Latest / Becker’s Healthcare Digital Health + Health IT / Scaling AI Responsibly in Cancer Care

Transcript

Scaling AI Responsibly in Cancer Care · Becker’s Healthcare Digital Health + Health IT

0:00
Hello, everyone, and welcome to Becker's Healthcare podcast.
0:02
I'm Scott King. Thrilled today to be joined
0:04
by two very special guests.
0:06
First, we'll start with Rick Peng, digital ventures
0:08
lead in Memorial Sloan Kettering Cancer Center's office
0:12
of entrepreneurship
0:13
and Commercialization.
0:14
And we also have Natalia Somerville,
0:16
director of decision intelligence at Memorial Sloan Kettering
0:20
Cancer Center. Thank you both so much for
0:22
joining. I know we have a a lot
0:23
planned for a for a great,
0:25
discussion on all the innovative things you're up
0:27
to over at Sloan Kettering. How are you
0:29
both doing?
0:30
Doing good. Very excited to be here. Thank
0:32
you so much for having us. And, yeah,
0:34
definitely,
0:36
very happy to share some of our experiences.
0:39
Yes. No. Thanks so much to to both
0:41
of you for joining, and and we'll go
0:42
ahead and get started here with this, important
0:44
discussion.
0:46
Now now, Rick, I'll I'll get started with
0:47
you. I know, you
0:49
know, MSK is is often seen as a
0:52
leader in applying AI in clinical settings.
0:56
How is AI currently being used across
0:59
the system, and where do you where are
1:01
you seeing the greatest
1:02
impact in regards to patients today?
1:06
Yeah. So AI is definitely being applied across
1:09
the board here at MSK.
1:11
It's being applied in not only clinical care
1:13
but also in the research space.
1:16
It's,
1:17
in the clinical spaces, it's it's used in,
1:21
use cases such as assisting
1:23
clinicians in their visits with patients. Key examples
1:26
are partnership with companies like Abridge that help
1:29
us, manage
1:30
a transcription documentation
1:32
of, patient visits.
1:34
In the research space, we we do a
1:37
lot of internal innovation,
1:39
as far as, AI,
1:42
development models that can predict responses to therapy,
1:45
and that assist in the the development and
1:48
discovery of new,
1:50
therapeutics in oncology.
1:52
We also partner with external companies in this
1:54
space as well to co develop new AIs
1:57
that can be applied in these situations. And,
1:59
of course, we have a large network of
2:01
pharma partnerships that we that, can leverage these
2:05
technologies to accelerate the development and bringing to
2:08
market new drugs that serve unmet needs in
2:10
cancer.
2:11
And of course, on top of that, we
2:13
also apply
2:14
AI as tools across the enterprise,
2:17
to help, staff at MSK,
2:20
outside of just clinicians and researchers to to
2:23
do their jobs as effectively as possible as
2:25
well. So really across the board.
2:27
Yeah. Appreciate the details there on all the
2:29
AI usages, Rick. Now, Natalia, I know you
2:33
you co create MSK's
2:34
AI governance
2:35
operating model, which is obviously a huge accomplishment,
2:38
but what problem was MSK trying to solve
2:41
when you were developing the framework?
2:44
Yeah.
2:45
So a couple years ago,
2:48
here at MSK,
2:49
our,
2:50
leadership
2:51
created what was called the AI task force,
2:55
where kinda given all the movement and,
2:59
progress
3:00
happening,
3:01
AI and health care,
3:03
our leadership wanted to understand what are the
3:05
important aspects and what should we be looking
3:08
for, what are the valuable use cases. So
3:11
one of the work groups was governance.
3:14
And, that work group was led by by
3:17
our
3:18
chief informatics officer, doctor Pitstetson,
3:22
at the time.
3:23
And, during that group, I was supporting the
3:26
lead,
3:27
and we had conversations with
3:30
important,
3:32
players,
3:33
let's say within the hospital, like
3:36
legal and,
3:37
clinicians who have been working ethics
3:40
and,
3:40
technologists.
3:42
And, that's where there was obviously
3:45
kinda an understanding between everybody that
3:48
we are starting to deploy more and more
3:51
AI,
3:52
and we need to be able
3:55
to make sure that it's done safely,
3:58
and also that we can,
4:00
scale it up because we totally understand that
4:03
if maybe a couple years there was a
4:05
couple models deployed,
4:07
this is
4:08
growing exponentially now. So we wanna be able
4:11
to do that,
4:13
responsibly and support
4:15
that innovation
4:16
and that,
4:17
scaling up of AI technologists.
4:20
Absolutely.
4:21
And, Rick, I want to ask you about
4:22
the AI governance as well. You know, in
4:24
in a high stakes environment like cancer care,
4:26
why is governance
4:27
it's not optional when it when it comes
4:29
to,
4:30
AI adoption,
4:32
and why is that?
4:33
Yeah. I mean, I I I think there
4:35
there's obvious,
4:36
you know, ramifications
4:37
on, you know,
4:39
how patients receive care, how research is conducted,
4:43
not only in terms of, like, the clinical
4:45
impact on patients, but also,
4:47
you know, everything from,
4:49
you know, ethics and, scientific integrity as far
4:52
as all that work that we do. And
4:54
I and I think, you know, broadly, like,
4:57
the the need for governance in AI is
5:00
particularly
5:01
important,
5:02
as we think about, you know, organizations like
5:04
MSK are very large. There's a lot of
5:07
different stakeholders involved,
5:09
including external collaborators.
5:11
And I think, you know,
5:13
AI requires
5:14
connectivity across all these different groups, whether they're,
5:17
you know, at the points of, applying the
5:19
AI, at the points of developing the AI,
5:21
at the points of supplying the inputs to
5:23
AI, including,
5:25
data,
5:26
where, you know, we really need all these
5:28
pieces to be, you know, working in coordination
5:31
with each other to make sure that what
5:32
comes out the other end
5:34
is,
5:35
created,
5:36
in a way that,
5:38
optimizes its intended performance as well as as
5:41
being applied responsibly,
5:44
across all the different groups that need to
5:46
be using it.
5:47
Natalia, what qualify what what qualities I'm sorry.
5:50
Should health systems look for an AI governance
5:53
partner to ensure responsible
5:55
adoption while still supporting
5:57
some innovation?
5:59
Yeah. I think it's definitely flexibility,
6:03
because
6:04
as we have learned and you just mentioned,
6:07
we have to be able to balance those
6:09
two goals,
6:11
be able to deploy AI responsibly, but also
6:15
making sure that
6:16
we are not seen and we are not,
6:19
blocking
6:20
any any innovation
6:22
performed by our clinicians. Our clinicians do amazing
6:25
research, and that is
6:27
partially why our institution is so well known
6:30
because of the research they do. So we
6:32
wanna make sure to encourage that research.
6:36
So the way we we partner and the
6:38
way we we see these processes
6:41
is making sure that there are different paths,
6:44
that if there is an AI
6:47
model or AI system
6:49
that,
6:50
the governance
6:51
committee is looking at a time that we
6:53
have different paths, that we are flexible,
6:56
understanding that the the research initiatives
7:00
may have a faster path. They are working
7:03
towards
7:03
proof of concepts and, developing the science,
7:07
versus tools that may go into workflow
7:11
workflows,
7:12
right away,
7:14
let's say predicting mortality
7:16
and we wanna deploy that within two months,
7:19
then we definitely wanna have a different path,
7:22
much more focused on reviewing the details.
7:26
So, yeah, I believe that that flexibility
7:28
on adjusting
7:30
the processes
7:31
and,
7:33
our AI life cycles
7:35
across the different types of implementation
7:38
is is key.
7:41
Natalia, how does MSK's
7:43
AI governance
7:44
operating model help ensure that AI tools are
7:47
safe and
7:48
accountable and aligned with clinical standards?
7:52
Mhmm. Yeah. That's that's a great question.
7:55
So over the last,
7:56
I wanna say two years,
7:59
we developed,
8:00
these processes
8:02
and,
8:03
an AI life cycle. So this is a
8:05
tool that,
8:07
kinda
8:08
ground the grounds
8:10
the
8:10
AI
8:12
research and deployment
8:13
that happens.
8:14
An AI life cycle is the the,
8:18
the the sequence of steps that you follow
8:21
such as,
8:23
you have an idea, you have to validate
8:25
that the idea,
8:27
has,
8:28
you know, value within the organization,
8:31
and then design, deployment, testing, all all those
8:34
different steps.
8:36
I do wanna call out that, actually, two
8:38
years ago, our our life cycle was,
8:42
our original life cycle. We did use,
8:45
Duke Health Systems life cycle. They are very
8:48
advanced in this area, and then we modified
8:50
it what makes sense for us. But so
8:53
what I was mentioning is this life cycle
8:56
allows us to create steps
8:58
depending
8:59
on what stage the AI system is.
9:03
So for example,
9:04
if the AI system
9:06
that we are,
9:08
looking into or,
9:11
governing at the moment is
9:13
in, pilot mode, then there are certain requirements
9:17
that we ask. We ask for,
9:20
the pilot,
9:22
plan to be very explicit and presented to
9:24
the committee. We're asking for a monitoring plan
9:28
in advance of the deployment
9:31
versus
9:31
if, the AI system is earlier. Let's say
9:35
they're just kind of finished design and going
9:37
into development.
9:38
There's other sorts of questions that we we
9:41
ask. So depending on the life cycle,
9:44
we have these different touch points
9:47
with the model owners and model sponsors
9:51
to make sure that it goes through,
9:55
in a in a safe way, and we're
9:57
able to
9:58
suggest any modifications
10:00
as early as possible if needed.
10:03
And and, Rick, what risks do health systems
10:06
face if they deploy AI tools, you know,
10:08
without a clear strategy or framework in place?
10:12
Yeah. So I think, like, for for health
10:14
systems, if AI isn't
10:17
isn't deployed in a way that's strategically directed,
10:21
I think there's there's a couple levels of
10:23
risks.
10:24
One is,
10:26
an operational
10:27
risk of,
10:28
have impacting the adoption of these technologies.
10:32
If there isn't a clear idea of what
10:35
workflows or what stakeholders,
10:38
need to pick this up and, you know,
10:40
build it into practice,
10:41
it can really limit,
10:43
the extent to which the the AI is
10:46
being applied in the first place.
10:48
I think, ultimately, longer term, it can impact
10:51
the ROI of these types of solutions, especially
10:53
if there, you know, if there's significant investment
10:56
in them, whether it's through purchasing external
11:00
solutions, or,
11:01
you know, investing in internal development of capabilities
11:05
if if there isn't the right adoption or
11:07
if it's not being applied in the most
11:08
impactful use cases
11:10
and reduce your ROI, which I think, you
11:12
know, can,
11:13
not only,
11:15
hamper the the the impact of the particular
11:19
AI solution in question, but also just AI
11:22
across the board as well.
11:24
And and one special consideration for for organizations
11:27
like MSK as well is, like, MSK definitely,
11:30
you know,
11:31
it's it's point of pride that we are
11:33
one of the leading cancer centers in The
11:35
US and in the world. And so I
11:37
I think, you know, application of AI also
11:39
needs to be done in a way that
11:42
fully takes advantage of our the differentiated
11:46
expertise that MSK has.
11:48
I think a lot of times where that
11:50
comes into play especially is,
11:53
you know, in terms of using AI solutions,
11:56
ensuring that there's,
11:58
the appropriate level of governance, but also input
12:00
the development
12:02
of those AI tools, whether it's through leveraging
12:05
a,
12:05
MSK's,
12:07
unique know how expertise of its
12:09
clinical and research experts, to really to really
12:11
to really, you know, infuse the AI solutions
12:14
that we do apply,
12:16
with the with the sort of secret sauce
12:18
at MSK so that, you know, when applied,
12:19
it really reflects the level of of quality
12:19
that,
12:21
applied, it really reflects the level of of
12:23
quality that, MSK has come to be known
12:27
for in the world of cancer care and
12:29
research.
12:31
Yeah. Thanks so much, Ray. And maybe diving
12:33
into that that secret sauce,
12:35
Natalia, that the process with AI over at
12:38
MSK, you know, you know, even after tools
12:40
are approved and deployed, the work's not done.
12:43
Right? So why is continuous monitoring
12:46
of AI systems essential?
12:48
Yeah. That is it's it's very important as
12:51
you mentioned, and, I the way I talk
12:53
about monitoring is actually twofold.
12:57
One is monitoring,
13:00
the the the safety. You know? Has there
13:02
been any adverse events, for example? Because,
13:05
as
13:06
as Rick was mentioning, there's,
13:08
kinda several
13:09
aspects that could potentially go wrong. So we
13:12
wanna be able to monitoring
13:14
any, AI adverse,
13:16
adverse,
13:17
events,
13:19
as well as kinda the the technical
13:21
model itself. Is there any drift?
13:25
For example, if there is a pattern change
13:27
in the data,
13:29
then we need to be able to acknowledge
13:31
that in the model if it doesn't do
13:33
it automatically.
13:34
So depending on the type of model, we
13:36
need to be able to,
13:39
see that every
13:41
quarter, every month. It it really depends on
13:43
how the model is is structured.
13:46
At the same time, the second part that
13:48
is also very important that sometimes
13:51
it's talked about less actually than the technical
13:55
monitoring
13:56
is precisely what Rick was mentioning. We also
13:59
want to make sure that the tool is
14:01
actually being used.
14:02
And beyond being used, that the tool is
14:05
bringing the expected impact.
14:08
If we are deploying,
14:10
an AI tool
14:12
to support,
14:13
for example, discharges,
14:15
the the nurses with their discharge efforts,
14:19
and the original goal was to reduce the
14:22
patient length of stay such that they can
14:24
go home as soon as possible,
14:27
and enjoy kinda more, comfort settings,
14:30
then we wanna see that the length of
14:32
stay is actually decreasing.
14:34
So we wanna be able to see those
14:36
metrics that we are targeting,
14:40
as well as usage. We don't wanna be
14:42
maintaining
14:43
tools that nobody's using. So those are the
14:46
two aspects, and they are typically monitored,
14:50
differently and in kinda different
14:53
platform and visualizations, but both are
14:56
key to be able to say that the
14:59
AI is being used properly.
15:03
And, Rick,
15:04
government regulation is always obviously a big topic
15:08
in health care, and it doesn't escape AI.
15:11
So with AI regulations
15:12
now evolving at the state level, how does
15:14
an organization like MSK
15:16
streamline compliance without slowing progress or innovation?
15:21
So I I think a big part of
15:23
that is,
15:24
creating visibility across all different stakeholder groups. Part
15:28
of that is, on an operational level, the
15:31
creation of the appropriate governance committees with, multidisciplinary,
15:35
multifunctional
15:36
representation,
15:38
which, Natalia and her colleagues have really spearheaded
15:41
over the past few years here at MSK.
15:44
A a part of that as well is
15:45
just facilitated
15:47
through through technology solutions that, enable us to
15:50
do this on an efficient basis. And so,
15:54
having the tools for these,
15:55
these these individuals
15:57
within these committees to effectively communicate, have the
16:00
right information at their fingertips so that when,
16:03
there need to be conversations or decisions around,
16:06
the the these AI solutions that that's handled.
16:09
And I think, like, at MSK, I'm not
16:12
directly embedded in these functions, but we also
16:14
have very robust functions around
16:17
regulatory compliance. And I think a big part
16:19
of it is MSK
16:22
overall, you know, prides ourselves in, you know,
16:24
staying abreast of all the latest developments, not,
16:27
in in all the regulatory and legal spaces
16:30
that impact the work that we do. And
16:31
so,
16:33
our our processes are, and our our stakeholders
16:36
in those functions are, ever evolving to to
16:38
meet
16:39
the the the the latest requirements,
16:42
and make sure that we we stay not
16:44
only up to speed, but also think ahead
16:46
in terms of where we need to be,
16:48
not just this year, but also next year.
16:51
Yeah. And if I if if that's okay,
16:52
I can, jump in as well as as
16:55
definitely, as Rick is mentioning, the the
16:58
the other aspect is also kinda
17:00
being
17:01
involved with and in communication with other organizations
17:05
like us
17:06
that are also working and implementing their governance
17:09
processes,
17:11
and part of CHI. For example, MSK is
17:13
part of CHI,
17:15
and that is a way that we try
17:17
to,
17:18
keep,
17:19
keep present in kinda the latest developments
17:22
and,
17:23
as well from the
17:25
government or legal aspects,
17:27
any
17:28
new technology developments.
17:30
I do believe that it's kinda important that
17:33
all these organizations that we are,
17:35
sharing with each other and kinda progressing together
17:38
in this way.
17:40
Yeah. Yeah. As you as you share what
17:42
you've learned through this experience, Natalia, like, what
17:44
lessons do you do you think from MSK's
17:47
approach to AI governance do you think can
17:49
be applied to other
17:51
hospitals or or health care organizations?
17:54
Yeah. Definitely. I love this question because
17:56
one thing that,
17:58
is often
17:59
also
18:00
not talked about is actually how
18:02
operationally
18:03
heavy
18:04
is
18:05
to
18:07
manage the govern the the governance process.
18:10
Because as I mentioned at the beginning, the
18:13
AI, kind of the the AI
18:15
implementations and systems and development is,
18:19
growing exponentially.
18:20
But
18:21
if we want to have a touch base
18:23
with every model, every system that is medium
18:26
to high risk, there's a lot of operational
18:30
tasks that need to be done, you know,
18:32
kinda check with the model owner, have them
18:35
register their systems
18:37
and verify they are pilots and perform risk
18:41
assessments. And so it's actually operationally
18:43
very, very heavy.
18:45
And,
18:47
so partnering
18:49
with somebody who can help with that,
18:53
those sort of operations
18:55
can be very, very helpful.
18:57
So for organizations that are beginning to look
19:01
into this AI governance systems, I will definitely
19:04
recommend two things. One is take a look
19:07
at what has been done already.
19:09
Many many of us has have published
19:12
some of our systems. We are also always
19:14
happy to talk about that, kinda implement some
19:17
of the processes,
19:20
and guidance that other organizations have done already
19:23
and also look at platforms that can help
19:25
with that operationalization
19:28
of the governance.
19:30
That is typically
19:32
bypassed
19:33
at the beginning and,
19:35
that can really delay
19:37
any sort of meaningful
19:40
scaling that you can do with your governance
19:42
process.
19:43
Well, Rick and Natalia, thank you so much
19:45
for joining the podcast and just for a
19:47
great discussion
19:49
on MSK's, you know, innovative use of of
19:51
AI. I look forward to working with you
19:53
both again soon.
19:55
Thank you so much for having us.
19:57
Yeah. Thank you so much, Scott. It was
19:58
great to be here.