How Machine Learning Enhances Healthcare
i’m going to talk to you a little bit
today
about machine learning for health and
as you have heard i do machine learning
i focus on artificial intelligence and
making these models
and you might be thinking why would you
do machine learning for health
how does that work isn’t machine
learning for self-driving cars why would
we want that as part of health
this is the far future that you’re
talking about
but actually machine learning and
algorithms are
a part of health care right now whether
we’re talking about the statistical
models that doctors currently use as
risk scores in the intensive care unit
or we’re talking about more advanced
high-capacity models that are being
trained
to understand what sort of risks are
relevant for patients and what sort of
treatments might be needed
machine learning and algorithms are a
part of health care
now you may be uh wondering well how are
you doing this
why do we have machine learning in
health and and how does it happen
generally how this works is we get
clinical data from
practice and from knowledge by practice
i mean
doctors they practice so if you could
look at clinical records from a hospital
from a clinic and see what sort of
treatments are given
how patients are interacting with a
healthcare team you could learn from
that practice
right but then there’s also knowledge
maybe we don’t just want to learn based
on how doctors are practicing
we may also want to look at the
knowledge that’s been generated
randomized controlled trials
rcts papers that are written textbooks
right we could learn from both of those
sources
once we have that data we can train
these simple statistical models
or more advanced high-capacity models
and then
we can predict things important clinical
events
forecasting treatments that a patient
might need those are really important
for
health care and progress so that’s what
could happen
in the future that’s what we’re working
on this is the goal of my group
what’s happening right now because the
data that we’re
getting to feed these models to learn
these rules
to train these algorithms that’s data
that we’re getting right now
and so the question that you should be
asking if we’re learning from
current practice and knowledge is how
much
do you trust your doctors because we’re
learning from clinical practice
we’re learning from doctors who are
burnt out
because we’re asking them to do a huge
set
of tasks not just provide clinical care
and burnt out doctors feel that they do
not have
the time to be empathetic
to their patients and that’s the kind of
practice that we’re learning from
and doctors who are burnt out and feel
that they don’t have time to be
empathetic
are also they’re just human humans
even humans with ethics training and the
best of intentions
are biased and so it’s been well
documented
that the biases of society are reflected
in the health care that we provide
so this is what we get from current
practice right now no technology
right what about knowledge
one of the big sources of knowledge is a
randomized controlled trial
where you give one set of people a
treatment and see how well
it works in that population however
randomized controlled trials are very
rare because they’re expensive
so only 10 to 20 percent of the
treatments that are given today
are based on randomized controlled
trials when we do have them
they are biased only six percent of
asthmatics
treated today would have been
eligible for the rct’s use to design
their treatments
not only is the knowledge that we have
rare and biased sometimes it’s just
wrong
so more than ten percent of the three
thousand plus
top journal papers from gemma the new
england journal the lancet
or what we call a medical reversal a
paper comes out that says that thing
you’ve been doing
common clinical practice the thing we
think is correct that’s wrong
we need to do something else
so neither practice nor knowledge are
perfect right now
without any machine learning without any
technology
and this is understandable because right
now think about what happens if you’re a
clinician
and a patient comes into your practice
sumana is having trouble breathing
how are you supposed to decide what kind
of care she needs
you’re going to think about prior cases
that you’ve treated
and then try to assign an appropriate
treatment but the issue is that humans
are actually reasonably unique and so if
i’m trying to assign a treatment for
sumana based on
patients who are like her i can go check
the medical record
but prior work has shown that even if
you take 250 million patient records
from
four countries and combine them a large
proportion of patients with common
conditions have unique
treatment pathways meaning nobody else
had a similar treatment history
and for common conditions like diabetes
depression and hypertension you would
not assume
this would be true but 10 percent of
diabetics
11 of those with depression and almost a
quarter of those with hypertension have
unique
treatment pathways what that means is
sumana comes into your clinic
and says can you give me a treatment
based on patients who are like me
and you would say no there is nobody
like you in 250 million patient records
so obviously we can’t use the practice
and the knowledge that we have now
perfectly and so my goal
is to use the complex data
that exists about a person to train
high-capacity models
and try to provide better indications
for doctors to provide
care to their patients and so if that’s
what’s happening now without technology
maybe we can use technology to improve
practice and knowledge so let’s go back
to sumana
she’s having trouble breathing which
goes into the intensive care unit what’s
the first thing we might want to do
we want to treat her there’s many
different interventions that we could
use
and so the first thing we might want to
do is plan our practice
right and so what we’ve done is we’ve
taken
the data from over 30 000 patients in
the hospital
and tried to learn what is the optimal
practice
how do we predict the interventions that
are needed for sumana
given all of this data so that we can
plan her
care and that’s important because when
you use these deep neural networks
to understand a patient’s history you
can combine their notes
their labs their vitals all of their
demographic information and every six
hours have a new
updated risk for what sort of
interventions might be necessary
and now that we’ve planned care right
tried to improve our practice another
thing you might think is
if sumana is still having trouble
breathing i need to send her down to
radiology
i think she might have a collapsed lung
a pneumothorax
but when you send a patient down to
radiology
what’s happening is maybe it’s late at
night
this radiologist has looked at hundreds
of images
they now have to try to generate this
knowledge make this note
but they’re tired and they’ve seen so
many of this
these kinds of images and it’s a very
tedious task
so another thing that we’ve worked on is
going from hundreds of
thousands of x-ray images
to the radiology reports that would have
been written about those
images and that’s important because if
we can
use high capacity models like
convolutional and recurrent neural
networks
to make clinical text that is both
accurate and human readable
we’re allowing a doctor to make better
decisions about care
while removing a tedious part
of their day another thing to think
about
after we’ve improved this practice and
we’ve provided more knowledge
is exactly how we’re doing that so
sumana is a minority
how do i ensure that the treatment that
she’s receiving
is fair and
remember our prior slide humans are
biased
right and we’re learning from human
generated data
so as an example we took some text
and asked a neural network to complete
the blank
and here the only thing we changed was
the patient’s race
so 71 year old caucasian patient patient
is in blank condition was dnr
do not resuscitate in the nursing home
if i change the race
in this sentence it shouldn’t change
what gets filled in the plank right but
it does
and it does because this model was
trained on real data
generated by people with real biases
and this is important to establish
because when we train high-capacity
models
we need to be able to audit them and
demonstrate for example
how well they do on patients of
different protected classes
different ethnicities different genders
and
importantly it could be that this model
performs better
than most doctors in your clinic
right the average performance even
though it varies across these protected
attributes
could be better but whether we’re
comfortable deploying an algorithm like
this
that could improve care but at different
levels
for different people this is an ethical
question
this is a societal question this is not
something that the machine learning
algorithm can answer
for you we have to make that call so my
group broadly focuses
on creating actionable insights in human
health we focus on creating models that
are healthy
so can we train models that work well in
a medical setting
what kind of healthcare is healthy once
you’ve trained a model that works well
can you then use it to interrogate
healthcare and establish what works best
for people
and finally what kind of behaviors are
healthy because ideally we don’t want to
just recommend optimal treatments
improving practice and knowledge we’d
like to keep you out
of the hospital entirely thank you
you