No one who signs up for a clinical trial wants to be placed in the placebo group. Placebos are like our savings accounts, our daily workout, why we wash dishes: the end result justifies the means. That doesn’t mean we have to love the process. In fact, the fear of being placed in a placebo group is why some people avoid clinical trials all together.
To really know if a drug is working, researchers have to study groups of similar people, where the only difference is the drug being tested. That way, any differences in outcome can be clearly attributed to the experimental treatment. This is achieved through randomization – randomly assigning clinical trial participants to groups that receive the treatment, or groups that don’t, also known as receiving a “placebo.”
“The purpose of clinical research is to provide clear answers to patients, physicians and regulators alike,” says Erik Pulkstenis, Ph.D., vice president of data and statistical sciences, AbbVie. “Historically that’s been done using randomized placebo-controlled studies because of the comfort they provide with respect to interpretability.”
But what if, in some situations, we could substitute a traditional placebo group with a “synthetic control arm” that includes pre-existing data?
The question at the heart of most clinical trials relates to determining the treatment effect for one agent versus a placebo control. In the case of synthetic control arms, placebo arms are modeled using information that has previously been collected, instead of patients receiving a placebo. This data comes from sources including historical control data, real world data, or the generation of a companion data set from other sources to serve as a comparator.
There are already situations where the placebo can be avoided altogether. In oncology, for example, approvals are sometimes based on single-arm trials, where everyone receives the experimental therapy.
“When we’re looking at tumors, the response rate for untreated patients is known to be essentially zero, reflecting that tumors do not shrink on their own. As a result, if all patients in a trial are given a treatment, tumor reduction of sufficient magnitude and duration is believed to indicate treatment effect and can support approval,” Pulkstenis explains. “In this case the ‘control’ group is based on what we know historically about the course of untreated disease and the lack of placebo effect in this setting.”
Could this new system make the need for control arms redundant? The short answer, says Pulkstenis, is unfortunately no.
“Randomization is very powerful and stands as an almost sacred principle when it comes to clinical trials,” Pulkstenis said. “It serves as the foundation for all analysis that is performed, and the resulting conclusions from the study. Conclusions are only valid in the presence of truly similar groups, so leaving the safety of randomization is not something to be taken lightly.”
To illustrate the inherent risks of synthetic controls, say you are researching whether parachutes improve outcomes for people who jump out of planes. You know, based on past data, that the survival for people without parachutes in planes of a certain height is zero. If five people with parachutes jump out of a plane and all survive, it’s a clear sign that parachutes vastly increase someone’s chances of survival.
In this parachute example, the synthetic control is all the historical data we have on those unlucky people who didn’t use a parachute. But here’s the critical part - people in the control group have to be very similar to those in the new study for the results to be accurate. If, in your group with parachutes, people jumped at 10,000 feet and in the historical control group, people jumped when the plane was stationary on the runway, it would be a non-valid synthetic arm.
However, there are limited situations where the benefits of a synthetic arm might outweigh the risks: if a disease is rare, and traditional trial methods may be prohibitive; areas where control group performance is well characterized historically, and results are generally consistent from trial to trial; or cases where there are clean objective endpoints that are easy to measure.
Machine learning helps researchers access huge amounts of multi-dimensional data that had once been too big and complex to use effectively. Could machine learning be used to generate a suitable control group?
“I hesitate to say there is anything machine learning cannot do because it has proven itself to continually surprise us and it is feasible that machine learning could draw a suitable control group from large, well-characterized sources of data,” Pulkstenis says. “However, it can’t help with unmeasured or unknown data, which quite often can drive important differences in a population. This would be like teaching machine learning to play a game but not telling it all the rules.”
The concept of synthetic control arms is gaining traction in the pharmaceutical research community; even the FDA is interested in the possibilities it presents.
“Five years ago, there wasn’t much talk of technologies like synthetic controls or telemedicine or wearable sensors to collect data and today, at least at AbbVie, a digital component is considered for every clinical trial we put together,” says Rob Scott, M.D., chief medical officer and vice president of AbbVie’s Development team. “There is no doubt these technologies will redefine how trials are run by including better data that is more relevant to patients and researchers.”
AbbVie currently is running two trials designed in close partnership with regulators that include components of synthetic control. Using advanced analysis methods to characterize existing data, researchers have cut control groups in half, supplemented by information already learned about control performance. The new Development Design Center is also exploring the enrollment of matching virtual control subjects through access to real world contemporary data.
Like so many other innovations, synthetic control arms aren’t about completely replacing established ways of doing things, but about making them better. After all, dishwashers came along to more efficiently clean dishes, but we still sometimes get down and dirty with the dish soap. Checkbooks can be balanced via mobile apps, but there are times when a pen and paper get the job done.
“Synthetic control arms aren’t the solution to all of the challenges facing randomized trials, nor do they realize the full promise of real-world evidence in drug development,” Pulkstenis says. “But in certain situations, they might be a better option.”