How a Multi-Platform Randomized Clinical Trial Impacted the COVID-19 Pandemic
Coordination and Preparation in a Pandemic
At the start of the COVID-19 pandemic, clinical researchers faced an urgent, global challenge: derive meaningful, actionable answers about effective (and harmful) therapies for hospitalized COVID-19 patients amidst a rapidly evolving disease landscape. Three major platform trials: REMAP-CAP, ATTACC, and ACTIV-4a each launched to interrogate therapeutic dose anticoagulation in this patient population.
REMAP-CAP (Randomized Embedded Multifactorial Adaptive Platform) was designed in 2015 specifically to be “ready for a potential pandemic.” Its protocol included a “sleeping strata”, a pandemic domain that could be activated instantly in response to a respiratory outbreak. The trial enrolled patients not only in traditional community-acquired pneumonia, but rapidly adapted to COVID-19, switching on its pandemic structures in early 2020. Patients entered the trial in defined clinical states: moderate (hospitalized, not yet ICU-level organ support), or severe (hospitalized, with ICU-level support such as ventilation, vasopressors, or ECMO). The endpoint: organ support-free days through day 21, an ordinal outcome ranking in-hospital mortality as the worst result.
ATTACC began as a Canadian-funded platform investigating the same core question, but with a distinctive focus: moderate patients were stratified by D-dimer level, distinguishing high versus low in recognition of coagulation's mechanistic importance in acute COVID-19. ACTIV-4a, an NIH-funded US trial, adopted similar design and stratification features, and shared overlapping investigators with the other platforms. Each platform offered overlapping but non-identical approaches, but unified around therapeutic vs prophylactic anticoagulation for both moderate and severe groups.
Initially, the three platforms ran independently, each addressing the urgent question, should hospitalized COVID-19 patients receive therapeutic anticoagulation? However, simultaneous, uncoordinated results would risk confusion, delay, and contradictory publications. Post-hoc meta-analysis was inadequate: “the three trials are now talking and figuring out, ‘What do we do in this scenario?’” The solution: create a prospective, joint analysis—a new entity called the “multi-platform randomized clinical trial.” As described, “the three efforts all decide to pool their data together into a single analysis. They’re going to create a joint analysis plan where they all sign on board prospectively. They will not read out separately... they’ll publish together... combine their data together.”
Technical Design: Bayesian Adaptive Methods and Prospective Stratification
Agreement on technical and statistical norms was essential. The primary endpoint harmonized across platforms: organ support-free days, an ordinal, time-bound composite spanning both survival and liberation from ICU support. All three platforms adopted the same endpoint for COVID-19 in their pandemic domains.
Potential stratification of effect was prospective: severe (requiring ICU-level support), moderate with high D-dimer, moderate with low D-dimer, and moderate with missing D-dimer (modelled but not triggering results independently). All three platforms agreed to the combined analysis plan.
A critical design element was the single Bayesian hierarchical model used for primary analysis: “dynamic borrowing” occurred within moderate strata, and potentially across moderate and severe states. The design was “a two-tiered hierarchical model where the moderate D-dimer levels can borrow from each other because they're all within the same disease state... and then the moderate and severe effects can borrow if they're similar.” Covariates like region, site, age, and calendar time adjusted for the pandemic’s geographic and temporal heterogeneity.
The joint statistical analysis plan finalized August 29, 2020, formalized Bayesian adaptive decision thresholds for each pre-defined group. Rules: “superiority if the probability that the odds ratio is greater than 1 ... is greater than ninety-nine percent. Futility if the probability is greater than ninety-five percent that the odds ratio is less than 1.2.” “Harm is a ninety-nine percent chance that it has an odds ratio less than 1.” Monthly adaptive analyses assessed these triggers.
Operationally, “all three of these trials combine their outcome data together into… an unblinded statistical analysis committee.” Results and interim efficacy reports were presented by this committee to all three independent Data Safety Monitoring Boards (DSMBs), which met simultaneously to hear the unified analysis—a logistical achievement.
Outcomes: Speed, Clarity, and Scientific Precision
The first adaptive analysis in November 2020 yielded no triggers. By December 19, with 1,207 patients in severe state, futility was triggered: “meaning ninety-five percent chance or higher that the effect has an odds ratio less than 1.2.” All randomization to therapeutic anticoagulation in severe state was stopped—publicly disclosed. Enrollment in the moderate state continued.
By January 22, 2021, adaptive analysis showed “superiority for therapeutic anticoagulation in the moderate state... both high and low D-dimer groups. 2,200 patients... superiority greater than a ninety-nine percent.” This triggered immediate stopping of randomization for moderate state as well.
The final analysis made use of Bayesian shrinkage. As outlined: “the Bayesian hierarchical model is shrinking. Now, they're shrinking within the moderate state, potentially, and across moderate and severe. By the differential conclusion, you can guess, of course, that one was futility and one was superiority. The component that borrowed in the moderate state between D-dimer levels did shrink those values together, enabling the conclusion... that therapeutic anticoagulation is beneficial in the moderate state.”
Key statistics: 98.6% probability that therapeutic anticoagulation is superior in moderate state, with credible interval lower bound at 1.03 (point estimate ~1.27); in severe, 95% probability of harm, odds ratio ~0.83, 99.9% probability of futility. Without pre-specified heterogeneity, “If you would have pooled the data... the odds ratio... would have been 1.03... with the bottom of the interval .85, the top 1.22. It would have reached a conclusion of ... no difference. Therapeutic anticoagulation doesn't matter.” Only the prospectively stratified Bayesian model revealed likely harm in severe, benefit in moderate.
Results were published “side by side” in The New England Journal of Medicine; “one model is run, and the results are presented in two different papers.”
Guidelines followed: “Guidelines adopted these results. There's a lot of communication to guideline results that these weren't post hoc, but prospectively set up in the SAP... Because of that, therapeutic anticoagulation is given to patients in moderate state. It is not in severe.”
Rethinking Platform Trials for the Future
This multi-platform randomized clinical trial stands as a model for scientific coordination and analytic rigor. With joint protocols, real-time adaptive Bayesian modeling, and synchronized operational execution, REMAP-CAP, ATTACC, and ACTIV-4a delivered credible answers “faster and more effectively.” Had they not joined, “very likely these conclusions would have been months later... leading to even more confusion.” This approach set new standards for urgent clinical trial collaboration, scientific clarity, and improved patient care.