Building External Control Arms with Synthetic Data

In summary

• In clinical trials, recruiting a traditional control group can be slow, expensive, or ethically problematic, especially in rare diseases and hard-to-reach populations.
• Historical, External, and Synthetic Control Arms are three distinct approaches to building a comparator without a real second randomised arm; Synthetic Control Arms are those based on statistical and artificial intelligence models.
• The ability to generate synthetic patients is not enough to guarantee a good control arm: the quality and relevance of the source data, the comparability of populations, bias management, the representation of rare events, and clinical-statistical validation all matter.
• A reliable SCA requires representative data, transparent methodology, bias control, validation against clinical endpoints, statistical robustness, privacy protection, governance, and regulatory consistency.
• External and synthetic control arms can make some studies more efficient, inclusive, and ethically sustainable. Still, they do not automatically replace traditional controls: their value depends on the rigour with which they are designed, built, and validated.
• Aindo supports pharmaceutical companies, CROs, and research groups in building and validating synthetic control arms, with a certified infrastructure that never exposes personal data.

When the control group becomes a sensible topic

Imagine building a new-generation commercial aeroplane. To check how it performs during a violent storm at high altitude, you would not build a second identical aeroplane, fill it with real passengers, and deliberately launch it into a hurricane to see how it behaves. Instead, you would use a state-of-the-art flight simulator, fed with very accurate real weather and physical data. Synthetic Control Arms are the flight simulators for clinical trials.

In many clinical trials, assigning real patients to a control group with placebo or standard of care is not just a methodological choice; it is also a choice involving human, economic, and time cost. In rare and rapidly progressing diseases, or in pediatric and geographically dispersed populations, recruiting a sufficient number of patients for a traditional control arm can turn out to be slow, expensive, and, in some cases, ethically difficult to justify, especially when a promising experimental therapy already exists.

These constraints have long driven researchers and sponsors to ask whether it is always necessary to build a second comparative arm entirely from scratch, or whether existing data, coming from previous studies or real-world clinical practice, can be leveraged to build an equally solid comparator. It is from this question that external control approaches arise, including synthetic control arms.

Historical, External, and Synthetic Control Arms: what are the differences

When considering alternatives to the randomised control group, it is useful to distinguish among several families of approaches, often used interchangeably but conceptually different.

Historical Control Arms use data from patients treated in the past, typically from previous clinical trials conducted on the same pathology. External Control Arms broaden the perspective by also including other observational sources external to the ongoing study, such as electronic health records, disease registries, or real-world data (RWD). In both cases, these are actual observed cohorts, simply not randomised along with the experimental arm.

Synthetic Control Arms (SCAs) are the next step. Instead of limiting themselves to reusing observed cohorts as they are originally, they employ statistical and artificial intelligence models to build synthetic comparators from the same real clinical data. Not all generation methods are based on AI in a strict sense, but machine learning is what makes it possible today to generate virtual cohorts of sufficient size and quality even starting from limited initial samples.

If built with rigour, a synthetic control arm can in some cases allow a traditional comparative design to be transformed into a single-arm study, where all real participants receive active experimental therapy. However, this is an option whose feasibility is subject to the quality of available data, population comparability, study design, and regulatory acceptability, and must therefore be evaluated on a case-by-case basis, not assumed as a default.

AI is part of the solution

This is where the central point lies: the simple technical capacity to generate synthetic patients does not, in itself, guarantee the quality of a control arm. The reliability and usability of an SCA depend on the quality and relevance of the source data, the comparability of real and synthetic populations, bias management, the ability to correctly represent rare events, and the clinical and statistical validation of the results.

The architectures at a glance. Several machine and deep learning architectures can generate synthetic data while safeguarding the privacy of real patients: among the most widely used in this context are Conditional Tabular GANs (CTGAN), Conditional Variational Autoencoders (CVAE), PrivBayes (based on Bayesian networks), and Private Aggregation of Teacher Ensembles GANs (PATE-GAN). They differ in their ability to handle complex clinical data, the diversity of the records generated, computational efficiency, and protection of the original information. None is universally superior to the others: the choice depends on the nature of the available data, the characteristics of the population of interest, and the specific objectives of the study.

Knowing these differences is useful, but algorithm choice remains a component of the process, not its core. The true value of a synthetic control arm does not lie in the generative model itself, as much as in the ability to transform available data into a credible, representative comparator adequate for the specific study. Empirical validations published in the literature have already shown that, when this work is done with rigour, synthetic cohorts can achieve survival curves and key clinical outcomes statistically overlapping with those of real control populations.

However, two fundamental technical challenges remain. The first concerns rare events: generative models are mathematically incentivised to capture the most common pathways to minimise training error, so a rare and potentially lethal adverse event risks being overlooked, producing a virtual cohort that appears healthy but lacks the outliers necessary for rigorous safety benchmarking. The second concerns bias: if the historical datasets used for training underrepresent specific populations, for example women or certain ethnic groups, the generative model mathematically amplifies those same gaps. About 78% of participants in global genomic databases are of European origin, and models trained on these data can systematically miscalculate genetic drug metabolism rates for populations of East Asian or African ancestry; similarly, because women can present significantly different liver enzyme activity compared to men and are historically underrepresented in clinical trials, models trained on legacy data risk generating data that is not representative of the real population.

Therefore, there is no one-size-fits-all model for every synthetic control arm; every trial is a case unto itself, requiring the selection of the right model and proactive verification of hidden pitfalls, such as temporal confounding related to evolving standards of care or systemic demographic imbalances.

Requirements of a reliable SCA

The critical issues described above translate into a set of concrete requirements that a synthetic control arm must fulfil to be considered reliable. What is needed is relevant data representative of the population of interest; a transparent and documentable methodology that allows generation choices to be reconstructed; active control of demographic and clinical biases; rigorous validation against study-relevant clinical endpoints; statistical robustness, including on distribution tails; privacy protection for the real patients on whom the model was trained; clear process governance; and consistency with the regulatory context of the study.

These elements should not be understood as an abstract checklist, but as concrete criteria that allow a transition from a technical data-generation exercise to a comparator that is truly usable in a clinical trial, including for regulatory submission purposes.

This is also the direction taken by the EMA framework on Real-World Data (RWD) of March 2026 (EMA/503781/2024), which sets clearer expectations regarding the quality, traceability, and regulatory usability of evidence based on real data, including evidence used to build external and synthetic comparators. Aligning early with these criteria is now an integral part of building a reliable SCA.

How Aindo can support with its services

Aindo supports pharmaceutical companies, CROs, and research groups in building and validating synthetic control arms starting from data from previous clinical trials, registries, and real-world sources. This experience also translates into participation in application projects, including an initiative dedicated to Kawasaki disease developed in dialogue with AIFA and multi-site European initiatives aimed at validating the use of synthetic data in shared clinical and regulatory contexts.

The work does not simply consist of generating virtual patients. To build a credible comparator, Aindo prepares and integrates available data, checks its quality, manages any missing information, and, when necessary, applies rebalancing techniques to reduce underrepresentation of specific patient groups. The goal is to obtain a control population that is sufficiently representative and comparable to the one receiving the experimental treatment.

Aindo then selects the most suitable methodology for the specific study, considering the population involved, clinical endpoints, data availability, and the level of protection required. Validation verifies not only that the synthetic data reproduces the main statistical characteristics of the real data, but also that it preserves relevant clinical information for planned analyses and is fit for the specific use.

Data protection is integrated into the generation process. Aindo’s models learn the statistical characteristics and relevant relationships present in the source data to generate new synthetic records, without creating copies of individual real patients. Specific privacy tests also allow evaluation of the risk that original information could be reconstructed or associated with data subjects. This approach allows synthetic arms to be developed in compliance with applicable data protection requirements and can facilitate collaboration between organisations, reducing the need to share individual data directly.

Aindo’s activities are embedded in a formalised system of quality, security, and data protection, supported by Europrivacy certifications (aligned with Article 42 of the GDPR), ISO 9001 for quality management, and ISO 27001 for information security. This allows secure multi-site clinical collaborations, even internationally, without the legal friction of transferring raw patient data.

Aindo also specifically supports the construction of control arms for studies on rare, underserved, or hard-to-recruit populations: when patient populations are small, geographically dispersed, and highly heterogeneous, a large randomised control group may simply not be feasible, and a methodologically rigorous synthetic arm becomes a concrete alternative to make the study achievable. By providing credible, AI-generated synthetic comparators, Aindo enables sponsors to strengthen comparative effectiveness analyses, reduce patient exposure to placebo, and safely accelerate access to transformative treatments.

Towards more efficient trials

External and synthetic control arms can contribute to making some clinical trials more efficient, inclusive, and ethically sustainable, particularly when recruiting a traditional control group is difficult, slow, or problematic. However, they do not automatically replace traditional controls and should not be considered a methodological shortcut: their value depends entirely on the rigour with which they are designed, constructed, and validated, from source data quality to methodology transparency, down to the timely verification of bias and rare cases.

The advantage, when the methodological prerequisites are met, is concrete and measurable: a synthetic control arm can substantially reduce recruitment time and costs, avoiding the need to enrol and follow an entire second group of patients. The benefit is therefore not only ethical, but also economic and temporal: fewer patients exposed to placebo, fewer resources tied up, and a faster path between a promising therapy and the patients who need it.

Used in this way, synthetic control arms are not a way to bypass scientific and regulatory standards of clinical trials, but an additional tool to accomplish them faster, more cheaply, and with greater respect for the patients involved.