Digital twins can support clinical trials in different ways depending on the context of use and how their predictions are incorporated into the study design and analysis. The term “digital twin” can encompass different types of patient-specific predictions and applications, and its meaning depends in part on the data used to train the underlying model. Training data define the population, outcomes, and relationships the model learns and, in turn, what the resulting digital twin can represent. In randomized clinical trials, digital twin predictions can provide prognostic information that improves statistical efficiency through covariate adjustment, ie., PROCOVA. In single-arm trials, they can contribute to external control arms by predicting outcomes under a control condition when a concurrent randomized control is absent. These applications have different statistical and regulatory considerations, underscoring an important point: digital twins are not a single statistical method, but a technology that can support multiple approaches to clinical development. 

The PROCOVA story is one of scientific innovation in a highly regulated space: drug development. Developing a novel statistical methodology built on advanced machine learning was only half the challenge. The other half was demonstrating that the innovation could be trusted, validated, and adopted within the rigorous statistical and regulatory frameworks governing clinical trials. Rather than treating regulatory requirements as a hurdle, Unlearn approached them as an integral part of the scientific process: grounding the methodology in established statistical principles and demonstrating that gains in statistical efficiency did not require compromising Type I error control or introducing bias into treatment-effect estimation. The result has been a progression from scientific innovation to EMA qualification, to alignment with FDA guidance, and now to growing adoption by pharmaceutical and biotechnology companies through both external partnerships and internal development. 

PROCOVA illustrates an important lesson in drug development: the boldest innovations are not those that circumvent regulatory science, but those that advance alongside it, earning confidence through evidence and, over time, shaping clinical research in ways that endure.  What began as a methodology introduced by Unlearn has evolved into a growing body of regulatory evidence and broader adoption, both through direct partnerships and through companies developing internal capabilities. This is how a sound methodology should spread: it gets validated, it gets used, others learn from it, and the whole field benefits.

A Simple Idea With Profound Implications

The core principle behind PROCOVA is elegant: if we can accurately predict how a patient will progress before the trial begins, that prediction can be incorporated as a covariate in the statistical analysis. A more prognostic covariate can explain more of the variability in patient outcomes. That, in turn, can increase statistical power without requiring additional participants or potentially allow a trial to achieve the same power with fewer participants. This is not a new statistical concept. Covariate adjustment has been part of clinical trialists' toolkit for decades. What is new is the source and richness of the prognostic information used for adjustment.  

Unlearn trains machine-learning models on historical clinical trial data to generate digital twins: comprehensive, longitudinal predictions of participants’ outcomes across multiple measures and time points. From these predictions, a prognostic score can be derived and prespecified as a covariate for use in the trial’s statistical analysis. Importantly, the AI model does not determine whether the experimental treatment works. Randomization remains the foundation for estimating the treatment effect. The prognostic score helps explain outcome variability, allowing the randomized comparison to be conducted more efficiently. This distinction has been central to the regulatory story.

In 2022, the European Medicines Agency (EMA) qualified Unlearn’s PROCOVA methodology as an acceptable statistical approach for use in the primary analysis of Phase 2 and 3 randomized controlled trials with continuous outcomes. Engagement with the FDA subsequently reinforced that the approach can operate within existing regulatory principles, including the FDA’s May 2023 final guidance on covariate adjustment in randomized clinical trials. 

That is one reason PROCOVA represents a relatively pragmatic application of machine learning in clinical development. When appropriately prespecified and implemented within a randomized clinical trial, it can preserve Type I error control and unbiased treatment-effect estimation while fitting within familiar statistical frameworks. In its simplest form, PROCOVA can be understood as an ANCOVA analysis using a highly prognostic and optimized covariate.

From Methodology to Broader Adoption

The broader biostatistics community has increasingly engaged with PROCOVA, with researchers across academia, industry, and regulatory settings examining the methodology, its theoretical properties and its applications. Statisticians across pharmaceutical companies and academia have published evaluations, applications and extensions of prognostic covariate adjustment reflecting growing engagement with the methodology beyond Unlearn. Research has also extended the framework to longitudinal endpoints through PROCOVA-MMRM, as well as approaches for binary outcomes and Bayesian settings. This is an important stage in the evolution of any methodology: researchers test its assumptions, explore its boundaries, and extend it to new problems.

The range of disease areas is broadening as well. What was first validated in Alzheimer's disease is now being applied in ALS, Huntington’s disease, Parkinson’s disease, and increasingly across therapeutic areas such as oncology and cardiometabolic disease, including obesity, where sufficient historical trial data are available to train robust prognostic models. We are still relatively early in the adoption curve. But the trajectory is clear: prognostic covariate adjustment, when implemented rigorously and transparently, is increasingly part of the conversation about how sponsors can design more efficient randomized clinical trials.

The regulatory frameworks are in place. The statistical theory is solid. The empirical evidence supporting meaningful efficiency gains continues to accumulate. Published analyses have reported control arm size reductions in the range of up to 33%. The question is no longer whether this approach works. It is how quickly the industry will move to make it standard practice. The fundamental aspiration of these technological advancements is simple: to make clinical trials more efficient in ways that translate into meaningful benefits for patients.

The Next Regulatory Frontier: External Control Arms

The next frontier is single-arm trials, where digital twins are used not to adjust a randomized clinical trial, but to help construct an external control arm. This represents a natural extension of the technology, but also introduces a different set of statistical considerations. Unlike PROCOVA, model-based synthetic controls do not come with a built-in guarantee of unbiased treatment-effect estimation, making rigorous validation and careful characterization of potential bias central to advancing this approach.  

Unlearn’s recent technical work is exploring this challenge through doubly robust estimators, combining outcome models with propensity-based adjustment using an augmented inverse probability weighting (AIPW) approach. Compared with relying on an outcome model alone, this approach is more robust because it combines two complementary methods for estimating the treatment effect. Initial evidence has been generated through reanalyses of amyotrophic lateral sclerosis and Huntington’s disease clinical trial data, providing a foundation for continued evaluation. 

The path to broader regulatory acceptance remains open and more complex than for PROCOVA. Unlearn has identified three potential routes: building precedent trial by trial; pursuing a dedicated FDA pilot or demonstration project; or developing a disease-specific initiative, potentially led by a patient advocacy or research organization, to build and validate an external-control dataset that could serve as a reusable resource within an indication. Unlearn intends to collaborate on the latter two approaches, which could provide a more systematic path to regulatory clarity than relying on case-by-case precedent alone.

Looking Ahead

PROCOVA demonstrates what becomes possible when machine learning is introduced into a randomized clinical trial while preserving the statistical guarantees regulators already understand and trust. External controls pose the harder question: how much further can those same technologies take us when randomization itself is absent?

Answering that question will require more evidence, more collaboration, and continued engagement with regulators. If the evolution of PROCOVA offers a lesson, however, it is that lasting innovation in clinical trials is built not by asking regulators to accept a black box, but by making the methodology transparent, testable, and grounded in the scientific principles that guide clinical evidence. 

To learn more about how digital twins enable more confident, better-informed decisions in clinical trials, visit our Evidence page, and if you talk to us about your study, please reach out to our team.