Rebecca Flournoy:
Welcome to today's forum on AI in healthcare. My name is Rebecca Flournoy, and
I'm a senior health policy leader at the Kaiser Permanente Institute for Health Policy. And I will
be your emcee for the day. So, at the institute, we host events and publish information to shape
policy and practice on important health topics, exploring ways to promote high-quality, equitable
healthcare and healthy communities. And we are so excited to see all of you here today. I
think many of
us are here because we see the really exciting potential that AI in
healthcare could bring. The way that AI in healthcare could help us promote patient care
and population health, reduce burnout among healthcare professionals, and make healthcare
more efficient, accessible, and affordable. And I think at the same time, we also recognize
how critical it is to be really thoughtful about how AI is designed and implemented and monitored
from the start. What needs to happen to ensure that AI and
healthcare is well designed and
well-governed? How can we ensure that AI is safe, useful, trustworthy, and equitable? We are going
to dig into these questions together today. We will start the day out with a fireside chat to
hear about current uses of AI in healthcare, bringing in some examples from Kaiser Permanente.
We'll then shift to a session on the future of AI in healthcare, bringing in some additional
current examples, but also looking to what's on the horizon, and some important c
onsiderations
to keep in mind. We will then break for lunch. And then in the afternoon, we'll have a session on
practice and policy and ways that we might address risks while also harnessing benefits. And then
we'll hear some closing reflections at the end of the day. I know some of you submitted questions
in advance. And I want to thank you for those and let you know that the moderators and speakers
have seen those, and we will do everything we can to incorporate those into the presentatio
ns
and the discussion. And then we'll also have a little time at the end of each panel for a
couple of live audience questions as well. Before we begin a few housekeeping items. We are
recording the event today, so please silence your devices. And if you need to make a call, please
either step outside or you can see one of our staff members to see if a room might be available
for use. To access information on the event today, you've probably noticed that there are cards at
your tables that
have information on connecting to Wi-Fi, and also a QR code that will take you
to a website with the agenda and speaker bios. If you're using social media to share posts
on today's event, you can mention us using the handle @KPIHP, and hashtags like AI and
healthcare AI. And then a final note here in the event space. Restrooms are located behind this
screen, down the hall, and to your left. And then emergency exits are located behind you through
the doors where a lot of you came in this mo
rning. So, now I want to transition to our first
conversation. I want to welcome up to the stage Tony Barrueta and Vivian Tan for our
opening fireside chat. Tony will introduce Vivian in just a moment. And Tony, as you come
up, I will tell the audience a little about you. Tony Barrueta is senior vice president of
government relations for Kaiser Foundation Health Plan and Hospitals. And in this role,
he oversees our legislative and policy efforts, leading a team of legislative advocates and
policy professionals. He dedicates -- excuse me. He directs the development of her public
policy positions in support of the organization, our members, and the communities we serve. And
I will turn it over to you, Tony. Thank you. Tony Barrueta:
Great. Rebecca, thank you. And let me add my welcome to all of
you. It's really wonderful to see so many people come out, you know, here at the Center for Total
Health. This has been a wonderful space, really, for the Institute for Health Policy to
put on a
number of different health policy forums. I think, you know, we've gotten restarted again
now after the pandemic. And prior to that, we just had this amazing run of really interesting
sessions. And I think we're back on track again, Rebecca, I think it's fair to say, with this
as well. So, I just want to welcome all of you. So, I am really happy to be able to have
this role today interviewing Vivian Tan, my good friend. Vivian has been with
Kaiser Permanente for 16 years now. So,
she's just barely not new
anymore at Kaiser Permanente. [laughter] You know, and she is really one of our key leaders
within the organization who is delving deep into issues of AI, machine learning, and all of the
others. And she has been -- and she has a really unique perspective within the Kaiser Permanente
organization. Because in addition to her current role as vice president of Strategic Information
Management and Global Relationships for Kaiser Permanente, the Global Relationships has
her
really representing Kaiser Permanente at the World Economic Forum, where there's a tremendous
amount of this work going on and a lot of policy thinking and development that's going on
there. She's had a number of roles within the organization that have her really right
at the intersection of senior leadership, engaging with the most interesting things
that are going on at Kaiser Permanente. She has been, you know, a key person in
the organization. She and I kind of grew up together in
some respects, providing support to
senior leadership on key initiatives and public policy. And still, she works very closely
with our wonderful chief financial officer, Kathy Lancaster, on a variety of issues, and
our CEO, Greg Adams, who takes a tremendous interest in this subject as well, and I know
has all kinds of thoughts on this subject. So, she's currently leading a lot of the
work that has been transformational in terms of what we're doing. And what
anybody who knows anything abo
ut Kaiser Permanente would know is it is a very
interesting and complex organization, and there's a tremendous amount of creativity
that goes on in a whole range of different places. And it's really -- I think Vivian's role is in
part to keep track of all of that stuff that's going along and figure out how to get it all
rowing in a common direction for the benefit of the people who we serve, our members,
and the communities in which we operate. So, Vivian, I'm delighted to be
able to be he
re with you. Vivian Tan:
Thank you for having me. Tony Barrueta:
I think it's fantastic because, you know, a key part -- I was just chatting with
Marilyn from BPC. And one of the things that's really necessary in this area is just kind
of explainer stuff and one-on-one stuff. And you know, can we give a basic overview of what's
going on? So, I'll ask -- I'll pose this question the way I pose all of my questions. As you've
heard me say many times, I'm not a healthcare person. I'm a History ma
jor. So, what do I need
to know? What can you tell me about the basics of AI and what's really happening right now, in
a way that a History major would understand it? Vivian Tan:
Well, thank you, Tony, for that question to start off things. And welcome
everybody. Tony calls himself a History major, but he's much more and far from that. Maybe
we'll start with a primer. Some of it might be very rudimentary for you. And then for
others, hopefully, it's helpful. So, Tony, I always like to try t
o think of an analog or a
parallel. And if you think about AI like human intelligence, right, and you think of the phases
of human understanding and then sense-making and then actually being able to respond to it,
AI has, like, has -- you can think about the same components. So, in the understanding, you
know, like human beings, we have five senses, right? There's a lot of AI technology to help us
process a lot of structured and unstructured data. So, when you think about text and images an
d video
and you hear things like NLP or Computer Vision, that actually helps. You can use AI techniques to
ingest a lot of information. The second, you know, kind of again, likened to human intelligence is
the sense-making part. And you're trying to use the data that you've collected to make sense
of it and find patterns in the data. And you apply machine learning techniques, deep learning
techniques, reinforcement learning techniques. They're supervised and unsupervised. All these
are buz
zwords, but at the end of the day, you're really trying to figure out if the
information has patterns to predict outcomes, to diagnose things, to actually say that
this happened or did this not happen. And for those kinds of things, for a while, they
were very narrow in nature, right? You build a model, it could play chess, but it cannot do
anything else, or it can play the game of Go as many of you know with the Google Brain. And
now, we're in this phase where it's maybe the evolution of w
here it's going to self-generate
information and content. And this is the rise of, you know, foundational models that are
based on texts. If they are based on text, they're called LLMs. They are basically
transformer-based models using, again, mimicking the brain in terms of how you think
about -- you have layers and layers of processing. And you hear buzzwords like prompt engineering
or rag or fine-tuning, and all of these things actually just help us make sense of a lot of
data. Basicall
y, if you look at OpenAI, you know, ChatGPT, it's been trained on every piece of
information out there in the internet. And it's --people are speculating. And actually, we
have a Microsoft expert here. Maybe he can verify this [laughs]. The people are speculating that
it's actually trained on 100 trillion, you know, tokens of information. This is humanly not
possible to process. And it's very -- it's become very sophisticated in terms of predicting the next
word, but it essentially is a nex
t-word predictor. Tony Barrueta:
So, it's sort of a term -- I mean, AI is sort of a term that
covers actually a lot of different things. Vivian Tan:
A lot of -- yes, yes. Tony Barrueta:
So, what types of things are we really focusing on? What types of things should
people be aware of that -- where this is moving? Vivian Tan:
Yeah. You know, I have a little chat up there and I know it's
very busy. It's got a lot of things on it. And maybe I want to tie AI back to healthcare. And
you know, AI
is an enabler at the end of the day. It's a tool. You don't do AI for AI's sake. So,
I want to maybe tie a feature of healthcare with the application of AI. So, the first feature
of healthcare is that at the end of the day, people are heterogeneous. They are
not the same. And we are -- you know, I think one of our key leaders always says
healthcare is about people caring for people. So, on one hand, people are heterogeneous
that we're caring for, and the people that are providing that care
are experts,
right, and there's a lot of shortages, right? We don't have infinite resources. So,
you really want to match. You want to segment your populations and really match them with
the right care at the right time at the right place. And that takes -- that can be powered by
AI and that drives value. So, that's, you know, an example where we're applying AI to
healthcare in a way that drives value. Tony Barrueta:
So, that's a staffing support issue, basically. Vivian Tan:
It can be. Yes
, it can optimize staffing. It can optimize
supply chain. So, I'll give you another example. Healthcare is a dichotomy in a way, right?
We have long lead times. If you want a nurse, you can't get a nurse, you know, on the spot,
on-demand. If you want some drug that needs to be shipped from another part of the world. So,
there are long lead times. And yet it's also a real-time business where if you don't use that
bed, you don't use that exam room, you know, you use it or lose it, right? So,
that dichotomy
of long lead times and also the need to actually optimize things real time really, again, is
another place where AI would -- using prediction and prescriptive analysis, you can actually
match and optimize for your health system. So, we have a lot of examples of that in Kaiser,
where we actually have a real-time app that's running in all 40 of our hospitals that give
all our clinicians and operators, you know, up to the minute -- okay, no, it's up to five
minutes [laughs] whe
re it's refreshed. They have it at their fingertips on the go. They
can see everything that's happening. And we predict sensors. We predict staffing both for the
current shift and two shifts out, and beyond that. Tony Barrueta:
So, just so I -- because you told me about this the other day and I'm like blown away
by this. So, this is really real-time examination of the population currently in the hospital
or I guess, scheduled to go to the hospital. Vivian Tan:
Yes. Tony Barrueta:
And AI is d
oing what with that? Vivian Tan:
So, AI is helping with the throughput management. And you know, if you actually have
good throughput management, you also improve the access and also the experience. So, it's kind of
like it makes us more efficient, which has a cost impact. And it also improves the experience.
You know, no patient, no member wants to be waiting for something to happen. So, we predict
and highlight the bottlenecks and constraints in the system so that we can act on it, you
kn
ow, in the moment and get things resolved. Tony Barrueta:
So, it's sort of -- I was thinking of this in terms
of is it -- it's kind of predicting what this current patient population is
going to need a couple of days from now. Vivian Tan:
Yes, in the same shift and two shifts out. Tony Barrueta:
Oh, wow. Okay. What else is there? [laughs] Vivian Tan:
And we refresh that prediction on the hour every hour
with new information that we have. Tony Barrueta:
That's remarkable. Okay. What else? Vivi
an Tan:
So, you know, there are so many things. So,
we have about 300 deep learning models in production and 3,000 machine
learning models in production at KP. Tony Barrueta:
Okay. Tell me what this means. Vivian Tan:
Segmenting everything from -- with breast cancer screening and doing
segmentation there, predicting high risk for suicide in certain populations. We have things
that predict COVID for a long time. We have things that are really looking at -- we also
are using ambient AI and th
ings to support documentation and reducing clinical workload. So,
there are so many different experiments going on in different regions. I feel like Dr. Ainsley,
is she in the audience? She is leading a lot of amazing work in the area of imaging. And I
know she's going to come on in the afternoon. And hopefully, she can highlight some of the
innovations going on in that space as well. Tony Barrueta:
So, I think in trying -- a lot of this is trying to figure out, okay,
we've been using algor
ithmic processing for a long time. What actually -- what's different
now? So, I mean, we've -- for many years, we've been sort of the early flu warning
system for the state of California and through that data and we do reporting. What's
probably advanced now beyond what it used to be? What's the difference between what was
happening before and what's happening now? Vivian Tan:
Yeah. So, I think there are some things that haven't changed.
We're still doing value-based care. That is, you know
, kind of enabled by data and
analytics. I think with generative AI, the difference is the speed and the scale. And
I think it's both a good thing. And there are probably things that come with it. And I
know in the afternoon, Dr. Yang is going to -- where's Dr. Yang? He's going to lead us in
a conversation about benefits and risks of AI. And you know, I think with generative AI,
there's like half glass full and half glass empty. I think the big difference is the
speed and the scale. Becaus
e suddenly, I have not seen this industry and
this space move this quickly -- Tony Barrueta:
Right. Vivian Tan:
-- within such a short period of time. And the scale in which things can be
done actually is, you know, sort of in the order of magnitude. And the amount of data that's
being processed is mind-boggling, to be frank. Tony Barrueta:
Okay. So, when I first started reading about AI six or seven years ago, it was like
the impending apocalypse and all of that stuff. [laughter] So, there
are things to worry
about. I don't think that is the thing that's first to worry about
in this context, in healthcare. So, what do you worry about that's in
the kind of shorter period of time? Vivian Tan:
Yeah, I think for healthcare, you know, the bar is higher for healthcare, and actually, rightly so.
The bar is higher from a data privacy perspective. The bar is higher from a data security
perspective. The bar is higher from an algorithmic perspective. So, we have to actually, you know,
both govern and regulate ourselves in, you know, to the highest standard. And then we also have
to work with regulators and policymakers, all of you in this room, to ensure that we have the right
public, private, you know, collaboration as well. So, you know, I don't know if it's a worry,
but I do hold it as a key responsibility and obligation. Because you know, healthcare
data is sensitive. The application of that data is sensitive too. And actually, with
the executive orders and stuff, yo
u know, it's also very clear where regulation is going,
where we have to ensure there's transparency. And we have to show that our data and models
are not biased. And we have to explain it. So, they are, I think, a lot of things that are
actually exciting as well about this space. Tony Barrueta:
So, and I know we'll have a lot of discussion of this later
today. But you know, policymakers and people who are observing this are really focusing
on the equity issue. And I think the concern, you
know, the immediate concern that we see,
whether it's with the Federal Trade Commission or others looking at it, is will the use
of AI exacerbate existing equity issues, or -- which is certainly possible. Is there an
opportunity to actually make progress on some of the problems with health equity that we haven't
been able to make as much progress on them? Vivian Tan:
Yes, I'm really excited about that. And I see KP being a leader
because we've taken really very intentional steps to create t
he right data that is not biased
and is very inclusive and equitable. So, I'll share an example from World War II because
you're a history buff here. So, [laughs] I'll do an example. So, in World War II, they had all
these planes. And they were trying to figure out what part of the plane they had to reinforce,
you know, to prevent it from being shot down, right? So, they started collecting data on
the planes that were coming back, and they tried to figure out where the holes were, right?
B
ut there was a huge flaw in that data because the plane -- some planes weren't coming back at
all, right? And the places where they were most vulnerable was actually the engine, not the wings,
which, you know, which the data was showing. Tony Barrueta:
Right. Vivian Tan:
So, one of the key things we really have to do with health equity data is to
make sure that the data is correct. The data is robust. It's not biased. It's inclusive. It's
equitable. Because with AI, the data actually drives
the model. So, if the data is wrong
or not complete, the model is not good. So, one of the things that we're really doing is when
we do any work now, we actually look through the lens of health equity. We actually check if the
model performs well when we look at subsegments by race and ethnicity. And we're very thoughtful,
you know, in terms of what models we use if it doesn't, you know, and we actually don't use
those models if it doesn't prove to be equitable. Tony Barrueta:
Wow. So, you'
re fortunately -- oh, yes. We need to get --
getting a few questions from the audience. So, if there are questions -- please help me ask these
questions because as you can tell, mine are not as sophisticated as yours are likely to be. So,
please, think about some questions. I was -- you know, I think there's so many different types
of opportunities that are in front of us. First, I'm really pleased that you're in the spot
where you are to help us think this through. Where do you think other
organizations will
be learning from what we're doing up front? What opportunities do others have that maybe
they've a greater opportunity than we do in terms of the data they may have access
to or their models differing? I mean, maybe the other way to ask is, is
there something about our model that makes this a particular opportunity
or a particular hindrance to doing? Vivian Tan:
Yeah. So, I'll say a few things. So, the first
is because KP is an integrated system, our data sources and wh
at we call data
domains are very rich and deep. So, we have our health plan data. We have clinical
data. We have consumer data. We have real-time data. We have retrospective data. So, one of the
things is the breadth and the depth of our data. And actually, we've got data because we've been in
existence for how long now, 77, 78 years. We have longitudinal data. And in the area of genetics,
we have family history from cradle to grave, right? And we have it for generations. So, our
data is -
- oh, and then we have a tumor bank with, I don't know, now like 60,000 different samples.
So, we have such rich data that it's really amazing to be able to look at all those assets and
figure out how to leverage it for the, you know, benefit of our members and patients. So, that's
one, I think, unique differentiator for us. The other is that actually we have done
a lot in partnership with other systems around interoperability and data exchange. And
we've also -- I wouldn't say we're unique
here, but we've learned a lot. And we
are hoping that we continue to learn and share and collaborate with
many health systems across the US. Tony Barrueta:
Any questions? Yes. Kathy Curran:
First, thanks for putting the session on. I don't know if people
can hear me. Thank you for putting this session on. I'm Kathy Curran with the Catholic Health
Association. It's a really important topic and really important work. And I wanted to
follow up on your comment about, you know, making sure we g
et the AI usage right from an
equity perspective. But how do you do that? Because the image that comes in my mind is if
you don't know that you don't know how to spell, then the dictionary is no good to you because
you don't know to look in the dictionary. So, how do you -- like what are some practical
examples of how we can know what we don't know when these applications are being developed
and the data sources are being chosen? Vivian Tan:
Yeah, that is such a good question. And at KP, we
have a very practical seven-step approach which
starts with the framing of the question. You know, there's a lot that you can do to ensure that
your question is not biased, right, and you frame it in a way that doesn't introduce bias. The
second which I talked about was the data piece. It's so important to step back and investigate if
there's something critical that's missing in your data set. Whether it's, you know, whether it's an
actual, you know, segment of the population, or whether i
t's a dimension of feature or labeling of
the data that you need to really concentrate on. And this is where we have taken a very
multidisciplinary approach where -- and we really go and work with the subject matter experts.
The people that are closest to the population, the people that are doing the work day to day,
the people that understand the data, you know, in a great depth, right? So -- and we look at
it from multiple kind of places. And finally, when the model is built, there are a
lot of techniques to actually check if the model is working well. And
you can actually explain it. So, whether it's looking at counterfactuals,
[unintelligible], attention models, there are ways to actually investigate the
model to make sure that the model is working. And the final thing I would say
is actually, most of the time, [laughs] when you first build your model with the
data, the model is actually pretty good. But what happens over time is there is this thing
called model drift.
As your data changes, as you get new data in, your model doesn't
get as good at predicting whatever outcome or, you know, prescribing or making
a recommendation on something. So, managing model drift with
continuous model management is critical. I could not stress
this over and over again because, you know, people spend all their time doing the
upfront work. And then once it's in production, I think there's a little bit of a culture of,
you know, forget, right? So, you put it in and then y
ou forget about it, where that
continuous management of it is critical. Tony Barrueta:
So, we got to keep -- so, there's actually a critical human role in
making sure that this doesn't go off the rails. Vivian Tan:
Absolutely. Actually, maybe if there's a chance to put the next slide on. I
try to keep things very simple. So, this is my ABCD chart of the ingredients it takes to actually
have AI at scale probably in any organization. So, first, you -- and I'm not going to go in any
sequence n
ecessarily. But you have to first have, you know, infrastructure, right, in terms of cloud
infrastructure. And I know we use Microsoft Azure for a lot of our data and analytics. And we have a
leader from that organization here with us today. But the infrastructure without the talent is also
not helpful. So, you know, you have to actually hire and retain very good talent. And then with
the data that we have curated, which we have a lot of in this organization. And then applying
the AI techni
ques that I talked about. You know, those are the ingredients of success, if you will.
There are probably other things. But you know, these are sort of the basic four ingredients
that go in to help an organization scale. Tony Barrueta:
Awesome. Well, Vivian, as always, I learned a ton whenever
I talked to you. And fortunately, you're here all day. So, people have an opportunity
to buttonhole her as the day goes on. So, let me ask you all to join me in thanking
Vivian for starting out this m
orning. [applause] Vivian Tan:
Thank you. Thanks for having me. Rebecca Flournoy:
All right. Thank you so much, Tony and Vivian, for all those great examples and
making it very real. I really love that session. And now, I am very excited to welcome our first
full panel to the stage. This next conversation on the future of AI and healthcare will be
moderated by Dr. Ainsley MacLean, who is chief medical information officer and chief AI officer
for the Mid-Atlantic Permanente Medical Group. Dr.
MacLean oversees a team of software
engineers, physician informaticists, and clinical analysts in the design
and implementation of technology, AI, and telehealth enhancements. She is the
founder and co-executive sponsor of the Permanente Medicine Artificial Intelligence
Center of Excellence. And she is certified by and on the board for the American Board of AI
and Medicine. She is also an associate medical director overseeing radiology teams and an
ultrasound program that spans 20 differe
nt specialties. Dr. MacLean, thank you so much
for joining us, and I will turn it over to you. [applause] Ainsley MacLean:
Yeah. Thank you so much. Thank you. Well, good morning, everyone.
It's so wonderful to be here. I really want to thank the KP Institute for Policy for putting
on this conversation. At this moment in time, we're at such a point of excitement looking
forward, and conversations like this are critical in ensuring that everything is done
well for our patients. So, thank you
so much. So, I'm going to introduce our first
panelist. And I'll let you know a little bit about how this is going to work.
So, each of our panelists who are amazing, are going to actually give some brief comments and
an overview of their opinions on what the future of AI and healthcare holds. And then we'll
sit up here together and we'll have a little conversation. And then we'll have an opportunity
for some of you to ask questions as well. So, I hope it's really engaging. And again, we're
going
to be now moving from the present to the future. So, our first panelist is Dr. Junaid Bajwa.
He's the chief medical scientist for Microsoft Research, and he's going to be talking about an
exercise of our imagination. Our next panelist is Dr. Julia Adler- Milstein. She's the chief of
the Division of Clinical Informatics and Digital Transformation and director for the Center for
Clinical Informatics and Improvement Research at UCSF. And she's going to be talking about AI
at UCSF Healt
h. And our final panelist is Dr. Judy Gichoya, who is an associate professor of
radiology and informatics and an interventional radiologist at Emory. She's going to be talking
about the future of AI in healthcare, capacity building for an AI-enabled future. So, please
join me in welcoming our panelists to the stage. [applause] Junaid Bajwa:
I might stand here if that's okay. I'm tall enough as it is. Good morning, everybody.
My name is Junaid Bajwa. I'm the chief medical scientist at Microso
ft Research. I'm a practicing
physician in the U.K.'s National Health Service. And I'm delighted to be here. My sincere thanks to
the KP team for the kind introduction. Feel free to connect with me on your social poison of choice
if that could be helpful. Very happy to engage in conversations. Before I start, we've spoken a
little bit about these new tools. And without making any judgments, how many of you have used
any of these tools in the last year? Just a show of hands. Who hasn't ever
used any of these? Okay,
so a few. How many of you have used them for work? Okay. Interesting. Okay, fine. That's a
nice baseline for me to start with. So, one of my challenges was to think about actually,
how do we just ground everybody in what we potentially will begin to discuss a little bit
about today. So, I spent most of the last 20-odd years across the payer-provider, regulatory, life
sciences, and now big tech side of healthcare. So, one might argue that I just never stop doing
rot
ations. And the other argument is I just wanted to make my mom really proud and make sure
that I do lots of different things with my life. So, I sit on the board of the MHRA, which is the
U.K.'s equivalent of the FDA. I'm on the board of a large academic medical center in the U.K. known
as UCLH. And the thing that really drives me is impact at scale. And really thinking through
not just health equity but population health, reducing the cost of healthcare, but also
improving the experience o
f care for those on the receiving side of healthcare, but also
on the delivery side of healthcare, too. And Satya Nadella a few years ago, who's the
CEO of Microsoft, said that AI is technology's most important priority and healthcare its
most urgent application. I think the context for this is really some of the supply and
demand challenges that we have globally. So, we have an increasing aging population
with increasing complex comorbidities, and we have a massive workforce
crisis in mos
t parts of the world. The World Health Organization, pre-pandemic,
estimated that the world will have 14 million less doctors, nurses, and pharmacists on the
planet than what society would need. I would argue that post-pandemic, we're closer to about 25
million less doctors, nurses, and pharmacists by 2030 than what society will need. So, within six
years, we have a massive challenge around who will be delivering care for this aging population with
complex co-morbid needs. And I teach at a
range of institutions. So, in the Boston ecosystem in
the West Coast as well. And this kind of role of technology, if you kind of take this graph on
the left-hand side -- and I apologize for the font size being very tiny -- just shows the rise
of technology across all industries over time. And on the right-hand side, something that Vivian
was referencing around the size of models that now exist and the scale that exists to inform these
models. So, thinking about data, compute power, and the
parameters. And many of you would have
heard of GPT or other similar models. And we'll talk a little bit about what they might mean in
a moment. But what's really interesting is how they're now democratized. Everybody has access to
them. Many of you in this room have raised your hand, may have played with them to say, write
me a poem, or help me make a travel itinerary. And it's almost quite magical when you
start looking at these and you start to interact with these tools in novel ways. B
ut
what we're seeing is language understanding, reading comprehension, image recognition,
speech recognition, handwriting recognition, surpassing human performance now
around kind of the 2020 mark. And it's probably best if we demonstrate
this through a series of some examples. So, if you were to go to Copilot today and type in
something like this. So, what is metformin? Many of you will probably know what metformin is.
But the response it will give you is this. And it will tell you that m
etformin is a prescription
medication to treat type 2 diabetes and provide you some additional data. If you carry on the
conversation underneath to say, can anyone with type 2 diabetes take it? It will start to
tell you about who can and who can't take it. So again, this is just something that's freely
available today. Any citizen in the U.K., in the U.S., or otherwise can begin to play
with these things. So, it's actually getting quite complicated in terms of the information it's
providin
g. If you carry on the conversation even further to say, well, actually, are there
alternatives to metformin because actually, I'm one of those people that have a liver disease
or kidney disease? It'll give you even more information around what the possible alternatives
might be for you. And many people in this room who treat patients with diabetes will probably
understand that. Actually, that's relatively reasonable in terms of what it's providing.
But all of that can be massively complica
ted. And you might say, actually, well, that
seems massively complicated. I just don't understand what to do. And not only
will it try and empathize with you, but it will recognize that you feel overwhelmed
in your language. It will recognize that you feel overwhelmed and then ask you to
think about a holistic way and ensure that you engage with your holistic, healthcare
professional team in tailoring the need to you. So, it's safety nets, if you will. It will
reason behind your words as y
ou carry on the conversation. Out of curiosity, because
many of you are kind of on the policy side, what's the average health literacy,
the average health reading age within the U.S. or the U.K.? Give me a
number. Shout as loud as you can. Female Speaker:
Sixth grade. Junaid Bajwa:
Sixth grade. So, how old is that? Female Speaker:
12. Junaid Bajwa:
12. When you think about your medical communication, how much
of it is targeted to a 12-year-old? It's an interesting question, right?
And I ask
that because actually, the first time I ever played with this
was when my 12-year-old came home with the geography homework about 22 months ago. So,
I had an early access to some of these tools. And his homework was, write a 500-word
essay on the microclimate of the school, comparing the field to the playground and the car
park of the school in London. And being the heroic father that I am, I could have chosen to go to
a search engine of choice, type in microclimate, then understand actual
ly is a microclimate
different for grass versus mud versus tarmac. Being very lazy, I wrote down my prompt which was
write me a 500-word essay on the microclimate of a school in northwest London, breaking it down
into introduction, methodology, and conclusion, but put it in the voice of a 12-year-old. And
when it put it in the voice of a 12-year-old, not only could I understand it
and explain it back to my son, but actually I could have a better conversation
with my son. I never gave him t
he answer because that would be very bad parenting, but
I could explain it to him on his terms. [laughter] And so, what I've now done in my own clinic
is to challenge myself to say if I'm having conversations with patients about diabetes or
heart failure, how do I meet them on their terms and have better conversations? And when I'm
at the MHRA, and we look at the risk alerts that we get for drugs, often 15 scrolls worth
of information, might we be able to actually communicate that better? A
nd these tools are
fantastic at summarization and communication. Many of us who have done medical school
or nursing school, we often learn through clinical vignettes. So, if you take a clinical
vignette like this, which is from the USMLE, talking about a 12-year-old girl with a
set of symptoms and some early diagnostics, we're often given a multiple-choice
question. You could take this entire prompt, put it into Copilot today, and it will give you
the answer. It'll tell you what the answer
is, and it will rationalize to you
why it is that answer as well. So, it's educating. Not only giving you the answer,
but you can prompt it to then educate you as to why you think that answer may well be true,
and why it possibly can't be any of the others. But remember, that's just a clinical vignette
talking about a 12-year-old girl with a set of symptoms. What if we push the model a little
bit and remind ourselves that actually, each of these models have not been trained on
any medical
knowledge. They've been trained on the digital exhaust of the internet, and they
have begun to reason in the background. So, the same model is passing the bar
exam, passing USMLE examinations, passing MBA exams in the background,
and no specialized training. Imagine if we said to it, what do you think
the girl in this problem might be thinking and feeling? That requires a significant
level of reasoning underneath. And these have not typically been the types of models
that we would have be
en able to have access to in the past. But it will give you a very
reasonable answer. As an AI model, I cannot directly assess a person's thoughts and feelings,
but I'll make an educated guess. The 12-year-old girl in this scenario might be feeling worried and
scared. She may be concerned about missing school, and it's essential for the healthcare team to
address her concerns and provide reassurance. There's been some papers that have been published
over the last year around the empathic re
sponse that these models are giving, arguably more
empathic than some of the practitioners that practice medicine today. Let's go a little bit
further even than that. So, the girl's name is Meg. If you were Meg's doctor, what would you
say in order to provide comfort and support? And it will give you a very, very reasonable
script to go in. And when in this -- certainly, in the U.K., and I'm conscious in the U.S.
too, we have lots of role play that takes place. And that role play can be ver
bal, but
you could end up having chat-based roleplay. And if you extend this out and you play with this
role play, you could ask the model to assess how well you have done versus other benchmarks. And
you could say, "I am training to be a doctor. Please role-play the role of Meg. I will be
the doctor, and let's have a conversation on it." And so, you'll see how these models are
very, very different to any types of technology that we've used in the past and before. And
again, freely accessi
ble, and available. What I would say -- and there was one little
bubble in Vivian's chart, which is really, really important, is thinking about how we
deploy these tools in a responsible way. So, the exam question that I focused on at Microsoft
is how might we transform the practice of medicine with trusted, reliable, human-centered AI? And I
say that very deliberately. It's not just about AI. In healthcare, it must be trusted. It has
to be reliable, and it has to be human-centered. And the
re are various tests that you have
to pass for trust, tests that you have to pass for reliability. And when we're building
models, we think about each of these things. So, on a bedrock of transparency and
accountability, assessing for fairness, reliability and safety, privacy and
security, and inclusiveness. Most life science trials that exist on the planet are not
representative of people that look like me. I'm a 6'4" South Asian guy from northwest London.
Most trials that exist for chole
sterol drugs, diabetes drugs are not for me. Therefore,
generally, 40-year-old white men, they're not designed for people like us. We
already have a bias in our healthcare system. I think that if we were to have a conversation
about the bias, challenge ourselves, and move that we are pushing for heterogeneity of
data in the development of our models, thinking very carefully about how we deploy
these; thinking about the end user impact; and push ourselves to be much more inclusive
around da
ta, where I think organizations like Kaiser can really lead the way. How
might we be better at addressing some of these things? This will not be a magic
bullet, but we will get there over time. And if we can identify issues with the outputs,
measure them, and mitigate them will be way, way better off. Many of you have
seen blood test forms like this. These models are pretty smart. These
models can take your blood test form, and you can ask it a question and it will
give you a response. Say
, "In simple terms, explain this pathology result to me."
Saves massive amounts of time certainly in my primary care clinic where I have to do
42 patients a day, 10-minute appointments each. And what you'll begin to see is what we didn't
do at the beginning, but we do a lot more now is ground the answers with reference points
within the internet. So, each of the references you can see underneath, and you can go back and
validate against them to reduce the hallucinations and confabulations t
hat many of you may have
heard of in the past. A lot of that is what's happening on the tech space. There's research
saying, how might we be able to actually begin to engage with images in a different way?
So, what can we do with two-dimensional and three-dimensional images? And I'm sure
colleagues will talk in depth about these. But potentially having a conversation with
the images, if you're in an ER or in an HD or an ITU setting, and you can't wait for
the radiologist to give you the op
inion and the review, could you figure out if there's
a pneumothorax here? Could you figure out whether the tube has been placed in the right
place? Can you figure out if there's a temporal dimension to the tumor? And it's changing
over time. And we're getting better at that. In the U.K., this is Dr. Rajendra [spelled
phonetically] on the left-hand side. We engage in a seven-year piece of research with him to think
about prostate cancers and radiation oncology, radiotherapy treatment for pr
ostate cancer.
It typically takes him from 30 minutes to 3 hours to contour through every single CT
scan, identify a tumor versus nontumor. So, in a typical day, he could plan six
patients' worth of radiotherapy. We've applied open-source AI models to that. That
takes now 13 minutes. So, six patients an hour can now have radiotherapy planning. Which
means how many more patients will now have potential access to radiotherapy planning for
the future? And it's an open-source model. He's now a
dapting that to use it for glioblastomas,
head and neck cancers, lung tumors, and others. So, we live in a world with
massive amounts of information, increasing complexity. We also have a workforce
challenge around supply and demand. I think that the doctors of the future and the nurses of
the future and the pharmacists of the future, the healthcare professionals of the future
that have access to these models and work with these models hand in hand, may
be safer practitioners than perhaps
practitioners that don't use them in the
future is what I would put it to you. I have a very short video. It's like 45 seconds.
Would that be okay to play? So, I'm not sure if this is going to work as well, but we'll give
it a go. There's a short video on the left-hand side. And I'll try and talk through if it doesn't
work out as best as I wanted to. Can we try? Oh, here we go. So, this is the model. This is
demonstrating that these models can "see and hear." So, this is a model demonstrati
ng
-- yeah, demo what you can do with a bike. So, this individual is trying to say, "Help
me change my bike seat." Move up and down a little bit. And it'll give you the instructions of
what you need to do just based on the image. So, recognize the bike and it recognizes your
challenge. And it says to you that "If you can give me a more specific picture and give me
the manual, I'll help you out even further." And you might say, "Is this the lever?" It'll say,
"No, that's not the lever. That
's the bolt." And it will say, "Help me identify what you need to
do next." So, you take a picture of your toolbox, your Allen keys. You take the manual. And it will
identify where in the toolbox you need to go, which is the Allen key that you need
to use, and what you need to do next. And it does this without any specialized training.
I show this because actually, these are tools that will be ubiquitous in our day-to-day lives. They
are not specific to healthcare only. These tools are now
democratized around us. It's as if we're
living in the iPhone moment, the internet moment. And the question for us, I think, is how do we
responsibly deploy these? How do we responsibly adopt these and really think about the change
management that needs to take place for us to leverage these in a responsible, equitable
manner for the future? Thank you very, very much. [applause] Julia Adler-Milstein:
Thank you. Thank you so much for having me. Two terrific sessions already,
and I'm excited
to continue. I will say that I also have a 12-year-old. And so, I think any
chat GPT output that doesn't include the use of the word "bruh" with at least five years
is definitely not doing its job and will need some fine-tuning. So, that's the easy case where
you can easily verify the quality of the output. What I will be talking about, I think,
unfortunately, is that the pathway to the future does involve some more complexity
that is really happening on the front lines of healthcare delive
ry organizations that
are having to make the decisions of are these tools ready for prime time, and what
does prime time mean in terms of safety, effectiveness, equity, et cetera? So,
I think you're already seeing a theme of talking about the future really is about
talking about where we are in the present. And I'm going to sort of double-click down
and give you a perspective from our health system. What does this actually
look like on the front lines in terms of how we are approaching
th
e challenge of where the tools are today, and whether they are ready for
wide-scale use in our patient population? So, where we started this journey
was really the need for a vision, right? What are we trying to do with
these tools? We don't want to just implement technology for technology's sake. We
really want to ensure that they are solving our real-world problems and doing so in ways that
are trustworthy, which again, is, you know, a word that we all kind of intuitively understand.
But
we have to operationalize that. What does it mean to be trustworthy? And how do we develop
processes that ensure that the AI does meet principles of trustworthiness? That includes
some of the things that you've already heard about today in terms of longitudinal AI and
impact monitoring. So, I'll talk about that. And the other piece is AI at scale that
it's really easy to say we have, you know, 2,000 models implemented. What we're really trying
to do is to say we don't want 2,000 AI pilots.
We really want to figure out which are the tools
that are most valuable and get them to enterprise scale so that they can impact our entire patient
population. So, we are thinking a lot about how do you start with a pilot, rapidly determine whether
it's doing what it's supposed to do, and then have a plan for scaling that so it does get to the
enterprise level. So, those are really the two dimensions that we are emphasizing trustworthiness
and ability for the tools to get to scale. So, we
think a lot about this three-horizons
framework in terms of whether we are using our resources around deploying AI effectively. I
really like this because I think it's intuitive to think about what is your mature business,
what is your rapidly growing near-tum business, and what is an emerging future business
opportunity. Thinking about gas-powered cars, electric-powered cars, and the future of
autonomous vehicles. And the point is, we need to be investing on all three horizons,
right? We
need to make gas-powered cars better. We need to make electric vehicles better. And
we need to be designing for this future state. So, what does this mean for us in our
health system? It means that right now, we are just trying to make our current
fee-for-service healthcare system better by, you know, implementing AI for operational use
cases. Again, you heard about some of them already this morning. At the same time, we're thinking
about our rapidly growing business. How do we think about
generative AI for augmenting clinician
intelligence and really bringing together the best of human and computer intelligence as
that exists today and in the near future? And at the same time, we can start to
see a model for AI-driven virtual care, right, where we can get a sense for what might
happen with a given patient trajectory. Start to anticipate some of the lab tests they may need
and send them for lab tests before they even come and see a physician. So, such that interaction
is rea
lly fully optimized even before the patient touches our system for a particular issue. And
again, there's probably even more opportunities for that future state that we're just still
trying to get a handle on. So, we are, again, trying to design strategies that will allow
us to invest in AI on all three-time horizons. It takes a team to do this work well. I don't
think we have the answer. Is this the right team? Are we the right expertise? Is it the
right number? We are really all learning
this on the fly as we do it. At the heart of our
operation is our chief health AI officer. I think you see a growing number of health
systems that have these roles. They are the sort of data science and technical team, as
well as people with expertise in different sort of domains of clinical informatics. And they
oversee our governance process, our technology, and then they're really supported by people in
these other roles. So, my division is really the group that's being brought in to thi
nk about
evaluation and impact and measures of impact. And then we have our sort of clinical systems
health IT team. We have people who are really thinking on sort of that third horizon in
terms of like the research capabilities, doing more of the precision medicine work. And
we have a newly named chief research information officer that's, again, really pushing us
to that third horizon. While my group, even though we do research, is a little bit more,
you know, current state. And then of c
ourse, all the IT infrastructure and security and
our chief data officer that's really trying to work to ensure that we have that solid data
foundation that you already heard about today. So, I think a lot of what we're going
to learn about how to do this work well is who is the team that needs to be at the
table? How do you think about the expertise, how they work together? What are
the structures and functions? Our evolution over time has really started with,
like, what is the technical
infrastructure that we need to be able to rapidly
deploy different types of AI models? We're an epic shop. And so, we started
with Epic's cognitive computing platform and pretty quickly realized that it was not
going to be sufficient to meet the needs of AI at an academic medical center. So, we built
basically our own platform that we call HIPAC, that allows a much broader set of data
to feed into models in real-time. You can again see sort of the basic architecture
of it and how it integr
ates with Epic, but also has some unique features that, again,
allow us to deploy and monitor our models. Where are we today? We, frankly, even though I
think have been at this a while, we are not in the thousands of models deployed. If we look at
sort of where we are at an enterprise level scale, we're still in the early days. And we are doing
sort of a mix of models that are coming from our electronic health record vendor and some models
that we have homegrown or self-developed either by
people within our health system or by
researchers. A lot of these are really still focused on those sort of operational use
cases, things like the capacity management that you heard about this morning, you know,
predicting the use of blood products. So again, sort of clinically adjacent and relevant,
but they're not predicting diagnoses or sort of directing patients to, you know,
to certain parts of our health system yet. We're always evaluating new models, again,
mostly driven by the need
s of our health system, but with some capacity for what our researchers
and frontline clinical faculty are interested in. We have evaluated and turned off models
that we thought were problematic and have written about sort of that process of trying to
understand enough about some of the models that were available to us for us to determine whether
they met our threshold for sort of being ethical. And some of them, we felt like they just weren't
at our institution's ethical benchmark for sort
of equity of deployment. Again, it's a longer
story, and happy to talk about it at lunch. And then again, a lot of models are coming through our
pipeline. So, again, just to give you a real-world sense of sort of, you know, where we are in terms
of number of models that we're able to look at. We also have now GPT integrated, a secure
-- fully secure version. So, no data goes to GPT. So, it's sort of our internal product that's
integrated with HIPAC that I described before. So, what that me
ans is that if you want to use
a generative AI model as part of the model that you're deploying, it's sort of fully
integrated with the full architecture that then flows down to our EHR. Right now, we
mostly have GPT-4, but are anticipating adding sort of [unintelligible], and you
know, we ideally want to move, frankly, to more open-source models that we can, you
know, work with in a broader set of ways. If you look at, you know, the breadth of
applications of healthcare that we hope to to
uch with AI, it really is, you know,
moving from that horizon 1 to the horizon 2 framework. We're having a really interesting
meeting in a couple of weeks about how should AI support our medical education mission, right?
So, should we have virtual AI coaches for all of our trainees so that as they are going through
the process of training -- again, this is across a breadth of different trainee types, they can
have real-time feedback on their performance. Right now, it's very labor intensive
. They tell
us that they feel like the feedback is delayed, and they don't get enough of it. So, is
AI our solution there? So again, maybe, you know, six months from now, we'll be able
to tell you where we landed there. But again, just a fascinating breadth of conversations. So, we are evaluating use cases across all these
different domains. Again, we talked about AI Ambient Scribes. How can we triage high-risk
patients? Anyway, you can see down the list. I just want to keep going to get to
this last
section on this issue of the day around sort of AI fairness. There was a question that
came in. I saw an advance of sort of what can happen at the state level. And I
will say that a lot of our work got kicked off by an inquiry that came from our
California attorney general that asked us, how are you assessing your racial and ethnic bias
in the algorithms you have deployed? And we said, "Well, give us a few months and we will let
you know." And really, what we realized is that we
had to start with even what do we have
deployed? At that point in time, we didn't even have a sort of centralized way to know all the
different models that existed at UCSF Health. And so, we had to sort of come back
to centralize that function and then stand up this whole governance process. So,
I do think even just the inquiry can be very powerful at the state level to know that
someone's looking and wants to know what you're doing before we even get into sort of how
do you regulate this
and all of that complexity. So, as I said, we have adopted a set of
trustworthy AI guidelines. We decided to adopt the HHS guidelines because we think that they
really do emphasize the key dimensions. Again, so much complexity in how you operationalize
these trade-offs between different dimensions. I don't know that we've got it right,
but these are the conversations we need to be having amongst health systems to
understand how they're thinking about this. What we have stood up is what we
call our AI
Governance Committee. They are the final sort of gatekeeper for what models get deployed to
our patients. And so, I sit on that committee, a bunch of different people with different
expertise. And the models -- every model, whether that's research or an operational model
or an EHR vendor model -- come through the same governance process. And we sit down,
and we basically look at the models. We start with just discovery. Like, is this model
trying to solve a real problem? Have t
hey thought about the entirety of the solution? Not just
the model has a high-predictive value but like, do they understand how to integrate
it into workflow? If we implement it, are we sure that the interventions that will
come because of it are equitable? So, this whole notion that hasn't come up yet today of,
you know, patient-positive interventions, right? So, if we're going to predict a patient
no-show, you can either double-book that slot, right -- which is negative [laughs] for the
patient and if you have an already biased model, it's going to further worsen disparities -- or you
could say, "If a patient is predicted not to show, that we will provide transportation," right?
And that's a patient-positive intervention. So, we're really trying from the start to make
sure that we're thinking about implementation of these models holistically and sort of how
humans will use them and interact with them. We then do the development and evaluation,
retrospective evaluation, pro
spective evaluation, moving into pilot or RCT, and then the
adoption and ongoing monitoring. And what's important about this is it means we're touching
these models many, many different times. So, this is a huge amount of work just to even
take one model through this process. And so, what we're trying to figure out now is how do we
resource this if we want to be able to put 10, 20, hundreds of models through this process,
right? It's a huge amount of investment. Anyway, you know, lots of ch
allenges to doing
this. I think we're still learning how to do it well. It's something we're trying to do in
collaboration with a broad set of partners. So, if these names aren't familiar to you, I do
recommend familiarizing yourself with these groups of organizations that are trying to
come together and develop best practices. And as we head towards the end, I'll just say
that, you know, I keep coming back to making sure that the models are good and making sure that the
humans are good at
using the models. And we really have to think about those two pieces together. So,
yes, we have to think about algorithmic vigilance, algorithmic drift, but we also have
to think about clinician vigilance. Is it really realistic to think that if
a model gives a clinician bad output, that a human's going to be able to recognize that
and capture it and prevent that from getting to the patient? So, we're really trying to
bring these two together and think about measures and methods for both a
lgorithmic
vigilance and clinician vigilance. So, happy to talk more about our strategy
in the break, but we'll pass it off. [applause]
Judy Gichoya: Good morning, everyone. So, nice segue to talk
about the future looking at capacity building, especially with clinical vigilance, a lot of
pitfalls. Very tough on me because, I say, I work in four areas. I'm an associate professor
of Radiology and Informatics at Emory University. And one of them is bias and fairness, which I'm
very passionate
about. And I'm like "Well, maybe I should have talked about my work about that,
or the second one, which is capacity building." And so, here are my disclosures. So,
2016, this one godfather of AI said we should stop training radiologists. And this
had actually an impact. A lot of students became less interested in pursuing
Radiology as a specialty. Thankfully, you know, it's 2024. There's not anyone
who has enough radiologists to work for them. And he actually said he's leaving Google
too
because of these existential risks of AI. I show you this just because we
have a lot of hype even now, again, with the large language models that have
an impact. They have an impact early from training. You do have to really invest a lot
in making sure that the right message gets out, which is actually getting a little more
difficult. But I want to show you why doing this work -- again, I'm just following around
the same theme that it's very, very difficult. So, I hope some of you or most
of you have
heard of these two studies, one that shows pulse oximeter doesn't work well for dark-skinned
patients, and also the temporal thermometers also don't work as well. And, you know, some of
the people who wrote this work are actually very good friends of mine. We do this health
equity research using data science together. And I asked them "How has your practice
changed?" Nothing. You know, they know it doesn't work -- these researchers in this space.
We know the problem of hidden h
ypoxemia where you can send patients who really don't have enough
oxygen just because of the tools you're measuring. And it's really difficult to translate to
practice. And so, it's that theme that I'm going to use to inform sort of like what we -- what
should we be thinking about the future for us? Well, here's an example from a mammography
experimental test. What they did is that they had 50 mammograms, and they -- you
know, they changed the output of, you know, some of the mammograms. An
d then they give
them to read as of different experience. And it turns out that when AI is correct,
this -- unfortunately, if you're colorblind, it's going to be the last bar there. But the red
bar is your least -- your inexperienced reader. So, if you give a correct AI output to
your inexperienced reader, you improve their performance. And, you know, you raise them
up to almost the expert level. But if you give the wrong explanation -- in mammograms, you can
either say the patient has can
cer, or they really don't have cancer, and you say they have cancer;
or the other way around, that they have cancer, but you say they don't have cancer. And it shows
that you even bias your expert performance. And this sort of -- one of the big futures is
understanding how does AI and the human -- in case we're not anticipating replacement
-- hold hands together for an impact, remains one of the biggest challenges. And
here's an example in psychology. They are able to get through a beat for
these sort of
experiments where they can lie to the reader. And one of them here is just they run
three experiments with an imaginary use case that is saying one patient has disease,
one doesn't have disease. And what they show is that when you -- one of the experiments
always has an error, a systematic error. Forty percent are always incorrectly classified.
And they show, when you give AI assistance, the error rate goes up rather than when
you just let the human work by themselves. And t
he second phase, they start off, you know,
giving one group -- the blue one has an error and -- has -- uses AI, and the down one doesn't.
And it shows that you start off using AI and then you don't. And those participants showed that
when you use AI, your error rate is high. When you stop, it goes down. And then they flipped the
experiment. For the people who did not use AI, they used the AI in the second phase.
And they showed that the errors persist. I think this study -- it wasn't really
saying, "Hey, it's pneumothorax," but to me, it was very concerning because it shows
that the automation bias can persist even when the AI is withdrawn. We've seen sort of a
system that has actually pulled out AI. And so, to me, again, we -- you know, I had
to still sneak in a bias slide here. But it shows that we have a very, very complex
system, as we start to think about what the future deployment of these systems is going to
be. You have to decide, are you going to think about how you
deploy, or are you going to think
about how you build? We had some comments that, you know, "We are building a fair dataset." I
can tell you that that's impossible, you know, the model evolution or how you deploy your
system across the complexities of organizations. So, my talk is around capacity building
I've been involved in many, many years. The first time when Geoffrey Hinton caused
a drop in the radiologist's interest in AI, I went back and did a lot of outreach to try
to think, how
do we educate? Today, I run the AI elective at Emory University. And what I can
tell you is that this field is changing so fast. Today, when I look at my computer scientist versus
two years ago, they rarely rely on the computer scientist when its residents because they have
copilot that helps them even program. So, domain expertise is becoming very important. And so,
I think one of the biggest takeaways if I think about the future, even if I don't have a crystal
ball, is to think about whos
e voices are in the room. And to me, that's what policy is, as someone
who's just learning about how to be in this space. And if you think about epistemic harm, it's "the
contribution of health equities that is rooted in knowledge itself, its formation, shape,
set-up, and effectiveness. And it's embedded in the knowledge, in the ways in which it's
created and used to exert power." And really, it's around who gets a seat at the table?
And if you see, this is not an imaginary case about whose
voices is represented. It's
mainly the CEOs of the big companies. And they're very far removed, in my opinion,
still from the realities of healthcare. And so, we see quite optimism that we are going to
change how they burn out. We're going to make sure the patients get healthcare better. But the person
who's left behind is really where -- you know, who -- the person who's taking care of the
patient. And, you know -- and I'll show you that sometimes we think about the mental
models of how
we think about our world, how it should and could be, the way it is, the
way it's according to data, and the way it's according to model. And now, we have according
to AI. And the person who's determining how this world is according to AI is very different
from the person who's caring for the patient. And so, some of the examples, for example, social
problems, can be fixed by for-profit data-driven solutions. We've had a prior presentation saying
maybe we're thinking about open source. We a
lso have to think about closed models that are
open for use that encourage innovation. We see quite a lot of abstracting how much you
eat, but they don't really think about the underlying complexity of obesity or access
to -- you know, to the correct food. And so, we see this a lot of technology being
postulated as the solution for everything. We also see this focus on, you know, efficiency.
Today, we want to solve doctor's burnout. But I can tell you, soon, it's going to be how many more
patients can you treat and, you know, squeezing and squeezing and squeezing the efficiency model.
I can tell you that we have to think still about, what is the value? And also, big -- you know,
thinking about that investment in technology will augment human intelligence more than
investment in our fundamental pillars, for example, education, culture, you know, and
the institutions. And so, it's this gap about where should we put our limited resources,
and how that is going to change the fut
ure? And so, when I think about can we really -- what
have I learned in my years working in this space, it's really been around learning with
others. And that's -- I can try to tell you what I -- what my thought process has
been. And it's this concept that we realize it's very difficult to go through school to
change curriculums to really learn, "Okay, the Kaiser data, what are pitfalls," unless
you have people working on those datasets. And so, for a long time, we've ran this
Datathon. So
, we bring different experts to work on the datasets. And I can tell
you some of the realities tend to be "Oh, I didn't realize that my recording of the,
you know, arterial blood gas five minutes later -- because it's not automatically captured
-- has an impact on how the data are analyzed, you know." So, it's this gap. But it turns
out, with the technology that is getting there, that programming is really -- can
be done by high school students. We have to really rethink about how do we
th
ink about educating and localizing concepts of fairness, accountability, transparency, and
thinking and moving to this HASTE camps for, how do we think about this concept of
discussion around policy? And, you know, this is an example of one that was done in Uganda.
I can tell you, as someone who does fairness work, I only knew that I'm Black when I moved
here 10 years ago into the U.S. And so, fairness has a very, very different
concept based on where it's done. We know that in Europe, ther
e's actually a
penalty for social risk classification and social scoring systems. We know that in our
country, there's a big debate about the woke and the not woke. And it is getting very,
very tough to do this type of work. And so, we have to be really intentional in terms
of how we're thinking and actualizing and conceptualizing what we do and think about
policy. So, I started off with 2016. Thankfully, I still have a job. And the dance are dropping
off, but also trying to tell you what
has informed really is looking for advice, how should we
be thinking about even how we fashion policy? And I want to leave you with a call
for you to emphasize on learning, learning from the past mistakes. We see a lot
of technology being slapped on, for example, EMRs, which all the doctors and the nurses
hate. We think about how do we empower this learning workforce and learning healthcare
system if we're going to think about policy that doesn't come 10 years later? And then it's
this emp
hasis on this rapid learning cycles. How are we -- by the time our paper is
published, that's almost like, you know, two or three years down the line from
the intervention. How do we encourage especially learning from failures and
regarding -- especially from the past? If you think about most of the innovation
has been focused on a very small subset, if we are going to envision a future that
cares for everyone, we're going to need to do a lot. And so, thank you, again, for the
invitation t
o be part of this presentation. [applause] Ainsley MacLean:
Right. A big round of applause once again for all of
our panelists. Thank you so much. [applause] So, I'm sort of still processing so many messages
that came through really strong and powerful, and I'm sure the rest of you are too. I think
I'll start off by distilling down a lot of what you said and just asking, if there's
one takeaway you have for our audience, as they go off and do all the
important work they're doing related to
implementing AI in healthcare,
what would that be? I'll start -- yeah. [laughter] Junaid Bajwa:
Start with me. Thank you. So -- can I have two? Ainsley MacLean:
Yeah. You can have two [laughs]. Junaid Bajwa:
So, one is -- so, I think about two-by-two matrix, if you
will, around risk versus complexity. And if I think about healthcare all up, I think
the clinical facing use cases that we consider are generally high-risk, high-complexity
work. And like colleagues have described, like Julia and
Judy were talking, there's
a lot of work that needs to be done if we think about the implementation of those.
But the business of healthcare may afford us the opportunity to build a muscle on
low-risk, low-complexity kinds of work. So, I wonder how much productivity we might
be able to unlock if we think about the impact of these tools in the finance part of our
organizations, the HR parts of our organizations, and focus on boring, mundane activities
that might unlock hidden potential. So,
that would be one kind of
axis for us to think about. I also would just encourage people to play.
Think about what you need to do to build your own experiences and muscle. Don't immediately
deploy them into clinical patient-facing tools and services, but just experiment with these
models and tools and see what value you can have. Don't think about it as prompt engineering
but think about the act of just posing beautiful questions and seeing what the response
might be and focusing on your
own curiosity. Ainsley MacLean:
I love that advice. And that concept is one I've heard a lot recently,
really trying to focus on the lower risk, but really high-impact areas. And then before --
I think sometimes we're tempted to kind of dive right into that slightly higher risk area which
can be more challenging. Thank you so much. Julia? Julia Adler-Milstein:
Yeah. I think my key takeaway is that we have to do this work
collaboratively. And there are many forces that push you to not be coll
aborative, right? In
some ways, the work is so hard and complex that, like, we can just spend all our time trying
to figure out how to get it right for UCSF and really miss the important opportunities to have
conversations like the one we're having today. And so, I think it's to really sort of push
yourself to, like, come to these events, talk to, you know, other organizations
like yours. Because I think that is really where the rapid learning and the sort of
avoiding common pitfalls will
happen is by doing this collaboratively. And I worry
that we don't have enough structures that are forcing that to happen. And we have
sort of all the pressures just to stay in the four walls of our own organizations
optimized for what we're doing locally. So, that's, I think -- you know, just
continue to come to these -- you know, there's a lot of different coalition's that
are trying to put in place those collaborative structures. So, just, you know, go to one
more of those than you migh
t otherwise, because I think that is really what
the key to the future is going to be. Ainsley MacLean:
What great advice. I was struck hearing UCSF's journey, how similar it was to Kaiser Permanente.
So, it was really refreshing. Judy, how about you? Judy Gichoya:
So, in addition to learning, which is what I would say is -- my reaction to your comments is
you may not have a choice really in what's coming down if you see sort of like every week there's
a new policy document or new guidance w
e see now every time I go to work. When the Joint Commission
is coming to visit, there's even an overhead like the Code Blue announcement that the Joint
Commission is here today. And the Joint Commission is now interested in what you're doing with this.
We see, for example, the AG of California is here. And I think that can get very, very overwhelming.
So, my thing is to not be afraid to fail and to just get started. And usually, I say sometimes it
just means picking a simple thing. It does
n't need to be, again, patient-facing and just touching,
because it shows you so much about your own internal organization. For example, which deputy
should you deploy? Even just that simple thing, you'll find out that there are so many
bureaucracies and partnerships between AWS, within Azure, within every cloud provider.
That's going to spend years for you to solve. So, if you think about your small
experiments as learning and helping you get to a pillar that is going to help
you to be su
ccessful, I think, ultimately, you're going to be successful. And that also
means looking in your own divisions to see what people are doing. Sometimes you're
very blinded to what medicine is doing, and yet Radiology is miles away. And that
can be an easy way to just get started. Ainsley MacLean:
I love that. And I loved what you shared also about Radiology.
And I think that the lessons from Radiology are really a lesson for everyone in the room who's
in a profession that's been told that A
I might ultimately replace them. I think Radiology
is now the leading medical specialty in publications within artificial intelligence. And
I, too, was in the room many years ago and was told that everyone in the room should stand up
and wave goodbye to the table of radiologists. [laughter] So, I always like to say we're
still here, right? Sorry about that. [laughter] So, such great answers. So, I guess I'm
going to get a little bit trickier. And maybe I'll start with you, Judy. When
you'r
e assessing a project to implement, an AI project, how do you determine the
ROI? How do you determine if this is the right investment we make, whether it's a
pilot that we're going to scale for Emory? Judy Gichoya:
So -- and this is the problem of being a researcher. Because you can do so many
grandiose things without -- and kill them without consequence. My work has always been really to
think about how AI works, does it really bring value? So, anything that I -- I stopped moving
from diag
nostic models quite early in my career. Although I'll tell you that I'm back there for
one reason, is because I felt that -- and really, even the research now shows that you don't really
get a lot of efficiency gains when you use AI. You save radiologists maybe an hour or two. Like,
it's really not worth all that investment. And so, the thing that I have come to when I pick
a project, I do obviously have a lot of students, so I look at the learning opportunity. But also,
when I think about
the value for healthcare, I look at what is going -- what is it going to
do? And living in this gray area which is our AI world is something that everyone in this
room has to be. And the reason for that is, for example, one of my lessons this year is that
all the metrics that we use for AUCs, they're useless when it comes to a radiologist, how I use
AI. And it took me a long time to come to that. Because what I want to know is, if I'm looking
at this chest x-ray, is the AI usually right or
wrong? And that's almost like a positive
predictive value or a negative predictive value, which are never ever reported. And I would
never use those to buy. I want to know, if I'm going to hedge and say, correlate
clinically, how forceful am I going to come down? And so, to me, I feel that the
personalized metrics are a little off. The second area, as someone who works in
health equity, is that the metrics are all over the place and really don't translate
to outcomes. So, I love the word c
linical vigilance. Because if you give me a model that
performs at 56 percent ejection fraction versus 55 percent ejection fraction, to me, they're
kind of a little bit the same. It doesn't make sense to keep tuning and tuning because I'm
not going to change my management. And so, that bridge to outcome and that
bridge to personalization is how, today, I'm thinking about it, but that's
also an evolving process over these times. I said that I'm going to come back to
the predictive model. I
just came back from home in Africa, and the quality of care
still needs to get better. And in that place, I really should be building predictive
models and more predictive models. Ainsley MacLean:
Great. Thanks, Judy. Julia, how about you at UCSF? How do you determine whether to make
an investment in a new technology or use case? Julia Adler-Milstein:
Yeah. I mean, I'll agree that, like, we are not doing traditional ROI calculations
because I think we don't have the investment piece well-def
ined [laughs], and we don't
have the return piece well-defined. We're still trying to figure out just what are the right
metrics to assess. But I think we are starting, as has come up before, with like the use
cases that are in your sort of high-value, low-complexity bucket, like what we call keyboard
liberation, I'd say is mostly our focus and where we're just convinced that there is a return on
investment because we know that our clinicians want to be doing less documentation. And
we've
seen the performance of the tools, and I think are pretty convinced that
they are good enough for scale deployment. So, I think, right now, it's -- we are --
it's almost more of a sense [laughs] that there is an ROI there rather than like
a traditional calculation. And I think it'll get more and more complicated as we
move into some of those, sort of, you know, second and third horizon to think about,
what are the right metrics to -- you know, to assess, in particular, the benefits?
I thin
k we'll get a better handle on the cost. But I think that benefit one is tricky
because there's lots of different dimension of benefit. There's also going to be risk in
there. So, I don't see that we're going to use sort of hard and fast ROI calculations and
thresholds for decision making in the near term. Ainsley MacLean:
Julia, could you talk a little bit more for our audience
about what those tools are to really help with physicians? Because some of
them may not be as familiar with them.
Julia Adler-Milstein:
Yeah. Absolutely. I mean, I'd say the two that we are focusing on again at this scale benchmark
are, like, Ambient -- you know, basically, an AI scribe. So, it is something that sits in
the exam room, listens to the interaction between the clinician and the patient, and then generates
the draft note for the clinician to review, edit, and then that goes in. So, it's still a human in
the loop. It's not sort of fully autonomous. But again, we've seen just even in the last
year that
the tools are, I think, maturing quite quickly. I'll also note we're going to pilot two
different tools to see which one our clinicians like better and are partnering
with another health system that's going to pilot one of the same tools and then a third
tool. And our goal then is to put together the results of the two pilots to basically be
able to do an A to B to C comparison. So, again, this is where I think the
collaborative work can really be valuable. The other thing worth
turning on is GPT
response to inbox messages. We know that our clinicians are mostly feeling crushed by the
volume of patient messages that they're getting, and we don't want to turn that off. But if we can
make those a little bit faster and more efficient by offering them a pre-populated drafted response,
then that I think is a win-win. But again, are they going to take the time to edit it? Will
they catch a hallucination? You know, these are some of the risks of turning on these models.
So, we're also trying to sort of, you know, validate and make sure that they are sort of
clinically ready across a breadth of scenarios. Ainsley MacLean:
I love that. And our radiologists have been using sort of GPT-generated
responses in their impressions of their radiology reports. And one of the things that we hear is
just how much that cognitive burden is really decreased when they're using that generative
AI for just a small part of the report. So, after a long shift, they're able to,
you know,
go out to the gym, drive their kids around, whereas before they would have just crashed in
their bed [laughs]. So, I was curious, Julia, what sort of response have physicians been giving
to this technology work you've been implementing? Julia Adler-Milstein:
Yeah. I mean, I think it's exactly what you said. It's not necessarily a
huge time reduction -- I think you mentioned this too -- but it is less cognitive work. And
it is more empathetic. And, you know -- so, it's -- I think o
ur sense is that -- you know,
that it is relieving some of the burden. Again, we're really early on in the journey. And,
you know, we're going to measure and monitor, like, which clinicians are continuing to use
it. You can turn it off and stop using it. So, I think once it's sort of out there at
scale, we'll really understand, you know, who's finding it beneficial, who's not? And
like, what explains that? But I think for now, there's a lot of interest in the tools. I'll
just note, I'm per
sonally very curious because some of our clinicians have started to say,
"Well, the GPT might say, 'I recommend you come in for a visit,'" right? And then does
that shift the physician's propensity to say, "Actually, yes, I agree. I think that patient
should come in for a visit," or not. I mean, I think there's going to be a lot of,
sort of, nudges and other pieces of these. So, again, it's just why, when we talk about
the -- thinking about the outcomes and impact, you have to take just suc
h a broad set of
measures. So, they could tell us it is, you know, less cognitively burdensome,
or it isn't saving time. And like, we're going to have to put
this all together and say, like, "Is this the win [laughs] that we want or
not?" And it's very complicated to kind of make that final call in terms of, like, whether
the tool is doing what we thought it would. Ainsley MacLean:
Wonderful. And I do think it's interesting how the areas you focused on, as been
seen with KP, are the areas
that have the biggest bang for the buck for our people. Because I do
feel that, you know, our people -- and I'm hearing that from you as well -- are really our greatest
resource. Wonderful. Thank you. Junaid, talk about -- you're looking at it from a different
angle, right, not necessarily care delivery, but for Microsoft. How does your company determine
where that -- those investment dollars should go? Junaid Bajwa:
So, from a health and life sciences perspective, I sit on the R&D
side of
Microsoft. And what we don't do is say research for hire. But we do think about,
actually, what are the biggest problems and challenges that we could try and address? And
so, if I think about that exam question on how might we transform the practice of medicine
with trusted, reliable, human-centered AI, you could do lots and lots of things, right?
You could choose to work provider centric. You could choose to work in drug discovery, in drug
development, or across the entire continuum. And w
hen we think about where we're going
to spend our time, resource, and focus, it has to be on problems, ideally, that nobody
else has solved before. So, we think about the horizon three category that Julia was
referencing. And we think about those deep, deep problems that can have transformative
impact and move the needle on an industry but can also have real world impact. And
almost think about almost a public value proposition if you will. And if we were to
do this, what would be the tota
l impact on society that aligns with the broader mission
around empowering everybody to achieve more? So, there are some things I could probably
share -- perhaps over lunch -- certain things I can't talk about publicly just yet. But we get a
chance to really explore, with these novel tools, what the art of the possible might be? We kill
way more things than we actually then deploy into practice. But things like that radiotherapy
planning tool, for example, do something that could tangibly m
ake a difference from a
3-hour interaction to a 30-minute interaction, adds both productivity and reduces the
cognitive load, but also has a deep and meaningful impact on cancer potential outcomes
is -- are the kinds of work that we like to do. Ainsley MacLean:
Wonderful. And I'll put it back to you. What are you most excited
about as you look at this next year? What are, sort of, the big things we
should be on the lookout for? Junaid Bajwa:
So, I -- can I go longer than a year? Because --
[laughter] -- I think in a year's time, I would hope
that organizations such as yourselves, Emory, UCSF, have become a lot more confident
on where these tools can really work and where actually you need more evidence to deploy
them in a meaningful way but if we kind of extend it out a little bit. People talk a
lot about this notion of precision medicine. I think we live in a world where we practice
medicine on a law of averages. We discussed earlier the challenges around equity, fairness,
and others. Can we push ourselves to be better on that? And could we move to a future
where we have more precision diagnostics, more precision therapeutics, more precision
ambient intelligence that allows my practice of medicine to change and allows me to
actually practice in the future that allows me to be much more precise over
time and way less focused on averages? And if I had access to information when -- so,
I work in primary care. It's very different to secondary care settings. But i
f I knew not only
what's in my medical record, but I knew more details around genotype, phenotype, environment,
social circumstances, if you will, I think I could really offer better care to the patients in public
that I serve as a consequence of that. And I think that's the future that we might be able to enable
if we make the right choices now for the future. Ainsley MacLean:
So, you're talking about within genomics and
pharma and targeted treatments? Junaid Bajwa:
All of the above. Ainsl
ey MacLean:
Okay. Great. I like that. Julia, how about you? What are you most excited
about? And you can go beyond a year also. Junaid Bajwa:
[laughs] Julia Adler-Milstein:
Yeah. I mean, I feel like at the most conceptual level, I am excited for the
moment when our care teams, our patients say, like, "We love this technology [laughs].
It is making healthcare better." I mean, I just feel like we've been
talking about the promise for now like decades. And it felt -- it's
felt like it's contin
ued to fall short. And I just -- I really do think that in the next
few years, we're going to have that breakthrough moment where the technology is going to start
working for us better. It's going to make, you know, the whole process feels more efficient,
safer, you know, like it's really putting our people, as you said, at the center. And it just
-- I just think that will be such a satisfying moment because we've been putting in so much
time and effort, right? So much data has gone into th
ese systems. It feels like what are
we really doing with it? Is it adding value? And I think we're going to start
to see it really feel tangible, that value moment. And then we're going
to get to do so much more exciting stuff because we all sort of won back the hearts and
minds of the people. And they will then say, "Okay, I do love my EHR as much as I love my
iPhone," right? Like, that will be a huge moment. Ainsley MacLean:
Wow. Julia Adler-Milstein:
And, you know, I'm not saying that's
one -- Ainsley MacLean:
I'll retire on that day [laughs]. Julia Adler-Milstein:
-- to two years. But like -- I mean, I really do feel like we're -- it's starting
to feel like we're flipping towards that. And I just think that's going to be such an
exciting moment. Again, a lot of hard work still to get there. I don't want to trivialize
it. But since you asked what I'm excited about, it is really that moment in the sense that
it is finally starting to be on the horizon. Ainsley MacLean:
I lov
e that. I'm getting chills as you're talking about it because I also feel that coming. Judy,
how about you? What are you most excited about? Judy Gichoya:
I think two things. One is we're going into an election year. And
I think we've seen quite a lot of generative AI uses beyond healthcare. And I think they're going
to serve as a big lesson to the potential and the harms of this technology. So, we're going
to work what the Hype Cycle does. I think, as of today, I saw, for example,
the Meta
toolkit that was done to show this is AI-generated, did not
work, was broken down in a few hours. And it's going to really show the complexity of
working in this space, not just for healthcare, for this general business. And that's going, I
think, to ground or move some general policies, for example, health AI literacy for the
population, even if it's at a sixth-grade level. And so, I think we're going to see those societal
changes. And that's going to be good to allow the hard work to con
tinue, for example, in mission
critical areas like defense and healthcare. And then the -- personally, I'm very, very
excited about the potential of using AI as a hypothesis generation ad when I think
about, if I had all the data at Kaiser, that's what I would do. And the reason for
this is that AI has amazing abilities to pick up trends that are not very obvious to us, that
will take us years to come up with hypotheses, not just for patient care, but just our practice.
I'll give you an ex
ample. As someone who does Radiology AI from a chest x-ray, today, I can
tell you the image-based age of the patient, the risk of the patient, their cardiac ICD codes,
their healthcare costs with amazing ability and even a mark of kind of where they live when you
think about the area in the provision index. And those things, if you have to hire the
biggest workforce to do that underwriting for you, it would be impossible. It has tremendous
opportunity for harm. Because you can say, "I'm not
going to cover this group of patients."
But if imaging only, there's more to an image, and we can harness that from AI. This is, to
me, much more important than ever augmenting a radiologist. And I hope that we're going
to see some of this bold population-based use cases and opportunistic screening as
something that I'm very excited about. Ainsley MacLean:
I love that. And I'm really excited about the role of AI within women's
health, especially in the breast cancer screening end-to-end so
rt of diagnosis. So, thank you so
much. So, I think now we'll open it up to our audience and see if we have any questions
from you all. I'm sure you have several. Amy Andersen:
Well, thank you so much. This is an amazing forum. I'm so happy to be
here. Amy Andersen, I'm with Oracle Health. And I just want to thank the clinicians and researchers
for grounding this discussion where it should be. I spent a lot of my time within Oracle, you
know, engaging with our leaders around sort of first p
rinciples. It's about -- this is
-- I've loved how Vivian mentioned, you know, people caring for people. And that's really what
we need to focus on. I think what I learned from this forum particularly is the importance
of really looking at the delivery of care, how data can inform, and then where do
we -- where do we apply our priorities for AI? We hear this all the time from
our customers. Where do we start, right? And I'm really passionate about looking at
sort of how can we enable the s
tandard of care and practice, which we know
is evolving all the time? And I think the highlights that you have mentioned,
particularly around the clinical experience, is what are those roadblocks, right?
So, we see with personalized medicine, in particular, the increasing use of
NGS, right? However, what we know from clinicians is, those NGS reports are not
necessarily helpful in and of themselves. So, you know, one of the things that we're doing
is working with some innovative companies t
hat are applying, you know, AI to those tests in order
to deliver insights that are based on, you know, evidence. So, I love this. I love kind of thinking
about, you know, being practical innovators and also ensuring that we have the ethical guide
-- you know, guidepost for us. So, thank you. Ainsley MacLean:
Thank you. So, I'm not sure if there was a question, but you could
-- I would just latch on to one word you said, which is roadblocks. So -- and I think that I do
want folks to leave h
ere knowing that sometimes there are roadblocks almost all the time in
terms of implementing them. And I don't know if I could put that to the panelists because
I think it's a great question. What are the biggest roadblocks you've seen in terms
of -- and obviously, you're going to come at it from a different angle, but in terms of
trying to get this technology out to be used. Junaid Bajwa:
So, maybe I can answer it from some of the
NHS experience, if you will -- Ainsley MacLean:
Okay. Junai
d Bajwa
-- right? Ainsley MacLean:
Wonderful. Junaid Bajwa
So, implementing any change is extraordinarily difficult, right,
from the simplest thing that you have to think about doing. And even if you believe that
the right -- it's the right thing to do, do all of your stakeholders believe it's the
right thing to do? So, we've -- at the hospital I sit on the board of, we implemented
Epic, pre-pandemic, massive investment, really good change management. And the change
management investment ac
tually paid massive dividends because it took the clinical workforce
and the non-clinical workforce alongside. And I think where I've seen things fail is when
the change management hasn't been considered, and you just say, "Hey, deploy this
thing, and you'll reap the benefits, won't you?" And they haven't taken
the entire multiple tribes that exist within any given healthcare
system along that journey with you. But making -- understanding that change is
hard and understanding that you have
to get stakeholder buy-in and understanding that you can
start off with some relatively small, meaningful, impactful projects in a view to getting on to
more complex things is probably the journey that I would encourage people to consider. And
the NHS has multiple, multiple instances of where things have failed over time. So, learn
from others and experience from others too. Ainsley MacLean:
I love that. So, change management, absolutely -- Junaid Bajwa:
Yeah. Ainsley MacLean:
-- critical w
ith this. Julia, how about you? What roadblocks do
you see being sort of the most -- Julia Adler-Milstein:
Yeah. Ainsley MacLean:
-- significant? Julia Adler-Milstein:
I mean, I frankly would say I don't think that there are true roadblocks in
terms of things that are stopping us. I think what they are, are speed bumps. And we're trying to
decide how much to slow down as we see them. So, probably the hardest things are,
like, safety and equity, right? As we look at the tool, like what is the
actual performance on safety and on equity that is good enough to deploy, right? And
then you have to just have really complex, nuanced conversations about, like, what is
the current state, right? Because it may be inequitable but as long as it's better than
the current state, like, is that good enough? Ainsley MacLean:
[affirmative] Julia Adler-Milstein:
And so, I feel like you get very quickly into sort of ethics and
ethical discussions. And then I will just stay on our AI Governance Com
mittee. It feels
like, are we really well equipped to make these decisions? You know, I'm not an ethicist
and -- you know? So, I just think that -- Ainsley MacLean:
[coughs] Excuse me. Julia Adler-Milstein:
-- that is where -- it's like how sort of slow or quickly to move is really, to me, what feels like
the hardest and scariest part of these decisions, right? Because it's sort of one -- you know, one
wrong word and a GenAI message that goes back to a patient could really be life or death,
and like we
are the final arbiter of whether that model should go forward or not. So, the stakes feel very high.
But yet, if there's broad value to that tool, we don't want that, you know, possibility
of one thing going wrong to stop us. So, it's -- I'd say, it's those that are -- to me,
feel like the biggest challenges right now. Ainsley MacLean:
Thank you, Julia. Judy, how about you? Judy Gichoya:
I'm going to just respond to this as opportunity. What's
the opportunity today? And first of
all, you should have a team, and you should staff it
and provide the resources that you need for it. I think we are missing the patient voice. And
this is because we're trying to reinvent how to communicate results to patients. And I believe
that there's quite a lot of opportunities to have an engagement. For example, you show me that
GPT will be used for a rare disease diagnosis, then you know the realities of patient privacy
or anonymizing someone with a very, very rare condition is near
ly impossible. Yet those are the
same group of patients that can benefit a lot. We see a lot of patient advocacy groups
for those patients with rare diseases, but they're completely, you know, not at the
table. And I think that if we can engage them, I think we're going to mitigate some of the
risks downstream, because we will, you know, be very forward with what worked, what failed.
And I think it's just a blind spot to most of us. And then the other surprising thing for me is
-- as watch
ing this deployed to systems for Radiology at Emory, is the human factors. And I
think we underestimate what that will look like. For example, one of this is triage algorithm. We
know outpatients are long, long, long. You don't want to be sitting on a patient with a bleed
when you should have read the study early. But showing people where they missed, where they
-- you know, and documenting them has a tremendous impact even how you deploy the system. Because is
this discoverable? Are you go
ing to start to say, "Judy misses a lot of these cases all the time.
Maybe she's not a good radiologist," when they come to review me in my annual feedback? You
know, you're going to say, "Who else knows?" And that can have such negative consequences that
some people say, "We don't want you to deploy it." So, the technology, I do believe, with the answer is we'll get better. But the
human factors and the human side are really, really behind and is underestimated in
the success of this techn
ology deployment. Ainsley MacLean:
Thank you. I think you can see why these are all such
seasoned leaders. We went from roadblocks to change management,
to speed bumps, to opportunities. [laughter] All right. Great. And any other
questions? I see a couple of hands. Neil Carpenter:
Sorry. I'm trying to figure out how to turn this on. Okay. Sorry
about that. Neil Carpenter, Redesign Health, and Tau Ventures. Let me just kind of change the
level of question, if we can, for just a second. A lot
of how we've been talking about this is from
sort of inside the care delivery system, out. So, let me ask a very different question, is if
we sort of took a step back to a societal level -- because this is like game-changing technology
and said -- as a society, how do we need to think -- what is something that healthcare system
could do differently with this technology that'd be beneficial to society, which is kind of above
the clinician, above a specific patient, above an administrator, a
bove a P&L? What is maybe one
thing that it could do in the next 5 to 10 years? Ainsley MacLean:
Well, that's a very easy question. So, I'm going to turn -- [laughter] Judy, I'll start with you,
just to switch of the -- Judy Gichoya:
So, this is society -- I would say that when I flew yesterday through Atlanta Airport, one
of the busiest airports in the world, I saw a lot of digitalization, digital ID, CLEAR, TSA Pre. I
think we're underestimating displacement of jobs. And I think we should
be looking at what that
future is going to look like. Not replacement, but there will be -- I can tell you that in
five years, when I go through TSA to boarding, that there'll be very, very minimal human
contact. It's very clear that you can see that this is going to happen. But we are
not thinking about how to educate today our students to live in this world where they can
search online and get the answers. What are those important skills to build the next generation
workforce is what I w
ould spend my money on today. Ainsley MacLean:
Wonderful. Julia? Julia Adler-Milstein:
Are we getting signaled? Am I going to -- Judy Gichoya:
[laughs] Julia Adler-Milstein:
-- escape with our time's up, or should I go ahead? Judy Gichoya:
[laughs] Ainsley MacLean:
You know what, I think that's the perfect place to stop, TSA -- [laughter] -- since we're all dealing with
the TSA. All right. Great. Well, thank you once again for all
of you for being, I think, so honest really with the problems y
ou're
dealing with and also being very brave and paving the way for the rest of us. So, thank
you. A big round of applause for our panelists. [applause] Rebecca Flournoy:
All right. A huge thank you to Dr. MacLean and our panelists for such an interesting
conversation. We are going to transition now to lunch, which is going to be served in the very
back, buffet style. We encourage you to bring your lunch back to your tables. Or for those of you
not seated at a table, there are tables near t
he buffet area. And please do be back in your seat on
time. We will start right at 12:30 p.m. Thank you. All right. So, our final panel today focuses
on addressing risks and harnessing benefits of AI in healthcare through policy
and practice. I'm so excited about this panel. And I'm very excited that Dr.
Daniel Yang will be moderating. And we've also asked him to share some closing
reflections after the panel as well. So, Dr. Yang recently joined Kaiser
Permanente as our new vice president
for Artificial Intelligence and Emerging
Technologies. And in this role -- [applause] -- yes, people are clapping. We're
very excited. Yes. In this role, he's responsible for ensuring oversight for
all AI applications for the organization across clinical operations, research,
education, and administrative functions. Before joining Kaiser Permanente, he was a program
director at the Gordon and Betty Moore Foundation, where he founded and led a philanthropic program
on diagnostic excellence
. In this role, he helped to establish several public-private partnerships
to promote responsible use of AI in healthcare. He created infrastructure to support development,
implementation, and evaluation of diagnostic AI algorithms. And he advanced research methods for
evaluating clinical impact of AI in real-world settings. Dr. Yang is also a practicing internal
medicine physician. And he completed his residency at UCSF and a fellowship in Healthcare Systems
Design at Stanford. Dr. Yang, I
'll pass it to you. Daniel Yang:
All right. Hi, everyone. First of all, I'm really excited
to just be here at this event. I feel like it's a homecoming in many ways. I'm able to
connect with all my KP colleagues I've been seeing on Teams in these little boxes.
And to see them in person is remarkable and really just understand the national
impact and footprint of this organization. And too, I just see so many familiar
friends and colleagues that I've gotten to know over the years, in the
au
dience, on the stage. So, for me, this is really kind of bringing these two worlds
together to talk about a topic that, I think, we're all passionate about, which is ensuring that
we're deploying AI in a way that's responsible. I'm really excited to be joined on the
stage by three people that I admire that bring complementary perspective and expertise to
this topic, particularly as we think about the risks and benefits of AI and what that means for
policy and practice. The theme for the dis
cussion is really, how do we move from what we know we
should be doing, to actually getting it done? And so, let me introduce my
speakers. I'm not going to embarrass you the way that Rebecca
embarrassed me with my whole bio. [laughter] I'm just going to provide some, you know,
key bullet points to provide context of where you're coming from, and the expertise that you
bring. So, first of all, we have Tom Romanoff, who is the director of the Technology Project
at the Bipartisan Policy Cente
r. Now, prior to working at BPC, Tom led IT initiatives
for several federal agencies and advised executive leadership on the impact of new and
emerging technologies on government operations. Second, we have Laura Adams who is special
advisor at the National Academy of Medicine, where she leads the development of
an AI code of conduct. Prior to NAM, Laura was the president and CEO of
the Rhode Island Quality Institute and has been a longtime expert in
healthcare data interoperability. And l
ast but not least, we're joined by Dr.
Maia Hightower who is co-founder and CEO of Equality AI, which helps data scientists
develop fair and unbiased algorithms to eliminate discrimination in ML models. Prior
to Equality AI, Dr. Hightower was a physician executive at multiple academic medical centers.
Most recently, she was EVP and chief digital transformation officer at the University of
Chicago. And she was also the chief medical information officer at both the University
of Utah and als
o at the University of Iowa. So, just a quick note on how I plan to run this
panel. It'll be slightly different than the last panel. I'm going to bring up each panelist to
provide some remarks. But as I'm bringing them up, I'm going to channel the thoughts and questions
what I think the audience has in their mind. And after each of the panelist finish their remarks,
I'm going to open it up for the other panelists to kind of function as a quick reactor panel. And
then after we get through ea
ch of their remarks, we'll just open it up to a more general moderated
discussion and finally, for an open Q&A. So, Tom, you'll be our first speaker.
And so, as you're coming to the podium, here's what I think the audience
is thinking about in their head. The current policy landscape for AI is messy.
We've got guidance from the FDA on AI-enabled clinical decision support tool. We've got
the final rule recently released from the ONC on predicting decision support interventions.
We've got th
e White House Executive Order on AI. We've also got state-level efforts. We heard
about the California Attorney General letter. So, can you help us clarify the current policy
and regulatory landscape for health AI? What are the key issues? And what can we learn
from AI policy approaches in other areas outside of healthcare and in other settings
outside the United States? So, Tom, please. Tom Romanoff:
All right. Good afternoon. Some easy questions there, and I
will do my best to address the
m. But also, fully aware that it's after lunch. So, if you
need to take a nap or anything like that -- [laughter] -- you know, I won't fault you for it. But,
you know, just beginning with, you know, who I am. I'm the director of the Technology
Project. And to answer Dan's question about the landscape, it is messy. We have seen a lot of
interest in AI over the last couple of -- last couple of months, couple of years. And, you know,
in Congress, we're looking at around 43 different AI-related
bills, a number of executive orders
that are up there, all applied to artificial intelligence. And we haven't even started talking
about the state-level legislation framework. What we're seeing is that, you know, AI has been
around for a while. These folks that have been working with this technology know that there's
not -- you know, there's new products out there, but the technology itself is fairly known in
the software development world. But there's new applications, and we're just now
catching
the attention of policymakers on how those applications are being used in ways that, you
know, frankly, have some serious concerns. So, when I say that AI is not new --
just to put it in context. Okay? So, 1956, the concept about AI and the mathematical
equation that led to what we have today was first conceptualized. I'm not going to go through all
of these. But you can see from 1956 to today, there's been a number of different advances. And
it's always been that way with this te
chnology where the technology advances a little
bit, and then people start to freak out. Are we having a superhuman intelligence? Is --
how's this going to put everybody out of a job? It also doesn't help that sci-fi has oftentimes
pointed to artificial intelligence as the villain, right? So, I can't get through an AI presentation
without mentioning the "Terminator." And, you know, in some ways, you know, you
can conceptualize a path forward towards, you know, AI being able to do a
lot tha
t we think of as sentient, but the clear path is not there yet.
And so, whenever somebody asks, "How many years are we away from superhuman artificial
intelligence," I don't even put it on a timescale. But what I do want to say is that what we are
seeing now means that we can't go back to the way that things were pre-ChatGPT or pre-large language
models. The genie is out of the bottle. Pandora's box is open. And, you know, putting it also in
the context of where we were on a global scale, y
ou know, in the last two or three years, I
pointed to state-level legislatures and some congressional AI bills that are out there. You
mentioned some of the healthcare-specific ones. But if you can see it, there's been a lot of
talk -- this graph goes all the way back to 2016 -- around how to strategize around AI,
how to regulate it. 2016 is when a lot of the militaries have all agreed that they
would not use autonomous weapons against each other. And some of those militaries
included Chin
a and the United States. But what I'm -- what I want to convey here is that
no one is really pointing to these agreements or these publications and saying, "Look at the
history here. Look at how much we've already been thinking about AI." By and large, these are
all forgotten. And I will put -- my organization's efforts in 2019, we put out a national strategy.
No reporters are calling me up about [laughs] the national strategy we put out in 2020. It's
okay. I'm a little hurt by it, but it's
fine. [laughter] And so, you know, when one of the speakers
was talking about, how do we localize AI, I think this is a really good example. Because
the other thing I want to convey is that, you know, AI is a reflection -- in this case
with ChatGPT -- of language and how we use language. It's -- and language often has kind
of that anthropological effect of, you know, reflecting how, you know, certain cultures think,
or religions practice whatever it might be. So, localizing AI is going to
be exceptionally
difficult. Because on a global scale, we don't have common thinking about how to regulate this
AI, or what is the common definition of fair, or what is the common definition of equity and
access? And I say that as a pragmatic point of view, not as a kind of general -- you know? I
can point to what I think of as fair in practice, but I do realize and recognize that
there's other definitions out there. And so, with that, I just want to also point out
that, you know, we are i
n a new era, right? And so, what is different now? So, this thing called
Moore's Law where you have exponential growth in compute, in memory, and all that fun stuff
is leading to a lot of what we're seeing in kind of these incremental advances in AI. At the
same time, if you look at the graph on the right, we're seeing increased connectivity. IoT
devices are nearly universal in the United States. And they're growing -- they're
expected to continue to grow through 2030. And so, we have a lot
of data out there. And
all that data is -- from a human perspective, really difficult to recognize
those patterns to create, you know, predictions. But AI is an excellent prediction
machine. That's also a really good book if you want to read something on that. And it's
also something that I like to convey, is called exponential growth being very
difficult for humans to understand. And so, while we're talking about this -- how to
regulate and how to make it safe in healthcare, today, it's
already grown by leaps and
bounds. And exponential growth is a concept that humans just really can't wrap their
heads around. So, you know, a lot of -- so, a good example again is ChatGPT. So,
when it comes to exponential growth, it's always -- they're finding new products for
it. They're finding new ways to make it learn. Somebody mentioned about, you know, having ChatGPT
write a poem or whatever it might be. Well, the other example I want to use is ChatGPT,
when it first came out -- I do
n't want to know how many hands go up on this. But how
many of you actually went in there and said, "Let me see if it can do my job a little bit
better than I can do it"? Right? Let me see if it can write an outline for something
I want to write or whatever it might be. And then you got the product, and it wasn't
that great. There was something you needed to go fix. Like, "Phew, I've got another two
years until [laughs] the technology catches up, and it's better than mine." Right?
But that
two years is coming. And, you know, what I like to say in terms of
job disruption in the space is that what is cognitively very hard for human beings
is very easy for AI to do and vice versa. What's cognitively very easy for us
to do is very hard for the AI to do. And so, that's why you're starting to see people
use it as a first draft, and then go in and use their experience to try to build it out. And
along those -- that point, Goldman Sachs came out with a really scary report, if you ha
ven't
seen it, that, you know, current iteration of AI will raise GDP by 7 percent, but it will
also impact over 300 million jobs over time. Phew, I only have two minutes. Let me hurry up.
All right. So, problem not solved. All right. So, what does this mean? So, if -- in the
health equity space and this idea of, you know, data bias, people have been
working on this for a long time. If you're following the big data trends
and how to use it in the medical space, data bias was big in that. A
I is kind of
magnifying those issues. It's going to amplify a lot of that because it can do things
at scale that we haven't been able to do. And so, along those same lines -- and I tell
this to policymakers a lot -- is that if you cared about data bias in the healthcare space
or equity in the healthcare space or social safety nets to mediate technology disruption,
then you probably care about it as well in the AI space. Nothing has really changed there.
It's just really going to continue t
o amplify. I'm hurrying it up because I only have
1 minute and 20 seconds. Sorry. So, now, getting into the policy conversations, are
the policy solutions justified in some of these scary things? Well, you know, we
have -- policymakers have some bright lines that they don't want to cross,
mis and disinformation election space, you know, making sure that it doesn't exaggerate
existing biases. And in the context of that, we often find ourselves looking
at what other countries are doing. So,
the European Union has been called the
Silicon Valley of regulations. And that's because they're going ahead, and they're going
to go regulate. They put out the AI Act. And then in United States, that has led to this debate
around what do you do around innovation where, you know, folks will argue that innovation
can't thrive in regulatory environments? But then what do you do about things like deep
fakes and bias and job issues? And then finally, at the same time, we're having this big deba
te on
what competition with folks like China means. And so, all of these things are pieces that
policymakers are trying to put together. In the healthcare space specifically -- I'm
going to run over a little bit if that's okay. I just want to talk -- you'll see that, you
know, there's kind of two different worlds. Patients are scared of AI in some situations.
They don't want to see it rolled out too quickly without the safety norms in there. But at the
same time, healthcare is viewed as on
e of the biggest areas of disruption for this technology,
one of the biggest opportunities for cost savings. And so, as you see these things play
out in the industry, the idea of trust is really hard in this space. Because
take, for example, facial recognition, right? Facial recognition is something that they
rolled out in the U.K. specifically, and it did not do well with people of Color or people of
East Asian descent. It was very inaccurate. And that idea of AI -- facial recognition is AI
-- being inaccurate persists today. Even though the technology has continued to excel, once you've
lost that trust, it's really hard to gain it back. And when I say you've lost that trust
in the facial recognition space, I like to say it's Kleenex versus facial tissues,
right? AI is AI to anybody who's not, you know, in this field practicing it. And so, if there's
a group that doesn't trust it because of the outputs that happen in facial recognition, they
probably won't trust it in their
healthcare space. And so, the idea that we need to regulate
across these things, it's dangerous because we see AI as kind of this one big thing
that we need to tackle, but in reality, there are very specific industry considerations
that need to be taken into account. There are some very smart people that need to step up in
those industries to make sure that they are, you know, providing the context needed for
the AI in that space and how it's being used. And so, what I usually tell folks is
, you know,
regulating by use case is really easy but it's also really hard to do in terms of long term. And
you also need to address these problems early. So, with that, I've run over a little
bit, and I would like to thank you. [applause] Daniel Yang:
All right. So, first, let me open it up for reactors --
reactions from both Maia and Laura. So, feel free to -- if you have any thoughts that
you wanted to provide, I'd love to hear them. Maia Hightower:
Yeah. How far away are we from some m
eaningful regulation? Tom Romanoff:
So, comprehensive AI regulation is something that we talked a lot about last year with the
AI Innovation Forums. Those have largely lost a lot of steam, and a lot of members on the House
and the Senate side have been given permission to engage folks on their committee specialization or
their bosses' specific areas of interest. And that means that a lot of the AI legislation potentially
happening will be around some of those use cases like mis and disinform
ation, deep fakes, things
like that, instead of a comprehensive type of approach that, you know, potentially was being
floated when they did the innovation side. Now, it's still not dead yet, but, you know, I
don't see it happening in the 118th. And that's always dangerous to make [laughs] predictions
about what Congress is going to do in this space. But I think smart money is not in the
18th, maybe the 19th, but we'll continue to see states regulate in that space, including in
the healthc
are applications. You'll continue to see the federal government push out, whether it
be rule changes, executive orders. There's one that happened last week around privacy which will
have an impact on AI. And so, you'll see that. The problem with that approach
is that with the executive order, it's subject to a new executive coming in and
saying they don't want to go that direction. It's really expensive to change directions
in this space. And then at the state level, it's really hard to man
age
technology with the patchwork. Laura Adams:
Tom, I thought you did a fantastic overview of that. And
it does give us a sense of the complexity and especially when we think about trying to
harmonize this globally. Because I think what we understand across the globe is that
the threats don't stop at the border, and I think we don't want innovation
stopped at the border either. So, I think there's an incentive for us to get out
and do that and to act collectively, globally. And I guess I'
m seeking some advice from
you. I've just been asked to do a -- to chair a global innovations group
out of a U.K. regulatory science and innovation network there. Who would
you invite to the table for a global innovations view to assist the U.K. and also
cross-fertilize for the rest of the nations? Tom Romanoff:
So, I have specific names I'd be happy to talk to you about. [laughter] Laura Adams:
I was hoping that was the case. Tom Romanoff:
But I would say in terms of experts, some of the bi
ggest ones that I would -- so, standards and
definitions are something that in the tech space, especially when we're talking about harmonizing
across international borders, just doesn't exist yet for the AI. And so, folks like NIST and
ISO, anybody who's kind of working on and has experience in trying to establish these technical
standards, I think is critical. It's -- speaking of after lunch, it's really boring, but it's
critical to happen. And so, I'd invite that group. The other big one,
I think, especially in this
iteration of technology and large language models, is the IP and copyright people. Because open AI
is getting sued every other week on these things, and Google also has some pending lawsuits on
that. And I just don't think they've figured out what the technology's use of the data that
it needs to consume means in the apparatus that we have in terms of IP and copyright.
And I actually think that's going to be, long term, really difficult for scaling up
the use a
nd deployment of this technology. Laura Adams:
Thank you. Daniel Yang:
Great. And I love how in this format, I'm actually
outsourcing my job to the panel. So, thank you for asking great questions.
So, let's move on to our next speaker. And we'll have an opportunity to dive deeper
into all these topics after your presentations. I wanted to bring up Laura Adams, senior advisor
at the National Academy of Medicine. And, Laura, as you come up, here's what I
think the audience is thinking. So, we
just heard from Tom about the messiness in the
policy landscape. If you take it one step down, there's also messiness when it comes to
guidance documents, codes of conduct, best practices that are coming from both
within government and outside of government. Just to give you a few examples, we heard about
the NIST AI risk management framework. We have a trustworthy -- sorry, blueprint for trustworthy AI
from CHAI, Coalition for Health AI. We also have some guidance documents from the World
Health
Organization, et cetera, et cetera. And so, one, what do these guidance documents
tell us? And using a clinical example, you know, we worry about alarm fatigue
in the hospital, particularly in a place like an ICU where you've just got beeping
constantly going on. You just drone it out. Are we at risk for AI guidance fatigue? You know,
are we at risk of just hearing so much of what we should be doing that we just stop caring about
it? And so, help me make the case for why the NAM AI
Code of Conduct is going to cut through the
noise and how it differs from these other efforts. Laura Adams:
Sure. Thank you so much, Daniel. It's such a pleasure to be here. I'm delighted to be
able to have an opportunity to interact with all of you. I want to take that first question of what
does it say that we're developing all of these? I think the good news on this front is everybody
recognizes that there's enormous promise, and there's also significant peril. So, I
wouldn't want to be
rushing headlong into this thinking there's only an upside to it. I also
wouldn't want to be focusing only on the downside. So, I think the good news is that people are
trying to make sure that AI does not break bad, because we know that this would be a
crushing blow to something so -- with so much promise for the future. On the other
hand, I think because of these concerns, we keep seeing -- and, in fact, it was the
genesis for the Code of Conduct project in the beginning as we were appro
ached by
people that said, "Gee, for crying out loud, every day we turn around, there's a new
guideline, principle, framework that's put out." And so, the good news is that everybody's
putting out guidelines, principles, and frameworks. And the bad news is the same. Because
we have, now, fragmentation. And in healthcare, we seem to have this penchant for fragmentation
where we will go down our own silos and, as Julia said earlier today, that idea of
feeling we want to work in our own silo
and get it right there without thinking
of the impact on others in the system. And so, the Code of Conduct project is if we
can think about harmonization of data standards, harmonization interoperability, why can't
we think a little bit about governance interoperability? For me, it was so important
when I was beginning to do the living laboratory in Rhode Island to see if we could link up
all of the data sources there clinically and put it into a central repository. It was
really exciting
for me to be able to do that. But we knew we needed an ironclad privacy
framework. So, we worked with opponents like the ACLU and the Coalition Against Domestic Violence
who came in and said, "You have a breach in this system. It's like slitting a feather pillow in the
wind. You'll never get that information back." So, they were so interested in building an ironclad
privacy framework. So, we did. And then I couldn't wait to share our data with the rest
of the states in the nation. I turned
around. We made out pretty well in
the AI -- in the -- rather, the HITECH Act. We got $27 million in
90 days in just our first three grants, and we were off to the races building our health
information exchange. But we spent way too much of that funding that we got on attorneys trying
to reconcile privacy frameworks before we could either exchange any shred of data. So, can
we think a little bit about, what is that? So, the Code of Conduct project was to take a
look at what are the commona
lities here? What are we sort of in violent agreement on? Where
are the gaps? What are the things we need to be paying attention to? But that's only the first
element of the Code of Conduct. The other three elements that are, I think, critically important,
maybe more than the harmonized principal sets, were, can we distill this now into
something memorable? Five or six commitments. Almost think of it like the Ten
Commandments but not ten, just about six. We based that on complex adaptive sy
stems theory
where we can -- we saw in that theory that very, very complex behaviors in the world are governed
by very simple rules. We see that all the time. I saw it in quality improvement in my background.
Whether it was central line infections, we want to decrease perioperative
infection rates. Whatever it was, it almost always boiled down to do about
five things, do them consistently for every single patient, and you'll drop
the bottom out of those complications. And so, we started th
inking, what are those six
simple rules? So, you'll see them come out on the 29th of this month in the Code of Conduct
commentary publication for public comment. You'll see the harmonized principles. By the way,
there's a heavy overlay with the learning health system. I think with AI, if we didn't think
we needed a learning health system before, we know we need one now, because this
is an all-teach, all-learn moment. I was thinking about Judy's remarks
and how she and I spoke at lunch, and
she mentioned being a humble learner. And I
think that ought to be our goal and aspiration for our own conduct and behaviors is being humble
learners. We have a lot to learn, and we will for the foreseeable future. There won't be a time
when I think we tackle this, and we're done. So, when I think about what it means,
I want you to know who's behind the Code of Conduct. The CEO of Mayo, so we have
Gianrico Farrugia, is one of the co-chairs; Bakul Patel, who is the Digital Health
Strategy
global lead for Google; and Roy Jakobs, who is the CEO of
Royal Dutch Philips in the Netherlands. And you'll see here on our AI Code of Conduct --
if I can get this slide to change. I like to do this without names first to see how many faces
you can recognize. I know you're going to say, "Hey, there's Andy Bindman, chief
medical officer from Kaiser." Yes, there's Andy Bindman, and he even goes as
far as not just to serve on our steering committee but he's serving on one
of our work groups
that I'm going to explain a little bit about in a moment.
But I'll give you a minute to take a look. That's Peter Lee from Microsoft on there
too. Eric Horvitz, also from Microsoft, on there. Suchi Saria, brilliant mind in AI,
just absolutely amazing. Grace Cordovano, one of the most articulate, outspoken, and effective
patient advocates you'll ever come across. I was astonished watching Grace at our first couple
of steering committee meetings. Because we might be talking along the lines of
things, and we got
insular and more insular and more insular. And she would make one or two statements, and everyone's
head would snap around to what it meant to be. We were talking about something that might
be important to patients and families, and she cut through all of that by
saying, "Listen, patients and families, family members, they want their
family member to live. Patients, they want to live. They want to be able
to raise their children. They want to be able to live out their l
ives. And in doing
so, here's what they need." Crystal clear. So, I love this steering committee. I'll
give you the names on it. There we go. So, you can take a look at the names on the steering
committee. We wanted Sanjay Gupta because we wanted to be able to get good communications
advice on this. And what we're finding here is that we have strong equity and ethics experts that
have guided our work every step of the way. We, too, are worried about workforce. Kedar Mate
and his work at IH
I, quality improvement. I'm looking across the landscape of those
organizations that I'm interacting with now that are putting together their work on AI, and I
think, "Hey, those of you that made investments in quality improvement, in plan, do, study, act, in
that idea of small-scale tests of rapid change, you're way ahead." You made a smart
investment because, as we heard today, so much of the success of AI has to do with
can you implement it in your own setting? So, the Code of Conduct in
itiative goals
are that governance interoperability, want to be able to -- can we have a common
language to start with so that we're not trying to reconcile all of these? And by the
way, we do recognize that in working with these coalitions that we're working with, the
NIST risk management framework, for instance, is an exquisitely well-done document. It's
42 pages though. And if you are a sort of community health system CEO and you're trying
to look at that document for guidance, I might
recommend -- well, maybe I wouldn't go this far,
but you might want a cardiac defibrillator nearby. [laughter] Because you are going to have
palpitations when you think, "Oh, my god, there is no way I can put 42
pages of detailed requirements. I can't build that into my system." They're going
to stop breathing. So, we think that some of these publications are suited for some
audiences. Some are not suited for that. So, we're working with a number of
coalitions. The Coalition for Health AI,
many of you know that that coincided
today with their announcement of their board of directors. They're formally
incorporated now. We're super excited. They're going to be putting together a network
of certification labs for AI. We're excited at the National Academy of Medicine because
they will certify to our Code of Conduct. So, this is what it looks like when we
begin to align. The most wonderful thing about us running a little bit scared about
AI, at least having the wherewithal to kn
ow what we don't know or at least have
an inkling about what we don't know, is that we're coming together in big time,
big ways to learn. We're doing more and more. I think about Steve Jobs' admonition to
us about when we've got really difficult problems or we've got complex things to
work through, the most exciting thing to think about here is collect as many dots
as you can. The more dots you connect, the more dots that we can pull together and get
general pictures and come up with creat
ivity. So, these are the groups. The Light Collective
is that group of organizations that are rare disease advocates. They are patient
voices. And they are helping us put together the translation of what this looks
like when you take the Code of Conduct down to another level. We can't stop at the Code
of Conduct. We've got to translate it into, what does it look like from those perspectives
so that we're creating an interstitium, a connective tissue, and we don't just
proceed ahead without
regard for each other? So, that's what we mean when we're talking about
the roles and accountabilities component of that. And, boy, are the consumers giving us an earful.
They're telling us exactly what those behaviors look like if you really are including them as
patients. They're even developing a scorecard to say you can score yourselves, and, by
the way, we want to score you as well. The last thing that I'll say is that
it translated into behaviors. And then, once again, the other thin
g that we're looking
at here is when I think of bias and my concern about a digital divide, another equity divide, I'm
concerned not only that the algorithms are biased, but I'm concerned that those with the
most resources are now galloping ahead. Most Americans get their care in midsize
hospitals, in community health centers, in rural hospitals, critical access
hospitals, in these places where they're not as well resourced. We want
to, in this Code of Conduct project, describe what it mea
ns to build the resources
in the center so that all of us can proceed, so that we don't, once again, prevent the patients
being served by those entities from getting what it is that they need out of AI. Because, frankly,
they probably need it more. Thank you so much. [applause] Daniel Yang:
All right. Thank you, Laura, always so eloquent. And while you
didn't say it explicitly, I think one of the advantages of the NAM AI Code of Conduct
versus these other efforts is having you. And, you kno
w, really, you talk about being a dot
connector, and I think you're one of the best dot connectors out there. So, appreciate
your and the NAM stance of really being a humble learner and kind of building on the
great work of many other efforts before it. So, let me transition back to
this reactor panel model. So, Tom or Maia, feel free to ask your
co-panelists or share any remarks. Tom Romanoff:
[clears throat] Excuse me. So, I'm just curious about the -- maybe I missed
it, but the patient
input side of things. Laura Adams:
Yeah. Tom Romanoff:
How are you capturing that? Laura Adams:
Ooh, thank you so much for asking because I ran out of time,
and I wanted to be able to tell that part. The Light Collective is a collection of
patient advocacy groups. And so, they are in the middle right now. And I don't know that
I mentioned that Daniel was the first funder in, of five funders on our Code of Conduct. He saw the
vision for it immediately. And California Health Care Foundation, P
atrick J. McGovern Foundation,
NIH, Epic, all followed. And we were so pleased. You also funded a group called The Light
Collective, and that's Andrea Downing. And she's a brilliant spokesperson for
patient advocacy. Andrea Downing found out that their BRCA3 breast cancer group
that was putting together all of this deep sharing going on in Facebook, that there
was a flaw in Facebook that was allowing their very intimate information to be
channeled to other sources. You know, we think of Ca
mbridge Analytica and places like
that. She was the first one to discover that flaw. Andrea is no dummy. She's really smart in this
field. And so, what she has done with Gordon and Betty Moore Foundation funding is to put
together the guidebook for what it looks like when you are actively engaging us. So, it's
built upon -- they did a patient-led research guidebook which was for researchers to say,
"We're going to tell you all the different ways in which we want to be involved and be
partn
ers in research. We're going to tell you what it looks like when that behavior
is poorly done. We're going to put it on a continuum until we can show you what it's like
when you really do it extraordinarily well." They're doing the same thing for their AI
rights and what they want their roles and responsibilities to be in governance of AI, in
deployment of AI. And we're taking that guide and bringing it right into the Code of Conduct.
So, a section of our final capstone paper will be that a
pplication of what consumers have told
us. "This is how -- these are the behaviors. This is what it looks like to apply this and
to have us feel as if you've done us justice." Daniel Yang:
I would just add, if Andrea was in the audience, probably
the first thing she would say is, "Pay patient advocates for their time and
attention." So, we definitely want their input, but oftentimes, we don't think about
reimbursing them for their effort. Maia Hightower:
Yeah. Speaking of which, that was my
question, is how do we
then translate these lessons learned on amplifying the voice of patients on a
governance level to the very local health system level or even our policymakers? How
do they do that using that same capability? Laura Adams:
Yeah. I would cite Mark Sendak's work out of Duke University. He's
got a collaborative governance model paper. Go get it. Look it up, S-E-N-D-A-K, Sendak, Mark
Sendak. And Suresh Balu also co-authored that paper. They have done one of the best pieces
of
work I've seen in actually bringing the voices of the local entities around. Because as much as I
appreciate, applaud, and anticipate working with the Coalition for Health AI, we can certify
an algorithm or a model at a certain level. We're almost -- not almost. We're always going to
need to take that same algorithm -- all algorithms are local; all AI is local -- and have it
fit and tuned to the local environment. And that includes the local voices. We're working
now with indigenous nat
ions, and they have put together such ironclad principles about
how they want their data used and done, how they want their artifacts regarded,
adopted by the Smithsonian Institution. And so, I want to see us begin to permeate all
of the -- when we write an implementation manual, when we write an implementation guide,
I want it to talk about not just what you and your own silo should be doing but
what you ought to be doing in relationship to the other people. And I want that
written all th
e way through from the highest levels of governance down to
the local community health center. But I think if we can help them develop those
guides, give them templates and models for that, nobody knows their patient populations
like some of these places like a community health center. They have some of the best
opportunities in the world to make this and tune this toward the true patient needs. So,
I'd say that that's what we're looking toward, and we've got a lot of work to do
ahead, tho
ugh. I understand that. Daniel Yang:
All right. Well, thank you. So, moving on to our last speaker,
last but not least, Dr. Maia Hightower, CEO and founder of Equality AI. So, Maia, the
question that I have for you is, you know, the theme throughout the day was really around
inequities and risk of algorithmic bias. We know it's a critical issue. We know it's an issue that
policymakers are aware of. But sometimes it can feel like there's this massive chasm between
acknowledging the problem a
nd actually doing something about it. So, can you illuminate
us on what policymakers, what health systems, what everyone should be doing to ensure that
the AI tools we're using are fair and equitable? Maia Hightower:
Absolutely. That's what I'm going to talk about
today. It's about being pragmatic, really making sure that when you
are implementing your AI systems, that you have a strategic alignment. And that
strategic alignment includes health equity. So, AI and health equity are not two s
eparate silos or
different parts of your strategy but intertwined. And the second is really around governance.
No matter where you are in your governance structures, you can actually repurpose
your existing governance structure. You don't have to pull in, you know, a ton
of different experts as long as you use what you've got, and then grow. So,
AI governance is extremely important. And then the third pillar is measurement,
whether you're using an external source to measure or internal cap
abilities. But
you have to measure the impact of what you implement to ensure that you actually are
achieving the outcome that you seek. So, that's what I'll talk about. And
[laughs] I'm going to stand over there with my little pointer because it's
actually easier for me to stand over here. But I did want to -- for each of us to
take a moment and pause and think about the last time you saw your doctor, the last
time maybe a loved one was in a health care system. Maybe you were in an exam r
oom,
and you had one of those, you know, cute, attractive robes on that exposes your
derriere. And your doctor was probably sitting in front of a computer typing
away. Maybe now there's an ambient system. But how much did you trust that the AI behind
the curtain actually was personalized for you? How much did you trust that your loved one in
the emergency room or in the operating room, that the AI behind the curtain actually
applied to their specific circumstances, the data, the personaliz
ation,
their genomics, their labs? And for the vast majority of us, you
know, given the right circumstances, we each are at risk of being an outlier. We
are each at risk of not being in the middle of the bell curve that that particular
model may have been trained on. And so, that's why, you know, making sure that we
are measuring bias that we are addressing is so important because we have a long way
to go. AI truly has the potential to be our pathway to personalized precision medicine if
we do it right. We are at an inflection point. We've talked so much about the promise of
AI already. Again, it really does have the opportunity to improve quality of care, drive
down costs, and create these amazing experiences, amazing experiences for our patients, amazing
experiences for our providers. We just had the whole panel talk about how we're so close
for the EHR actually to be a joy. Never in my career did I think that would be a possibility
since I've spent my whole career wearin
g like a flak jacket when it came to the EHR. But
to be so close, yet we have so much to do. When it comes to health equity, you know, Martin
Luther King said it best. "Of all forms of injustice -- inequity, injustice in health care
is the most shocking and inhumane." And in the last 50 years, what have we done? Have we done
anything since Dr. King said this statement? No. We continue to have the same inequities we had 50
years ago today. So, do we have this opportunity with AI to instead o
f widening the chasm,
to actually close it? And I do think we do if we implement our AI systems responsibly with
health equity intertwined in our AI strategies. And the reason why we know that there is a risk
of widening health disparities is because we have caused harm with AI already. We have widened
disparities already with AI systems. Example, Obermeyer et al., many of you probably are
familiar with the work of Ziad Obermeyer, but he was able to measure the impact of a biased algorithm
that was deployed in hundreds of
healthcare systems. Probably at this point, has been calculated to have been exposed to
80 million people that decreased resources, case management referrals to Black patients
compared to White patients. Black patients, while being equally sick, had 50 percent
less referral rate for case management. The good news is that it was detected and
mitigated, that we do have that capability of detecting a bias within a model and
fixing it. We no longer have to have
a model be deployed to 80 million people or 80
million exposures before we recognize that harm has been done. And so, bias occurs.
And again, this is sort of repetition from earlier today. We had that beautiful
example from our colleague from Emory. But bias occurs across the AI life cycle from
those that have the ability to ask a question, so problem formulation. Not all
problems have equal voice. If we asked our community members what is the most
important problem they want solved from
AI, it probably isn't to improve their revenue
cycle -- rev cycle management or auto denials, right? Yet that often is where a huge amount
of AI resources is currently targeted, is in rev cycle, in cost reduction, in process
improvement, which is also very important. But bias occurs from the moment of problem
formulation to the real-world data that is embedded within our EHR, to how that data
is acquired, to the way that the model is developed. There are thousands of decision
points in mod
eling in and of itself. In the Obermeyer example, the error that was
created was called a labeling error. So, they used cost as a proxy for risk. And we
know systematically that African-American patients spend less health care dollars
while being equally sick. And so, that was the actual error that occurred. So,
that occurred during the modeling process. The data in itself, the data set actually had
plenty of higher-quality labels that would have produced -- that ultimately produced a bette
r
model to the way that the model is evaluated and to where -- how it's deployed. We had
an example earlier today about a punitive, a negative model versus a positive
one, where a very neutral model on an AI -- a very neutral model on no-show
prediction, you can either be punitive as a health system and double book or you
can be assistive as a health system and provide resources that actually addresses why
somebody might be at high risk for no show. So, bias can occur throughout the AI lif
e cycle,
and the great thing is that bias mitigation can occur throughout the life cycle. And so, when
it comes to different mitigation techniques, we have a combination of social methods. So,
some of them do not require any technology in and of itself. You can start with diverse teams,
with AI governance, you know, with our regulatory environment, with applying some of these
principles that we have talked about by my panel members, all the way to more technical approaches
where you can ac
tually dissect the data set, you know, the model itself, and measure for
precision and performance by subpopulation. You know, for that same instance
of, say, you go to the doctor, and you want to know if it's personalized to you,
you could actually ask the folks at Kaiser for the performance of the model by subpopulation.
I heard that this morning, which is fantastic. Is it accurate if you're African-American? Is
it accurate if you're Asian-American? Right? And then the fairness of it. Is
the distribution
of resources equitable across subpopulations? Are the mammogram referral rates that this
model may be triggering equitable? Does it match the demographics of the people that that
population is serving? So, those are some of the very technical and can get way into the weeds
on the technical approaches to bias mitigation. But what that leaves, is this challenge, right?
There is -- when you talk to healthcare -- when I talk to healthcare leaders, they're feeling
this overwhel
ming sense of competing priorities. "Where do I start?" And that's really where --
again, three simple steps: your AI strategy, making sure that your organization is prioritizing
what's important, including health equity; and with that role is your AI governance, making
sure you have an AI governance system in place; and then holding your AI governance
system accountable through measurement. You can measure through an audit. You can
measure through technical methods. There's a lot of ways o
f measuring, but measurement is
so important. So, the way that we approach it at Equality AI is really helping health systems
find that alignment by domain and making sure that their AI strategy really does align to
what's important for that healthcare system. And then AI governance, there's a lot of different
frameworks for AI governance. Personally, NIST and ISO, yes, it may be dizzying. If
you think 42 pages for NIST is extensive, what about the 200 pages for the
ISO standard [laughs]?
But these are -- there are more simplified versions.
You just talked about Mark Sendak's work. That is the Duke work over on that bottom
right-hand corner. And then Obermeyer et al., Ziad Obermeyer has his own playbook, the Algorithm
Bias Playbook that Booth UChicago has published. And then the model, when you audit a model, you actually audit by two different
methods: the technical approach, but also process. So, did the model go through
the AI governance process as it was intended? So, we
each have a role to play
regardless of what hat you wear. If you're a healthcare administrator, your
role is to make sure that from the top, that your AI strategy includes health equity,
that you're appropriately resourcing your teams to provide AI governance policy
accountability within your system, and that you're actually checking to make
sure that those processes are working. If you are a doctor, if you're a clinician, it's
being a part of those AI governance committees, making sure y
our voice is part of the
solution; if you're a patient advocate, that you too are part of AI governance
and that your voice is being represented; if you are a policymaker, making
sure that we have some guardrails, we have some foundation. Because right now,
it is very confusing for healthcare systems. And so, I'll just -- the call to action
really is straightforward; again, AI strategy aligning with your health equity goals,
governance, and then measure. So, that's it. [applause] Daniel Ya
ng:
All right. So, let me open it up. Laura, Tom. Laura Adams:
I loved it. I have -- what I think I heard you say is that
AI is the problem and AI is the solution -- Maia Hightower:
[laughs] Laura Adams:
-- with regard to its ability to detect and to help mitigate bias. So, the very problem
is sort of creating -- well, it's not creating. It was created before. But it scales and exacerbates
the problem can also be part of the solution. Can you say, Maia, more places where you think AI
will be
the solution to some of these super intractable long-term problems that we've had very
little success in actually, you know, addressing? Maia Hightower:
Well, if you were to ask the folks at Microsoft and Google, they'd say the environmental
issues. But the same -- like one of the biggest challenges of AI and big data and high-performance
compute is the incredible amount of energy. We don't talk about -- we haven't talked about the
incredible amount of energy it consumes. But the argument i
s that, as well, that there may
be some innovative solutions that AI is able to generate that may solve some of our climate
concerns. Not saying that it's true [laughs], but besides health equity and healthcare in and
of itself, precision performance medicine. Tom Romanoff:
And also, water, consumes a lot of water as well. Laura Adams:
Yeah. Maia Hightower:
Yeah. Tom Romanoff:
So, I'm, you know, very cognizant of the lack of trust that
some companies have with policymakers in terms of the vo
lunteer solutions that are
out there. How do you make this enforceable? Maia Hightower:
Yeah. The only way that you can make it enforceable is through regulation. I do think
you talked about the use case approach within healthcare. There is an incredible opportunity
for HHS and the various HHS agencies to use the executive orders that they've been given to
actually implement regulations that are very targeted. So, ONC has done a great job where they
now -- their proposed rule has become a f
inal rule by which EHRs are required to have certain checks
and balances in place when it comes to AI systems. It will be -- the FDA is moving in a similar
direction when it comes to software as a medical device. They have guidelines right
now. And many of -- much of pharma are already adopting some of these guidelines when
they are using software as a medical device. A lot of healthcare tech companies want to be
FDA-certified as a marker of good citizenship, so similar to, you know, in env
ironmentalism,
to be a good environmental partner. And so, I think that there is an opportunity at
that very targeted HHS agency level to have meaningful regulations
sooner rather than later. Daniel Yang:
All right. Well, we're going to transition to a moderated
Q&A and then an open Q&A. But actually, I wanted to -- because, Laura, you stole one of
my questions around the positive uses of AI for addressing inequity as opposed to exacerbating
it. And for me, my favorite example is actually
also from Ziad Obermeyer and Emma Pierson
around predicting pain from an x-ray image of a knee film. And so, everyone knows Ziad's work
around the algorithm around case management that was biased. But he also did another study that I
loved, which is they simply trained an algorithm to predict a patient's subjective report of
pain, which they happened to be able to collect. And what they found was that human radiologists
were very biased in identifying or accurately assessing pain in Black p
atients. It's partly
because they were trained in kind of severity levels of osteoarthritis based on, you know,
traditional cohorts of patients in the past. And the AI algorithm was not biased by those same
training approaches. And so, they were just much better at accurately predicting subjective pain
in Black patients than human radiologists were. And so, in the same way, I mean,
the AI is intrinsically -- you know, it's as biased as we want it to be.
And so, a lot of it is in the proble
m definition. But there are a lot of examples
there if you're training towards outcomes as opposed to training towards human
interpretation that you can actually get less biased results. But let's transition,
or people may have thoughts on that comment. All right. Okay. So, Tom, I have a question
around regulatory burden, and Laura, you actually mentioned this in your remarks, this concern that,
yes, we need regulation. We need guardrails here. But if the burden is a little bit too high,
i
t's one thing for Kaiser Permanente -- I mean, we've got a lot of people. We've got just
tons of expertise when it comes to compliance, when it comes to legal functions, when
it comes to the technical expertise. So, you know, we're welcome. We're happy to kind
of, you know, to excel in those environments. But I worry about what Laura was describing.
You know, care oftentimes happens in the non-, you know, Mayos or Stanfords or Dukes of the
world. How do we provide policy solutions that work
for them that doesn't actually kind of
create the system of AI haves and have-nots? Tom Romanoff:
Yeah. It's a great question, and I think it kind of gets to the heart of the debate currently
with this. You know, the folks that are incumbent dominant players in this space and the folks
that represent industries that are already highly regulated, they know how to approach these things.
They know how to either create products that meet compliance or, you know, that are catered towards
the re
gulatory environment that they're in. And the smaller folks and the ones that are
startups, they don't have either the capital or the expertise on staff. And so, that
is a debate that happens quite often is, you know, how do you make sure that you
are addressing some of these issues around, you know, very real negative
externalities while at the same time, not favoring the big incumbents and the
regulatory folks that are in the space? And I think that's really hard to do. You know,
I strug
gle with that question quite often, but I also struggle with the
conversation of, you know, what we would -- what would happen if
we don't do anything, right? And so, if we completely ignore this idea and kind of
embrace nonregulation as the fuel for innovation, well, then we get some negative use cases
around, you know -- I'm not going to name names, but insurance companies' unit for end-of-life
care, whatever it might be. And that was a really bad outcome that I think drives the policy
q
uestions around, "Well, we can't do nothing." One area that I think we can look at is a
lot of tech companies, they don't want to be the ones that come out and say, "This is
what we need to do in this space," same with, you know, healthcare, I imagine. Though I'm
not a health care guy. And they want to codify what they're already doing because they build
processes and teams around what they're already doing. And if they can standardize that,
then that means that they can standardize it for
a lot. They can minimize some of
the risks that they have in that space. And I think there's something to be said about
that when it comes to the voluntary commitments that a lot of folks have already talked about. The
White House came out on what NIST is doing and the huge amount of support that's happened there.
And so, there are kind of these indicators of, you know, different regulatory practices that
won't inhibit further development of this. Daniel Yang:
Laura. Laura Adams:
I would lo
ve to see -- when we think about when we needed EHRs to
get implemented, I mean, Economist magazine had ranked us second only to mining. Healthcare,
second only to mining for lack of investment in health IT. You know, we like technology.
We can look at every molecule of the human body. We just couldn't get your lab test across
the street even if your life depended on it. So, when I think about the approach the
federal government used at that time, it was to ask ONC to put together, and we p
ut
together this in the HITECH Act. Every single solitary physician provider out there got up
to $63,000 per provider to do the transition, to do what it took to get up to speed,
to understand, to acquire technology, and do those things. We also had things put
together under that same act called the regional extension centers. You know, these are like in
agriculture where you go out and you help small farmers understand the newest technologies,
the newest science to go on with things. I th
ink we need to begin to replicate that and
put that back in place for all of the places that do not have the resources. Create regional
collaborations, create vaults and banks where we can start to see how algorithms are functioning
in critical access hospitals, create communication nexus where they can begin to share. I loved what
Julia said about, "We're working with some other healthcare facility over here. They're going to be
testing and trying this one. We're going to test and try this
one over here. And then we're going
to combine our learning." Let's do that at-scale, and let's do it directed toward those that
do not have the resources to do it. I think we can replicate the regional extension
centers and really take a run at this. Daniel Yang:
What about you, Maia? Maia Hightower:
Yeah, I love that idea of recreating the HITECH Act because I was a primary care doctor
in private practice when we implemented the HER, and I got one of those cool checks. Like,
it was like
literally like $30,000 -- [laughter] -- at the end of the year. And I was
like, cha-ching. I was already over my margin for the year. When you're in private
practice, you know, every dollar counts, so definitely recall that. But I do think
that there is plenty of opportunity to, again, align, say, HHS with whether its
meaningful use ONC and meaningful use type of approach or just plain CMS and
CMMI having some sort of incentivized projects and innovation projects around
appropriate, respon
sible AI implementation. I think there's plenty of opportunity for both
some sort of carrot versus stick approach and definitely starting with carrots. That has
always been very helpful in healthcare is when they start with carrots, incentives. And
then, you know, by the time 10-15 years goes by, the penalties start kicking in. But there's
plenty of frameworks that HHS and ONC and CMS can use as examples and frame a
similar type of policy and incentives. Daniel Yang:
My next question actual
ly goes back to you, Maia. The last time that we met,
you were still at the University of Chicago -- Maia Hightower:
Yes. Daniel Yang:
-- as the EVP and chief digital transformation officer there.
You've since gone, you know, and full time into this startup. And you talk about the importance
of AI governance and getting this right. And so, it's a bit of a meta question, but can you
simulate the conversation between Maia, the physician executive that has tight margins
that are just coming ou
t of COVID and, you know, were told to cut costs, and then, you know,
Maia, the startup CEO that's trying to sell into University of Chicago of why we need
this. What does that conversation look like? We could probably, I don't know,
simulate that in ChatGPT as well, but -- [laughter] Maia Hightower:
I think the conversation really is, "Okay, what's keeping you up at night?" And
I've spoken to many healthcare IT leaders, and what's keeping them up at night is
the loss of focus on health equ
ity. This sense that we've lost -- that the demand
for increased productivity and decreased cost has taken away from the overarching
mission. If you ask healthcare IT leaders, those that were in the throes of the pandemic that
had this huge rallying cry with purpose around the pandemic and health equity, now, it feels
like we've made this about shift and no longer is equity and quality and all the rest of the
mission so important, but we have to drive down costs and increase productivity. W
e need another
tool that's going to help our providers, you know, see more patients. That's not -- sometimes
it's fun as a wellness tool, but in the back of the head is like, "How are we going to pay for
it," which usually means increased productivity. So, what I -- my conversation to the executive
me is wouldn't there -- there is a possibility for aligning your health equity and quality and
patient experience mission that has always been part of our mission with AI, without the AI
being a
detractor and just about increasing productivity, decreasing cost, and potentially
the cost of some of our employees, that there is this bright side if we make these choices around,
you know, setting up our AI strategy in the very beginning to align with health equity, that we
continue to -- that we have AI governance, which AI governance is just a fancy word for clinical
decision support governance. You already have clinical decision support governance. Now you just
need a few extra exper
ts. But it's not that much different. And then really making sure that you're
measuring. So, the conversation that I have is -- Daniel Yang:
That you're having in your own head. Maia Hightower:
-- in my own head is how do you get the joy back for the poor IT people? Everyone talked about the
joy of medicine for the clinicians. Believe me, our IT teams are so tired. They are -- they've
been, you know, adapting to change for the last four years nonstop. The clinicians were the
heroes, but in t
he back end, the IT teams were basically duplicating, creating whole hospitals,
whole health systems digitally. And so, of course, I got to hear that because these are my team
members, and yet nobody was, you know, waving the hero flag for our IT team. But they, too,
have been through a lot in the last five years. Laura Adams:
I saw a meme yesterday. It was clearly maybe an 80-year-old
gentleman's picture, and it said, you know, "IT is the fun, exciting place to be," you know,
like Mark, 28
, you know, years old. And clearly, this was a picture of an 80-year-old
man, and they were assuming this was a 25-year-old person making this statement.
So, there was the picture about that. I think you're bringing up something
really critical that we would be remiss to close out a conference
like this without talking about, how are we incented? Because I have heard
people say we can't wait to use Nuance, Abridge, whatever the ambient -- picking it
up and being able to reduce the provider
time. And the providers are whispering, yeah,
so they'll add more people to our panel. We're not going to save more time, spend more
time with our patients. This is not happening. The thing why I will drop everything and speak
at a Kaiser conference above all is because you have the payment model figured out for your
patients. I wish I lived near where my family could be cared for by Kaiser because I worry
so much about our incentive model. The reason I worry about it is because it isn't a
ligned with
the patient's best interest like Kaiser's is. You're self-contained. You're integrated. You've
got it all, payer all the way through to delivery. When I think about -- I was in Rhode Island, and
at our Health Information Exchange, we were able to build capabilities onto that system where
we could identify and notify a provider in real time in the primary care setting that your patient
has gone in or out of any hospital in any state, any part of the state in Rhode Island or in
a
nd out of any ED. They could give us their high-risk panels and say, "Track these people
so our care managers know in a nanosecond." We watched people be able to be healthier, be kept
out of the hospital, get things prevented, and one of my CEOs in that state shut it off. And so, when
I went to see him, I sat down. I said, "Hey, John, you shut off the system. Why did you shut it off?"
And he said, "I had to, Laura, because it works." He said, "The system keeps these patients
out of the hosp
ital." And he said, "Frankly, I run the whole health system, and I can't
keep my hospital open unless those people are sick and admitted." He said, "So, I've got
to have them sick, so I can have that revenue from the admissions, so I can keep the other
services going." He didn't say that with a blank look on his face or with a half-smile on
his face. He said it with anguish on his face. And my sense is that we do have a health incentive
system here that doesn't set us up to do the best outs
ide of Kaiser. Because you, again, have your
incentives aligned. The rest of the world doesn't, and I don't think AI is any solution for that.
I think we've got to take that on head on. Maybe AI can help us with that, but it's no magic
bullet for it. We've got to summon the will to move toward value-based payments models
a whole lot faster than we already are. Daniel Yang: I can look at Tony and like, "We're definitely
inviting Laura to the next panel discussion." [laughter] Laura Adams:
We
ll, I'll drop everything and show up because I want
everybody to be aligned like you are. Daniel Yang:
Thank you. Well, let me open it up to the audience Q&A. Female Speaker:
I think it's come up a few times today. Like, can we get access to
broader, better data sets that can be sort of available to anyone that wants to build
a model and train a model on it? I know it's just one of the steps along the way, but
it feels like a really important one. And so, could you just comment on the pathwa
y to that
endpoint? Like, what would it take for Kaiser to share all their data with anyone that wants
to develop a model on it? And is regulation the only way to get there, or do you see any sort of
private sector models that might get us there? Daniel Yang:
I mean, maybe I guess I'll start in my KP role. You know, what really attracted
me to KP, and I think many people here as well, is -- I mean, we do really have quite remarkable
data sets. Vivian spoke about them. And one of the things
that I learned about KP is that
not only do we have deep and broad data, we have longitudinal data. You know, the average
Kaiser member stays with us for 11 years. And so, when you're really trying to look at
outcomes, we've got that outcome data. So, one is, as Laura mentioned, there's
another event, the Coalition for Health AI and this whole concept of quality assurance
labs. And, you know, how do we build the wire cutter or the consumer reports for AI in an
independent kind of third-par
ty validator service? When I look at the organizations
that are lining up to fill that role, I see the usual suspects, you know, Mayo,
Stanford. Probably UCSF is on that short list. I think KP is a great place to play. You
know, I need to make the case internally that we could. But when I think about a lot
of the issues around diversity of data sets, we've got that. You know, our 12.5 million
members across eight states and the depth of the data we have looks a lot more like
the rest of Am
erica than Palo Alto does, you know. And Stanford delivers great
care, but it just -- it doesn't look like the care we deliver in Oakland or
the populations that we care for there. So, you know, I -- but Kaiser is just one
healthcare system. I think there are others, you know, that similarly reflect that diversity.
I would love to see county health systems and rural hospitals contributing data. And
so, if I put my Moore Foundation hat on, we were funding in that space. And
so, just to give
you one example, we gave a grant to Contra Costa County, their
health system, to make their EKG data available for public use. And it took one year for
the data use agreement to get signed. So, of our grant dollars, not a single dollar
was spent because they've never signed a data sharing agreement. They didn't even know
the lawyer that had the authority to sign it. And so, I do think that there are huge almost
infrastructural gaps around making data available. It's easy to talk about it f
rom the stage
that health systems should be contributing their data. I think the challenge really is
that, you know, the academic medical centers, there is very much an intrinsic motivation for
making their data AI ready, for doing research on it. We don't have that necessarily in the
other care delivery systems. And so, I'd love to see us move forward, but I do think that there
are big structural issues that stand in the way. Laura Adams: I think that it's going to become infinitely
easie
r to get a hold of data sets soon, but only if you have money. Peter Lee, at a
recent conference that we had with Harvard, was on the panel with me, and he said, "Here's
what I see for the future. Hospitals haven't had a very good opportunity to monetize the data
that they have." And that was pre-AI. All bets are off now that this is going to become suddenly
super valuable data for those. It will become, from the hospital's point of view, "Oh,
my God, do I have a revenue stream here? And I
think that -- I think back to Michael
Millenson's point that if data is the new oil, then patient rights and privacy is climate
change. And I think we're coming up on an era where we're going to see a shift
in that for money, and I'm very, very worried about the imbalance and how that
will play out. Because I do think that's going to be suddenly an issue for patient privacy
rights where organizations are going to be willing to give up because they're dying for
the resources that they need.
And we've got a lot of conversation to be had in that arena
before that future comes, and it's soon. Maia Hightower:
Absolutely. And I think we need to remember that each data point in healthcare is
a digital representation of a person's experience. Laura Adams:
Yeah. Maia Hightower:
So, who really owns the data and health systems were just fiduciary,
you know, managers of patient -- individual patient data. And so, until we figure out how to
actually adequately protect a patient's privacy
or the patient's wishes when it comes to their
data, I think we're going to have, you know, whether you call them structural barriers,
but I also think that they are protections, right, for who really is the owner of
that data when it's your digital twin. Tom Romanoff:
Yeah. I'm not a healthcare guy, so I might butcher this.
But, you know, I was thinking about that in terms of digital twinning and, you know,
the case of Henrietta Lacks and, you know, the fact that the family took generation
s to get
a retribution for that. What does that mean in the space of a digital health world where, you
know, it might not be one person. It might be multiple -- thousands of people that ultimately
have their data used for some long-term outcome. Daniel Yang:
Yeah. I think the flip side though, you know, one of my colleagues at KP, Vinnie Liu,
loves to say, "Not an ounce of data wasted." And so, you know, our patients entrust us with
this health care data, but they also trust that we're lear
ning from it and that we're
using it for good purpose as well. And so, I think most of our patients, again, if we
were to survey them, would say, "Please, like we want to advance care delivery. We want to
advance the state of the art. We want better care for ourselves. I want more personalized care.
Use my data to make my care more personalized." So, I think there's always a double-edged sword.
We certainly care about privacy. We certainly care about data rights. I also think that, you
kno
w, our patients are expecting us in some ways to really make sure that we're leveraging
this precious resource for greatest impact. Maia Hightower:
Absolutely. And that would come through consent
as well as transparency. Daniel Yang:
Yeah. Maia Hightower:
Right? So, if you have the systems in place, then you can
definitely ensure that you're appropriately using the data that has been consented for
that use. And discovery is part of that. Laura Adams:
I totally agree. It's like donating blood
[laughs], you know. It's just doing
good for the system and doing good for the whole. Daniel Yang: All right. Well, with that,
we're going to close. Oh, okay. Male Speaker:
I'm just going to make one comment on this. Daniel Yang:
Yes. Male Speaker:
I really appreciate all of the comments around this last question
and, Julia, your question about this, which is a really, really important policy
question. We've had a number -- we've had a large number of sessions in this space that
the instit
ute has put on, around drug pricing, and at the central -- and there's a whole
bunch of really identifiable reasons why we have problems with drug pricing. And at the root
of a lot of that is the failure to think through the economic model of technology transfer
and what's happened through universities to private ownership of the goods that come
out of that to how it's been on the market. I would suggest, Daniel -- and maybe you and
I can both have this conversation with our colleagues -- t
here's policy to do in
this space. How do we take this good, which is this data, and how can we make sure
that it is used exactly the right way and doesn't create a kind -- you know, a private
ring fence that then winds up being things where people start not doing things because they
haven't solved the intellectual property problem? So, you know, how long it takes, contracts to be
signed, I mean, these are all the same problems. And it's something that all of us, I think,
can work on toget
her and think -- if we think of it as a policy problem before somebody puts a
solution forward that creates a reality we'll be living with for another 50 years. So, I think
excellent question, lots to do in this space. Daniel Yang:
Thank you. All right. Well, great comments, seems like maybe a potential
topic for the next IHP meeting. But I want to thank my panelists here for your
great remarks and insights. Thank you. Laura Adams:
It was great to be on the panel with you. [applause] Daniel
Yang:
All right. So, it's me again. I've been asked to give some
closing remarks for today. Thank you. I mean, first of all, I just want to thank all of our
panelists, our moderators, our organizers for hosting such a wonderful and insightful event.
You know, I've got AI in my title. I'm literally paid to think about this all day long every day,
and yet I've learned so much, really, from all the people here. And it reminds me of what Julia
said earlier on. You know, sometimes we need to kin
d of stand outside of our own silo because
we have so much more to learn from each other. So, I've been working in clinical AI for the last
decade, but I've witnessed, and I think all of us have witnessed an incredible transformation in
just the last 12 to 18 months. For many years, AI has been focused on very narrow but important
clinical use cases, for example, trying to identify diabetic retinopathy from an image.
But when you talk to patients and providers, these were often viewed as ju
st incremental
improvements in care, not transformative. And this was evident when you talked to
our providers. When you talk to providers about clinical AI, the general
reaction that I got was one of skepticism. And many providers even felt
threatened by AI. They felt that it was threatening their sense of professional
competency, their sense of autonomy. But something really changed with the advent of
generative AI. Suddenly, we have general purpose AI tools that are showing impressive p
erformance
across a very wide range of use cases. Moreover, these tools are starting to solve some of
the quality-of-life pain points that have always -- that clinicians have always struggled
with, things like administrative burden. And what surprised me most is the physician's
change in attitude toward these tools, from one of skepticism to one of delight. And
now, what I'm facing is one of incredible demand. Now, the promise of these AI
tools to me is, paradoxically, they make medicine m
ore human again. These
ambient scribes that we talked about allow providers to turn their attention away
from the computer screen and back to patient faces. It gives providers more
time at the bedside or in the exam room. And so, hearing these benefits, you might think
that the natural approach is to move forward as fast as we possibly can. But I think
that to move faster and to move further, we have to proceed with caution and vigilance, especially now. Why? Because I believe
in these too
ls, and I also believe that nothing slows innovation more than patient harm
or algorithms that deliver -- that promise more than they deliver or biased algorithms that
mimic our history of inequities or biases. Now at KP, I have an incredible privilege and
responsibility to help lead our organization's program on responsible AI. For us at KP,
responsible AI is a journey. It's not a destination. And just like everything
at KP, we strive for excellence, not just compliance. And so, what does
excellence in responsible AI look like for us? First, it starts with culture, and it starts
with system guardrails. So, I come from the world of patient safety and healthcare quality.
And if there's one thing that I've learned from that environment, it's that to err is human.
And so, we have to apply that same approach to our AI algorithms. Our algorithms,
like humans, are intrinsically fallible. There is only one way to prevent harm
and to prevent bias in health AI, and that is to never d
eploy it at all. But there's
a risk in that approach. If we all believe in the transformative positive benefits of AI, we have
to acknowledge that there is risk in deploying AI, but there's also risk in not deploying AI,
the opportunity cost. The only way that we can manage this tension is if we move forward
with a culture of responsibility that travels from frontline providers all the way to the most
senior executives. We also approach responsible AI as an organizational muscle. We underst
and
that our providers are extremely busy. And so, they look to us to do rigorous QA testing
on these algorithms to identify risks and mitigation strategies, to provide actionable
education on how they should be using these tools, and to create system guardrails that prevents
harm from happening in the first place. But we also realize that our upfront risk
assessment and mitigation strategies just provide preliminary guidance. We also have
to learn by doing. And in order to do that, we nee
d the discipline and the infrastructure
to monitor the performance of algorithms we've already deployed, looking for drifts
in performance, the introduction of new biases we hadn't thought about. We need the
discipline to quantify the real-world impact of our AI algorithms, what I describe
as these algorithms' return on health, and with the infrastructure to flag and
remediate patient harm before they occur and to learn from these system errors, what
Laura talked about as a learning health
system. So, as I said before, the work for responsible AI
is never done. It's a journey, not a destination. And so, in closing, I invite you all to join us
on this journey. If I've heard one thing today, it's that the stakes are too high for each
of us to be doing this on our own. I really hope that industry, policymakers, health
systems like KP and others can partner and share their expertise and their perspectives
as we think about national frameworks for advancing responsible AI. We've
got so much
to gain from learning together. Thank you. [applause] Rebecca Flournoy:
All right. Well, thank you so much for sharing those reflections. I think you're
leaving us all inspired, which is fantastic. So, for me, today's conversation, I think, is really
a reminder that there is so much work still to be done, but it also makes me really glad that there
are so many thoughtful leaders bringing their insights and perspectives to the table. And so,
we look forward to continuing that con
versation. And on that note, I want to say a huge thank
you to all of our speakers and moderators today. Such a fascinating -- so many fascinating
conversations. Thank you also to our IHP team, Ceci, Renee, Nicole, Nani,
Krysten, and Ben. Thank you. [applause] And thank you also to our forum advisory
group for all their really helpful advice and assistance. And then thanks also to
the team who helped us host today's event, including consultants Maryland
[spelled phonetically] and Wendy, th
e Center for Total Health, Spark Street Digital,
Kian [spelled phonetically], and Ridgewells. And, of course, a huge thank you to all of
you in the audience. So, thank you. Before you leave, we hope that you will
take just a couple minutes to take the event survey to give us some feedback. It's
very helpful for future events. It's on the event page that you access through a
QR code. And if you need that again, it's up on the screen. And then we will be
posting a full recording of today's e
vent. So, we encourage you to share that with others
who may be interested but weren't able to attend today. And this concludes our
forum today. Thank you, everyone. [applause]
Comments