Multi-Platform RCT

In this episode of "In the Interim…", Dr. Scott Berry details the design, execution, and results of the multi-platform randomized clinical trial (mpRCT) pioneered during the COVID-19 pandemic. He describes how REMAP-CAP, ATTACC, and ACTIV-4a—each developed independently—pooled data prospectively for joint analysis to address therapeutic anticoagulation in hospitalized COVID-19 patients. Scott outlines the operational rigor required to harmonize endpoints, establish monthly adaptive analyses, and stratify patients by disease severity and D-dimer level. He examines the unified Bayesian hierarchical modeling approach, dynamic borrowing across strata, and the process for simultaneous DSMB reviews coordinated across all platforms. The mpRCT framework enabled real-time, evidence-based adaptations and rigorous distinction of treatment effect by patient subgroup. Results were incorporated into clinical guidelines because prospectively specified analysis revealed benefit for moderate patients and futility or harm for severe patients—findings that would have been missed by standard post hoc pooling.

Key Highlights

• Integration of REMAP-CAP, ATTACC, and ACTIV-4a under a prospectively unified analysis plan.
• Primary endpoint and stratified patient subgroups defined in advance.
• Monthly adaptive analyses using a shared Bayesian hierarchical model.
• Simultaneous oversight by joint statistical and DSMB committees.
• Superiority of therapeutic anticoagulation in moderate, non-critically ill groups; futility and possible harm in severe patients.
• mpRCT model established a framework for future global multi-platform trials.

For more, visit us at https://www.berryconsultants.com/

Creators and Guests

Host

Scott Berry

President and a Senior Statistical Scientist at Berry Consultants, LLC

What is In the Interim...?

A podcast on statistical science and clinical trials.

Explore the intricacies of Bayesian statistics and adaptive clinical trials. Uncover methods that push beyond conventional paradigms, ushering in data-driven insights that enhance trial outcomes while ensuring safety and efficacy. Join us as we dive into complex medical challenges and regulatory landscapes, offering innovative solutions tailored for pharma pioneers. Featuring expertise from industry leaders, each episode is crafted to provide clarity, foster debate, and challenge mainstream perspectives, ensuring you remain at the forefront of clinical trial excellence.

Judith: Welcome to Berry's In the
Interim podcast, where we explore the

cutting edge of innovative clinical
trial design for the pharmaceutical and

medical industries, and so much more.

Let's dive in.

Well, welcome everybody
back to In The Interim.

I'm your host, Scott Berry.

All right.

We cover a number of different
topics on In The Interim, uh, w- some

recent sports editions and all that.

Today I'm gonna go back a little bit, and
I'm gonna go back to a really interesting

effort of multiple clinical trials coming
together for questions during COVID.

This effort, uh, was given a
name called the Multi-Platform

Randomized Clinical Trial.

Now, I'm gonna come back to the name.

It's a, it's a really important
part of the story actually.

But let me set up the story, and I
think it will make a little bit more

sense as to why that name is, is such
an important part, uh, of the story.

Okay.

A great deal of people, time, and
effort went into this, so I'm gonna

give some, some explanation of this.

I'm gonna walk through a number of
people from here at Berry Consultants,

Roger Lewis, Lindsey Berry, uh, Liz
Lorenzi, Mark Fitzgerald, Michelle Detry,

Anna McLaughlin, Christina Saunders.

We have huge numbers of people
globally working on this.

I can't possibly mention everyone, um,
uh, in this, but a, a huge effort and

I, and I think you'll, you'll hear why.

Okay.

So there were...

A long time ago in a place far, far
away, there were, there were three

different trials, and I'll, I'll, I'll
try to set them up briefly and then

describe the intersection of them.

So plat...

Yeah, you, you, you're probably aware
that platform trials played this enormous

role during COVID for finding therapies
that did and did not work for the

treatment of patients with COVID-19.

Vaccines were a whole different issue,
but for therapeutics to treat COVID,

the Recovery trial, the Principal
trial, the REMAP-CAP trial, the, the

active programs, the US NIH, the US,
um, uh, Operation Warp Speed efforts.

There was ICE by COVID, the Together
trial, Solidarity, DNDi, ATTACK.

So multiple platform trials
made huge impacts during it.

So I'm gonna describe three
of them somewhat briefly.

REMAP-CAP, which you're gonna hear
if you tune into this, uh, uh,

podcast for, for more interims.

You're gonna hear much
more about this effort.

Amazingly, we are on episode
sixty-two Two or three, and I've

not talked about Remap-Cap yet.

It's, it's an incredible effort.

In part, it's just such a, um,
uh, enor-- it was such an, an

enormous and very cool effort.

I, I haven't quite figured out
how to do podcasts about it.

So this is, uh, Remap-Cap.

Remap stands for Randomized Embedded
Multifactorial Adaptive Platform.

It was built in 2015.

It was built for a potential pandemic.

Yes, 2015, and you know COVID-19
was largely a 2020, uh, uh, disease.

So Remap-Cap was built in 2015 and started
enrolling patients in multiple countries

that had community-acquired pneumonia.

The idea was that if there's a
pandemic, it's very likely it's going

to go through intensive care units.

It's going to be something that looks
like community-acquired pneumonia,

and the trial started enrolling with
multiple therapies that are very

interesting in and of themselves in
non-pandemic community-acquired pneumonia.

And a master protocol was built, and
it built something called a sleeping

strata that was-- We called it the
pandemic strata, and it could adopt

and say, "We now have a pandemic.

We are enrolling that pandemic,
and we already have therapies.

We already have a database.

We already have sites enrolling.

This is a platform trial ready to go."

And in fact, that's exactly what happened.

So Remap-Cap adopted the sleeping
pandemic strata in February of 2020, very,

very early, and one of the domains...

Well, uh, uh, uh, let me
go backwards a little bit.

Uh, Remap-Cap from the beginning
adopted two stratifications.

Now, stratifications in that trial
are really, really, uh, different

than what you m- might be used to.

That, that implies that therapies
are modeled f- prospectively,

that they may have differential
efficacy in those groups, and it

was severe state and moderate state.

All of these patients are hospitalized.

Moderate state are hospitalized but
are con-- are not considered severe.

It's sort of the complement of severe.

Severe state is that you're
hospitalized, and you have ICU-level

Organ support, cardiovascular
or respiratory organ support.

Largely ventilator, vasopressors, uh,
ECMO, uh, uh, ICU-level organ support.

You are severe state.

Moderate state, you're hospitalized
in the ward, but you don't have

severe-level, um, um, organ support.

The primary endpoint is for the
pandemic strata is organ support free

days, and that's through 21 days, and
it actually had a neat aspect of it.

So this is an ordinal endpoint where
mortality is the worst outcome.

We, we refer to that as a minus one,
and that was actually 90-day mortality.

If a patient died through 90 days,
they are considered a minus one.

Actually, I, I wanna be...

I wanna, uh, make sure I'm clear on that.

It's in-hospital mortality, and that,
that extends through 90 days of exposure.

So if the patient dies in their hospital
visit, they are considered a minus one.

And then it's ordinally the days
they are free of organ support, given

they survive and leave the hospital.

So zero is the second-worst outcome.

It means you survive and left
the hospital, but for 21 days you

were not free of organ support.

And then one, two, three, all
the way up to 21 days, you were

free of organ support, which
can't happen in the severe state.

You start on organ support.

There, there were many domains
that were started in this, and

this is a multifactorial platform.

They were enrolling antibiotics, steroids,
uh, from the beginning in, in CAP.

As soon as the pandemic came, they
started investigating steroids, macrolide,

antivirals, immune modulation, uh,
interferon, Anakinra, tocilizumab,

sarilumab, convalescent plasma.

And then they adopted very early
in March a domain that was looking

at therapeutic anticoagulation.

And the notion here is that the,
the cytokine storm at the time was

discussed as the, the, the coagulation
was part of that within the body

that was causing the severe disease
of COVID Now I'm a statistician, so

please don't take this as medical, but
that's largely in my statistical mind

what was going on, and the idea is
that would therapeutic anticoagulation

heparin be beneficial for patients?

Would it improve their survival
in getting off of organ support?

So a domain was adopted with two arms,
standard dose thromboprophylaxis,

so, uh, uh, the, uh...

Consider this low dose, which is
given standardly or therapeutic

dose, a high dose of anticoagulation.

Is that beneficial?

And so it started randomizing
those two options.

Patients and doctors were not blinded
to which arm they were in, uh, uh,

within this setting, but that was
adopted by the, the REMAP-CAP trial.

It started enrolling, sorry,
in April of twenty-twenty.

ATTACK is a, uh, Canadian-funded, and
at the time, a Canadian-funded platform.

The initial goal of this was to
investigate exactly the same question,

therapeutic anticoagulation versus
prophylactic dose anticoagulation.

Enrolling moderate and severe in
Canada, actually, this extended a

little bit into the US, South America.

So this was a platform for
investigating the exact same question.

They also, interestingly, in moderate
disease, they stratified by D-dimer level.

D-dimer is a biomarker that i-it has
some indication of high levels of fibrin

in the blood, es-essentially suggesting
coagulation is going on, intravascular

coagulation, and the notion is this
thought that this might measure,

again, as a statistician, levels of
coagulation, the cyto-cytokine storm.

These may be patients that particularly
have differential effect due to

therapeutic dose anticoagulation.

So their design stratified
severe disease and then within

moderate, high and low D-dimer.

So really three different groups the
trial could come out with a conclusion.

Originally, their endpoint was
a twenty-eight-day, are you

alive and free of organ support?

As we're gonna see, they start
to work together, and they adopt

the same endpoint as REMAP-CAP.

ACTIV-4-A was one of the...

A third platform now, was a, is
an NIH-funded trial investigating

It ended up investigating multiple
things, but the first thing to

investigate is therapeutic dose heparin
versus prophylactic dose heparin.

There are multiple investigators
involved in these trials that are

involved in, in, in the same trials.

Uh, me personally, I was a statistician
for each of these three trials.

So I had been working on REMAP-CAP
since two thousand and fifteen.

I still work on REMAP-CAP today.

Uh, ATTACK, we had worked with the
Canadian investigators, uh, uh,

Patrick Lawler, Ryan Zarichansky,
and, um, and many others in that.

And we were working with
them on their design.

And then when ACTIVE4 came around,
we worked with multiple other

statisticians on the design of this.

And that was the first
question for them, therapeutic

anticoagulation versus prophylactic.

Again, hospitalized patients.

Organ support-free days
is the primary endpoint.

And now this is somewhat of the third to
be adopted, and at this point, there's

already discussions about these three
p- platforms doing very similar things.

So they adopt organ support-free
days as the primary endpoint.

They are going to stratify
differential conclusions by

moderate and severe, and also within
moderate by the two D-dimer levels.

They adopt...

By the way, all three of these
trials adopt a Bayesian approach.

And maybe that was correlated to me
being involved in these, but they

all three adopt a Bayesian approach.

So the que-- the, the, the landscape here
is this is globally considered a very

important question, and there are these
three trials that are all separately

addressing the exact same question.

So what happens when one of them
reaches a conclusion, maybe the other

one doesn't reach that same conclusion,
but maybe it has po- similar data and

papers come out at different times?

It was thought to be, uh, somewhat messy.

Okay, we can do a meta-analysis
of the three trials.

They don't carry the same weight.

Their perception of them
is somewhat different.

And so the three trials are now
talking and figuring out, "What

do we do in this scenario?"

The most important part of it is each
of the three trials working separately

are going to take longer to come
up with an answer to this question

in the middle of a raging pandemic.

Twenty twenty Where the pandemic
is globally raging, and this is

an important question, all three
are gonna take longer separately.

So the investigators, the funding
groups, everybody comes together and

say, "We're gonna work together, and
we're gonna actually do it in a way.

We're gonna create something brand
new, and we're gonna create something

new, and it's gotta be differentiated
from a meta-analysis because it is

different than a meta-analysis."

Statistically, it has components
of that, but that this is...

We're gonna call it a multi-platform
randomized clinical trial.

It's a randomized clinical trial.

The three efforts all decide to pool their
data together into a single analysis.

They're gonna create a joint
analysis plan where they all

sign on board prospectively.

They will not read out separately, so it's
not that you're gonna see a publication

from REMAP-CAP and then the MPRCT.

They all agree they'll publish
together, they'll com- they'll

do their adaptations together,
they'll combine their data together.

This is a randomized clinical trial.

It's just that these three trials all take
their data from all of their global sites.

They put it together prospectively
in a, in a combined analysis.

Hence, we called it the
multi-platform, rather obviously

the three platforms coming together.

It's a randomized clinical trial, and
we wanted to make sure that message got

across that this is not a meta-analysis.

This is the prospective primary
analysis of all three trials.

There was quite an operational effort
to this, as you can imagine, that all

three of these trials combine their
outcome data together into a, an, an

unblinded statistical analysis committee.

That committee carries
out the primary analyses.

They create a efficacy report, and then
they talk to the three different DSMBs.

Each of these platforms have a DSMB,
and they did that simultaneously.

So these three DSMBs all come together
and meet together because they're

hearing the same analyses of the
combined data of the three trials.

Okay, so an incredible effort.

There is a statistical analysis plan
agreed to by the three platforms.

Uh, uh, that's finalized
on August 29th, 2020.

And, uh, ACTIVE starts enrolling in,
I, I, I think it was largely, uh...

Do I have that written down?

ACTIVE, uh, starts enrolling patients,
I believe summer, June-ish Within that.

So all the trials I have started
enrolling, no analyses have been done.

August twenty-ninth, the
analysis plan is finalized.

They're gonna do monthly
a-adaptive analyses.

The adaptive analyses which could trigger
in severe disease is its own group.

And then in moderate, there's
low and high D-dimer levels.

So there are three inferential groups
where we could reach differential

occlusions, uh, conclusions or the same.

What happened was early on in
the trial, they ended up with a

large amount of missing D-dimer
data for moderate patients.

So we had moderate patients
with low, high, and there

was a group that was missing.

We created our own group and
said it's missing D-dimer.

But it's really awkward to come up
with a conclusion, so we, we reported

on that strata, and it was part of
the modeling, but it couldn't trigger

separately because that would be
really awkward to say, "Well, if you're

missing D-dimer, here's what we think."

So those three other groups
could have adaptive conclusions.

Could conclude superiority if the
probability that the odds ratio

for proportional odds ratio model,
Bayesian model for organ support-free

days, this combination of mortality
and days free of organ support, if

the probability of superiority is
greater than ninety-nine percent.

We're gonna do monthly analyses.

If it's above ninety-nine percent,
we're gonna call that a trigger.

We're gonna say it's futile if
the probability is greater than

ninety-five percent that the odds
ratio is less than one point two.

That was considered a reasonable
clinical threshold where we're gonna

say the effect is a-at best small.

And harm is a ninety-nine percent chance
that it has an odds ratio less than one.

So that's the design.

Single Bayesian analysis combining
all these patients together.

A critical part of the primary analysis
is a Bayesian hierarchical model that

allows dynamic borrowing, and the model
was really a model that borrows moderate

patients, low and high, and missing
D-dimer are part of a hierarchical model.

The mean of that model is in a
hierarchical model with the severe effect.

So it's this two-tiered hierarchical model
where the moderate D-dimer levels can

borrow from each other because they're
all within the same disease state of

moderate And then the moderate and severe
effects can borrow if they're similar.

Okay.

Covariates adjusted in the model, regions,
sites, age, critically important, time.

Uh, there are no
non-concurrent controls here.

Uh, all randomized concurrent
controls within them, but we adjust

for time because it's such an
important thing within the pandemic.

It-it-- Both for disease variations,
but also because of, of, uh, times of

surge where maybe outcomes are different.

Okay.

In the first adaptive analysis,
November 20th, twenty-twenty, no...

I-I'm blinded to all of this.

Uh, no triggers are met.

We're told to continue.

Um, the DSMBs all meet,
no triggers are met.

On December nineteenth, six days
before Christmas, twenty-twenty,

adaptive analysis number two occurs,
and a severe state triggers hit.

Within the hierarchical model, there are
twelve hundred and seven patients are

met, and therapeutic anticoagulation meets
futility, meaning ninety-five percent

chance or higher that the effect has
an odds ratio less than one point two.

Those results are publicly disclosed,
but the-- no paper comes out.

Uh, randomization to severe is stopped
in all three platforms, and publicly

disclose this, this futility trigger.

So patients immediately in all
three of these global platform

trials stop randomizing to
therapeutic anticoagulation.

Now, the, the, the message is
continue in the moderate state.

So all three platforms are still
enrolling in the moderate state.

We don't know the answer yet.

Adaptive analysis number
three happens January 22 2021

At that analysis, superiority
for therapeutic anticoagulation

in the moderate state is met for
both high and low D-dimer groups.

2,200 patients, just over 2,200 patients
go into that analysis, and superiority

greater than a 99 chance therapeutic
anticoagulation is beneficial in

patients in the moderate state And
it's the same conclusion for D-dimer.

vary by D-dimer.

Randomization is stopped in
each of the three platforms.

Those, uh, results are announced.

The, the data aren't disclosed to
the level of typical publications.

That comes out.

So I, I, I, I'll, I'll come back to that.

So two papers are published in the New
England Journal of Medicine, and they

are published side by side, and they come
out in the summer of twenty twenty-one.

And you can, uh...

easy to find.

Therapeutic anticoagulation with
heparin in critically ill, that's

what they refer to the steer--
severe state, patients with COVID-19.

And then therapeutic anticoagulation
with heparin in non-critically

ill patients with COVID-19.

Back-to-back papers in the New England
Journal of Medicine report out on

this, and the fascinating thing
about that is the data are the same.

One model is run, and the results are
presented in two different papers.

And, um, the result of each is
done separately, yet 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 model, uh, learned that the effect
was differential and didn't borrow much.

But yet the component that borrowed
in the moderate state between D-dimer

levels did shrink those values together,
enabling the conclusion to happen that,

uh, uh, therapeutic anticoagulation
is beneficial in the moderate state.

Now, these conclusions, th-this effort
coming together, the whole idea of this,

if they would not have come together,
it's very likely these conclusions

would have been months later and may
have been differential because they

didn't combine them together, leading
to even more conclusi-- more confusion.

So they all come together.

They read out.

The patients are combined together.

It's done faster and more effectively
by the three groups doing this together.

Now, it was an incredibly
incredible operational exercise

to have this done all together.

Huge amount of time and effort done
to have this happen, but it made

a huge impact in the disease Now,
in the moderate state, the final

primary analysis comes out with a
ninety-eight point six probability

that therapeutic dose anticoagulation
is superior in the overall group, and

it's ninety-seven and, uh, ninety-three
and ninety-seven in the three groups.

In the Bayesian hierarchical model,
the estimate o-overall is about a one

point two seven, one point three one,
one point two two, one point three two.

So one point two seven is the,
uh, uh, moderate state estimate

with the bottom of the c-credible
interval being one point zero three.

So statistically significant
with that ninety-eight point six

probability, it's beneficial for
patients in the moderate state.

Now, in the severe state, the probability
is ninety-five percent that it is harmful.

That-- And the odds ratio
is point eight three.

We set up the odds ratio, so an odds ratio
less than one were negative outcomes.

Increased mortality and organ
support through-throughout the scale.

Ninety-nine point nine percent probability
it was futile, less than one point two.

The top of the ninety-five percent
credible interval is one point zero three,

exactly where the moderate state bottom
of the interval was, one point zero three.

Ninety-nine point nine probability
of futility, ninety-five

percent probability of harm of
therapeutic dose anticoagulation.

So this effort was unbelievable in
coming together, the impact of this.

The result is an amazing thing
of these trials together.

And let me sort of...

The, the, the role of borrowing
within these analyses and the

role of prospectively identifying,
these were not post hoc subgroups.

These were the analysis plan started in
2015 actually, in the Remap-Cap trial

and adopted by all for this domain.

If you would have pooled the data
and said, "We want one conclusion

in COVID-19," the odds ratio,
pooling those groups together,

adjusting for all of the states and
covariates, would have been 1.03

Essentially 1 with the
bottom of the interval 0.85

the top 1.22

It would have reached a conclusion of
futility, uh, a-and and no difference.

Therapeutic anticoagulation doesn't
matter If you would have pooled them

all together in these two groups.

Remember the conclusion through
the Bayesian hierarchical

model was harm, likely harm,
in severe, benefit in moderate.

Guidelines adopted these results.

There's a lot of communication to
guideline results that these weren't post

hoc, prospectively set up in the SAP.

Because of that, therapeutic
anticoagulation is given to

patients in moderate state.

It is not in severe.

Unclear what happens if they
don't do this differential

analysis, hierarchical modeling,
uh, of these results within this.

Okay.

Uh, I, I-- and again, go to
these papers and look them up.

Uh, uh, it was amazing effort within that.

The, the, the, the take-home of this
is a brand new entity, multi-platform

randomized clinical trials.

We are talking about this in
a number of other efforts.

So this continues to go on that there's
multiple separate funded platforms that

are investigating the same question.

Rather than them, the two of them compete,
be the first one out, let's combine our

data together and have one conclusion.

It's a more heterogeneous patient
population, bigger data, faster results.

During this effort in COVID, it was
an amazing effort, amazing effort.

Differential HTE, the Bayesian modeling
completely changes the result around.

Benefit and harm rather than it doesn't
matter, which had been the answer

if that was not done in the trials.

All right.

I hope you enjoyed this
look back into COVID.

Um, we-- lots of people
jumper- jumping in in effort.

The, the scientific, uh, aspects of
this are very much worth revisiting,

even though COVID is, is not
nearly the, the, um, uh, medical

challenge that it was at the time.

Many scientific lessons, platform
trials, amazingly successful.

And here is a new entity called the
multi-platform randomized clinical trial.

You heard about those interim analyses.

We stay here in the interim.

Thanks for joining.

Until next time