A Visit with Dr. Derek Angus

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.

Scott Berry: All right.

Welcome everybody.

Back to, in the interim where we talk
about all things clinical trial science.

And I have a, uh, distinguished clinical
trial scientists, uh, with me today.

So I have Dr.

Derek Angus.

Who is a distinguished professor and
holds the Mitchell p Fink endowed

Chair in Critical Care Medicine
at the University of Pittsburgh.

Uh, also secondary appointments
in medicine, health policy and

management, clinical Trans Clinical
and translational sciences.

He's been the chair of the Department
of Critical Care Medicine since 2008.

Is, and since 2015, he is
been the director of the, uh,

U-P-M-C-I-C-U Services Center.

He's done everything in critical care.

Uh, he's in Pittsburgh, uh, uh, in
the Department of Critical Care.

He's also a senior editor at jama.

So Derek, welcome to in the interim.

Derek Angus: Thank you, Scott.

It's uh, it's my pleasure.

Scott Berry: Yeah, so
this is a, a, a space.

We want to talk about all things
clinical trials, and we, I think we have

several topics we'd like to go to here.

And the first topic, uh, I think is, is
interesting in terms of clinical trial,

trial science, and something I don't
know much about, but steroids for the

treatment of community acquired pneumonia.

And maybe we should set
this up a little bit.

Um, my experience and, and Derek and I
worked together on the REMAP CAP trial.

We've worked together on multiple things.

Uh, I, I consider Derek
a very good friend.

Uh, so I'll get that, that, uh,
conflict of interest outta the way.

But let's talk about steroids.

So the history of steroids.

It'd be better if you give this history of
steroids from Cape Cod through Remap Cap.

What do we know about using steroids
in community acquired pneumonia,

and why is it so controversial?

Derek Angus: It kind of tires me just
to think about the topic, but, um,

all right, let's, let's, so first
of all, steroids are incredibly

powerful pleiotropic agents.

They do all sorts of things, uh, that,
broadly speaking, in this space of

critical illness can include dampening
down the immune system and can

include some potentially beneficial.

Uh, effects on your cardiovascular
system by helping, uh, uh, retain

fluids help improve resolution.

But they're pleiotropic in that
they do all sorts of things that

are disadvantageous as well.

Uh, every medical student learns about
all the pros and cons of prolonged

steroid therapy and so forth.

So they are powerful agents that do a
lot of things and they're dirt cheap.

So if they were to work, uh, and
you knew how to give them, uh, they

could be incredibly advantageous
worldwide in all sorts of settings.

Um, in the space of
community acquired pneumonia.

So first of all, we're not necessarily
talking about the pneumonia where

you never come to hospital, but of
all the pneumonia that gets you sick

enough that you need to go to hospital.

That also overlaps with getting sepsis.

If you come in with a pneumonia and
you get organ dysfunction, you've

effectively met the criteria for sepsis.

And so when you talk about
steroids for pneumonia, you're also

talking about steroids for sepsis.

In fact, pneumonia is the
most common cause of sepsis.

So there's a big literature on trying
steroids in both sepsis and in pneumonia.

And when you do a sepsis
trial, uh, of steroids, half

of the patients have pneumonia.

And when you do a pneumonia trial, half
of the pneumonia patients have sepsis.

So they're sort of, they're joined
at the hip, um, back in the eighties.

Uh, people were giving huge boluses
of methyl prednisone, like a gram of

prednisone, and people thought that
would instantaneously bring people

back from the brink of otherwise dying
within minutes of profound septic shock.

And then a couple of big trials said.

Oh no, this is actually killing people.

You, you might think that you get
some temporary stabilizing effects

on the cardiovascular instability,
but actually that's such a huge dose.

It probably causes secondary
infections, et cetera.

And so then steroids disappeared
and for a long time people thought,

yes, steroids always help in the
short term, but as best as we can

tell, they're not actually helpful.

And then around the turn of the
century, the French, uh, really led by

Jali and Ann started to suggest, um,
we weren't using steroids properly.

Um, and he in particular started thinking

Scott Berry: And, and properly, is it
mean, the right patient or the right kind?

Derek Angus: Yes, exactly.

So he came up with you.

He came up with, you need to
use a different formulation in

a different subset of patients.

Um, so, uh, he started experimenting with.

Effectively giving a week of
hydrocortisone, not a day of methyl

prednisone at a much lower dose
of sort of total steroid dose.

He also combined it with food cortisone
because he argued that that would

also be beneficial and he restricted
it to these really sick septic shock

patients that he felt had not responded.

To fluids and, and, and, and who
were so likely to die of the sepsis

that you should give them steroids.

That if you, if you used a milder or
broader cohort, you might accidentally

give patients who weren't going to die
anyway, but now you might increase the

likelihood of harm with the steroids.

And he, and he particularly also
said there's even a subset in whom.

They're ex, we all make steroids and
it, it's a stress response hormone.

And so he said you should even be doing a
test to try to understand whether your own

endogenous stress response is compromised.

Uh, and if you can't mount a good enough
internal endogenous steroid response, you

would be particularly likely to respond.

Bottom line is he had a
spectacularly successful study.

Suddenly the international guidelines
were saying, oh, you should give steroids.

They clearly work in sepsis.

And then people started playing
around with broader patient

populations, again, like pneumonia.

Uh, they said, well, maybe the
pneumonia that's particularly

pro-inflammatory might benefit the most.

Once you know you have antibiotics on,
you should maybe also give steroids.

And there were a number of.

Studies that were sort of grumbling
along, somewhat beneficial or not.

But meanwhile, in the broader area of
sepsis, no one could repeat Ali's study.

The Europeans tried it in
something called Corticus.

Um, uh, the, the international
guidelines, which were published

every four years, all through the.

21st century.

If you pull up the steroid section, the,
the literature was changing slowly, but

every four years the guidelines would
be written differently depending on who

was in the room because this was, uh,
what I would call an opinion rich data,

poor environment, and, and, and people
just felt, they continued to feel like.

Uh, everyone had their favorite study.

If you wanted to give steroids,
you liked the Annane study.

If you didn't wanna give steroids,
you liked the Corticus study.

So then people said, enough is enough.

We need some bigger,
more definitive trials.

And two groups decided to do much larger
trials in septic shock, and they're much

bigger than any of the pneumonia trials.

So, I know you asked about
pneumonia, but I feel like in many

respects, the sepsis literature.

Is the more dominant.

So, uh, the French did another much
bigger study called APROCCHSS um, and it

was published in the New England Journal
of Medicine, and yet again, in their

hands, they knocked it out of the park
with a big improvement in mortality.

Meanwhile, the Australians ran
the trial called Adrenal with

several thousand patients also
published in the New England.

No Benefit from steroids.

It's like, are you

Scott Berry: A, any, any harm?

Any harm potential from adrenal?

Derek Angus: So Adrenal didn't
really show any harm overall.

Um, but one obviously wonders if
the French were successful and if

the int criteria weren't exactly
the same and they were recruiting

a, a narrower subset was adrenal.

Neutral because it was including
the people that just looked like the

French in the APROCCHSS trial, but
also including other patients that

in whom there was no benefit and one
canceled the other, if that makes sense.

Now, some of this is published, some of
this is not published, but people have

then tried to find within adrenal the
subset of patients that look like the

patients that were in APROCCHSS to see.

Well, is there at least some suggestion?

Uh, but thus far, no one can really, no
one can find within the adrenal trial

a subset that looked like APROCCHSS who
then have a similar effect to APROCCHSS

So people are confused along the way.

Uh, the French also did another
trial called Cape Cod that looked

quite like the APROCCHSS trial only.

It was just in pneumonia, slightly
lower, uh, severity of illness group.

There again, they showed
fantastic benefit.

So every trial that's run
with flu drug, cortisone based

outta France has been positive.

There's a lot of ity there.

I don't know.

I don't know whether it's about.

Being French or if it's about giving
food, cortisone, no one has tested

food cortisone outside France.

Uh, and every time they test
it, it's been beneficial.

Um, and then the French then also went
back to approach, um, and looked at

the patients in approach the subset
of sepsis patients who had pneumonia.

And there too, they found most of
the signal in APROCCHSS appeared

to be in the pneumonia patients.

So, uh, Cape Cod was positive APROCCHSS
was a positive and APROCCHSS was

most positive within the patients
that had sepsis and pneumonia.

So this definitely.

I mean, if you are in Fran, if there
was only the French literature, this

would all be tied up with a bow on
it, which is you would give steroids

broadly to hospitalized pneumonia and
broadly to sepsis, and you would expect

the largest benefit in the patients
who had sepsis due to pneumonia.

And we would sort of wipe our hands
and go on to one of the many other

unresolved questions in critical care.

Uh, but the problem is we cannot
generate an evidence base outside

France that goes in the same direction.

There are some other pneumonia
trials with non mortal endpoints that

are somewhat beneficial, but, uh.

Most of the positive signal is driven
by these French trials, and so even

when you look at meta-analysis,
I would say most of that is being

driven by the French experience.

The French are now leading another
huge consortium, uh, to study sep

uh, steroids, particularly in sepsis.

But again, many patients who have
sepsis due to the mor, I forget

the acronym for it, but it's just
been funded by the European Union

and it's been led by Jali Nan.

And one of the things they're
going to be trying to tease out

is, uh, well, let me step back.

So.

I've somewhat jokingly said there's
the French experience versus the

rest of the world experience,
but, but more pertinently.

I think what we can say is we don't
get consistent results and across the

different trials, we give the drug
slightly differently and we select

the patient slightly differently.

We could also be giving co
interventions differently as well.

We also have a slightly
different duration of follow up.

You asked earlier on about harm.

Steroids always run the possibility
of being beneficial on the short end,

but then giving steroids has some
unwanted sequelae that manifests later.

So it could also be the timing and
the nature of the primary end part.

So when the evidence base is consistent,
consists of variable, variable

drug and dose, variable patient
selection and variable endpoint,

and then you don't get a consistent
result, you're left saying, okay,

uh, what's driving the inconsistency?

Now, obviously the endpoint
is relatively easy to fix.

In all of the different trials, you
could collect multiple different

endpoints and then try to measure a
common endpoint across all the trials.

So regardless of what was primary,
if you could, if you had long

enough and rich enough follow up,
you could, you could standardize.

Um, so then let's go back to the two
other plausible domains of heterogeneity.

Um, there, there, there
are, I should say, before I

Scott Berry: Well, we, we should wait.

Let's touch on, let, let's
touch on a little bit.

But then there was C and
Cape Cod, for example.

I don't think they enrolled influenza,
so they stayed away from, uh, influenza.

But then during COVID recovery, REMAP
CAP showed strong benefit of steroids.

For the treatment of hospitalized COVID.

Derek Angus: Yes, yes, yes.

So.

Uh, Hey, it's Friday.

It's Friday afternoon.

What can I say?

I'm not firing on all cylinders.

I should totally have brought
that up that prior to COVID, I.

We were a bit on again, off again
with steroids and we were moving

increasingly towards, well maybe in
the sickest septic shock patients, in

part because of the approach trial.

But we'd got away from giving it
more broadly in pneumonia 'cause it

just wasn't a strong enough signal.

Cape.

The, the non COVID Cape Cod hadn't
yet been published in so on.

However, then during COVID, which
is bad pneumonia, suddenly it looked

like steroids were beneficial.

And in the Dexamethasone experience
and recovery, the, the benefit

was all in the sicker patients.

Uh, which again, that just
almost reinforces that it's not

just in bacterial pneumonia.

It may be in general that,
that they, this sort of.

The severe end of this syndromic
experience of getting threatened lung

function from some invading body that then
could be complicated by multiple organ

failure sepsis, regardless of whether
that's driven by virus or by bacteria.

It seems.

Steroids are really doing something.

People also thought one of the advantages
of COVID is that it's, it's a much more

homogenous insult that in some of the
other trials you can't, sometimes you

can't even work out, is it pneumonia?

Like a ton of pneumonia,
classic pneumonia trials.

Many patients have no identified organism,
and so one of the longstanding things

was, well, you didn't even put the
right patients in the trial, whereas.

The COVID trials people said, well
now we, these patients are in hospital

for a pathogen driven pneumonia.

They're not just in hospital
for a pneumonia complex.

That might sometimes actually just
be heart failure that looks on

the chest X-ray, like pneumonia.

And so that, and so then people
sort of said, well, you definitely

want to give it in COID and
what's more, when we have a true.

A classically selected population in
whom we have very high confidence that

they definitely have an infectious
pneumonia and especially when they're

severely ill and getting altered,
you know, distal organ dysfunction,

IE sepsis due to the pneumonia.

We seem to get this pretty strong
homogenous sense of benefit from steroids.

That's one argument.

Uh, there are definitely.

Critics of these trials.

But for those who like them, they would
say, not only should you give steroids in

COVID, but while we're at it, it's making
us increasingly think maybe steroids.

You know, if you like peanuts,
you'll love peanut butter.

And so now the guidelines are, they've
totally swung to be very positive.

Current guidelines.

They all.

Differ slightly in who they say you should
give it to, and even which kind of drug

you should give, whether you should give
hydrocortisone or prednisone, et cetera.

But I would say in general, guidelines
from the different infectious disease

and cial care societies, the WHO, et
cetera, have all started to become much

more, they're, they're all down on,
they're all pretty positive on steroids.

Scott Berry: so we then we were
involved in remap cap in non COVI,

where we hydrocortisone and we were
looking at, uh, stratifications of

patients by shock, by influenza.

Yes.

No.

And that result showed.

Reasonable probability of harm.

It showed worse mortality
in all four of those groups.

We didn't have a whole lot of
influenza, but, uh, it, it was

pretty convincingly negative, 90%
probability of harm in, in cap, uh,

shock or not, uh, patience, again, a
differential result to all of that.

Derek Angus: Yeah, I hated that result.

Um, that was incredibly disappointing.

Uh, it,

so

I, I think we would agree that.

That trial didn't run quite
as well as we would've hoped.

Um, it was an open label trial and we
had worked with site who said that they

have enough equip poison that if the
patient is randomized to control, they

would be happy not to give steroids.

But there was bleeding of
steroid use into the control arm.

It wasn't massive, but
you never know whether.

You know, if it, if it's,
if it's random bleeding, it

probably doesn't matter too much.

It's just slightly weakening
any efficacy signal.

But if the physicians have some magic
inside track on knowing just exactly who

they should treat, then you could imagine
that any signal you would've found is

getting wiped out by, uh, contamination.

Having said that, I don't
think contaminate, so I

wish there hadn't been that.

Steroid use in the control
arm, but the steroid use in the

control arm wouldn't explain the,
almost hitting the harm signal.

Like, uh, steroid use in the control
arm would, if, if steroids would've

normally worked, all it would've done
is reduce the size of the benefit.

It wouldn't have flipped
it in the wrong direction.

Now obviously you, as you start
to get smaller sample size,

some of this is just sort of.

Random error and so forth, but that, you
know, it made me slightly uncomfortable.

We also, as you know, we had a.

Uh, there was a transposition error
by one of our data vendors early

on in the trial that sent us a
data set that had been mislabeled.

We, we, as soon as we found that out,
we sort of fixed it all and they're no

longer a data vendor even in the trial.

But, um, that did mean that there was a
period when we were probably randomizing.

Uh, with a more distorted sort of
response, adaptive the response.

Adaptive randomization triggered
a more aggressive randomization

away from what would've been ideal.

And again, the main consequence there
was probably, you know, we had several

hundred patients in this analysis,
but they weren't evenly distributed.

We would've had more power if it was
more of a one-to-one distribution.

Um, and so that again, just introduces
another level of uncertainty.

Uh, but, but what I don't want to
lose in any of that is while none

of these things were ideal, we were
a long way from looking anything

like the Cape Cod experience.

Scott Berry: Yeah.

Derek Angus: In, in other words, no
matter how much you subset the trial,

no matter how much you tried to, is,
for example, in the early part of the

trial before there was any response,
adaptive randomization whatsoever, when

this was just like a straight head to
head, traditional one-to-one trial.

Even there, the trend that, uh, the
patient's getting steroids, their outcome,

the, the observed outcome rate was worse.

Um.

So is it all just the play of chance?

Uh, possible, but, but to your point,
um, uh, uh, the, the probability

statement nonetheless would say that
even, even accounting for chance, it's,

it's just highly unlikely that, um,
that steroids were strongly beneficial.

Um, and.

I wanna remind people that
the adrenal trial with 4,000

patients found no benefit.

So it's not as if remap
cap was an outlier.

Um, remap cap in many ways was one more
trial consistent with the many trials

that have not actually found a benefit.

Uh, and earlier on you
had asked about harm.

Um, there was.

It's like the old one, the
big European Cardus trial.

Um, they had thought that
patients appeared to get slightly

faster resolution of shock.

And by the way, in remap cap, it looked
like the observed rate of cardiovascular

dysfunction appeared to resolve sooner.

And that was also what
the adrenal trial found.

That, that the early
short term cardiovascular

instability resolved sooner.

But in Cardus, where they had done
follow up on secondary infections,

secondary infections, which, you know,
steroids can cause, uh, were higher.

There was a higher nosocomial
infection rate, uh, in the steroid arm.

And I, it is possible that.

Uh, something was going on in Remap Cap
where we might have actually been getting

some of the early cardiovascular benefits,
but the net effect for the, for the cohort

in general was that steroids were worse.

And so now we get back to,

Scott Berry: Yeah.

Derek Angus: uh, can you better

Scott Berry: Okay.

So let's let Larry, yeah.

Right.

So one of the things that was sort
of striking in the remap cap trial is

that mortality was against steroids,
but organ support free days, which

includes mortality, but it includes
time to resolution of organ support,

actually was better on the steroid arm.

Despite the fact that mortality was
worse, it, it, it has the earmark of

another particular disease that was
somewhat striking is the treatment of

acute stroke, where early on in the
treatment of acute stroke and the use

of endovascular therapy, they saw this
non-proportional effect that they, they,

they certainly improved the number of.

Really good outcomes.

They certainly increase that, but
they increased the number of deaths

and it was a sign that there's
differential heterogeneous treatment.

In fact, they were probably treating
people they should not treat,

and they were causing harm on
those, and they were treating some

that they absolutely benefited.

And it took a while before they
kind of figured the right patients.

And then they saw huge benefits in that.

So, uh, the, the steroids
has this hallmark of.

Heterogeneous effect
across the population.

And I know you, you touched on, maybe
it's the infection, maybe it's the host,

maybe it's the timing of the disease,
maybe it's the steroid themselves,

dexamethasone, prednisone, that, the
type of that there's so much here.

So it's got the earmarks of heterogeneity.

You know, should we be running trials
that we enroll a thousand patients and

do a single analysis of is it beneficial?

Do we need better trials?

So let's sort of turn it to where do we go
and if you could design the right trial,

you seem, you seem to not know which
patients, which treatments, which time.

Shock based therapy,
uh, a seven day dosing.

You know, where would we go if
we could fund the right trial?

What's the right trial design?

Derek Angus: Yeah.

So right before, so I would love to
talk about that, but can I just ask,

Scott Berry: Okay.

Derek Angus: just make a clarifying, so.

Uh, when, when you have the organ failure
free days, if you died, any effect on

your organ dysfunction is sort of subsumed
within the fact fact that you died.

But you did allude to this all being
explainable by the right patient.

But it is possible with some of
these early versus late outcomes.

That's just the timing of the measurement,
even within an individual patient.

That that wouldn't be the case
when it's organ failure free days

because if you died you're dead.

But, um, uh, we know in critical care, for
example, um, low tidal volumes, which, um,

is a strategy of ventilating the patient
in a way very differently from the way

that we did for decades, um, would almost
certainly make the patient look bluer.

More hypoxic in the short run because
greater expansion of the lungs

actually facilitates gas exchange.

You get better oxygenation, but it turned
out that the low tidal volumes absolutely

protected the lung and improved survival.

And so high tidal volumes makes you
look better in the long, better in

the short run, but then kills you and.

Scott Berry: Okay.

Derek Angus: and so, and, and with
all of these interventions when we

move into these combined endpoints,
do so at our peril because they

do tend to rely on conceptual
models that could be wrong about.

Whether the intermediate outcome
is on the path to greatness.

You know, you, you would think a short
term measure to improve organ dysfunction

is surely helping you get out alive.

But the answer is not so fast.

So that's one issue.

But, but, but, but assuming we've got
the outcome right, and the problem

is heterogeneous patients and you
don't know which ones they are.

Um, I mean, this is a classic issue.

If, if you, if.

Uh, uh, so I feel, I feel like, um, uh,
there were always what I would call the,

the, the hope and pray models, which is
what everyone did before, which is you

just enroll everyone and hope that you
got the right group, or that if there's

a bad group, it's not so bad, it doesn't
drag it down and all the rest of it.

And then we wanted to start
having smarter trials.

If you have the luxury of doing many
sequential trials, you could always do

one large trial, see a group in whom you
thought there was a benefit, and then

do a follow up trial in just that group.

Even there, you don't necessarily
prove that there's heterogeneity.

You know, that was, I, I remember, I think
it was the editorialist, very astutely

pointed out that that huge Jupiter trial
in the New England, um, when they were.

Uh, treating acute myocardial
infarction patients with some

sort of anti-inflammatory strategy
just in those with a high CRP.

The authors then said, oh, look,
we, we improved outcome showing the

inflammation hypothesis and so forth
in myocardial infarction and, and, and,

and the edit in the editorial that said.

For that to be true, you would've had
to have enrolled both the inflamed

and the non-inflamed and showed
that the drug worked differentially.

You only showed that in the inflamed
that the drug worked, but you haven't

ruled out that the drug could have
been working in the other group.

Uh,

if, if, if we don't really have the
luxury to do lots of sequential trials,

but want to be learning within one trial.

Then, uh, we, we can
do a couple of things.

As you know, in remap cap, we've
tended to have these a priority

subgroups, um, that, that you could.

If, if you think you have a rough sense
of where the domains of heterogeneity

may lie, uh, across both the axis
and where the cut point in the axis

might be, then you could start a
trial with some embedded rules to then

effectively, um, uh, enrich over time.

Uh, uh, you can also do
it in the other direction.

You could deen enrich you.

You could start with the subgroup
that you think it works best in, and

as long as you're getting a benefit,
then you could then widen your entry

criteria for a broader set later.

I, I think those kind of trials still
rely on confident science or hubris

Scott Berry: Yeah.

Derek Angus: whether you have.

About whether you actually
understand the axes of heterogeneity.

Um, and so, and, and so

Scott Berry: Well, well,
doesn't, doesn't this now.

Yeah.

Doesn't this now demand a, a, a, a
different approach where the, the

amount of people you're caring for in
critical care that have sepsis cap,

moderate disease, severe disease, that
we enroll a very large, try, a very

large population and try to learn Yes.

No in whom?

Derek Angus: Yes, so, so this is where
you, so let's step back for a second.

So.

We're, first of all, you would say
we are committed to cause and effect.

We're not really
interested in association.

Uh, we want to know if something
works, but what we've done is

we've begin and RCT tells you if
something works, but it doesn't.

But it's like causality with
a small C, causality with a

big C is, why does it work?

And in a way, heterogeneity
of treatment effect is more

juda perel esque, if you like.

You know, it's more getting at
not just explaining if the drug

worked, but why the drug worked.

In fact, I think it was one of Judah
Pearl's, uh, postdocs, um, Barne who,

Elias Barnum I think, who started
using Judah Pearl's sort of, um.

Causal inference language to try to
understand generalizability in RCTs.

If you understood why, uh, effectively,
if you understood all of the moderating

and median effects between the covariate
structure of who was enrolled and

what the intervention was, then in a
way from one trial you could predict.

Exactly what the outcome would be
in another trial, even if it was a

slightly different patient population.

You tell me the admixture of patients, and
I will tell you the likely mean effect.

What, what you're effectively doing
is you're saying, I understand the

relationship between the baseline
variables and the, and the randomly

exposed intervention, and I'll
tell you what the effect will be.

Um.

The problem here is you quite
quickly run outta sample size.

Um,

and, and I mean, there's no,
there's no real magic around.

I mean,

in the ideal world, you
could imagine every time.

You were faced with a clinical
decision around which there was

reasonable uncertainty, then
you would almost want to be, uh,

randomizing among the best options.

Uh, so if the entire trial experience
was like a huge reinforcement learning

problem where you were sitting in some
markoff state where you had a current.

Of the world where when a patient came
into a particular state, based on the

cova structure among all the different
actions that could be taken, you

would have, you know, a probability.

Uh, of the eventual best outcome rank,
order for all the different options

and of the 30 treatments you could have
given, it turns out that there's two

or three that all look reasonable, but
none of, but they're all quite close.

And then you could say, okay, every
time I see this, it's, it's reasonable

to randomize to any of those two or
three options, but it's unreasonable

to randomize to the other 27.

Uh, and then you could imagine if hundreds
and hundreds and hundreds of ICUs, for

example, were all contributing to this
ongoing reinforcement learning model,

that this state could be a known state for
the, the agent, the agent being the, the

collective body watching the accumulation
of knowledge until say, several

hundred patients have gone through.

And, and then at then at that point you
could update the, the probabilities and

it would turn out, oh, remember how we
said there were three choices that looked

reasonable when you were in this state?

Turns out one of them is a
dud and now we're down to two.

So you would, you would essentially,
this is now you'd move into a new

state where you had sort of updated
probabilities and you would run that,

and you could imagine that that's really.

You, you know, this notion
of these freestanding RCTs

Scott Berry: Yeah.

Derek Angus: that give these single
estimates of a patient population

with an average effect, uh, for a
single drug at a single dose that

seems is a total misrepresentation of
the degree of uncertainty in which we

exist, and it's failing to leverage.

All of the uncertain clinical decisions
that are made in regular care.

And so in a perfect world then I am
massively leaping over all sorts of

issues about research versus clinical
care, the Belmont Report, et cetera.

But, but a, assuming you had the right
ethical framework for this, uh, assuming

that everyone was read in on this, that,
that they all approved, you could imagine.

A large federated set of learning health
systems where all of their electronic

records are all sort of wired together
and you have sort of some sort of

asynchronous API that could sort of read,
write commands that was able to stream

the information so that just at the
moment that the clinician was encountering

the patient and faced with uncertainty.

You could then pull from the data.

Aha.

I'll tell what this moment looks like
is it's a patient with the following

baseline characteristics about whom we
have uncertainty on the right option.

And so that's like a mini part.

It's like a mini randomization question.

Um, and then if people understood.

Yes, there's legitimate uncertainty
and it's, it's within the bounds of

reasonableness to randomly choose,
uh, what treatment they take or even

to randomly make a recommendation.

I mean, you could still, if you
want, you could still say the

patient and the physician are allowed
to override the recommendation.

Um, and if they did that a lot, then
the trial is getting contaminated.

But if it turns out that they
say, oh, that's, that's not

an unreasonable suggestion.

We're both happy with that.

If that was happening 99% of the
time, then effectively you're

running these huge, large,
embedded, ongoing randomizations.

And if it turns out that among all
of pneumonia and sepsis, there are

a thousand different phenotypes.

Um, all with d and, and some of them
cluster into net benefit, some of the

net, you know, net harm, et cetera.

Um, we could be learning about
them over time if you, I mean,

these diseases are common anyways.

Have I got way too far out over my skis?

Uh.

Scott Berry: well, well, no.

Let's, uh, let's see where
you got, you got that.

Currently, right now, we're not
learning from any of these patients.

They're being treated.

They're, they're, they're, and so now
we're gonna tie this all together.

We're going to embed randomization.

And we're going to favor randomization
of treatments that are working.

So we're gonna care for
those patients better.

But it would give us the possibility of
learning about heterogeneous treatment

effects, the, the signatures that
benefit, uh, uh, this we're gonna enroll.

Thousands of patients
in this common disease.

You, you've described a learning
healthcare system, and maybe in the

world of ai, this maybe the thought is,
okay, AI is now deciding what to give.

It's learning from it,
it's reinforcing it.

I mean, this is, this is the future and
we kind of wonder why it's not now, but

the, uh, I, I love where you've gotten.

Uh, so we've gotten from, uh, these
randomized trials that are showing

differential effect from steroids
to a learning healthcare system.

And I want to come back and
talk about that because I think

that's more than just a idea.

Uh, I think we, we have some really.

Uh, great hopes of embedded
randomization learning healthcare

system, call it ai, call it modeling.

Uh, underneath that
is, is pretty exciting.

So, Derek, I, I, I'm, I'm
thrilled that you were able to

join us here in the interim.

Um, we do lots of interim analysis here.

Uh, so we're gonna, we're
gonna have you back.

Uh, to talk about this, to talk
about learning healthcare systems

and maybe we'll have resolution
about steroids in sepsis and cap

Derek Angus: We can only hope.

Scott Berry: we go.

Alright, so thank you Derek for
joining us here in the interim.

Derek Angus: thank you so much, Scott.