Adaptive Trial Design Takes Center Stage During Pandemic

By Deborah Borfitz

February 23, 2021 | The pandemic has had some “silver linings” for clinical studies and the overall U.S. healthcare system, according to speakers discussing adaptive trial design at the recent COVID-19 and Cancer virtual meeting of the American Association for Cancer Research (AACR). “We can’t go back to business as usual” was the parting remark of Laura Esserman, M.D., a surgeon and breast cancer oncology specialist at the University of California, San Francisco, mirroring the sentiment of her fellow panelists. 

COVID has laid bare inefficiencies in meeting population healthcare needs and the urgency to better integrate routine care and clinical trials, says Esserman, pointing to continuous improvement as the Holy Gail. “Getting to a learning healthcare system is where we want to be.”

If the level of collaboration and experimentation aimed at COVID-19 were to be redirected to other global health problems, “imagine where we would be,” remarks Robin Mogg, biostats leader with the Bill & Melinda Gates Medical Research Institute. Developing products to fight diseases like tuberculosis that disproportionately impact low- and middle-income countries, and improve outcomes in maternal and newborn health, is the mission of the three-year-old nonprofit biotech company.

“The clinical research landscape is almost unrecognizable from what it was before,” notes Derek Angus, M.D., one of the world’s leading researchers in the field of critical care medicine who is based at the University of Pittsburgh School of Medicine and UPMC. Many “theoretical applications” of platform trials have been deployed during the pandemic. “It has been excruciating but the learnings have been massive.”

The U.K.-based Randomized Evaluation of COVID-19 Therapy (RECOVERY) trial would never have so quickly enrolled more than 30,000 patients in anything other than a pandemic scenario, Angus continues, even with its more traditional study design. UPMC, for its part, worked with electronic health records vendors Cerner and Epic to rapidly identify and enroll nearly a quarter of its hospitalized COVID-19 population into the Randomized Embedded Multifactorial Adaptive Platform for COVID-19 (REMAP-COVID) study.

Angus, Esserman, and Mogg were all fielding questions from forum moderator Marko Spasic, M.D., medical director of the Parker Institute for Cancer Immunotherapy in San Francisco.


Platform Trials

Esserman is founder and co-principal investigator of I-SPY 2, widely regarded as a pioneer of the platform trial. It is a phase 2 trial targeting locally advanced breast cancer with neoadjuvant treatment. Over the past decade, 26 drug combinations have gone into the platform and several have subsequently replaced a therapy that appeared statistically futile, she says.

Platform trials require an unprecedented level of efficiency and teamwork to get to learnings sooner, Esserman continues, but can also make trialists victims of their own success. For example, drugs may work only for early-stage disease or disease subtypes for which there are as yet no biomarkers.

Health systems are also not focused on capturing a “single source of truth” and curating it in a way both clinicians and researchers can use, Esserman adds. Trial sites may also be highly resistant to moving to a central institutional review board (IRB).  

When the pandemic hit, a 20-site network newly formed for I-SPY2 was repurposed for the phase 2 I-SPY COVID-19 TRIAL to rapidly test repurposed and investigational agents in treating critically ill patients, she says. It has been “a challenge” to work with pharma and sites to arrive at a more efficient process for running trials.

At the AACR forum, she reports the study has launched at 18 sites with over 750 patients enrolled. The idea is to efficiently identify good therapies for handoff to phase 3 trials. The I-SPY COVID-19 TRIAL is focused on the ICU population, she says, which “may be very different than those who are mildly ill.”

Similarly, the REMAP-COVID study is a sub-platform of Randomized, Embedded, Multifactorial Adaptive Platform Trial for Community- Acquired Pneumonia (REMAP-CAP) that evaluates treatments specific to COVID-19. Angus says he started moving away from traditional randomized controlled trials (RCTs) more than a decade ago after realizing complicated problems involving therapy combinations, specifically sepsis and acute respiratory distress syndrome (ARDS), required a less simplistic approach.

REMAP-CAP was “not quite” ready for COVID, Angus says. Nonetheless, over 5,000 patients with suspected or proven COVID-19 have been enrolled and, among them, roughly 11,000 randomizations have occurred. The trial is currently testing 31 interventions in up to 12 domains, which calculates out to over 5,000 separate combinations, he notes.

The trial went live with 290 sites (now 296) in 19 countries, says Angus. The list includes third world countries such as Pakistan and Nepal.

An ongoing adaptive platform trial is also underway by the Bill & Melinda Gates Medical Research Institute but is targeting the mild COVID-19 population at high risk of disease progression, says Mogg. The focus is on repurposing drugs already on the market to enlarge the benefit for low- to middle-income countries and already has 450 enrollees.


Defining ‘Adaptive’

Adaptive trials and adaptive platform trials are two entirely different things, notes Angus. The former uses a “suite of tricks,” even in otherwise traditional RCTs, such as pre-specified rules for stopping boundaries, study arms that open and close, “response randomization” stacked toward certain subgroups—usually but not necessarily using Bayesian design—and interim data analysis.

Platform trials, on the other hand, “use a feature in a repeated way question to question,” Angus continues. Multiple interventions get evaluated in a perpetual manner, under a single master protocol with standardized procedures, to minimize the downtime between trials.

“The platform is pretty inelegant unless you use adaptations, and adaptations are a waste of time unless they are rolled out as platform trials,” says Angus. The “two-horse race” is still important to learn if one intervention is superior to another, provided there are only two treatment options. If not, it might better “nest” within an overarching platform trial.

Mogg concurs, saying “you can always adapt.” She personally does a lot of simulation before settling on a strategy for adaptation. Interim data should be examined in “two-horse races” to determine when it makes sense to stop early for futility or efficacy, or to re-estimate sample sizes.

“Pre-specification” is the key word, says Mogg. “You can’t decide to change the primary endpoint midstream … usually, anyway.” Adaptive trials have a lot of built-in flexibility—allowing, for example, to start with a broad question and then narrow it down.

“Bias is the key thing to be concerned about,” Mogg says. Forethought is needed to mitigate risks.


‘Islands of Uncertainty’

Bayesian statistics is the “science of decision-making under uncertainty,” says Esserman. “Isn’t that medicine?” It is good neither to keep exposing individuals to a therapy that does not work nor to withhold one that might be better for them, she adds.

Leaders in the field need to be challenging accepted beliefs, including the fallacy that “all trials need to be blinded,” Esserman continues. She questions how someone involved in a large trial could even keep track of treatment effects in patients across hundreds of study sites. The “learn-as-you-go” approach could be the way to ensure patients are managed well.

Cancer is not a single disease, yet trials are designed as though it were, she says. Immunotherapies, for example, typically do not work well for metastatic disease.

Trials that move quickly and more often incorporate exploratory biomarkers are needed to truly move toward “personalized medicine,” she says. Health systems currently move at a “glacial pace... it takes three weeks to [modify] an order set to change drugs.”

Angus sees three levels at which adaptations might be incorporated into trials, and there is plenty of room for improvement, he says. “[Even with] the things directly in front of us in design, we can barely pull it off. We cannot get out of our own way. Logistics do not look like a problem on paper [and] a lot of people have veto power in the change.”

More challenging level two adaptations involve “nuanced treatment estimation with biomarker subtypes” and randomizing populations according to the hypothesized effects of an intervention, says Angus. While much has been written in the literature about individual treatment effects, biomarkers, and the combination of biomarker categories, “we have not quite worked out causal inference in clinical trials.”

The third, seemingly idealistic level of adaptations would be when clinical learning spaces are using observational data in a causal framework. “We’d only randomize when we have to,” he says, “when we have islands of uncertainty.”

That future might be a tad closer than some believe. I-SPY2 is “focused on specific subtypes, things that work,” Esserman points out. The trial introduces the concept of a “dynamic control” and uses 4:1 randomization so participants get the best therapy for their disease subtype.

Combination therapies are not inherently patient-specific, she says. “They need their day in court alone.”

Post hoc analysis would also be unnecessary if disease heterogeneity (e.g., mild versus severe) were leveraged in clinical trials in the first place, adds Esserman. The more typical scenario is that the same clinical trial gets built repeatedly, each with its own contract and IRB. “We’re tripping over ourselves.”


Disease Severity

In working with the mild COVID-19 population, it was a regulatory necessity to follow the same methodology as would be used for severely ill patients, says Mogg. This meant using a more traditional group sequential design that involves sharing controls and adding and subtracting compounds midstream based on how well they are working.

“The FDA [Food and Drug Administration] has been amazing through the whole [process] … and a critical piece of why we moved so quickly,” Mogg says. The agency was “engaged and had good turnaround time.”

Moving forward, Mogg says clarity about the objectives of different studies versus the “confirmatory aspect of drug development” will be important. The tendency has been to be “overly prescriptive" in early-stage drug development regarding emerging SARS-CoV-2 variants rather than learning from incoming data if the variants in fact make a difference.

“We’ll start to know if signals are not as large [and] treatment doesn’t work as well,” Mogg says. “[Adaptive platform] trials allow us to pivot at the right time,” which is why Bayesian-stye trials are popular with clinicians.

Angus says he classifies hospitalized patients as moderate or severe, with further subtypes within those two categories. “Severity and time are not the same thing,” he adds, noting that only some of the “moderates” progress to a severe disease state. Frequent randomization can be a highly complex exercise, he says, but the alternative is to overly simplify the disease process.

He cites a multi-platform RCT of heparin using a common statistical plan. After a few thousand patients had signed up, the data safety monitoring board stopped the enrollment of severe patients for whom the therapy was “futile and arguably harmful” while also stepping up randomization of moderately ill patients into the active treatment arm, Angus says.

At interim analysis, heparin had a 99.5% probability of being beneficial to the moderates—and the finding was consistent across trials, he explains, “In Brazil, [investigators] thought randomization was broken so many [participants] were going into the heparin arm.”

“We say we are objective but most of the time we are not,” says Esserman. “We do not allow data to teach [us].”

Esserman works with the highest risk patients in I-SPY2 because of the breast cancer “screening dilemma” that can give a false impression of benefit by increasing the overall number of cancers detected but not the cases that will be potentially lethal. In I-SPY1, 85% of high-risk cancers were not screen detected.  “It’s a different disease. We need adaptive screening.”

In her view, “the [COVID-19] pandemic is a crisis because a small proportion [of infected individuals] get very ill and a substantial portion of those die. That is the problem… the disease… what is shutting down the country,” Esserman says. “We need to focus on that and do a signal-finding phase 2 [trial], and pipe things that work into phase 3 trials.”

The silver lining here, she adds, is that the infrastructure built for adaptive platform trials can “port over” to COVID-19. “It doesn’t matter if it’s COVID or cancer if you use systems thinking.”

That will of course take the cooperation of pharma, the FDA, and clinical trial sites, Esserman says. COVID-19 already has companies “digging deep” into their pipeline and working around the clock to simultaneously save lives and their livelihood.

“I hope we don’t lose that sense of urgency as we move to [disease such as] ARDS and breast cancer when COVID is over,” she says. “We need to focus on real problems, commit to finding signals, and moving forward fast, with renewed commitment to reimagined trials and reimagined care.”