What Digital Twins and Virtual Controls Can Replace in a Clinical Trial
Every randomized control group exists to answer one question: what would have happened to the treated patients without the treatment? Stratification and randomization make the control group a fair estimate of that counterfactual, free of systematic bias. If we could predict the counterfactual well for each patient, the control group would have much less (or no) work to do. That is why it is worth being precise about what digital twins and virtual controls are, and what roles they can play in clinical trials.
Three kinds of models, often given one name
"Digital twin" is now applied to almost any model that predicts a patient's outcome. We find it more useful to reserve the term for system-level models and to separate three levels of modeling:
Population-level models, such as a control event rate from a prior published trial applied to every new patient.
Patient-level statistical models, such as propensity matching, Cox models, hot-deck imputation, or machine learning predictions for each participant.
System-level digital twins, which model the biology itself, often across hundreds or thousands of characteristics.
The Framingham risk model, for example, is a valuable mathematical prediction built on a modest number of covariates. It is not a model of the cardiovascular system.
A validated digital twin changes the question
A system-level digital twin of the cardiovascular system would let a clinician change an input, such as LDL, Lp(a), or body weight, and see how that patient's risk changes over the next several years. If such a model were validated, it would answer the question a trial is designed to answer. Seeing the effect on the digital twin of a treatment would immediately characterize the impact to cardiovascular health and risk.
We are not there yet as a scientific community and getting there requires evidence that the digital twin works well. ARPA-H's CIRCLE program is a current example. It funds digital twins of the immune system for critically ill patients, with the aim of guiding immune modulation therapies and reducing ICU length of stay by 25%. The program starts with in silico testing, and the natural next step is a trial that randomizes patients between twin-guided care and usual care.
Validation also exposes the hardest problem: causation. Models built on observational data learn that patients with one characteristic do better than patients with another, not that changing the characteristic changes the outcome. The recent Lp(a) outcomes trial (HORIZON) that missed its primary endpoint raises this question directly – having high Lp(a) means having a higher risk – does lowering it change that risk? Even though lowering Lp(a) may still be shown to be beneficial, having a digital twin model that learns from randomized data – what pathways are causal – would be critical.
Prognostic scores as covariates
One use of patient-level prediction is already accepted. The European Medicines Agency has qualified an approach that includes an AI-generated prediction of each participant's outcome as a covariate in a randomized trial. Any error in the model applies equally to both arms, and a good prediction, creating a good covariate, reduces variability. No control patients are replaced, so this is covariate adjustment rather than a virtual control, and well-chosen baseline covariates may offer similar gains.
From one historical rate to patient-level predictions
Historical controls are virtual controls in their simplest form. Suppose a prior trial publication reported a 7% two-year MACE rate on placebo. If the treatment arm of a new trial showed 3.5%, that would be read as a 50% reduction. The 7% applies to every enrolled patient, regardless of who they are, the same 7%.
We have designed trials for years that test such assumptions as they run. A trial randomizing 4:1 between treatment and control can check whether new controls reproduce the historical rate. If they do, Bayesian hierarchical models or MAP priors borrow strongly, and the 4:1 trial behaves more like a 4:4 trial. If new controls come in at 3.5%, the historical information is discounted, and the randomization ratio can adaptively move toward 1:1.
What has changed is access to patient-level data from electronic health records, registries, and prior trial arms. With patient-level data, each participant gets their own prediction. One patient may carry a 9% predicted risk and another 3%, based on their own characteristics. If the new trial enrolls healthier patients, the virtual control model will reflect that and predict a 3.5% event rate rather than 7%.
Ask for a distribution, not a forecast
Most proposals return one predicted outcome per patient. We think virtual controls should return a full predictive distribution instead: thousands of imputed outcomes from a black-box model, or propensity-weighted outcomes rather than a single matched patient.
A distribution changes the analysis approach in a clinical trial. Each treated patient's outcome can be placed within their own predicted distribution, whether at the median or the 99th percentile. An outcome far from a precise forecast is strong evidence, while an outcome far from a wide forecast may still be close in standard-deviation terms. The trial can estimate disease slowing directly from these comparisons or simulate thousands of outcomes per patient to build a very large control group for familiar estimates such as hazard ratios.
Verify the model inside the trial
Asking regulators to accept that a virtual control model works without appropriate confirmation of its accuracy is a challenge. Enrolling prospective controls lets the trial show, with current patients, that when the model makes a prediction for a control, the control observations are realized – and can be calibrated. The same design can estimate a systematic bias, such as a model that is consistently off by a hazard ratio of 0.8 in both arms and correct for it. That gives the trial a virtual control arm that has been validated or adjusted for changes over time.
What this could mean for trial size
The potential gain is large. A trial powered at 100 patients per arm might need about 70 treated patients when each is compared with a validated predictive distribution, supported by a smaller set of concurrent controls. As virtual control models continue to improve, disease by disease, trial design has to keep pace. If the clinical trial community does not adapt, the organizations that own the data and the models will move ahead without it.
These designs ask more of regulators, and they also give sponsors and regulators a shared way to approve better treatments and improve patient outcomes.