Transcription
[Music]
and now on to our first
kino talk for today i am delighted to
have professor fei-fei lee from stanford
university
join us today it is typical to have
someone introduce the keynote speakers
and in fayette's case i find that to be
both hard and easy
it's hard because i will need her entire
one-hour slot
maybe more to talk about all her
achievements
fa is currently serving as co-director
of the human center air institute at
stanford
um having previously served as director
of the stanford ai
lab and also chief scientist of ai and
ml
at google cloud she has had a huge
impact on the
field of ai broadly and computer vision
in particular
where she made numerous contributions we
will probably be sufficient to just
mention
imagenet which changed how object
recognition is done
but there are many other research
projects that define new directions in
computer vision
and for instance visual genome activity
that and many others
i'm also really admiring fei fei for
everything that she has done for
diversity in ai
through the ai for all program which she
co-founded and which spans
now many institutions including michigan
i mentioned that introducing feipe is
both hard and easy
it's easy because fife doesn't really
need an introduction it's sufficient to
say
here is faithfully welcome pepe the
screen is yours
thank you radha and thank you michigan
really an honor to be here let me just
uh try to share my screen first thank
you again
and i i think when i was thinking about
your introduction i think the most
important piece
is that i feel like part of the michigan
family i spent a
sabbatical in michigan in the early half
of 2013
and just loved living in ann arbor
visiting uh
zingerman's and you know the north and
central campuses it's just a gorgeous
place so
too bad this is virtual but i do feel
very honored to be back to michigan
today
and i want to share with my colleagues
some of these latest work we're doing in
ai
and healthcare before i just even get
started i
want to show this slide to really uh
drive home that this is
quite a team effort everything i'm going
to
present is kind of a result of years and
years of collaboration
among computer science ai researchers
mainly my students and postdocs
as well as a number of incredible
clinicians and clinical researchers
throughout stanford medicine as well as
our
partner hospitals and health care
systems like
unlock utah's intermountain stanford
children's hospital stanford
adult hospital and the list is still
ongoing
i especially want to highlight the two
people in red boxes
professor arne milstein who is a pioneer
in thinking about health care this
country's healthcare policy healthcare
excellence research and quality
improvement as well as cost reduction
i owe everything of my health care
knowledge
to arne who introduced me to this
incredible field about 10 years ago and
then
for this particular talk i want to
acknowledge
my graduating phd student
albert hawk he is one of the heroic
ai students behind a lot of the work
as well as a main paper that i will be
going over
uh today so without their work i
wouldn't be able to uh
do this talk so it's important for me to
acknowledge them
so let's just get to the you know the
main talk
and we're here that uh because most of
us are passionate about healthcare
um and in the past 100 years modern
medicine has gone through
incredible advances whether you look at
drugs and
medical imaging medical devices genomics
you just name it
you know our increasing life expectancy
as well as quality
owe a lot to the improvement of health
care system
and also a personal note i happen to
have taken care of my
mother who has been very sick for almost
40 years and
all my entire adult life
i am the sole caretaker of my mom and
i've been through
every single healthcare you know
environment from icus to ers to
operating rooms to
to home care with her and it gives me
a personal perspective of how important
healthcare is and what kind of
incredible advances
there has been and and how all of our
livelihood
owe it to modern medicine and its
advance
but despite this important advances
it is still incredible to think in 2020
or 21st century medical errors are still
the
third leading cause of death and there
has been
extensive studies by nih since 20 years
ago
as well as researchers around the world
that studies
unintended medical errors you know in
the healthcare delivery system there
have been
a lot of important protocols written
about
how to avoid errors how to execute in an
intended way
to both standardize but also keep the
quality of care
but you know unintended
errors still happen in fact there's way
more
an order of magnitude and more
death and fatalities caused by health
care than say car accidents
in united states annually so this is a
big
issue and um and why
is it so you know why is is the error
rate so
so huge well partially it's because
healthcare is a very complex human
behavior endeavor
and it takes place in very complex
environment
and there are many different people and
roles
involved and you know eras can be
introduced
in so many ways whether it's procedural
or you know incorrect
coding or or wrong prescription
or sometimes you know hand hygiene
compliance
induced errors just so many ways
that the errors can be um introduced and
it's really
not intended but can have grave
consequences
so at this point for those of you who
are in ai especially in
the broad area of ai and technology
research
i hope they remind you of another space
that is full of complex human behavior
and that space is driving and
you know michigan is a leading
university
in in autonomous driving technology and
as well as robotics technology
so i think we have a lot of colleagues
who appreciate
how complex driving is and and a lot of
driving's
behavior boils down to complex human
behavior and
the potential um errors or unintended uh
problems with with driving so
about 10 years ago just around the time
i
met with arnie milstein
autonomous driving was one of the most
exciting
revolutionary ideas and it was going
from academia to commercialization
even though it's there's still a lot of
research and what
excites me as a technology about
autonomous driving was
the emergence of a bunch of technology
that's enabling a whole new
way of transportation the sensor
technology
the machine learning algorithm
technology the system integration
technology the gps technology all this
is reimagining helping us to reimagine
autonomous driving and that was the time
that arne and i
serendipitously got together and
really made that intellectual connection
between what we see in autonomous
driving
and the technology behind it and what we
hope to see
in healthcare that just similar to
transportation
healthcare is a complex
human behavior endeavor situated in very
complex spaces
from hospitals to homes that
to a large extent what's happening in
these spaces with human behavior
is still very much in the dark we don't
know much about
what patient is going through unless we
you know take their labs or stick them
in a
some kind of imaging device
or we don't know much about clinicians
behaviors
unless they chart and
charting means documenting what they do
on a
electronic hospital record or medical
record
but all this is actually affecting the
outcome of healthcare
so what we kind of uh boldly
envisioned is endowing healthcare space
with the kind of mba intelligence we see
in smart cars
and with smart sensors and machine
learning algorithm
and transform this so-called dark space
of healthcare
into a visible and visualizable space
that we can really try to
help improve the outcome of
patient care and roughly speaking
the transformation of this physical
space
involves in enabling it with sensors
and i'll get into the details of what
kind of sensors we're talking about
and then of course for those of you who
are deep in
machine learning ai research um just
data itself doesn't speak much um
much insight so we have to go through um
machine learning computer vision uh
technology to recognize what's going on
in uh with human behavior and then at
the end this
should shouldn't be a standalone
technology should be integrated into the
clinical data ecosystem so that's the
kind of broad picture
we imagined um fast forward um
almost a decade later we just published
a
relatively comprehensive nature review
article
on what we have learned on this journey
as well as
many incredible research that has
concurrently
happened in this short period of time so
for today
the rest of the talk will be just kind
of a summary
of what we touched in this nature paper
especially
two critical health care space one is
hospital space one is daily living space
i'll go over briefly
what are the what are the potential
opportunities to use
ambient intelligence to improve the
healthcare delivery
and also rada please just tell me if i'm
going to over time because i do want to
leave
time for discussion so thank you um
so let's just start with healthcare
space in the paper we actually touch on
several
health hospital space icu
inpatient unit operating rooms and other
clinical
settings in the hospital i'll try to
quickly go through just two
scenarios to give a give a flavor of
what's happening
and what we can use ambient intelligence
for so icu
i see you as an important area of
healthcare delivery
in fact um you know
i i don't know if most people know this
number almost
one percent of our national gdp goes to
icus along
and uh and the cost of
icu is in hundreds of billions of
dollars
every year and in this covert setting we
hear
a lot of stories coming out of icu
that underscores the importance and this
is where
our patients fight for their lives this
is where our clinicians
are just by and large overworked
but yet the the health care what's
happening icu is so critical to every
single patient and their family
so um in icu
setting there was one project we we
participated and spearheaded which is
the mobile patient mobilization project
it turns out if we can properly
make patients move in icus in a proper
way
evidences say they recover better they
get out of icu faster
and they recover from the grave
situation uh better so patient mobility
is really important um
how do we ensure patient mobility and
how do we know the quality and frequency
of patient mobility is proper by and
large right now it's direct observation
and
nurses and clinicians charting and this
you can see
introduces a lot of issues charting
itself is erroneous
and it's also labor intensive a lot of
clinicians
complain about spending more time
charting than spending time caring for
patients and also charting is very brief
it doesn't have that nuances that we
we hope to see and there's a lot of
errors that can be introduced
so we worked with two hospitals utah's
intermountain and stanford
children's hospital to put sensors
into the patient rooms and
the orange dots you see are sensors
in you in a floor map of a icu
unit in the intermental hospital in utah
and then
what you see in a blue video is the
quality of data we capture
through a special kind of sensor called
depth sensor
if you're not familiar with depth sensor
i hope to draw a connection between
xbox video games and and the sensor
in fact this is a um
lidar base sorry it's an infrared based
sensor that um
that captures depth information in the
in the uh environment it does not
capture rgb pixels so it's actually
naturally a good sensor to try to
respect and protect privacy in the
clinical setting
and it's obviously noisy and
what we can do is to take these videos
and then use machine learning algorithms
to learn what kind of
human behavior has taken place in this
particular setting
for mobility we care about activities
such as
getting patients out of bed sitting
patients in a chair
getting patients back to bed getting
patients out of chair these
are clinically relevant mobility tasks
and then we learn to label them by our
ai algorithm behind the scene
and then what we can do is to start
collecting data insights
we not only can can in computer visions
technically
called action detection in long videos
rather than just action classification
of an
entire video so we can get into this
kind of
action detection um tasks where
in say 24 hour stretch we can
collect data insights about how patients
have been moved
over that uh over that uh um temporal
stretch
and um um this is just
to show you it's getting uh finer and
finer and we can actually identify the
the
the people um and hopefully in the
in the future the rose and the actors
and
this kind of quantitative
data insights is is
um the first step to provide a very
continuous objective measure of patient
mobility in icu
which has never done before it as i said
it related
relied on clinician charting which is
extremely sparse
and the specificity and sensitivity of
the
quantitative results are really really
high which
really is important for for our
clinicians
and of course in the ongoing work we're
going to extend this to
more finer grained human activities
including turning oral care putting on
compression stockings to avoid pressure
ulcers
and so on and then we can we can
tabulate these kind of data insights and
uh
um and provide clinicians the important
uh data for our patients so that was
one um icu project
another related icu or inpatient project
is hand hygiene
um you know eight years ago we started a
hand hygiene
project and and frankly not too many
people cared
but when kova 19 happens hand hygiene
compliance suddenly becomes such a front
and center
issue not only in hospitals but also in
our homes
in fact hospital-acquired infection is
one of the leading killers
of patients nationwide not just during
coffee but
just during entire healthcare times
and almost a hundred thousand patients
die
of hospital acquired infection every
year
three times more as car accidents and
so what have we been doing to ensure
hospital hand hygiene compliance
mostly through human observers or secret
shoppers
and as you can see it's subject to so
much
issues right it's very sparse it's
time-consuming it's expensive it's
subject to
human biases people have tried rfid
as a way to you know
automate the system but rfid is very
coarse it doesn't really observe human
behavior it's a proximity measure
and it also disrupts workflow if
clinicians have to
have to you know go to scan their their
badge
so again we installed some sensors
in this case again in utah's
intermountain hospital and the
children's hospital in stanford
this is what the physical depth sensor
looks like
and this is the kind of data we could
can collect and you can see that uh
clinicians stopping by before entering
the patient room
to to perform hand hygiene protocol
and what we do is to take the video data
and go and devise a
convolutional neural network action
detection algorithm to try to locate
both temporally as well as spatially
where the hand hygiene
prop behavior to take place
there are some technical challenges that
i'll just gloss over
including how to um
[Music]
combine multi-sensor views and this is a
combination of tracking and and and
tracking issue as well as recognition
and we have to
find a common uh space to project
our uh uh our people
in the in the video so that we can
combine the multiple views
across a large hospital or or
inpatient space and here's the
the detection result that a person
was about to enter the room and the the
uh
sensors and the algorithm behind
shows when they do the hand hygiene or
they don't
and so on and then of course we can also
collect big data analytics
over time and over people and this is
invaluable
insight for optimizing the location of
hand hygiene dispensers
for our clinicians um of course if you
are familiar with this space who says
hand hygiene
is not just about entering room it's
about five points of the five moments
of interacting with patients before
touching patients before touching
patient bed and all that
so um again without uh belaboring
um delivering this system i just want to
say that we're doing additional research
on the five moments that who has a
protocol
of hand hygiene that gets into the finer
understanding of
hand hygiene compliance and and this is
some preliminary work to
show in the patient room where you know
potential touch points happens and and
how people are behaving
so um that was a quick uh uh
overview of icu i'll just quickly go to
the operating room and try to speed up
operating room again very complex
environment very complex human behavior
there are a number of work done by other
groups
that are already looking at ambient
sensors
this is a group that are looking at
human 3d pose
in in
in a surgical environment in the hope of
training and tracking surgical staff
and and let me just fast forward this
but in addition to people people also
interact with
objects and objects is actually a big
deal tracking them and knowing where
they are
in surgery room and some objects can be
tracked by rfid and
barcodes but many objects cannot and
these are the objects that are actually
more prone to errors including getting
lost in the patient body
so um you know nurses have to spend a
lot of time
counting them and at the station next to
the
next to the patient and
we have started to pilot an idea of
using
um depth sensors and cameras to help
automate this counting process
i'm gonna warn the audience of the next
slide
it doesn't have blood but it does have
real life
surgery um so if you're screaming about
it
uh here's the warning okay in another
um in another uh
work in working with surgeons we
actually go
into the laparoscopic um videos
and try to track uh surgical
instruments uh inside the body and and
by tracking that
we can learn oh sorry this is just the
algorithm
system flow we can learn how the
the the usage of different
uh different surgical objects this is
actually critical
information keep in mind this has never
ever been
observed without this technology
surgeons just
um train through human apprenticeship
and and go through this um you know
more like a artful training rather than
objective training and this
is a very good quantitative information
for for surgeons in training as well as
for
surgeons to understand what happens and
what are the instruments
used uh in a surgery so this is also
ongoing research
um okay so i'm gonna switch gear
and quickly go through the daily living
space
the paper has more details about
hospitals i'll just invite everybody to
read on your own
daily living space is another space that
has
a lot of implication to our health care
and well-being
i'm personally passionate about senior
care
and but also physical rehabilitation
mental health are all
important part of daily living space
just to get into senior care independent
living
is really really important to our
seniors
and independent living involves
activities of daily living
and there is a standard list of what we
call adl
adl impairments indicates a lot of
potential diseases
and decay decline cognitive neurological
and other ones
and it has tremendous physical
as well as physical mental implications
and typically caregivers are the first
to identify
impairments but we do have a world
global dilemma coming up is that our
population is aging
the ratio of caregiver to seniors is
um is
declining we have fewer and fewer
capable
care givers compared to the number of
seniors
and and this industry in general has
other issues that is not well um um
well established that we need to help
our care givers
and one way to help our caregivers is to
go through smart sensors and mba
intelligence there is no way i want to
say
upfront we we do not believe in
replacing clinicians and help help
caregivers but we really believe in
helping them and helping our singers
and wearables is actually a good
supplement and there is a lot of good
research coming up and commercialization
but there are a lot of issues wearables
cannot
address it's inconvenient seniors don't
necessarily wear them all the time
and when it comes to nuanced human
activities wearables are limited
so we're piloting this as smart sensor
ambient intelligence
idea in a senior home facility in san
francisco this is my heroic
graduate students going to the senior
homes installing
both steps and thermal sensors by
themselves
and those of you who are not familiar
with thermal sensors you can see that
the
on the right is thermal sensor data on
the left is depth
they capture quite complementary
information which is really useful for
adl understanding and
our algorithm can start to understand
the
behavior of seniors whether they're
sitting standing walking sleeping
can do a real-time detection of false
which is
extremely life costing and and money
costing in senior care
and then this is the kind of opportunity
we cannot
put into important data analytics
for um for
um different adls that can become
really important clinically relevant
data for caretakers and clinicians
one of the example here to show you more
fine-grained understanding is sleep
pattern
we can actually with this kind of
technology we can go through
different kind of sleeps and and
sleep data on a daily basis
okay there's more in senior care we are
actually now starting a new project in
30 day readmission for seniors
it's been a exciting journey in the
in the time of covet but we're excited
this project will get started
but let me switch gears and just at
least
touch on physical rehabilitation because
i think this
is so relevant to seniors as well and to
show you what this technology can do
of course everybody knows physical
rehabilitation is so important it's a
big industry it impacts a lot of people
uh things like gait analysis is um
is kind of bread and butter for many
many rehabilitation
and treatments but most of these current
physical rehabilitation is done in the
labs and
patients have to get to the hospital get
to the clinics
it involves a lot of overheads
and it's uh also sometimes you have to
use your motion capture systems it's
very expensive and uh
and and and uh cumbersome can you
imagine
if there's just a camera a camera at
home that can start understanding
the gestures the gates the the the
walking
uh details of our patients it'll really
transform
physical rehabilitation and a lot of
people are starting to pioneer that
in the paper we um you know cited a
number of work
in understanding human gesture uh human
uh or using uh depth sensors
to observe um you know different kind of
physical
activities and we are also
doing gait understanding and
humor a human posture and pose
understanding using using depth sensors
and hopefully this can start uh um
in light enlightening the possibility of
uh moving some of the physical
rehabilitation to homes
uh these are just some quantitative
results
okay so um we also did some work in
mental health but i'm not gonna have
time to talk about this
in the last two three minutes i do want
to touch on a
really really important aspect of
ambient intelligence
it's its societal and ethical
implications
um in at stanford we strongly believe in
human-centered ai research
and a large part of this means that we
have to take into consideration of
ethics into the design of our technology
not as an afterthought
and this is where we work with a lot of
bioethicists
legal scholars philosophers and
actually ethicists who
are experts in different angles of this
problem
and we meet regularly with the computer
science team and clinician team
to discuss uh various issues i'm not
going to be able to go through
all the details but i will touch on
categorically some of these issues
privacy of course is a huge thing there
are technical ways to
ensure privacy protection including face
blurring
including video down sampling including
human body masking
these are all examples there are also
ways to you know
make sure there's privacy in data on the
data front
like differential privacy to add noise
to
individual record in machine learning
we're exploring federated learning
to uh so that data don't get neutr
centralized and
can facilitate edge learning we also are
looking at homomorphic
encryption um it's incredibly slow
but uh and and how we can uh put
stronger privacy guarantees and there is
also physical privacy
of devices from devices all the way to
central servers central servers
but of course privacy goes beyond just
technical solutions there
needs to be regulatory considerations
there are other issues in this research
or this technology legal issue
is another ones for example this is just
an example of
psychotherapist patient privilege but
ambiance sensors opens new challenges
like can the data become potential
witnesses
and we are working with legal scholars
or at least providing
our technical insight to legal scholars
who are looking at these
issues and hopefully forecasting and and
getting ahead
in in terms of uh
thinking about laws of course bias
is a big issue and fairness and and the
world
the world of ai has woken up to bias
there's bias introduced in data sets for
example
gender skin color and and so on there's
also bias in the
inclusion of different demographies we
care a lot about you know how to sample
patients and of course
regionally we have bias and we have
challenges but at least we want to pay
attention
to these issues um
i'm just gonna oh and then there is you
know
algorithmic uh transparency
accountability
uh there are nice research research has
happened such as model cards
in in in the past couple of years that
are useful tools to ensure algorithmic
transparency and accountability and of
course there is ethical issues
in the scientific research itself even
before the mass
application of this technology and we we
are
taking meticulous care into the
different steps
of our scientific research protocol from
you know irb uh human patient human
subject consent
uh all the way to data storage data
annotation
and so on so i i
i can spend days to talk about ethical
issue but i think the overall message i
want to deliver on the ethics front of
ai research related to ambient
intelligence is the importance of
multi-stakeholder approach
from again bake the ethical design into
the technology before a single line of
code is written
and that's what we have been trying to
do to collaborate with policy
leaders ethicists legal experts
clinicians patients
and of course technologists in this
multi-stakeholder conversation
so that's all i want to share with
everybody today
it's just a very uh
superficial summary of what we feel
so excited by this new wave of ambient
intelligence technology
in ai and healthcare and
what i just want to say is at the end of
the day healthcare
is about humans caring for humans
ai is here to really help enhance
humanity not to replace
the many many hours and days and weeks i
sit in icu or surgical room with my
mother
makes me realize how important and
irreplaceable
our caretakers our clinicians our nurses
are
and our role as technologists
is to make their work better is to
enhance their work
give their time back to the patients and
together to improve the health care
system
thank you so much
so happy thank you and this is great
very insightful and very inspiring
and i see there are a number of
questions um
so there is one question about the
multiple
sensors being used specifically the
question talks about the icu monitoring
where
is there a scope for multimodal data for
instance
including wearables or other devices
that the nurses or doctors could
could use yeah i mean as a technologist
absolutely for those of you who are who
are deeper in computer vision research
you can see that there is a really
exciting new
area of research integrating audios and
videos recently and here in clinical
setting i think multi-modality
does include wearables as well as
potentially audios and and
adapts in thermal in fact in
neurological disorders for example early
detection
of stroke let's say i think audios is
critical right so
it depends on the situation but short
answer is absolutely
uh there is also a question
again with respect to the different
sensors and different
angles um you spoke earlier about the
analogy between the
medical space and the
driving space um and so looking
there there is for instance slider being
considered as one option
and cameras and another option so if you
could speak a little bit more
both about um what do you think about
these
two different approaches um with respect
to the accuracy but also with respect to
privacy
yeah a good question i think it's
probably
pretty obvious to most of us that depth
sensor
and lidar they are not completely
privacy issue free but they go a long
way in protecting
privacy because we don't need color
information we don't see
faces and so on so from that point of
view
we actually chose to use
depth sensors since about 10 years ago
because we feel
it's just so much more respectful but
having said that we shouldn't just
pretend that there is no privacy issue
for example
gate identification is a way to identify
people and you can actually
de-identify the person just through
biological gait
analysis and that you can get that
information from depth and of course
if you're in people's home you know
whether you see the face or not it's
contextualized to identify
individuals so we have to be vigilant
about that
but on rgb camera
we have to go through a lot of good
regulatory guardrails
um to to ensure privacy
is uh is taken care of and the patient
consent and family concern
but i also want to say that there is
value there is a lot of clinical value
in using camera information because a
lot of diseases
and a lot of illness conditions the the
the signal come through the colored
pixels whether it's
nuanced movement patterns
or facial expression of pain or like i
said earlier neurological
early neurological condition detection
like
face drooping in in early stroke
patients
in ambulatory or mobile
medicine case you want that quick
detection and
pixel information is is useful so
so they're complementary i would say
i personally would try to push dap
sensor as far as we can
because of its its advantage in privacy
but work with clinicians and
patients and policy makers to enable the
possible use of cameras
and that that makes great sense and i
would follow up with a
question that has been on my mind for a
while with respect to
um ethical considerations including
privacy which you touched upon
so there is this concern that if you
account for personal information there
might be
unethical issues being raised but then
if you look all the way through to the
end application
i'm also hearing and from colleagues
in healthcare that it would be also
unethical to not help somebody who's in
need
so i would be curious to hear your
thoughts on this trade-off between
the ethical considerations say early on
in which kind of data to consider and
then
really to the end application like what
kind of
impact it could have or potentially
would not have if we would
stop working on this kind of um yeah
no rada i think you touched on really
important even deeply philosophical
point
and i think just by you talking to both
clinicians and patients
is the approach i would take it's a
multi-stakeholder approach
you know none of us whether as an
individual or
as technologist have the answer to all
this right
i i wear the hat as a caretaker of
my mother and there are many moments you
know
including now when i work from her
remotely
i want technology to help me
to to give me more insights
of how's her walking gesture today how's
her sleep pattern this week
and for me that information is
invaluable
arnie my my amazing partner always
remind me
and my team of technologists uh let's
not remember we're dealing with the most
vulnerable
one of the most vulnerable of our
population
and there are many ways we can help them
it's
really hard to give a sweeping statement
of what's ethical and what's not ethical
when the forces of privacy but also
health and life are are um
in play together so my my the way i have
been taking this
is first of all be very respectful and
humble
about the space we're working my with my
ai
students when they join the team the
first thing i ask them to do is to
shadow doctors
and and to just feel the human empathy
about the vulnerability of the
population we work with
and to start there take the
multi-stakeholder approach
listen to patients nurses legal scholars
i feel confident given the the
brilliance of human minds
and and if the will of our collective
you know uh community we'll find a good
way to balance technology and human
dignity
and and use necessary policy guardrails
um and technology smarts to solve these
issues
thank you and yeah this this is
something that like i said i'm currently
thinking a lot about and i think it will
require more thinking
from from many and balancing
the sort of the the trade-offs um there
is
another question regarding um the work
talked about
in senior care whether there is concern
that collecting data is infringing on
autonomy
and how do these people feel about their
independence and i will also add my own
follow-up to the question that was asked
to what extent will discourage people to
try to be
autonomous if there are all these
systems that are helping them and partly
doing their work for them
yeah so um really good question and
several
answers there for scientific
research that's what we're doing we're
not building
you know monetizable products for
scientific
research i think patient concern
and subject concern is a huge
part of this this process right like we
would
only work with seniors who understand
what this is about and who do not feel
this is infringing on their autonomy or
privacy and so on and and people are
different
not everybody is the same and their own
situation
is different and they need different
help so at least
we focus on working with the population
that
follows our human subject protocols
in terms of are they the the deeper
philosophical issue
as machines start to
kind of take responsibility of autonomy
what does that mean for
for humans i think this ambient
intelligence technology
is pretty far from taking over
autonomously because it's mostly an
observatory of observing technology
it doesn't actually enact on
um on behalf of people it cannot take
the pills for you
it cannot walk for you it cannot
exercise for you right
it's more a uh a way to
to gain healthcare insight
well-being insight so for you the
patient
as well as for your caretakers to try to
help optimize the the well-being so
from that point of view this particular
technology is
is not um i don't see a strong
danger of uh taking over the the sense
of autonomy
um and again the bigger um
context um even at stanford we often
talk about
i really really believe there's so much
more for ai to augment and enhance
people
rather than replace people whether it's
this particular healthcare we're talking
about or manufacturing or
or other space um of course there are
situations
ai or robotics can replace people
especially
safety security situation i would like
robots to
to replace people in dangerous places
but by and large this technology has
more
space to work on to enhance humanity
that's that's another angle to think
about this
yes and i i i very much agree that
that's where we can actually make a
difference
now we'll have to see in the futures um
and there one thought i have which will
take me to the to the next question
i i've actually thought about this
augmentation in the context of my own
children at some point i was thinking
about
how could i build technology that would
help small children be understood like
really sort of translate their
baby language into our language so they
are
understood and eventually i gave up just
because i figured that maybe that would
not encourage them to learn the language
that would help them communicate
yeah and there is a question um
which speaks about both um age people
which you talked about
and then um infant health care
um and i would be curious to hear about
your thoughts with respect to transfer
learning
are these algorithms that could be
reused between different population and
what are your thoughts with respect to
that yeah i actually think we have to
use
transfer learning few shot learning long
tail learning is a huge problem in
health care right like for example
senior fall
how much data can you collect of real
senior falling in their homes first of
all i hope none
second these are rare events but have
grave consequences and there's
no way um if you look at today's deep
learning algorithm and how much
supervision it needs
that we can collect enough data at least
not from
our budget and uh and uh
so i think um domain adaptation transfer
learning
um um fuel shot learning these are
situations these are learning methods
that are
so suitable for for health care i
actually
my ai students when we think about the
technical contribution we can do
i think this is a huge area because
almost
every single health relevant
activities we deal with is a long tail
event
and hopefully they are rema they remain
long tail event
and whether it's in the hospital or not
or at home these are ontario events so
we have to evoke these
you know the distribution
disparity you cannot hope for training
data and testing
inference time data to have the same
distribution so
really good technical area to to work in
in my opinion
thank you and just one more question to
wrap it up
um there is a question on what do you
think is the biggest button lack to
introduce ai into healthcare
domain right now there have been a lot
of advances as you highlighted
what do you think is the main button lag
is it
privacy ethical consideration legal um
yeah great question i i do think a lot
about this i
i think there's a couple of things that
i just want to highlight
the first one is public trust i think
a.i
right now as a member of the ai world
i'm concerned about ai's public image
and
and we can debate of about whether we
deserve that
but the the image is not that great
public trust is
very low in about ai and the technology
at in large you know think about the the
the
issues of bias you know there's a lot of
public outcry uh think about uh privacy
issues
so i think um working healthcare
we i personally care a lot about gaining
that
trust in the community and in the
multi-stakeholder
community that we work with and i think
as a
as a field collectively
we need to work very hard to gain
public trust and this is not just a pr
issue it's really about
how we conduct our our research and how
we
um you know help to move
uh the technological or combine the
technological aspiration
with the the the goal of
doing ai for good so that's one issue
the second issue is actually modernize
our
policy and regulatory measures i think
it's
really important that we work with
regulatory
agencies to modernize some of the
um some of the guard rails and and
and and policies surrounding data
machine learning
on one hand it's comforting to know
healthcare
is a highly regularized industry and
for good reasons and as a patient or
patient family
i appreciate that very much on the other
hand
technology has moved ahead for example
data siloing
and data sharing is a big
issue in healthcare research and
can we modernize our regulatory
measures so that while we keep
the protection of our uh patient
health record and data but we also
incentivize
innovation i think and that also will
help
public trust if we have the right policy
measures and guard rails we
we could hope to innovate within the
safe
boundary of trustworthiness and that's
also
very important so these are the two
issues
that uh at least rise to the top of my
mind
oh thank you um and thanks again for a
very insightful talk and
for a great discussion thank you to all
everyone in the audience for your
questions and i will now hand it over to
my colleague and collaborator
professor larian who will share the