Thank you Jeff in the introduction
I apologize my voice is a bit

lot of people have junk
over the past few weeks and

I'm one of those people so my voice is
a bit scratchy apologize in advance.

Today I really want to go through.

You know some of the big picture stuff
around health analytics around what sort

of current in the space in terms of
translating health analytics into

direct patient level care but I also want
to talk more specifically about how you

hear a students at Georgia Tech can
take advantage of a lot of the tools

that have been built up over the past
couple years here on campus and

through other broader open source projects
so over by the end of it you'll be able

to actually have very specific strategies
that you can apply to your own work and.

I want to get started this is the.

Logo Center for Health analytics and
informatics will tagline it's not a latte

which we believe very much that
there's still a lot of opportunity for

people to get more gauged in
health analytics as we go so

I wanted to start just by talking about
this idea of persuasion medicine and

what that reflects on a broader
scheme when you hear that phrase

because you hear it sometimes Certainly
if you go to academic medical centers

you go to campuses you'll hear a phrase
precision medicine generally speaking

what they mean is going from data will
talk about different types of data but

from all kinds of data in
a very broad spectrum and

processing it harmonizing it cleaning it
doing something to it that allows you

to then analyze it in a more sophisticated
way so that could be predictive modeling

will talk about other types of ways
that you could analyze data to arrive at

some type of evidence something that you
want to convey in order to hopefully

improve care improve patient spirits and
improve quality and safety and so

forth the tricky part is that it's not
just enough to build a model it's not

just enough to produce about.

And it's it's not just enough to publish
a paper even if it's for Katie or

another become friends.

In in health care you have to find a way
to deliver that to people where and

when they need it so there's a piece of
this that's about interoperability How

do you get information to where
you want it at the right time so

we'll talk about how you get those data
there in a form of decision support how

do you push that to a provider
to a patient to a team

in making a decision the last
part might seem redundant but

it's not it's once you've gotten it there
how do you actually drive a change in

behavior right how do you get a doctor
to order something or to not order

something how to get a patient to eat
the right thing take the right activity

it be adherent to their medications and
that's actually I think Jeff

mentioned the human computer interaction
part as a big part of it is how do we

employ psychology in a sense to drive the
behaviors that we're looking to achieve.

We're talking a little bit about the
origins of precision medicine as a concept

to talk about different types of help beta
talk about global research networks but

more specifically around
analytics patterns

talk about interoperability I know that
Dr Marc Bronstein gave a course or

a lecture here think there may have
been like a fire drill a minute or

something like that but in theory but
I'll just cover a few items that you might

have heard already and then talk
about Georgia Tech's own platform for

doing health data analytics so precision
medicine has a lot of definitions probably

none of them are right none of them
are exactly wrong but here's one

which is that it's about optimizing
the prevention the diagnosis and

the treatment of disease through more
precise definitions of genetic clinical

behavioral and environmental factors
that contribute to individual health

rights as a whole swath of things and
it could be for treatment it could be for

prevention other purposes put more simply
or at least graphically more simply.

You've got a group of patients and

today if you all out of this
maybe seventy five people in here

whatever it is you went in if you want to
see a doctor and you guys were having.

You know.

A cold then the doctor sort of more or

less would give you the same set
of medications or they pick from

a set of medications they might look at
a particular factor in your history and

in this case the treatment is for
say blood pressure and

all the patients just get a different
mix of drugs where the idea is

that in the future with a precision
medicine approach they would prescribe

the right medication not only in terms of
effectiveness efficacy but also in terms

of safety tolerability right some people
can't take drugs four times a day out

of you ever been prescribed an antibiotic
they have to take four times a day.

I've never successfully taken four times
a day ever I was just really hard to do

right but some people are really
fastidious about other people

the insurance differences it's expensive
for one not for the other other kinds

of things are tolerability and safety and
so forth so that's it is a procedure in

medicine you also hear this phrase
this phrase is personalized medicine

to be clear there is no difference
between precision medicine and

personalized medicine there's no material
difference this was a phrase that came up

first maybe about ten years ago and
the number of providers or provider

groups felt that you know doctors have
been providing personalized medicine for

years right were very personable you go to
see your doctor they say how are you doing

Jeff good to see you welcome in you know
housework center center so the idea of

being personalized been thoughtful about
individual needs is of course been part of

medicine for many many years precision
medicine is sort of getting a little bit

more of that finer grained approach
the route of precision medicine is in

Juneau mix so
one thousand nine hundred ninety which.

As many of you are not yet born but
anyway so one nine hundred ninety S.

was the beginning of this human genome
project this was a huge huge endeavor

to map the entire human genome and

they thought it would take decades they
thought that it would be when complete.

Revelent Torrie completely
revel atory we would understand

what caused the human disease because
we would have all the genes mapped

out and they were actually able to
do it relatively quickly you can see

the the pace of the human genome project
the first genome was actually completed

here you can see two thousand and one was
centrally the first fully sequence genome.

Pretty good and like everything Moore's
law type things the cost went down over

time and went down significantly to the
point where it really is under a thousand

and probably under I mean nowadays it's
really just a several hundred dollars

to get your genes sequence now there's
larger bits of the genome exome and

other parts but to get the standard
sequence that they got in two thousand and

one the price has gone down from one
hundred million dollars pretty well so.

Hopefully that's not
the Bitcoin trajectory but

it came down quite a bit over
the course of those years so

you can see that change for the problem
is that the reality of precision medicine

even with the changes in Juneau mix and
the capabilities

of the science have actually been a little
bit slow to proceed so you see these

these headlines some of which are you
know and in kind of blogs but

the second one is from like The New Yorker
which you wouldn't think be writing this

kind of thing in there saying here's a
problem with precision medicine essential

one of the doctors didn't come out
of training thinking about this but

the bigger problem is how do you
convince doctors to test for

something before there's evidence for
what works and

then how do you convince someone
to to pay or build a test

if they don't think anybody is going
to use it so there's sort of a chip.

And that problem if your company out there
as a great idea for a new genetic test.

Where's the financial
incentive to build that until

you know that there's going to die you
know sides it's a bit of a challenge and

him which is a is a huge conference
every year it's typically in February or

March it's all of the health I.T.
companies.

There's like fifty thousand people
it's either Las Vegas or Orlando.

I think like Magic Johnson is
speaking out this year like it's

always like this really like big and
like very very unusual set of like.

Kind of rock star type people and health
people and all these big parties and

stuff but the gist of it is how are things
moving forward in everything from health

analytics electronic medical
records population health and

all the other kind of things that you're
probably learning about in this course

one of the challenges
that they've identified.

Is that the data or to integrate these
data into the real world Veyron is

really tough so you can do a genome you
can capture fascinating information but

how do you move it to the right
place in order to be effective so

that's kind of a baseline idea of where
we want to go but also a little bit of

barriers to getting to that point so what
kinds of data might you be looking for and

hopefully over the course of this we're
probably what a couple months month and

a half into this class if you guys are
start to get a sense of the kinds of data

that you might encounter but I do want to
run through some of the broader notions of

health data what you might be looking for
so the classic at least a couple

classic times types of help data one is
clinical data right so the electronic

medical record doctor sits down that
enter your information they ask about

your allergies they ask about your past
history about your family put in the M.R..

Here's an example is some
screen shots from E.M.R.

as they look like like Windows X P or
whatever many of them still look that way.

Have a lot of compassion for
your doctor when you go and see them and

you're hacking away on your i Phone and
they're like working in this so

it's it's still though a challenge not so
much just because of the interface but

the new the notion of the data
can be tricky right so

I want to ask since it's ph I'm sure some
of you've got and you've allergies or

you're say your mom told you
had an allergy to penicillin or

something like that you can't take
it it's very common story but

a lot of times that might be because when
you were four you got some penicillin

you threw up your mom said don't
take any more penicillin so

you go to the doctor and you say I'm
allergic to penicillin but in reality

you may not have a true allergy you have
sort of a kind of adverse reaction but

there's a difference so if ever when
you're older and you get really sick

it's tricky because the doctors don't
know is this person really allergic or

you know can I treat them in a life
threatening situation so E.M.R.

data is can be rich but also tricky.

The same could be said of claims data so

claims data which I'm sure you have
a chance to come across in this course

that's the stuff that the insurance
companies produce right so

like every good system the part
dealing with the money is always

the furthest ahead in terms of
comprehensive this so the claims

data everything done everything ordered
everything captured goes into claims data

but there's a problem one is that they
sometimes reverse stuff so like something

goes in and then something is you know
comes out you don't always catch that

the other problem is that doctors often
just fill in the claims information

to I don't think they're doing it to get
paid but they're doing it to just kind of.

Go through the process of say
somebody comes in complaining of.

Feeling tired and the doctor
wants to order a thyroid exam so.

There they might circle fatigue
which would be great but

sometimes doctors will circle like rule
out hypothyroid is there other kinds of

things that end up with a code that's not
really what they intend to sell what they

put in their documents but it might show
up in the claims Another example is like.

A hip fracture so
if you look at a record and

you see seventy five year old woman who
has a hip fracture I can assure you that

she will have twenty more hip fracture
diagnoses over the next twelve months but

it's not that she obviously fractured hip
twenty times is that she goes back she's

follow up everybody just kind of
circles that hip fracture circle but

actually becomes tough to know what's
the original What's what's stablished and

so there's a lot of just trickiness
to work with claims data

the good news is it's voluminous
writes is a lot of information there

you'll see many many studies done using
these data so it's kind of a classic.

Story where people say like the guy
lost his keys and he's looking for

the keys on the lightened someone says
well you know where you lose your

keys that over there but this is where
the light is so it's the same thing people

study claims data because it's easy to get
there's lots of it is very satisfying but

it's also a little bit dangerous
in terms of the accuracy.

The next category is patient reported
data so this stuff is pretty cool it

has to do with like questionnaires to me
that seems a little bit old fashioned but

you go in some is doing a study or
a doctor says Can you fill out this report

and we're going to it's a validated
instrument music done studies to make

sure that this thing kind of does what it
says it's going to do so here is the Ph.

Questionnaire which is to blur into
small on this screen to read but

I will tell you that question number six.

Says.

Are you feeling bad about yourself or

that you are a failure who have or
have let your.

Well for your family down so like to me
there's two things to think about question

over sex number one is like that's like
a super nuance concept right there is not

like a snowman code I see nine killed like
one twelve that means that number two

is like a really depressing question like
When you read it you know to me I would go

from like a zero to like a two just from
reading that question because like well

now that I think about it you know maybe
my family is kind of you know whatever so

point is that questionnaires are
interesting they probably have a little

bit observation bias like Hawthorne effect
kind of stuff where you respond to things

based on a question but they're much
more nuanced right they're much more.

Sophisticated at capturing particular
elements that are hard to get in other

ways All right so now comes like
all this fun stuff that we're

doing all over Georgia Tech that people
are doing all over the country all over

the world right this is how you capture
patient generated data this is sensor data

this is high velocity data coming from
had been bits of cardiac monitors and

shoes and Under Armor and whatever you
have there's data flowing off of it and

people are trying to figure out not only
how to use it not only what to do with it

but how how to incorporate it right there
so much information I can assure you

that if you were to strap
her a monitor on to your.

Your head or your your chest and

you said your doctor hey good news I'm
going to stream my heart rate to you so

you will to keep track of it day in and
day out they would say thank you very much

I will be shutting that off immediately
because they don't time to look at it but

at the same time there's probably useful
information if somebody is a diabetic and

they have contact lenses now
they can read your blood sugar.

So how frequently should that be passed
to a provider right how it should be used

the good questions not yet solved but it's
a really important exciting area to work.

So now moving a little bit from the person
level to public health level All right so.

The national death index has a list of
people who have died it's not always up to

date but who's died little bit about the
causes where possible the more Morbidity

and Mortality Weekly Report comes out
of our friends down here at the C.D.C.

population levels you can understand for
example the flu has been really bad

this year it's peak there's actually
unfortunately been a number of deaths from

the flu this year and so
when you're studying events that occur

over time in your in your big databases
you want to take into account what was

happening in the world right what was
happening in the population of that time.

This thing on the right sorry it's
also a bit blurry to read but

it's called the Camden hotspot in Project
and what they did was they pretty

simple were pretty cool they just
took the addresses of all the people

who were seen in emergency room in
Camden New Jersey over about a year and

a half and what they did was they
focused on the high highest cost

patient spy kind of block
by micro block building and

what they found was there was a set of
like three buildings that accounted for

like forty or like twenty or thirty
percent of all the expenditure of all

the patients and so it turns out Of course
not to your surprise that there's a lot

of places that are high risk environments
where they don't have what they need

in terms of support structures in terms of
stores in terms of treatment medications

other types of activities or they can
possibly project was very illuminating

the says that you can just do simple
things you can just count stuff and

you can understand what might be going on
in fact I don't have a slide on this but

there was maybe about three or
four years ago

somebody went to the morbidity mortality
week report they looked at reasons for

death in American in Americans and they
compared it with other countries people

from Canada from Australia from England
and other countries in a similar.

Socio economic bracket and

what they found was that you know people's
life expectancy was going up OK except for

kind of white middle aged men were
like they were falling off the curve

everybody else was going up and
they were falling off the curve and

so it turned out that a lot of
these people were dying were.

Dying due to liver disease due to
suicide due to opioid overdose

relative to other countries so these guys
literally just looked at these there were

a couple of professors at Princeton I
just counted up nothing fancy and

they sent it to the New England Journal
of Medicine and the story

this guy writes who ended up writing
a paper that he got he got a rejection so

fast he thought it was like an auto
reply came back within like within like

ten minutes your articles been rejected
like this is this is totally useless So

two things happened over the next couple
months one is it was accepted by the Jam

are or another really high in journal and
it up getting a lot of media attention

because of how relevant it was
to think about opioid addiction

alcoholism people losing their jobs
falling off the curve what's happening

social economically in America that
was part one so they got to publish

part two is he won the Nobel Prize not for
this but for something else like a week or

so later I think sorry that's cool
that's good there's a story there but

I'll leave it out for
the moment let's see.

This could be interested to just
make finishing a presentation.

So.

So you can imagine that this individual.

You know caught something very simple and
I want to emphasize that because a lot of

what you can discover using the techniques
in the help that are are very simple.

Social networks I probably not
a proselytizer about social networks.

At the same time I have a feeling like I'm
going to just get this i will all take.

The moment required to undo this
incredible number of things this for

my kids just flying in here on the second.

I don't think anything bad will
come up but just you know.

It seems high risk so I will.

I will turn this off
OK think it's done OK.

So people look at social media quite a bit
right and they think about can we use data

from Facebook from Twitter other sources
like patients like me to build to derive

information it's a valuable potential
source of in of discovery and research

I encourage it I think that most of
what's come out has been interesting but

not hasn't like really turned the tide yet
I think it may get better but

some of the think about and they're very
big sources now we talk about public but

non-medical sources of course Twitter and
Facebook are not medical but

other things like census data right what's
happening these tracks like hot spotting

Noah OK I know what Know US stands for
but it's the the National Oceanic and

Atmospheric Association and these guys
track the weather and like you can go

to know and on any day anywhere in
the United States you can find out

exactly what the weather was right in
the storms and all these things and

they make it very easy to find hurricanes
and tsunamis and all kinds of stuff and

it goes without saying though not
obvious to researchers all the time that

if you're going to study outcomes and
you're looking at like how are people

treated with pre-diabetes and
you've got you know Houston in there and

you've got hurricane Harvey happening
in the middle of your of your study

there's going to be some biases there's
going to be some things that change and so

to go to be aware of that is extremely
valuable and things like Zillow just

figure out you know social economic status
things like that can be very useful.

And then there's like the Super.

Scary stuff that's that's cool if
you're on the data side of things but

to be on both sides all scary right so
axiom and Nielsen collect a crazy amount

of data about you and your habits and what
you buy and where you where you go and

drive and are able to really understand a
lot of data about your behaviors which can

be connected and there's been some start
ups that have taken pharmacy data and

integrated with other source of data to
try to say who's likely to be adherent

is likely non-adherence and things like
that you may have heard about the medical

credit scores sort of for risk I'm sure
you've heard about that target pregnancy

prediction thing that
happened a few years back but

those kinds of ideas where people
are taking other sources of data and

have you got you guys heard about that
target pregnancy thing where they they

they found out she was pregnant
via a message from Target OK.

So that's one of things that's happening
in the outside sources of data so

we've got a lot of different sources
of data but what we actually do it

right a lot of things on the horizon
how do we take advantage of and so

I want to I want to start with this
sort of like high level triptych

of three kinds of things you can do one
is clinical characterization All right so

what happened to people
who did something right so

this was the guy the guys from Princeton
right they just said who died and

kind of what's up with that very valid
study the next kind of study is patient

level I'm sorry is population level
effect estimation that's a big word for

something like comparative effectiveness
that means if if these guys take drugs day

and these guys take drugs be who's going
to get better faster who's more likely to

be successful whatever study you want to
do these guys are enrolled in doctor son's

class these guys are all this class who's
going to get better grades whatever it is

that's that's a population level
estimation All right wow OK Well

we're still OK it's better
than getting all the.

TEXT But now it says your phone
is no longer connected so.

And the third one is
patient level protection so

what that means is that you're
really doing predictive

modeling that's the percentage in
medicine part that we talked about but

the reason I emphasize other pieces
that they're every bit as valid and

in some cases much more effective
as forms of guidance OK.

Alright so we're going to start with
this idea of a patient journey right so

how do you guys.

Looked at E.M.R. data or claims data or
anything over the course of this class

was like a rough show of hands if you
got into clinical data little bit

OK little bit like not OK So
let's imagine a patient's journey

through the health system so they come in
and they're diagnosed with some sort of

disease right they have a condition
doctor says you've got a condition X.

they might get treated for
that condition sounds good.

They may get an outcome right so
they might get treated for

depression and then they end up having
a another outcome that's unrelated so

they get diabetes or
that maybe is related they take Prozac and

they develop insomnia
right could happen so

it's an outcome of this is the other and
other conditions again lots

of other drugs they have procedures
done rites of procedures like.

It may be an injection of
something maybe a surgery but

also just going to the doctor's
office is called a procedure and

they have all kinds of measurements right
that they check labs that geez they check

all those kinds of stuff so this is
sort of like what a patient's journey

might look like in a simplified form this
is this is what's up with patient data.

One of the things that we do often is
we pick a time zero time zero is like

the thing that you cared about happened so
if I'm doing a study about a drug that

I would say the day the person got this
drug for the first time that's time zero

that's what I'm gonna look at obviously
if I'm looking at a different problem

I might say there's an observational
procedure whatever it may be.

And so then we have to part
with the baseline time

everything that happened leading
up to that and we've got the.

The follow up time all right so

not overly complicated we've got these two
different errors everything before and

everything after and those can be
useful in the sense that every database

the sort talk about what you do with that
baseline time and up at baseline time says

this is the stuff that I might
use to predict something right so

things that came before are things that I
can use to say I'm not going to predict it

here's the way this person has behaved
previously the fall of time is like how do

they end up what happens to them so this
is kind of how data sets are picked up and

if you have a big database with
a million patients in a claims all it is

is just a is just a million sets of these
of patients who are having these different

types of events so
there's a kinds of questions

that you might ask which gets back to
what we described before so you may say

What treatment did a patient choose
after diagnosis right so that's kind of

a descriptive What did they get or what
what patient which patients chose those

treatments descriptive how many patients
experience the outcome descriptive all

these questions very straightforward
not highly statistical but

extremely relevant and very very
interesting then you start getting

a little fancier with the probability that
I as an individual would develop a disease

and what's the problem Spirit's particular
outcome and again do treatments cause

outcomes and so forth so the level of
sort of statistical complexity varies but

the nature of the questions are very
similar so I describe those three

categories and I'm emphasizing this
again because I want people to.

In this course are thinking about data
you think I did in health care and

a lot of times people do jump to

you know I'm running a deep neural network
to discover the relationship between X.

and Y.

those things are very very powerful and
there's a lot of value in them but

you'll find that a lot of healthcare
organizations providers insurance

companies pharmaceutical companies
a lot of sponsors of this kind of work

are very interested in these other
sort of simpler pieces as well and

to go to do them at scale to do them with
higher quality brings a lot of value so

characterizing and predicting but
also estimating population differences

is very very powerful so just kind of
lay this out again imitate an example of

depression because we kind of highlighted
a little bit Some who was diagnosed

with depression they're treated
with a S.S.R.I. as a Sarai.

Sierra tone and we have taken him or
that's Prozac and Zoloft and Lexapro and

all these great medications that
make people feel better but

they can cause problems of time in this
case there's concern about risk for

a stroke Now this is a pretty cartoon
it's nice to imagine but the reality.

Is that it's much more complicated
when you look at actual patient data

there can be hundreds of different
diagnoses medications conditions

we've had a patient in Indiana
where I was previously and

we did a study of what's called
frequent flyers in the emergency room

Hey buddy guess what that
means I'm a frequent flyer.

Yeah exactly so there's so
there's medical Sturm and

slang for everything we have another day
I'll give you a talk on medical slang but

that's exactly right so there was a guy.

Who was in the emergency room who had
three hundred sixty seven admissions

to the emergency room in one year right
in one year we know there's only three of

these five days maybe it's a leap year
whatever the case is he averaged more than

one admission to
the emergency room per day so

people are very complicated their
histories are very complicated and so

cleaning and sorting out harmonizing
these things are extremely important so

we've got a lot of different patients
some as young are simpler and

some are much more complex
with longer scale durations So

to be aware of that to think about that
as an important part of our data so

I want to come back to this question
of depression just because it

it's a tricky thing to manage and
talk about how that as an example

could be used to reflect these
different types of analytic approaches.

So how should patients
with depression be treated

the code is Major Depressive Disorder
it just means depression so

in two thousand and ten and
it looks like an old picture but

like the reference to thousand and
ten there was a guy that came out from

the American Psychiatric Association
right pretty legitimate body.

And here's what they said hard to
read OK you can recount all right

they said there's a lot of medicines
there S S R I's there S N R I's are like.

Effects or and things like that there's
all kinds of different treatments

try Cycloset Or and
here's what they say choose.

They recommend choose
a medication largely based on

the following patient preference OK so
you go in to see the doctor and

you say well here's my situation what
should I take do you know that's like

the commercials right when it says
blah blah blah blah blah you know ask.

Dr this is right for you and then
the doctors have guidance is that says

make the decision in concert
with the patient right so

you might both show up and say I thought
I was supposed to ask you you know but

patient preference right nature of prior
medications safety tolerability other

medications or taking a cost
etc What they don't have really

is like specific guidance about who to
treat with what based on anything specific

right these are just kind of
these broad general guidelines.

It says for
most patients as Suraya center.

Is optimal That literally is a list
of like fifteen different drugs

right in that Senate so
it makes it very difficult to say so

the next question is how are patients
with depression actually treated so this

is a pretty cool study that we did that
I was fortunate to be a part of where we

went to a number of organizations and I'll
just fast forward one secs to show you.

You can see here in total.

On the right side let's say we've got I
don't know almost two hundred million

patients from around the world from
Korea from Japan from the U.K.

from Hong Kong and many from the U.S.

some of which were claims some of which
were E.M.R. some were in patients or

outpatient lot of variation and
what we looked at was we said earlier.

And this is kind of getting into a slight
weed a little bit we basically said like

OK we have actually had over two hundred
fifty million patients start with and

we were very strict about
it we said you have to take

have to have at least four
years in the database right so

a lot of people come in and out you go you
show up and you guys are here to attack.

You're here for a few years you go on
somewhere else your data gets fragmented

so we got pretty here said you need to be
in the database at least four years and

you have to have at least three years
of continuous treatment it's a lot like

you need to be we have documentation
you can take this drug right along so

when you went from two hundred fifty
million people it drops down to just

two hundred sixty four
thousand that qualified.

And then we produce this this kind of like
sunburst charge you can sort of imagine

the inner ring is what people got first
the second ring is where people got

second third rings with who got third and
you can see number one there's a lot of

heterogeneity to tell a premise pretty
popular that's Lexapro search or

lean is lofted cetera you know but you can
see there's a lot of variation people come

in maybe it's patient preference maybe
it's something else or people just let's

try this right based on my experience but
the other crazy part about it.

If you just look at the colors that will
tell you enough the other crazy part about

it is just looking at different
places the patterns of treatment

were vastly different I mean
Japan to Hong Kong inpatient or

outpatient New York City to
Indiana all different patterns

by health systems etc There wasn't really
any preferred first line treatment and

it was a leaven percent of patients

had a treatment pathway that was shared
by nobody else in their database so

you guys as as computer science folks
are definitely thinking about predictive

modeling right thinking about how do I
predict the next treatment for this person

what if I told you that you had we started
with two hundred fifty million patients

we went down to twenty four thousand
because we had restrictions and

eleven percent of them had no nobody else
was like them nobody else watch the movies

they watched or ate at the restaurants
they were took the drugs they took right

tough tough to to build those predictive
models so this is the nature of things.

So how do we actually do this and I will
do a little bit of a demo towards And

I think we'll have some time.

As we build a patient cohort on the show
you how we build a patient cohort

because you start getting into data
soon you'll need to know how to do that

build a patient coort take the medications
and explore the differences so

we'll talk about how to do that
as we go and we're talking about

how we look at patients at large scale and
explore all these things and

that's through a resource that you have
a Georgia Tech where you have data

where you've got to be all is great
stuff to talk about how you can do that.

The other thing just to know about
clinical characterization is that if you

look at a lot of treatments you look a lot
of patterns you can generate millions and

millions of statistics
about population and

a lot of people confuse data mining for
they feel like anything that looks at

data at high volumes is like data mining
like you're dredging for stuff but

I want to encourage all of you and I think
you guys are are the kind of audience that

probably knows from the beginning is
that you can do high quality work across

large amounts of data across
large sequences of analyses

it doesn't mean that you're reducing
the rigor of a study sometimes if you

go to a professional society or
in my case you go to the F.D.A.

and you say we just did a study we
did we did five million studies

looking at all these drugs and all these
conditions they say well these studies

they can't be any good right because what
we like is a study that was done and

was printed in this journal and you
know and we can show all the metrics and

all the approaches but

something about people doing millions and
millions of studies freaks them out so

I want to encourage you guys to think
about how you can do things rigorously and

demonstrate that rigor and at the same
time be able to do that scale because this

is the type of place that can do that so
a bit hard to read I won't I will get

too much into it other to say that
the different databases have different

variations which we kind
of highlighted already.

So looking at individual outcomes and
again this is just look at the kinds

of adverse effects that occurred with
certain treatments for depression

we go to the product label there would
be no has even ever read a product

label like you can say As for your mom or
grandmother or high but as anybody like

OK you've already a magazine right where
like there's a picture on one side and

there's like two people frolicking
in a field of flowers so happy.

And there's a name of something with a
bunch of X.'s and Zs and then you flip it

and it's got the small print of like what
the drug can do and the five thousand side

effects that can be cost right so
this is the drug label this is the list

of how it was how you take a medicine
the adverse effects of that treatment.

And.

So.

What would people guess is
the average number of side effects

on a drug label give or take

any guesses it's probably more than
one I mean they're not terrible or

I'll tell you this seventy
is the average seventy

is the average number of side effects
per drug level what you think is the.

What's the max was the highest.

No guesses I kind of gave you
help with that seventy but.

That was good guess that's what I would
say the highest is about five hundred and

forty and that is on a drug for restless
leg syndrome OK so restless leg syndrome

you might see commercials for used to see
a little bit more right so there it's for

Restless Legs it says commercials like do
you feel the urge to shift around when

you try to go to sleep and you might like
why I didn't think about it until now but

not to mention it maybe I do need this
three hundred dollars drug a month so

they're kind of looking for patients
no offense because a lot of people do

restless legs and it can be very very
uncomfortable that said the treatments

can be worse than disease so there's five
hundred forty different side effects

among them interesting Lee our
compulsive shopping actual side

effects you look at up the F.D.A.
label compulsive gambling is another one.

So you know there's tradeoffs right
legs versus gambling and shopping and

all these kinds of things is pretty
complicated to make these decisions so

we collectively including you guys now
because we're now all involved here is we

need to find ways to be able to understand
what really causes what how what

the tradeoffs are how to guide patients
effectively is a complicated stuff and

we have a certain obligation to do it so
here's an example I may not linger too

long on this one this is looking at risk
of stroke with depression medications and

basically the way that you can do this
analysis at scale kind of talked about

a bunch of stuff so I probably will
get too much into it other than to say

we're able to identify this risk pretty
quickly isn't cool tools that we have here

Georgia Tech that I'll show you how
those work to find this information.

But let's say for example that we that
we did get a call that issue about

compulsive gambling maybe or something
else the F.D.A. is like we understand you

guys Georgia Tech are getting really
deep into this we'd like some guidance.

And they've actually done just
that here's an example of.

Something that they've done that for so
kep are a is a medication for seizures.

OK medication for seizures and explain
what that is in a sec but the F.D.A.

put out a a request for

support they said we've heard
that Keppra the drug can cause.

The side effect and we'd like some
input from people the world so

here's what Kepler is also known as
love atrocity him it's one of the major

antidepressants is very common in fact
it is the most commonly prescribed.

Drug this is looking at for
seven to thirteen thousand

prescribed drugs which ones
occurred in what order and

you can see that Keppra is number one
that's pretty common lot of exposures and

seizure sores are quite serious and
very tricky to treat.

And to a D.M. is also pretty serious So
this is what it looks like.

It is a significant swelling of the face
and the lips the cheeks the lips

that is caused either by congenital
people are born with this reaction or

they it's a medication induced swelling
Now the amazing part is this is what

the patient looked like after
resolution of the symptoms so and

her hair color looks to
have changed as well.

But this is the extent with
which somebody can achieve and

that's how much well you
can get from injury so

it's obviously very serious it's serious
how it can affect the ability to breathe.

And it's considered.

A rare but serious treatment
a rare of it serious side effect

it's relatively uncommon However there are
medications on the market that cause it

the most well known are what's called ACE
inhibitors or blood pressure medications

like a sinner Pearl kept a pearl
they cause between two and seven

out of a thousand person years somebody
gets injured so it's out there and

it's a good opportunity for us to study
Keppra in this case so get into all of a.

The how's and why's that we did this but

suffice to say that there's networks
of collaborators as a reference before

Japan Korea Australia Well those
guys are still hanging out and

there's now at fifty some odd sites
that work together to answer questions.

So what we did is we said well let's write
up a study let's write up a plan and we

publish it we sorted it we open sourced it
everybody can contribute to the the plan

and we wrote some code we wrote some
of our code and we tested it and

we post it on get up you can go right now
and check out our code and say you guys

maybe picked the wrong analysis or
something but it's all out there.

And I will describe the statistics
the biostatistics but certain type

of thing called a new user cohort study we
looked at who got seizures who got injured

Simo we compared it to a drug that had
been in the market on the market for

about seventy years old drug called Fenton
and we announced the study there's like

a forum you could go to and we said we've
got this study who'd like to run it but

this is pretty cool if you just think
about that like open source nature

of science of open science because
typically you just couldn't do something

like this you want to do a study you had
to have this massive multimillion dollar

multi-site trial and
you just it was took years to do but

in this case we were able to put out
the code people would get their own eye or

be approval is to sort of you board
approval they would run the code they

would share aggregate data and we would
compile it and do the study so we had

about ten sites that ran it and I won't
get into the details of the results but

with a lot of people came back sixty
thousand Keppra users seventy five

thousand paratroopers twelve million days
it's entered cetera and basically what we

found was there was no difference
between this new drug Keppra and

what have been in the market for
seventy years so it doesn't mean that

it's impossible that it could cause it
does mean that sort of a drug that's been

standardly use for a long time there's no
difference between the rates of it and.

It is reassuring that at least there
isn't a particularly high level of.

Edge of the moving caused by this drug so
produce all these nice graphs but

the short version is everything's
within the range and

we we said it to this
journal Epilepsia and

what was kind of cool about this was that
they said wow this is this is pretty neat

you guys looked at you know you did
a well designed study that should be

interesting to clinicians right that's the
goal is that it's useful to doctors but

it was also done at scale and
done across all these sites and

done within like a month and done an open
source way so it's nice to know that you

can have a paper published and
reviewed by sort of the official.

Had Gemini of of hierarchy of
medical literature reviewers but

still have it done in this way that's very
open and flexible so this was a good study

we we then put this study out you can
see there's a lot of authors I guess

I was biased in that I talk about study
obviously I'm one of the authors here but.

What you can see is that there's
a tremendous number of people

who contributed I think that's really
a direction that you all can leverage and

take part in that if you come up with an
idea and literally any of you come up with

an idea and you write me and say hey I
have this thing I've been thinking about.

Can we study it then we can say that's
cool Why don't we write an idea

let's send it to these forums let's post
some code I get let's try to average or

not advertise but convince people why this
is an important topic and we can have one

hundred million patients be analyzed using
the code that we wrote and be put into

a study and then be part of contributing
evidence to to the literature so

there's something I think that is a really
neat opportunity for all of us OK So

we've talked about generating evidence
we've talked about how that's done a broad

scale I do want to come back a little bit
interoperability just because I don't know

how much the fire drill took up from
the time that you had that but just.

A few brief words on that
on interoperability so.

There's two kinds of interoperability
problems one is syntactic like how you

move the data around what is semantic
which is like what codes you use.

You might never have looked at this but

if you ever bought a pill of anything at
the drug store there's like a code on it

it's called A N D C could and
it's not everything you can you know buy

Tylenol you can buy whatever it's
on there and that code has a set of

numbers that refers to the manufacture
first the drug refers to the pill.

Strength but also to like
the number of pills in there so

if you buy a one hundred pack of Advil and

then you buy the A five hundred pack of
Advil it will have a different code so

super head if you like supply chain right
but if you if you want to know does Advil

or ibuprofen cause stomach
ulcers which it does spoiler.

You don't care if the person
took C.D.'s ibuprofen or

took add the new printer took it
from a five hundred pill bag or

a fifty pill you don't care but
you do care about the ingredient so

the problem is there's all these
different codes to say the same thing so

that semantic aspect is a very interesting
but a different story we want to talk

about in detail today but syntax how
you move stuff around that's addressed

currently we're getting there by
something called fire fast health care

interoperability resources will not
blow you guys away in its technology

it's basically a set of A.P.I.
It's basically a set of A.P.I.

is that reference all these
different kinds of things that occur

in medical environments so say you want
to get a list of the patients conditions

there's just a condition A.P.I. and
you just do a GET request to the A.P.I.

and it gives you the patient's
conditions pretty cool you can post

up you can add new ones that's the gist
of it and it's become more of a standard.

That you can use across the electronic
record systems comes back most

the most isn't Jason but it comes back
with these different data and metadata and

information about the patient so

we'll get in the resources here I
assume Dr Brownstein covered this

provides different data standards
I mentioned she you before about

the code for ibuprofen this is a code for
blood pressure by something called loike.

If you guys want to learn more about these
code systems and things like that in C.S.

sixty four forty we go
through it more detail but

essentially it's a standard laboratory
naming scheme or observation naming scheme

so loike basically allows you to
have a code that everybody shares

around some of our blood pressure and
then you basically can use this A.P.I.

as I said you put in the patient you
put in the number and off you go and

returns the results seems just normal but
actually it's taken us like thirty

years to get this work that and here's
an example where they do a query and

they've got a particular Lloyd code
a particular time zone or time window and

returns the results so
that's pretty awesome.

One of the challenges is that what do you
do once you get the information in and

how do you create applications around
it so there's a lot of thinking

that happened and it said Well
what we should do is if somebody.

Has a great idea for an app we're
just going to we're just going to

put that app into electron medical
records system the doctor can click it

when they need it right so
somebody is at risk for.

Someone has depression and you want to
know what's the right treatment for

this person and so you want to take
a predictive model and you want to

show the doctor the right thing well
the problem was is that the doctor had to

think like click on depression app to find
something that's just not a very realistic

workflow type of thing right so
then they invented something called C.D.S.

hoax which is also part of
the fire standard I want.

About all the different pieces but
if you take fire and C.D.'s books and

some crazy languages then things
get a little more dynamic and

you basically have a system that can be
triggered so the doctor orders medication

in this case a blood pressure medication
the doctor orders a lab the doctor opens

the patient record and it calls an A.P.I.
the A.P.I. does some logic and

it returns back either a little card
some guidance recommendation a treatment

order an app whatever it may be so C.D.'s
hooks is how we've sort of addressed

the connection interoperability to trigger
things again all these might seem fairly

simplistic but it is the current state
of the art I just came from the C.D.C.

this morning and people you know
were just like wow we can deliver

recommendations based on public
health guidance using these tools and

it sort of reflected the fact that
you Mars are just kind of catching up

with what's pretty standard
now in web technologies etc.

And this is an example of what
that C.D.'s hope would look like

you pop up an alert at the bottom
comes all recommendation OK.

Let's get this for the moment and
OK so last piece is around.

H tap the Georgia Tech health
data analyst platform.

And this is trying to take all
the pieces we've talked about so far and

bring them into a common.

Locally usable framework for it so
that as I just mention sounds good and

all these tools are cool I'd like to
make apps I like to analyze data but

how do I actually use it as
a student as a researcher

as anybody who's part of
the Georgia Tech community so.

We built something called the Georgia Tech
Help data analytics platform which kind

of takes all the pieces that we've
discussed and put them into one bag so

we can add this picture at
the beginning this is just like it.

Just like a technical overlay
That's not good Powerpoint

has gone down so
this is this is the tricky part of.

Serbia that to

see or read.

And see go to the bottom all things being
even could have been much worse OK.

OK.

So this Georgia Tech health data analytics
platform write this last piece so

how do you actually use the stuff that
we talked about as I said we sort of map

this idea of this continuum onto
a platform that we've built here so

this is the help they tell its
platform is that God tech dot edu.

And essentially what it does it provides
you tools that you can use for your own

analyses is still a work in progress for
piloting this semester but essentially it

provides data right so the first
thing is you need data you have data.

Some a synthetic somebody identified but

all of it is accessible
through two mechanisms one

of the a direct database credentials and
to be a fire A.P.I. so you can take

advantage of it you can build on top of
it you can run analysis on top of it.

And the second one simple enough has about
two and a half million patients worth of

synthetic data but still it's large enough
to do a lot of interesting analyses on.

We have analytic notebooks right so if you
want if you're sort of Jupiter notebook

type developer you can go and you can
build Jupiter notebooks you can share

Jupiter notebooks you can create
analyses to build with others and

a lot of the pieces on this sort of really
remodel a lot of user user interface

pieces are kind of getting fleshed out
this semester but it's Want to show you

some the core pieces we have apps right so
one of the last sections is

how do we create those applications that
we describe how to create fire apps but

also how do we create absolute people
explore their data more effectively and

I want to highlight one of those here
which is something called Atlas.

Which is the.

It's like a.

Sort of public health.

It lets you take the data that we have and
run analyses on top of it and

some of the show a demo of
that in just one second.

And just a little bit of how that works so
well point us up.

Go back into my main world here and

see what's blown up and
stuff but hopefully not too bad.

OK.

All right so this is an example of C.E.

but sorry we've got a new
a new your I think.

OK So this is an example here we have
these data sources here at Georgetown who

have multiple different data sources
I think I mentioned we've got simple

we've got mimic we've got other
things that you can read about and

this information for example you can
say well what kind of people do I have

in this database now where
all the My people are here.

And it says information like when where
people born male female what what that

ethnicity etc and you can explore all
these different things what conditions

they have what procedures they have etc.

But one of the neater things that you
can do is you can create a population of

interest so for example and
just got a few minutes but

I just want to show you conceptually
how you might do this so

I'm going to create a concept of
something that we think is interesting so

we've been talking a lot today about
depression or as a just a as a concept so

I'm going to I'm going to say I'm going to
make it a concept that called depression

and I'm going to add some stuff to
it I'm to look up in this book.

Sorry all the stuff that's blown
up here I'm going to search for.

I'm going to search for depression or
actually I guess I know major depressive

disorder seem to be my my thing here so
I'll search for depressive.

Why this is going on there OK so
depressive disorder

is something I see here it's pretty common
I can sort them by frequency it cetera but

I know OK depressive disorder sounds
like the one that I want to look for

I've got some other examples but

will start their major depressive disorder
cetera some of these different findings

All right I will go there's a lot of other
things that one can do a search for but

I basically have created a couple things
that I I know about from this group

is patients that have these different
conditions and then I can I can go back

and I can say that's pretty neat I want
to create a group of patients that I know

that's out that I know has depression so
I want to go here to.

This section I'm going to say well.

I don't know what's going on here other
than I've got a weird euro reference so

I will take the a different version.

And show you the same concept
is that you can go and

you can say I want to create
a group of people and

it says OK what you want to call these
people I'll say patients with depression.

And I can basically define that group.

And I can pick things like I'd like
people who have some type of condition of

depression it's a longer story to explain
all the details of how you do it but

I Basic it's I want people
have depression I want to add.

A depression concept from here
which may or may not exist because

I'm now in a different cycle but you can
see that there's a lot of different there

different therapies I've moved
to a public site which is why.

You know I can say OK I want
a trophy ablation patients and

then I can basically run those patients so
this whole

goal is to have a simple user interface to
create patients that you care about and

then you can run a bunch of analyses
on top of those patients so.

Over time it will have a separate session
for how you would do it you can build

patients you can define patient groups you
can do comparisons analyses predictive

modeling and all of that on top of these
patients sets so sorry about that little

hick up there but essentially what you
have is you've got here at Georgia Tech

access to data access to analog notebooks
that can use those data access to

tools that can you utilize this data and
access to fire servers on which you can

build applications that other people can
leverage not just here Georgia Tech but

across all of the different E.M.R. So
if you come up with a great operation and

you build it using our fire service here
they can run it all over the country and

all over the world and you can contribute
that to essentially App Store more or

less that allows that to work
here just a few examples

of some of the apps that were
built this past semester.

Payment integration OP operating room
utilization traumatic brain injury sets.

Detection System clinical decisions
port for side effects etc.

A lot of other coal and
other things that were built using this so

that's a lot of information
I know that I've covered but

I hope that you have a sense of the very
least of how the process works and

most importantly how you can begin
looking into using these data

afterwards attack how you can create and
run analyses and how you can really

contribute back to generating evidence
that improves health thanks very much.

For.

Any questions.

That's a great question

was about can a pharmaceutical companies
predict the side effects of a drug so

two parts one is yes based on the chemical
compound they can but another.

Even cooler part of that and this is if
you guys want to think about interesting

jobs for your future there's
some called drug repositioning

which to take an existing drug and
apply it to treat something new and

the way that they do that actually
is they look at the side effects of

certain drugs and they say this
drug must be activating pathway X.

So maybe we can use it to treat something
that another drug that activates that

pathway treats so you might wonder like
why in the world would would they do that

right like why not build a better drug but
it takes a decade to to come up with

a molecule analyze it get it approved for
treatment but if there's a drug

that's already out there that you can
discover works on another condition

you can go get a new label you can kind
of yeah so it actually can save literally

billions of dollars by doing exactly you
said taking a molecule and identifying.

Its side effects but
also potential repositioning so yeah I.

Recall.

Any other questions all
right good thank you.