A digital twin in a clinical trial is an AI model’s prediction of how a specific participant would fare without treatment. Used alongside randomization, twins can shrink control arms and trial sizes while keeping results valid. Used instead of randomization, a wrong twin can produce a wrong answer. The role you give the twin decides the risk.
“Digital twin” is one of those phrases that arrives with a picture already attached: a glowing virtual copy of a patient, running in parallel, living out the life the real patient would have had without the drug. It is a seductive idea for clinical research, where the most expensive thing in any trial is the group of people who do not get the treatment.
The real technology is more modest, and more interesting. In clinical trials today, a digital twin usually means a statistical forecast. You feed a model a participant’s baseline data, such as age, symptom scores, lab values and disease history, and it predicts what their outcome would likely be if they received placebo. Train that model on thousands of patients from past trials and registries, and the forecasts can be quite good.
The question regulators will ask, and should ask, is what happens when the forecast is wrong. The argument of this piece is that the answer depends almost entirely on one design choice. If the twin supports a randomized trial, a wrong twin costs you statistical power, and the result stays honest. If the twin replaces randomized controls, a wrong twin costs you the truth. Most of the confusion around digital twins comes from blurring those two roles.
What is a digital twin in a clinical trial?
In drug development, the term covers a spectrum.
At one end are mechanistic models, built from physiology and pharmacology, that simulate how a drug moves through the body or how a device interacts with tissue. The FDA has used computational modeling for years in areas such as medical devices and pharmacokinetics, and in April 2025 it announced a plan to reduce animal testing that leans on computational models among other approaches.
At the other end are data-driven prognostic models, typically built with machine learning, that predict an individual patient’s disease trajectory. These are what most people mean by digital twins in clinical trials. Unlearn, a San Francisco company, is the best-known developer, focusing initially on neurology, including Alzheimer’s disease.
The key idea is that a participant’s own twin serves as a personalized prediction of their control outcome. The question is how that prediction gets used.
How can digital twins shrink control arms?
The most established use is as a covariate, a piece of baseline information included in the trial’s statistical analysis.
Randomized trials already adjust for baseline characteristics. If you know a participant’s starting symptom score, you can explain some of the variation in their final score, which reduces noise and makes the treatment effect easier to detect. A digital twin compresses many baseline characteristics into one highly informative number: the predicted outcome without treatment. Adjusting for that prediction can remove a lot of noise.
Less noise means fewer patients are needed for the same statistical power. Since sponsors often want to keep the treated group size, the saving usually comes from the control arm, for example randomizing two or three participants to treatment for every one on placebo.
The European Medicines Agency looked hard at this approach. In September 2022 it issued a qualification opinion on Unlearn’s method, called PROCOVA, short for prognostic covariate adjustment. The opinion accepted it for primary analysis of phase 2 and 3 trials with continuous outcomes, noting that it is a special case of standard covariate adjustment and shares its properties of type I error control and asymptotically unbiased estimates in large samples. It also noted that the gain depends on how strongly the prognostic score correlates with the actual outcome. The FDA’s final guidance on adjusting for covariates in randomized trials, issued in May 2023, likewise supports prespecified adjustment for baseline covariates expected to be strongly prognostic of the outcome.
That regulatory language is the heart of the matter. Type I error control means the chance of declaring a useless drug effective stays where it should be. Randomization protects the comparison. The twin just sharpens it.
What happens when the twin is wrong?
This is where the design choice pays off.
Suppose the twin model was trained on patients from older trials, and the new trial enrolls a somewhat different population, with milder disease or a different standard of care. The model’s predictions get worse. In a covariate-adjusted randomized trial, the consequence is that the adjustment removes less noise than expected. The trial is less powerful than planned. It may miss a real effect. But treated and control groups were still randomized, so the twin’s errors are spread across both arms. It does not manufacture a false effect.
Now suppose the twin replaces the control arm entirely. Everyone gets the drug, and each participant’s outcome is compared with their twin’s prediction. If the twin systematically overestimates how badly untreated patients would do, perhaps because standard care has improved since the training data were collected, the drug will look effective when it is not. Nothing in the trial can detect this, because there is no randomized group to check against.
That is the difference between losing power and losing validity. The first is a cost to the sponsor. The second is a cost to every patient who later takes the drug.
The same logic applies to external controls in general, as our discussion of real-world evidence found: real-world and model-based comparisons are most reliable when a randomized anchor remains in the design.
A worked example: an Alzheimer’s trial with a disappointing twin
Here is a hypothetical phase 2 trial, with illustrative numbers.
A sponsor plans a trial with a continuous cognitive outcome. Historical data suggest its digital twin model correlates with actual placebo outcomes at about 0.6. Under standard covariate adjustment, that correlation removes roughly 36 percent of the outcome variance, since the remaining variance is about one minus the correlation squared. The sponsor therefore plans a trial about a third smaller than an unadjusted design, sized for 80 percent power.
The trial runs. The enrolled population turns out to differ from the training data, and the twin’s correlation with actual outcomes is only about 0.2. Now the adjustment removes just 4 percent of the variance. The trial was sized for a much quieter signal than it actually has. A rough calculation suggests its power drops from 80 percent to somewhere around 60 to 65 percent.
That is a bad outcome for the sponsor. A real effect has a meaningfully higher chance of being missed. But if the drug does nothing, the trial still has only its prespecified chance of a false positive. The error is honest.
Rerun the same scenario with the twin replacing the placebo arm, and the damage depends on whether the twin’s errors are random or systematic. If untreated patients in this population decline more slowly than the model predicts, every treated participant appears to outperform their twin. The drug looks effective. There is no way to tell from within the trial.
The practical lesson for trial designers: when you use a twin to cut sample size, build in a margin. Prespecify a check of the twin’s performance in the blinded pooled data, or size the trial for a lower correlation than historical data suggest.
The strongest case for going further
Advocates of digital twins make a serious argument that caution here is itself costly.
Every patient randomized to placebo in a serious disease is someone who may miss a benefit. In conditions like Alzheimer’s, amyotrophic lateral sclerosis or rare neurological diseases, many patients will not enroll in trials where they might get placebo, which slows research for everyone. If twins can safely reduce control arms by a third or more, that is a real ethical gain, and accepting only modest reductions leaves it on the table.
They also argue that twins will improve. More training data, better models and validation across many trials will make predictions more reliable, and at some point the risk of a systematically wrong twin in a well-characterized disease may be smaller than the practical harms of randomization.
This is reasonable, and the covariate approach already captures much of the ethical benefit while keeping randomization. The step from reduced control arms to no control arms is where the argument needs evidence that does not yet exist: a track record of twins predicting control outcomes accurately in new trials, prospectively, across changes in patient populations and standards of care. Until that record exists, the burden of proof sits with replacement.
What this means for sponsors, regulators and patients
For sponsors, digital twins as covariates are a relatively low-regret way to make trials smaller or more powerful, with a qualification opinion from the EMA behind the approach. The real risk is overconfidence in the correlation. Plan conservatively, prespecify everything and check the twin’s performance on blinded data.
For regulators, the key questions are the ones the EMA’s opinion already emphasized. Was the model fixed before the trial? Was its training population similar to the trial population? How was its performance estimated? Replacement designs deserve far stricter scrutiny than adjustment designs, and guidance should say so plainly.
For trial sites and investigators, twins do not change the consent conversation much when used as covariates. Participants are still randomized. But if control arms shrink, the chance of receiving the active treatment rises, which may help recruitment, a stubborn problem we explored in trial diversity as a staffing problem.
For patients and advocates, it is fair to ask how a trial uses modeling. A trial that uses twins to reduce the number of people on placebo, while keeping randomization, offers a better deal to participants without weakening the answer. A trial that uses twins to eliminate placebo entirely may offer faster access, at the cost of more uncertainty about whether the drug works.
Frequently asked questions
Are digital twins accepted by regulators?
In a specific form, yes. The European Medicines Agency issued a qualification opinion in September 2022 for PROCOVA, a method that uses AI-generated prognostic scores as covariates in phase 2 and 3 randomized trials with continuous outcomes. The FDA’s 2023 guidance on covariate adjustment also supports prespecified adjustment. Using digital twins to fully replace control groups has no comparable regulatory acceptance.
Can digital twins replace placebo groups?
Not reliably today. Twins used to replace placebo groups face the same weakness as other external controls: if the model’s predictions are systematically off, for example because care has improved since its training data were collected, the drug can look effective when it is not. Most accepted uses keep a randomized control group and use twins to reduce its size.
How much can digital twins reduce trial size?
It depends on how well the twin predicts outcomes. As a rough rule for covariate adjustment, the remaining outcome variance is about one minus the square of the correlation between prediction and outcome. A correlation of 0.6 cuts variance by roughly a third, allowing a proportionally smaller trial. Weaker correlations yield much smaller gains, which is why conservative planning matters.
Are digital twins the same as in silico trials?
Not exactly. In silico trials usually refer to fully computer-simulated studies, often using mechanistic models of physiology or devices, with no human participants for some questions. Digital twins in drug trials usually refer to AI predictions for real participants, used to support analysis of an actual randomized trial. The terms overlap, but the evidentiary weight is very different.
The takeaway
The most useful digital twin is one that can be wrong without harm. Keep randomization, use the twin to sharpen the comparison, plan for a disappointing correlation, and a twin’s errors cost only power. Remove randomization, and the twin’s errors become the result.
Regulators asking “what if the twin is wrong?” are not being obstructive. They are asking the only question that separates a clever trial from a misleading one.
Related on the GHC blog: AI Diagnostics: The Evidence Is Strongest Where the Question Is Narrowest and Fast Approval Isn’t the Danger. Slow Withdrawal Is..
Would you trust a trial that used digital twins to cut its placebo arm in half? What would you need to see first?
