
Amid the brighter spotlights being directed at data use in clinical trials, synthetic control arms represent a unique approach.
Synthetic control arms use statistical methods to create a virtual control group based on observed data from a multitude of sources. Their utilisation is especially important in indications with inadequate standard-of-care and in diseases with small patient populations. The US Food and Drug Administration (FDA) supports their use, though states their suitability in controlled trial design “warrants a case-by-case assessment”.
Given that the use of models and artificial intelligence (AI) in clinical trials is only increasing, it seems plausible that synthetic control arms could become an established approach for rare disease trials or studies where a placebo arm is not feasible.
Clinical Trials Arena spoke to Ruthie Davi, senior vice president for Statistics and Regulatory Science Innovation at Medidata to find out how synthetic control arms are impacting not just therapy advancement, but trial design more broadly.
This interview has been edited for length and clarity.
Robert Barrie (RB): What are synthetic control arms and when are they needed?
Ruthie Davi (RD): People mean different things when they say synthetic control arms. In a typical clinical trial, there’s an investigational arm and a randomised control arm. All these patients are treated prospectively. In certain situations, applying the standard of care treatment in that control arm prospectively is very difficult.
This is usually in severe diseases with inadequate standard of care, or in rare diseases where it’s just very difficult to recruit patients. And in those instances, it’s very attractive to use what we call a synthetic control arm that draws on historical data. It can be from any source, such as clinical trials, real world data, registry data, and others. It uses statistical methods to align the composition of that control arm at baseline with the composition of the investigation arm. And the reason we want to do that is so that we have a fair ‘apples to apples’ comparison. We simply compare the outcomes for the investigational arm with the outcomes for this synthetic control arm that we created. And if there’s a positive difference, then the investigational product is worth it.
RB: Is it correct to say synthetic control arms use synthetic data?
RD: It depends on who you ask. For me, it’s not synthetic data. Our synthetic control arm is built on observed data from historical clinical trials. And that’s an advantage because the quality of historical clinical trials data is so high compared to other data sources, like real world data, or even registry data. Using observed data is very appealing because it feels a lot like a randomised, controlled trial. People are comfortable with that. But there are AI applications where groups will use predictive modelling to take the baseline characteristics and predict an outcome for a patient and use that as the control arm or counterfactual for estimation of the effect of the investigational product.
RB: Both the FDA and EMA have released guidance on the use of synthetic control arms in the past decade. What does the regulation look like now for this trial approach?
RD: I would still consider it an area of innovation as far as the regulators are concerned, and although we’ve made great progress as an industry. Medidata has had successes with both the European Medicines Agency (EMA) and the FDA in using synthetic controls for regulatory work. Others in the industry have made advances as well. What we see with the FDA is that they’re most interested in rare diseases and severe indications without adequate standard of care, so those that have unmet medical need. We work a lot in the accelerated approval space for FDA, but then also in the confirmatory trial space.
RB: Do synthetic control arms have further use cases, such as impacting trial protocol and design?
RD: As I mentioned, we often target severe diseases for synthetic controls in the regulatory setting but there is an entire world outside of those indications that are applicable for non-regulatory work. We can impact the design of trials and make whole clinical development programmes more efficient. We had a use case where a customer used a synthetic control arm in ovarian cancer to make a decision to continue to the Phase II trial. And they, importantly, also used the treatment effect size that we estimated in that analysis to size their Phase II trial. With this more precise estimate of what they expected for their treatment effect, they were able to reduce the size of their Phase II trial. The estimation of the endpoints from a recent readout of that trial were very similar to what we estimated years earlier as a result of that synthetic control and the single arm trial. That’s an illustration of how using a synthetic control early in development can really save time and money in the execution of future trials and help you move the best products forward.
RB: How do you mitigate data bias in the arms?
RD: A legitimate criticism of synthetic control arms is what’s called unknown confounding. This means some difference in baseline characteristics between the investigational arm and the synthetic control arm, and we don’t know about it at the time. If there is an imbalance between the two groups due to prognostic factors, then the difference that you get will not be your treatment effect. It’ll be a biased treatment effect, favouring one group or the other depending on that factor.
This is the concern we hear from regulators most often. In terms of addressing this, the quality of the data that you’re using is important because then there’s a thorough collection of prognostic factors and less chance of confounding. The second thing is we have a series of what we call tipping point analyses where we make assumptions that there must be confounders in there and recalculate the treatment effect. The third one is actually very effective. We have designed synthetic controls with a different approach called a hybrid design. This is where we have a prospective investigational arm and a very small, randomised control arm, and then we supplement this with the synthetic control patients to bring that control size back up to the size of the investigation alarm.
When you do this, then you can make a comparison between the prospect of randomised control patients and the external control patients, and if the outcomes of those two groups look similar, then you’re more confident that there is no unknown confounding. The FDA has a lot of interest in this hybrid approach.
RB: Do you see synthetic control arms being a more established clinical trial approach within the next decade?
RD: I do think it will become more relevant and higher utilised in the next five years. I think we will continue to make slow but meaningful progress with the regulators. But what I would love to see is significant progress in the non-regulatory space. People haven’t quite realised the potential impact there in terms of, for example, better decision making and in the space of randomised controlled trials in phase II. These studies being done to determine the dose that’s appropriate for a pivotal trial. Often, they’re large trials though, and you can make those trials more efficient, reduce the number of patients, reduce the time that will take to execute the trial by using synthetic control in that space. There’s no regulatory bar there and so I would really love to see the industry embrace that in the next five years.