Latest / Cardio Pod: Heart of the Matter / How Can AI Enhance Recruitment in Clinical Trials?

Transcript

How Can AI Enhance Recruitment in Clinical Trials? · Cardio Pod: Heart of the Matter

0:06
Well, good afternoon. I'm Chris O 'Connor, Principal
0:08
Investigator of the Heart Failure Collaboratory,
0:11
and we're here with Heart of the Matter, the
0:14
Heart Failure Collaboratory podcast. And today's
0:17
topic at hand is how can artificial intelligence
0:21
improve the conduct and recruitment in clinical
0:25
trials? Very exciting time for all of us in medicine,
0:29
and we have a distinguished panel today, and
0:31
I'm going to go around the room and ask. our
0:33
panelists to introduce themselves, and then we'll
0:36
jump into the topics. Dr. Cunningham. Hi, my
0:41
name is John Cunningham. I'm a heart failure
0:42
cardiologist at Brigham Women's Hospital in Boston,
0:45
interested in how we can leverage AI to improve
0:48
heart failure clinical care and clinical trials.
0:52
Dr. Martin. All right, Trey G. Martin. I'm a
0:57
heart failure transplant cardiologist at Cleveland
0:59
Clinic's main campus. and Director of Heart Failure
1:02
Population Health for the system. Dr. Bott. Good
1:07
afternoon, everybody. I'm Ankit Bott. I'm a cardiologist
1:10
and clinical investigator at the Kaiser Permanente
1:12
Division of Research with an interest in digital
1:15
tools and implementation science as they apply
1:17
to clinical trials. Elise Shalowitz. Hi, good
1:22
afternoon. My name is Elise Shalowitz. I'm a
1:25
Senior Clinical Research Coordinator at the University
1:27
of Colorado. Dr. Sopka. Hi, I'm Mitch Satka.
1:32
I'm chief of heart failure and transplant at
1:34
Nova Shara Heart and Vascular. Thank you for
1:37
having me. Dr. Lala. Hi, Anulala. I am a heart
1:42
function cardiologist at the Mount Sinai -Fooster
1:44
Heart Hospital in New York, and I'm particularly
1:47
interested in seeing how AI can significantly
1:49
enhance clinical trial recruitment and increase
1:53
efficiency, precision, and equity. And Dr. Jane.
2:00
Hi, everyone. I'm Sneha Jain. I am a preventive
2:03
and general cardiologist at Stanford University,
2:05
and I spend a part of my time on AI development,
2:09
deployment, and evaluation within our healthcare
2:10
system. Well, fantastic. Well, we know that clinical
2:15
trials are cumbersome, they're long, they're
2:18
clunky, they're expensive. I don't even know
2:21
why we do them the way we do them. And yet...
2:24
There's hope. I think we think that artificial
2:27
intelligence, the use of technology to enhance
2:31
clinical trial conduct is particularly exciting.
2:35
And there's an enormous interest from the entire
2:37
ecosystem that participates in the clinical trial
2:41
world to see if there are solutions to increase
2:44
the value. That is the quality, increasing the
2:48
quality, reducing the cost of how we develop.
2:51
Therapeutics, one of the missions of the Heart
2:54
Failure Collaboratory. And John Cunningham, I'm
2:57
going to ask you to kick it off and say, how
2:59
can we conduct clinical trials better by using
3:03
technology? Well, we're all very excited about
3:08
the potential of AI in large language models
3:10
and clinical trials, because I think anybody
3:13
who's used ChatGPT has seen just the incredible
3:17
explosion in the technical side of And the ability
3:21
of AI language model tools to summarize large
3:25
amounts of unstructured information and distill
3:28
them down into something that individuals can
3:31
work with. And that can work from a recruitment
3:33
and adjudication and a safety lens. And so we've
3:37
been excited about the potential of AI in all
3:39
areas of this, particularly in event adjudication
3:41
for safety and efficacy. So thinking of ways
3:45
that we can transform the unstructured data from
3:48
patient notes. into efficacy outcomes, safety
3:52
outcomes, as well as identifying patients who
3:55
may be eligible for participation in the trial.
3:58
So I'm really excited about all these areas and
4:01
eager to hear from the panel on sort of the individual
4:03
things that you're doing at your own institutions
4:05
to bring this technology from sort of concept
4:09
to actual implementation and patients. That's
4:14
fantastic. I mean, that's on the downside of
4:18
the clinical trials. Dr. Jane, what are you working
4:21
on at Stanford that excites you about AI, artificial
4:26
intelligence, and the way we might improve the
4:29
ecosystem? Absolutely. Thank you. Just a couple
4:34
of weeks ago, we announced a project we're working
4:36
on called Chatter, C -H -A -T -E -H -R. And it's
4:40
our way of talking with the patient's EHR across
4:44
all kinds of data, both structured, unstructured,
4:47
their notes that are in the EPIC system as well
4:49
as in Care Everywhere, which is EPIC's pathway
4:52
for looking at notes outside of the specific
4:55
healthcare system, to be able to identify and
5:01
find information about a patient in a way that's
5:03
so much more conducive to a busy clinician or
5:07
a busy study coordinator than typical search.
5:10
And you can imagine with this system, it can
5:12
become really easy to identify across very many
5:16
patients who meets inclusion -exclusion criteria,
5:19
the best way to contact them. And then it provides
5:22
a platform and foundation for that kind of work
5:25
within clinical trials, but then also just care
5:27
in general. And then if possible, after the podcast,
5:30
we can share some links if people want to learn
5:33
further. But that on the technical side is what
5:35
we're working on. On the non -technical side
5:38
or, you know, everything else that's needed to
5:41
help make AI as usable and as effective in our
5:46
systems as possible. For clinical trials, we're
5:49
really focused kind of on three areas. One is
5:51
how do we think about digital strategy and data
5:54
infrastructure? All of a patient's data that's
5:56
important does not live in one place. So how
5:58
do we think about that? Two, how do we think
6:00
about the business model of clinical trials,
6:02
which right now is very focused on clinical sites,
6:06
not really on the patient? And then third, what
6:09
do we need to think about in terms of the right
6:11
evaluation and monitoring of AI tools over time
6:14
to make sure that they are safe, they're effective?
6:17
What do we need in terms of regulation, not just
6:20
on a government level, but also within our health
6:22
systems and our clinical trial platforms? Well,
6:25
that's very exciting and really novel. And I
6:29
think we're going to be looking forward to hearing
6:31
more about these initiatives from Stanford to
6:35
move things along. Trey, tell us a little bit
6:38
about where you think AI could fit into the clinical
6:41
trial ecosystem and improve it. Yeah, you know,
6:46
I think that when you think about the conduct
6:48
of clinical trials, I think like every domain,
6:52
in healthcare, people are being asked to do more
6:54
with less resources. And so I think there's a
6:57
lot that we have to think about in terms of the
7:00
shortage of research coordinators, for example.
7:03
You know, if they are given a list of MRNs, for
7:07
example, of patients who could be eligible without
7:10
any of the justifications to go along with it,
7:13
they're not going to use it. And I've actually
7:16
had that experience more on the structured data
7:18
side in the past. And so... If what we need is
7:22
a user interface that actually cuts down time
7:25
for research coordinators and provides valid
7:29
justifications using AI and LLMs in order for
7:34
it to be used. I think the opportunity is huge
7:37
because the way we still recruit for the majority
7:39
of our clinical trials is through the contact
7:42
with the PI or co -Is who are aware of the study.
7:45
Direct clinical contact. That's still the bread
7:48
and butter of clinical trial recruitment. for
7:50
most of the country. Now, how do you capitalize
7:53
on the broadening of the search using LLMs? You
7:58
actually need to come down in the time used in
8:01
other circumstances. And I think for the recruitment
8:04
standpoint, I think that's the screening portion.
8:08
And we've deployed a technology within our health
8:11
system, within our firewall, a proprietary LLM
8:15
called Synapsus AI from a company called Diania
8:19
Health that has been able in our system after
8:23
doing some mapping of where structured data lives,
8:26
be able to combine structured and unstructured
8:28
data. And we've used them to recruit for several
8:31
trials, including a study I'm involved with called
8:34
Depletor, which is amyloid fibril removal in
8:38
TTR amyloidosis. And what we found is that when
8:42
you deploy this technology, you can actually
8:45
increase diversity. racially and in terms of
8:48
identifying variant amyloid, but also find those
8:52
who may not necessarily be seeing our advanced
8:55
heart failure group. So both geographically and
8:58
racial diversity with increased recruitment.
9:01
So I think it's an exciting technology, but we
9:04
still need a lot more kind of ways to disseminate,
9:08
ways to operationalize the technology and make
9:11
sure that our research coordinators are brought
9:13
along the way. in terms of providing feedback
9:16
to the vendors. Those are terrific points. So
9:21
we can eliminate perhaps some of the biases that
9:25
are inherent in our current way of recruiting
9:27
by taking this more structured approach. Ankit,
9:33
I'm going to ask you about take us one step even
9:35
before the recruitment, and that's protocol design.
9:41
As you know, you've been involved in designing
9:43
protocols and we get into this big event of inclusion,
9:49
exclusion criteria. And before you know it, everything
9:51
is put together as a very complicated protocol.
9:54
Can we use AI, predictive analytics to say, listen,
9:57
historically, these are the patients that are
10:01
going to be excluded based on these factors are
10:04
included. These are the patients that are going
10:06
to drop out. because they couldn't adhere to
10:10
the protocol. Can we use that type of approach
10:12
to design better protocols? Yeah, thank you,
10:16
Chris. I think that's a wonderful point. I really
10:18
see the vision for AI across the clinical trial
10:22
ecosystem. Trey and John and others have really
10:25
mentioned how it might hyperpower sites to be
10:28
more effective, to use their site coordinators
10:32
more efficiently and recruit at a greater speed.
10:35
Also, I think clinical trials have the power,
10:37
now that real -world data sets are being constructed
10:41
at scale, the application of AI to those types
10:45
of data sets may better inform clinical trial
10:47
protocols, the ambient event rates with contemporary
10:51
guideline -directed medical therapy for many
10:53
of the conditions in cardiovascular medicine
10:55
that are moving more quickly. And then also from
10:59
a sponsor standpoint, sponsors are conducting
11:01
clinical trials across many sites. Some of them
11:03
are conserved. They use the same sites for many
11:06
of their clinical trials and using AI to power
11:09
both the analysis of which sites to pick and
11:12
which protocol amendments to make. If they made
11:15
a particular amendment to their protocol and
11:17
could see how that might change the number of
11:20
potentially eligible patients at sites they use
11:22
frequently, that could be very powerful for those
11:25
study leadership to make. a decision based on
11:28
data rather than our current structure, which
11:31
is less data -driven, and actually speed up the
11:34
recruitment in clinical trials and make in -trial
11:37
planning. So I see it as AI really being transformative
11:40
across the ecosystem from trial planning to trial
11:43
recruitment and, as John mentioned, trial adjudication.
11:46
I mean, those are really important points. If
11:49
you think about, and we've talked about this
11:51
in collaboratory meetings, the TOPCAT trial,
11:53
the Data Safety Monitoring Board. This emerging
11:58
data coming in from Georgia and Russia that didn't
12:01
quite look right, but was coming in really fast.
12:04
And the native eye of the human eye could not
12:08
pick up the patterns fast enough before there
12:10
was the washing of all these patients randomized.
12:13
And it really took down an important trial. Elise,
12:18
I'm going to ask you, you're the one that's on
12:20
the front line of all of us on this call. And
12:23
you're faced with... running and conducting the
12:27
monitoring audits and the monitoring sessions
12:30
when you're doing a clinical trial as someone
12:33
who's on the front line. How can AI, you think,
12:36
help monitoring in the conduct of clinical trials?
12:41
So much of it's labor intensive. It's like the
12:44
most expensive part of running a clinical trial.
12:48
Yeah, that's a great question. I don't know if
12:50
we'll be able to solve it, but I think that just
12:53
bringing forward the data, like we spend so much
12:56
time digging through charts and looking through
12:58
notes and trying to aggregate structured and
13:02
unstructured data and bring everything together.
13:05
So I think that if there's a tool that can really,
13:07
you know, surface everything in one, you know,
13:10
summarized format, we've talked a lot about that,
13:12
like the chart chat sounds so incredible for
13:15
that purpose, where you can kind of bring all
13:17
the points rather than kind of in your head,
13:19
bringing the notes and the structured data. into
13:23
one place, I think that's where it can have a
13:25
really great impact. Any other aspects from your
13:30
point of view, being so close to the front line
13:33
that you think AI could help with the conduct
13:35
of trials? I think with recruitment, we've touched
13:39
upon it a little bit already today, just, you
13:42
know, targeting those diverse and role and community
13:45
health system, you know, affiliates, because
13:48
we are only limited to those, like you guys said,
13:51
that the PIs, you know, really touch with, but
13:54
being able to disseminate, you know, through
13:56
BPAs or CDS alerts to providers who may not necessarily
14:00
be so in tune to the clinical trials underway
14:03
and kind of bring forward those patients who
14:05
may not necessarily be at the academic center
14:08
so that we can screen more broadly and more equitably
14:12
across the entire health system. Now, because
14:16
I write Dr. Sotka's paycheck, I can ask him any
14:20
question I want and, you know, Mitch, the concept
14:25
of not eliminating controls, but maybe augmenting
14:31
controls with the concept of digital twins. What
14:35
in the world is that, a digital twin? Can we
14:39
get to a point where we don't have to utilize
14:43
so much energy around control groups the way
14:46
we do now? I'm so glad you asked me, Chris. The
14:52
idea, and I think if we just take a step back
14:56
here, the potential with which AI is talked about
15:00
in general and in the clinical trial space is
15:04
often grandiose. I think if we look at this concept
15:09
of digital twins, it really comes from, as you
15:13
mentioned earlier, that it is really costly to
15:16
run a clinical trial. In the end, though, the
15:19
truth is that we still need the data that is
15:21
generated in a randomized fashion to understand
15:25
the effects of an intervention on a population
15:28
of people and whether or not that is beneficial
15:30
or harmful or really has no effect whatsoever.
15:34
And the truth here is that even when you're talking
15:38
about, I think, the use cases where AI has most
15:41
of its potential, when that comes to trial recruitment,
15:45
it still needs to be used as a screening tool
15:49
that then often requires manual confirmation
15:52
that the patients are correct. When you get to
15:56
this idea of a digital twin, meaning you create
15:59
a digital version of a patient that then you
16:04
can in silico express a treatment or not and
16:08
see what happens to them. And you can emulate
16:11
a clinical trial based on inputting vast quantities
16:15
of information, either known from clinical trial
16:18
data that already exists or from real world data
16:21
that is being collected. And you can pretend
16:23
that you know what would happen. That trial in
16:26
the current setting still needs to be performed
16:28
because we don't know. We don't have a detailed
16:32
enough understanding of. The physiology of the
16:35
medicine, of the interaction with the drug, of
16:36
the side effects, of the off -target effects
16:38
of a treatment to actually emulate what will
16:42
happen in silico and know what will happen in
16:44
the real world. So this idea of being able to
16:48
eliminate control groups, while a very attractive
16:54
ideal because it would be astoundingly cheaper
16:56
than our current clinical trial infrastructure,
16:59
currently remains unachievable based on our current
17:02
knowledge. and current understanding of how AI
17:05
works. Okay, so we're not ready yet, but you
17:10
haven't given up hope. Anu, where do you think
17:14
AI could have the biggest impact in our clinical
17:18
trial ecosystem? Yeah, thanks so much, Chris.
17:21
I'm so glad we're doing this because I feel like
17:23
we all are talking about AI and yet we don't
17:25
fully, we've seen glimpses of it work, but really
17:30
seeing it in full. fledged implementation in
17:33
our healthcare system still remains to be seen.
17:37
I would say Sinai, for example, has an AI institute,
17:41
a data science institute, and we're seeing pockets
17:44
of things come up. For example, like Epic Notes,
17:47
we have an automated generated reply to patients
17:51
when they send us messages, for example. But
17:54
it's still, like Mitch just pointed out, it needs
17:57
that human oversight. To me, what I'm most excited
18:00
about about, many people have already mentioned
18:02
this, is I oversee our clinical trial coordinators.
18:05
You know this well, Chris, because you and Mona
18:07
do such an amazing job with our interns. And
18:10
they are, especially with heart failure trials,
18:13
these patients are complex. There's multiple
18:15
comorbid conditions. There are multiple criteria
18:18
that need to be met, both that are found in unstructured
18:22
and structured sources. And the screening part
18:25
is really the pain point. And so I would say.
18:29
mitigating that process would be an ideal scenario.
18:34
And Trey, we heard, spoke about his experience
18:37
in this respect. And then also reaching those
18:40
individuals who are traditionally under -enrolled.
18:43
I think that's critically important. And then
18:45
I'm also really excited about this idea of AI
18:49
enabling deep phenotyping of certain subgroups
18:53
using multiomics, imaging, longitudinal data
18:57
to potentially define specific molecular and
19:00
physiologic subgroups that then we can learn
19:02
more from moving forward. So I think the promise
19:05
is tremendous. I am actually optimistic. The
19:08
question will be how things are implemented and
19:11
how they're implemented in a way that can be
19:14
translatable across health systems and not just
19:17
at Sinai or just... at Stanford, but what can
19:19
we ubiquitously use across all of our healthcare
19:22
systems such that we develop more of a uniform
19:26
way of implementing AI across regular care and
19:30
clinical trial design and enrollment? Well, that's
19:34
optimistic and I think hopeful. I mean, I do,
19:37
you know, EHRs, I think 70 % are, you know, are
19:42
utilizing Epic. So there's some... harmonization
19:45
of EHRs. And so tools that can go into the EHR
19:50
and identify patients. Dr. Jane, one of the concerns
19:55
we have here is that, okay, now there's a tool
19:58
to identify hypertrophic cardiomyopathy patients.
20:01
We're doing HC trials. There's a concern about
20:05
turning this on because we don't have enough
20:08
physicians. to see all those patients if we turn
20:12
it on and get this tidal wave of activity. And
20:16
then, you know, what are the liabilities? What
20:18
are the downsides of misdiagnoses, et cetera?
20:21
So tell us a little bit about your thoughts on
20:24
these tools that are doing, assisting in maybe
20:29
rarer forms of conditions that we recruit for
20:33
in clinical trials. I am so glad you brought
20:35
this up. I think that this is... This is critical,
20:38
right? Because everyone thinks that, oh, we can
20:40
identify more patients. That's great. But if
20:43
we don't have the capacity to actually help those
20:45
patients, what are we doing? So that's why part
20:49
of, I'm glad you brought up HCM in particular,
20:52
because we've been looking at this with the American
20:53
Heart Association. And part of what we're starting
20:56
to do now is create target product profiles for
20:59
health AI. in the same way that there are target
21:01
product profiles for drugs to try to meet clinical
21:04
and regulatory guidance or for public health
21:07
efforts with the WHO. And by doing these target
21:11
product profiles, we not only look at the technology,
21:13
but the workflow in which they're supposed to
21:15
operate. So for hypertrophic cardiomyopathy,
21:17
we mapped out the patient journey, if they were
21:20
identified with or without AI, found the bottlenecks
21:24
in our system. and identified how well a model
21:27
needs to work to actually move the needle in
21:29
terms of mortality or symptoms or whatever it
21:32
might be. This was for clinical practice. And
21:35
this blueprint, this target product profile,
21:38
also allows you to go to your leadership and
21:40
say, hey, if we turn this on, what's going to
21:42
happen if we move the lever of capacity, if we
21:44
hire more APPs in this area, if we increase our
21:47
echo capability? And I think that that's been
21:51
really powerful because we move from this reactive
21:54
nature where most of health care is right now,
21:56
where an AI tool is built, and then we try to
21:58
retrofit it into clinical practice and instead
22:00
start with what's the clinical reality now? How
22:04
can AI improve this and where are we going to
22:06
run into unintended consequences? And can we
22:08
build our business and healthcare delivery systems
22:11
around this? How can we innovate those systems
22:13
to actually enhance care, especially for these
22:16
rare diseases? So we've done this for HCM. We've
22:19
done this now for cardiac amyloid as well. And
22:22
so that's kind of how we're approaching it. Exciting.
22:28
Gosh, can we fast forward and get through all
22:30
the growing pains and get to the seamless clinical
22:34
trial that is much easier to do? When you look
22:37
at what the journey we've been on, these large
22:41
clinical trials take five years, seven years.
22:45
You know, you can only do so many of these in
22:47
a lifetime. And so we have to solve this problem
22:50
by being able to be simultaneous and doing many,
22:54
many trials in a shorter period of time. John
22:57
Cunningham, I'm going to give you the last word
22:59
from what you've heard. Summarize for our audience
23:02
what the hope is, what are the challenges? Well,
23:07
I think we've heard a number of examples of promising
23:11
areas and the sense that now the challenges are
23:14
less technical and more related to implementation
23:18
and getting this technology in the hands of...
23:23
practicing physicians and research coordinators
23:26
to really make this happen. I think there's a
23:30
number of different pieces of software that can
23:32
do interesting things and really penetrating
23:34
hospitals is the next step. And I think expanding
23:37
to some of these perhaps less glamorous but really
23:42
important clinical trial areas like monitoring
23:45
really offer some low -hanging fruit to decrease
23:50
cost, improve reliability, and avoid fraud, while
23:55
also starting with the task that's easily checkable
23:58
and really not dangerous to patients or trial
24:03
integrity at all. So I'm really excited about
24:05
that going forward. Well, terrific. Well, I want
24:09
to thank my colleagues here today for starting
24:12
the conversation. This is not one that will end
24:14
anytime soon, but we've started this conversation.
24:17
This is a major emphasis of the Heart Failure
24:21
Collaboratory, and I want to thank you here today
24:24
for being on this podcast, Heart of the Matter,
24:26
Heart Failure Collaboratory, and we'll see you
24:28
next time. Thank you.