harmonisation guideline on control groups dates to July 20, 2000, and the United States Food and Drug Administration page for that guidance still points back to the older framework. (database.ich.org, fda.gov) The new movement is not “replace trials with algorithms.” The European Medicines Agency said in a July 24, 2025 concept paper that randomized controlled trials remain the gold standard, but also said Europe still lacks specific regulatory guidance on when external controls are appropriate for pivotal or supportive evidence. (ema.europa.eu) The United States Food and Drug Administration moved earlier with a February 2023 draft guidance on externally controlled trials. That document says sponsors can compare treated patients with people outside the trial, including historical patients from an earlier period or concurrent patients treated elsewhere, if the design can still support evidence on safety and effectiveness. (fda.gov) Companies want this because modern trials are getting harder to run, not easier. Applied Clinical Trials cited a July 2024 analysis of 66,935 studies from Phesi that found 32% of Phase 2 trials in the first half of 2024 were terminated, which it described as a 56% increase over pre-pandemic levels. (appliedclinicaltrialsonline.com) The appeal is strongest where a normal control arm is hardest to justify. In oncology, rare disease, and some neuroscience studies, patients may be very sick, eligible populations may be tiny, and sponsors often argue it is wasteful or unethical to assign many people to a treatment that is expected to do little. (fda.gov, nature.com) But a digital twin only works if the old patients and the new patients are truly comparable, and that is where the fight starts. The European Medicines Agency concept paper says the main unresolved issues are methodological reliability and operational details, which is regulator language for “show us the data are close enough, measured the same way, and not tilted by hidden bias.” (ema.europa.eu) That is why the likely path is narrower than the hype suggests. The most plausible early wins are settings with a well-mapped disease course, strong historical data, and endpoints that are measured the same way across sites and time, rather than messy diseases where standard care changes every few months. (fda.gov, ema.europa.eu) The safety burden also shifts if approvals lean more on modeled comparisons. A 2025 perspective in npj Systems Biology and Applications says digital twins may help from pre-clinical work through post-marketing, which means the real-world check after approval becomes part of proving the model did not flatter the drug. (nature.com) So the story is not that regulators suddenly trust artificial intelligence patients. The story is that the United States Food and Drug Administration already has a draft framework, the European Medicines Agency has opened the door to a reflection paper, and drug developers now have a clearer signal that synthetic comparators may be usable in specific cases if they can survive the bias audit afterward. (fda.gov, ema.europa.eu)