AI for Goodパネル - プレシジョン・メディシン - SXSW 2017 - #IntelAI - #theCUBE (AI for Good Panel - Precision Medicine - SXSW 2017 - #IntelAI - #theCUBE)

字幕表動画を再生する

>> Welcome to the Intel AI Lounge.
Today, we're very excited to share with you
the Precision Medicine panel discussion.
I'll be moderating the session.
My name is Kay Erin.
I'm the general manager of Health and Life Sciences
at Intel.
And I'm excited to share with you
these three panelists that we have here.
First is John Madison.
He is a chief information medical officer
and he is part of Kaiser Permanente.
We're very excited to have you here.
Thank you, John.
>> Thank you.
>> We also have Naveen Rao.
He is the VP and general manager for the
Artificial Intelligence Solutions at Intel.
He's also the former CEO of Nervana,
which was acquired by Intel.
And we also have Bob Rogers, who's the chief data scientist
at our AI solutions group.
So, why don't we get started with our questions.
I'm going to ask each of the panelists to talk,
introduce themselves, as well as
talk about how they got started with AI.
So why don't we start with John?
>> Sure, so can you hear me okay in the back?
Can you hear?
Okay, cool.
So, I am a recovering evolutionary biologist
and a recovering physician
and a recovering geek.
And I implemented the health record system
for the first and largest region
of Kaiser Permanente.
And it's pretty obvious that most of the useful data
in a health record, in lies in free text.
So I started up a natural language processing team
to be able to mine free text about a dozen years ago.
So we can do things with that that you can't otherwise get
out of health information.
I'll give you an example.
I read an article online from
the New England Journal of Medicine
about four years ago that said
over half of all people who have had their spleen taken out
were not properly vaccinated for a common form of pneumonia,
and when your spleen's missing,
you must have that vaccine or you die a very sudden death
with sepsis.
In fact, our medical director in Northern California's
father died of that exact same scenario.
So, when I read the article,
I went to my structured data analytics team
and to my natural language processing team
and said please show me everybody who
has had their spleen taken out and hasn't been
appropriately vaccinated
and we ran through about 20 million records
in about three hours with the NLP team,
and it took about three weeks with a structured data
analytics team.
That sounds counterintuitive but
it actually happened that way.
And it's not a competition for time only.
It's a competition for quality
and sensitivity and specificity.
So we were able to indentify all of our members
who had their spleen taken out,
who should've had a pneumococcal vaccine.
We vaccinated them and there are a number of people
alive today who otherwise would've died
absent that capability.
So people don't really commonly associate
natural language processing with machine learning,
but in fact, natural language processing
relies heavily and is the first really,
highly successful example of machine learning.
So we've done dozens of similar projects,
mining free text data in millions of records
very efficiently, very effectively.
But it really helped advance the quality of care
and reduce the cost of care.
It's a natural step forward to go into the world of
personalized medicine with the arrival of
a 100-dollar genome, which is actually what it costs today
to do a full genome sequence.
Microbiomics, that is the ecosystem of bacteria
that are in every organ of the body actually.
And we know now that there is a profound influence
of what's in our gut and how we metabolize drugs,
what diseases we get.
You can tell in a five year old,
whether or not they were born by a vaginal delivery
or a C-section delivery
by virtue of the bacteria in the gut
five years later.
So if you look at the complexity of the data that exists
in the genome, in the microbiome,
in the health record with free text
and you look at all the other sources of data
like this streaming data from my wearable monitor
that I'm part of a research study
on Precision Medicine out of Stanford,
there is a vast amount of disparate data,
not to mention all the imaging,
that really can collectively produce
much more useful information to advance our
understanding of science, and to advance our understanding
of every individual.
And then we can do the mash up
of a much broader range of science in health care
with a much deeper sense of data from an individual
and to do that with structured
questions and structured data
is very yesterday.
The only way we're going to be able to disambiguate
those data and be able to operate on those data
in concert and generate real useful answers
from the broad array of data types
and the massive quantity of data,
is to let loose machine learning
on all of those data substrates.
So my team is moving down that pathway
and we're very excited about the future prospects
for doing that.
>> Yeah, great.
I think that's actually some of the things
I'm very excited about in the future
with some of the technologies we're developing.
My background, I started actually being fascinated with
computation in biological forms when I was nine.
Reading and watching sci-fi, I was kind of a big dork
which I pretty much still am.
I haven't really changed a whole lot.
Just basically seeing that machines really aren't
all that different from biological entities, right?
We are biological machines and kind of
understanding how a computer works
and how we engineer those things and
trying to pull together concepts that learn from biology
into that has always been a fascination of mine.
As an undergrad, I was in the EE, CS world.
Even then, I did some research projects around that.
I worked in the industry for about 10 years
designing chips, microprocessors,
various kinds of ASICs,
and then actually went back to school,
quit my job, got a Ph.D. in neuroscience,
computational neuroscience,
to specifically understand what's the state of the art.
What do we really understand about the brain?
And are there concepts that we can take and bring back?
Inspiration's always been we want to...
We watch birds fly around.
We want to figure out how to make something that flies.
We extract those principles, and then build a plane.
Don't necessarily want to build a bird.
And so Nervana's really was
the combination of all those experiences,
bringing it together.
Trying to push computation in a new a direction.
Now, as part of Intel, we can really add
a lot of fuel to that fire.
I'm super excited to be part of Intel
in that the technologies that we were developing
can really proliferate and be applied to health care,
can be applied to Internet, can be applied
to every facet of our lives.
And some of the examples that John mentioned
are extremely exciting right now
and these are things we can do today.
And the generality of these solutions
are just really going to hit every part of health care.
I mean from a personal viewpoint,
my whole family are MDs.
I'm sort of the black sheep of the family.
I don't have an MD.
And it's always been kind of funny to me that
knowledge is concentrated in a few individuals.
Like you have a rare tumor or something like that,
you need the guy who knows how to read this MRI.
Why?
Why is it like that?
Can't we encapsulate that knowledge into a computer
or into an algorithm, and democratize it.
And the reason we couldn't do it
is we just didn't know how.
And now we're really getting to a point where
we know how to do that.
And so I want that capability to go to everybody.
It'll bring the cost of healthcare down.
It'll make all of us healthier.
That affects everything about our society.
So that's really what's exciting about it to me.
>> That's great.
So, as you heard, I'm Bob Rogers.
I'm chief data scientist for analytics
and artificial intelligence solutions at Intel.
My mission is to put powerful analytics
in the hands of every decision maker
and when I think about Precision Medicine,
decision makers are not just doctors and surgeons
and nurses, but they're also case managers
and care coordinators and probably most of all, patients.
So the mission is really to put
powerful analytics and AI capabilities in the hands of
everyone in health care.
It's a very complex world and we need tools
to help us navigate it.
So my background, I started with a Ph.D. in physics
and I was computer modeling stuff,
falling into super massive black holes.
And there's a lot of applications for that
in the real world.
No, I'm kidding. (laughter)
>> John: There will be, I'm sure.
Yeah, one of these days.
Soon as we have time travel.
Okay so, I actually, about 1991, I was working on
my post doctoral research, and I heard about
neural networks, these things that could compute
the way the brain computes.
And so, I started doing some research on that.
I wrote some papers
and actually, it was an interesting story.
The problem that we solved that got me really excited
about neural networks, which have become deep learning,
my office mate would come in.
He was this young guy who was about to
go off to grad school.
He'd come in every morning.
"I hate my project."
Finally, after two weeks, what's your project?
What's the problem?
It turns out he had to circle these little fuzzy
spots on these images from a telescope.
So they were looking for the interesting things
in a sky survey,
and he had to circle them and write down their coordinates
all summer.
Anyone want to volunteer to do that?
No?
Yeah, he was very unhappy.
So we took the first two weeks
of data that he created
doing his work by hand,
and we trained an artificial neural network
to do his summer project
and finished it in about eight hours of computing.
(crowd laughs)
And so he was like yeah, this is amazing.
I'm so happy.
And we wrote a paper.
I was the first author of course,
because I was the senior guy
at age 24.
And he was second author.
His first paper ever.
He was very, very excited.
So we have to fast forward about 20 years.
His name popped up on the Internet.
And so it caught my attention.
He had just won the Nobel Prize in physics.
(laughter)
So that's where artificial intelligence will get you.
(laughter)
So thanks Naveen.
Fast forwarding, I also developed some time series
forecasting capabilities that allowed me to create
a hedge fund that I ran for 12 years.
After that, I got into health care, which really is
the center of my passion.
Applying health care to figuring out how to get
all the data from all those siloed sources,
put it into the cloud in a secure way,
and analyze it so you can actually understand
those cases that John was just talking about.
How do you know that that person
had had a splenectomy and that they needed
to get that pneumovax?
You need to be able to search all the data,
so we used AI, natural language processing,
machine learning, to do that
and then two years ago,
I was lucky enough to join Intel
and, in the intervening time, people like Naveen
actually thawed the AI winter
and we're really in a spring of
amazing opportunities with AI,
not just in health care but everywhere,
but of course, the health care applications
are incredibly life saving and empowering so,
excited to be here on this stage with you guys.
>> I just want to cue off of your comment
about the role of physics in AI and health care.
So the field of microbiomics that I referred to earlier,
bacteria in our gut.
There's more bacteria in our gut
than there are cells in our body.
There's 100 times more DNA in that bacteria
than there is in the human genome.
And we're now discovering a couple hundred
species of bacteria a year
that have never been identified under a microscope
just by their DNA.
So it turns out the person who really catapulted
the study and the science of microbiomics forward
was an astrophysicist who did his Ph.D.
in Steven Hawking's lab on the collision of black holes
and then subsequently, put the other team
in a virtual reality,
and he developed the first super computing center
and so how did he get an interest in microbiomics?
He has the capacity to do high performance computing
and the kind of advanced analytics
that are required to look at a 100 times the volume
of 3.2 billion base pairs of the human genome
that are represented in the bacteria in our gut,
and that has unleashed the whole science of microbiomics,
which is going to really turn a lot of our assumptions
of health and health care upside down.
>> That's great, I mean, that's really transformational.
So a lot of data.
So I just wanted to let the audience know
that we want to make this an interactive session,
so I'll be asking for questions in a little bit,
but I will start off with one question
so that you can think about it.
So I wanted to ask you, it looks like you've been thinking
a lot about AI over the years.
And I wanted to understand, even though
AI's just really starting in health care,
what are some of the new trends
or the changes that you've seen
in the last few years that'll impact
how AI's being used going forward?
>> So I'll start off.
There was a paper published by
a guy by the name of Tegmark
at Harvard last summer that, for the first time,
explained why neural networks are
efficient beyond any mathematical model we predict.
And the title of the paper's fun.
It's called Deep Learning Versus Cheap Learning.
So there were two sort of punchlines of the paper.
One is is that the reason that mathematics
doesn't explain the efficiency of neural networks
is because there's a higher order
of mathematics called physics.
And the physics of the underlying data structures
determined how efficient you could mine those data
using machine learning tools.
Much more so than any mathematical modeling.
And so the second thing that was a reel from that paper
is that the substrate of the data that you're operating on
and the natural physics of those data
have inherent levels of complexity
that determine whether or not
a 12th layer of neural net will
get you where you want to go really fast,
because when you do the modeling,
for those math geeks in the audience, a factorial.
So if there's 12 layers, there's 12 factorial
permutations of different ways you could sequence
the learning through those data.
When you have 140 layers of a neural net,
it's a much, much, much bigger number of permutations
and so you end up being hardware-bound.
And so, what Max Tegmark basically said
is you can determine whether to do
deep learning or cheap learning based upon
the underlying physics of the data substrates
you're operating on
and have a good insight into how to optimize
your hardware and software approach to that problem.
>> So another way to put that is that
neural networks represent the world in the way
the world is sort of built.
>> Exactly.
>> It's kind of hierarchical.
It's funny because, sort of in retrospect,
like oh yeah, that kind of makes sense.
But when you're thinking about it mathematically,
we're like well, anything...
The way a neural can represent any mathematical function,
therfore, it's fully general.
And that's the way we used to look at it, right?
So now we're saying, well actually decomposing the world
into different types of features
that are layered upon each other
is actually a much more efficient,
compact representation of the world, right?
I think this is actually, precisely the point
of kind of what you're getting at.
What's really exciting now is that
what we were doing before was sort of building
these bespoke solutions for different kinds of data.
NLP, natural language processing.
There's a whole field, 25 plus years
of people devoted to figuring out features,
figuring out what structures
make sense in this particular context.
Those didn't carry over at all to computer vision.
Didn't carry over at all to time series analysis.
Now, with neural networks,
we've seen it at Nervana, and now part of Intel,
solving customers' problems.
We apply a very similar set of techniques
across all these different types of data domains
and solve them.
All data in the real world seems to be hierarchical.
You can decompose it into this hierarchy.
And it works really well.
Our brains are actually general structures.
As a neuroscientist, you can look at different parts
of your brain and there are differences.
Something that takes in visual information,
versus auditory information is slightly different
but they're much more similar than they are different.
So there is something invariant,
something very common between all of these
different modalities and we're starting to learn that.
And this is extremely exciting to me
trying to understand the biological machine
that is a computer, right?
We're figurig it out, right?
>> One of the really fun things that Ray Chrisfall
likes to talk about is,
and it falls in the genre of biomimmicry,
and how we actually replicate
biologic evolution in our technical solutions
so if you look at, and we're beginning to understand
more and more how real neural nets work in our
cerebral cortex.
And it's sort of a pyramid structure
so that the first pass of a broad base of analytics,
it gets constrained to the next pass,
gets constrained to the next pass,
which is how information is processed in the brain.
So we're discovering increasingly
that what we've been evolving towards,
in term of architectures of neural nets,
is approximating the architecture of the human cortex
and the more we understand the human cortex,
the more insight we get to how to optimize neural nets,
so when you think about it,
with millions of years of evolution
of how the cortex is structured,
it shouldn't be a surprise that the optimization protocols,
if you will, in our genetic code
are profoundly efficient in how they operate.
So there's a real role for looking at biologic
evolutionary solutions, vis a vis technical solutions,
and there's a friend of mine who worked with
who worked with George Church at Harvard
and actually published a book
on biomimmicry and they wrote the book completely in DNA
so if all of you have your home DNA decoder,
you can actually read the book on your DNA reader,
just kidding.
>> There's actually a start up I just saw in the--
>> Read-Write DNA, yeah.
>> Actually it's a...
He writes something.
What was it?
(response from crowd member)
Yeah, they're basically encoding
information in DNA as a storage medium.
(laughter)
The company, right?
>> Yeah, that same friend of mine who coauthored
that biomimmicry book in DNA
also did the estimate of the density of information storage.
So a cubic centimeter of DNA
can store an hexabyte of data.
I mean that's mind blowing.
>> Naveen: Highly done soon.
>> Yeah that's amazing.
Also you hit upon a really important point there,
that one of the things that's changed is...
Well, there are two major things that have changed
in my perception from let's say five to 10 years ago,
when we were using machine learning.
You could use data to train models and make predictions
to understand complex phenomena.
But they had limited utility and the challenge was that
if I'm trying to build on these things,
I had to do a lot of work up front.
It was called feature engineering.
I had to do a lot of work to figure out
what are the key attributes of that data?
What are the 10 or 20 or 100 pieces
of information that I should pull out of the data
to feed to the model,
and then the model can turn it into
a predictive machine.
And so, what's really exciting about the new generation
of machine learning technology,
and particularly deep learning,
is that it can actually learn from example data
those features without you having to do any preprogramming.
That's why Naveen is saying you can take the same
sort of overall approach
and apply it to a bunch of different problems.
Because you're not having to fine tune those features.
So at the end of the day,
the two things that have changed to really enable
this evolution is access to more data,
and I'd be curious to hear from you
where you're seeing data come from,
what are the strategies around that.
So access to data, and I'm talking millions of examples.
So 10,000 examples most times isn't going to cut it.
But millions of examples will do it.
And then, the other piece is the computing capability
to actually take millions of examples
and optimize this algorithm in a single lifetime.
I mean, back in '91, when I started,
we literally would have thousands of examples
and it would take overnight to run the thing.
So now in the world of millions,
and you're putting together all of these combinations,
the computing has changed a lot.
I know you've made some revolutionary advances in that.
But I'm curious about the data.
Where are you seeing interesting
sources of data for analytics?
>> So I do some work in the genomics space
and there are more viable permutations of the human genome
than there are people who have ever walked
the face of the earth.
And the polygenic determination of
a phenotypic expression translation,
what are genome does to us
in our physical experience in health and disease
is determined by many, many genes
and the interaction of many, many genes
and how they are up and down regulated.
And the complexity of disambiguating
which 27 genes are affecting your diabetes
and how are they up and down regulated
by different interventions is going to
be different than his.
It's going to be different than his.
And we already know that there's four or five
distinct genetic subtypes of type II diabetes.
So physicians still think there's one disease
called type II diabetes.
There's actually at least four or five genetic variants
that have been identified.
And so, when you start thinking about disambiguating,
particularly when we don't know what 95 percent
of DNA does still,
what actually is the underlining cause,
it will require this massive capability of
developing these feature vectors,
sometimes intuiting it, if you will,
from the data itself.
And other times, taking what's known knowledge
to develop some of those feature vectors,
and be able to really understand
the interaction of the genome and the microbiome
and the phenotypic data.
So the complexity is high and because
the variation complexity is high,
you do need these massive members.
Now I'm going to make a very personal pitch here.
So forgive me, but if any of you have any role in policy
at all, let me tell you what's happening right now.
The Genomic Information Nondiscrimination Act,
so called GINA,
written by a friend of mine, passed a number of years ago,
says that no one can be discriminated against
for health insurance based upon their genomic information.
That's cool.
That should allow all of you to feel comfortable
donating your DNA to science right?
Wrong.
You are 100% unprotected from discrimination
for life insurance, long term care and disability.
And it's being practiced legally today
and there's legislation in the House,
in mark up right now to completely undermine
the existing GINA legislation and say that
whenever there's another applicable statute
like HIPAA, that the GINA is irrelevant,
that none of the fines and penalties are applicable at all.
So we need a ton of data to be able to operate on.
We will not be getting a ton of data to operate on
until we have the kind of protection we need
to tell people, you can trust us.
You can give us your data,
you will not be subject to discrimination.
And that is not the case today.
And it's being further undermined.
So I want to make a plea to any of you
that have any policy influence
to go after that because we need this data
to help the understanding of human health and disease
and we're not going to get it when people
look behind the curtain and see that discrimination
is occurring today based upon genetic information.
>> Well, I don't like the idea of being discriminated against
based on my DNA.
Especially given how little we actually know.
There's so much complexity
in how these things unfold in our own bodies,
that I think anything that's being done
is probably childishly immature and oversimplifying.
So it's pretty rough.
>> I guess the translation here is that we're all unique.
It's not just a Disney movie.
(laughter)
We really are.
And I think one of the strengths that I'm seeing,
kind of going back to the original point,
of these new techniques is it's going across
different data types.
It will actually allow us to learn more about
the uniqueness of the individual.
It's not going to be just from one data source.
They were collecting data from many different modalities.
We're collecting behavioral data from wearables.
We're collecting things from scans, from blood tests,
from genome, from many different sources.
The ability to integrate those into a unified picture,
that's the important thing that we're getting toward now.
That's what I think is going to be super exciting here.
Think about it, right.
I can tell you to visual a coin, right?
You can visualize a coin.
Not only do you visualize it.
You also know what it feels like.
You know how heavy it is.
You have a mental model of that
from many different perspectives.
And if I take away one of those senses,
you can still identify the coin, right?
If I tell you to put your hand in your pocket,
and pick out a coin, you probably can do that
with 100% reliability.
And that's because we have this generalized capability
to build a model of something in the world.
And that's what we need to do for individuals
is actually take all these different data sources
and come up with a model for an individual
and you can actually then say what drug works best on this.
What treatment works best on this?
It's going to get better with time.
It's not going to be perfect,
because this is what a doctor does, right?
A doctor who's very experienced,
you're a practicing physician right?
Back me up here.
That's what you're doing.
You basically have some categories.
You're taking information from the patient
when you talk with them,
and you're building a mental model.
And you apply what you know can work on that patient, right?
>> I don't have clinic hours anymore, but I do take care of
many friends and family.
(laughter) >> You used to, you used to.
>> I practiced for many years
before I became a full-time geek.
>> I thought you were a recovering geek.
>> I am. (laughter)
I do more policy now.
>> He's off the wagon.
>> I just want to take a moment and see
if there's anyone from the audience
who would like to ask, oh.
Go ahead.
>> We've got a mic here, hang on one second.
>> I have tons and tons of questions.
(crosstalk)
Yes, so first of all, the microbiome and the genome
are really complex.
You already hit about that.
Yet most of the studies we do are small scale
and we have difficulty repeating them
from study to study.
How are we going to reconcile all that
and what are some of the technical hurdles
to get to the vision that you want?
>> So primarily, it's been the cost of sequencing.
Up until a year ago, it's $1000, true cost.
Now it's $100, true cost.
And so that barrier is going to enable
fairly pervasive testing.
It's not a real competitive market
becaue there's one sequencer that is way ahead
of everybody else.
So the price is not $100 yet.
The cost is below $100.
So as soon as there's competition to drive the cost down,
and hopefully, as soon as we all have the protection
we need against discrimination, as I mentioned earlier,
then we will have large enough sample sizes.
And so, it is our expectation that we will be able to
pool data from local sources.
I chair the e-health work group at the
Global Alliance for Genomics and Health
which is working on this very issue.
And rather than pooling all the data
into a single, common repository,
the strategy, and we're developing our five-year plan
in a month in London,
but the goal is to have a federation
of essentially credentialed data enclaves.
That's a formal method.
HHS already does that so you can get credentialed
to search all the data that Medicare has
on people that's been deidentified according to HIPPA.
So we want to provide the same kind of service
with appropriate consent, at an international scale.
And there's a lot of nations that
are talking very much about data nationality
so that you can't export data.
So this approach of a federated model
to get at data from all the countries is important.
The other thing is a block-chain technology
is going to be very profoundly useful in this context.
So David Haussler of UC Santa Cruz
is right now working on a protocol
using an open block-chain, public ledger,
where you can put out.
So for any typical cancer,
you may have a half dozen, what are called sematic variance.
Cancer is a genetic disease
so what has mutated to cause it to behave like a cancer?
And if we look at those biologically active
sematic variants, publish them on a block chain
that's public, so there's not enough data
there to reidentify the patient.
But if I'm a physician treating a woman with breast cancer,
rather than say what's the protocol for treating
a 50-year-old woman with this cell type of cancer,
I can say show me all the people in the world
who have had this cancer at the age of 50,
wit these exact six sematic variants.
Find the 200 people worldwide with that.
Ask them for consent through a secondary mechanism
to donate everything about their medical record,
pool that information of the core of 200 that
exactly resembles the one sitting in front of me,
and find out, of the 200 ways they were treated,
what got the best results.
And so, that's the kind of future where
a distributed, federated architecture
will allow us to query and obtain a very, very relevant
cohort, so we can basically be treating patients
like mine, sitting right in front of me.
Same thing applies for establishing research cohorts.
There's some very exciting stuff at the convergence
of big data analytics, machine learning,
and block chaining.
>> And this is an area that I'm really excited about
and I think we're excited about generally at Intel.
They actually have something called
the Collaborative Cancer Cloud,
which is this kind of federated model.
We have three different academic research centers.
Each of them has a very sizable and valuable
collection of genomic data
with phenotypic annotations.
So you know, pancreatic cancer,
colon cancer, et cetera,
and we've actually built a secure computing architecture
that can allow a person who's given the right permissions
by those organizations to ask a specific question
of specific data without ever sharing the data.
So the idea is my data's really important to me.
It's valuable.
I want us to be able to do a study
that gets the number from the 20 pancreatic cancer patients
in my cohort, up to the 80 that we have in the whole group.
But I can't do that if I'm going to just spill my data
all over the world.
And there are HIPAA and compliance reasons for that.
There are business reasons for that.
So what we've built at Intel is this platform
that allows you to do different kinds of queries
on this genetic data.
And reach out to these different sources
without sharing it.
And then, the work that I'm really involved in right now
and that I'm extremely excited about...
This also touches on something that both of you said is
it's not sufficient to just get the genome sequences.
You also have to have the phenotypic data.
You have to know what cancer they've had.
You have to know that they've been treated
with this drug and they've survived for three months
or that they had this side effect.
That clinical data also needs to be put together.
It's owned by other organizations, right?
Other hospitals.
So the broader generalization of
the Collaborative Cancer Cloud
is something we call the data exchange.
And it's a misnomer in a sense that we're not actually
exchanging data.
We're doing analytics on aggregated data sets
without sharing it.
But it really opens up a world where we can have
huge populations and big enough amounts of data
to actually train these models and draw the thread in.
Of course, that really then hits home for
the techniques that Nervana is bringing to the table,
and of course--
>> Stanford's one of your academic medical centers?
>> Not for that Collaborative Cancer Cloud.
>> The reason I mentioned Standford is because
the reason I'm wearing this FitBit
is because I'm a research subject
at Mike Snyder's, the chair of genetics at Stanford,
IPOP, intrapersonal omics profile.
So I was fully sequenced five years ago
and I get four full microbiomes.
My gut, my mouth, my nose,
my ears.
Every three months and I've done that for four years now.
And about a pint of blood.
And so, to your question of the density of data,
so a lot of the problem with applying these techniques
to health care data is that it's basically a sparse matrix
and there's a lot of discontinuities
in what you can find and operate on.
So what Mike is doing with the IPOP study
is much the same as you described.
Creating a highly dense longitudinal set of data
that will help us mitigate the sparse matrix problem.
(low volume response from audience member)
Pardon me. >> What's that?
(low volume response)
(laughter) >> Right, okay.
>> John: Lost the school sample.
That's got to be a new one I've heard now.
>> Okay, well, thank you so much.
That was a great question.
So I'm going to repeat this and ask
if there's another question.
You want to go ahead?
>> Hi, thanks.
So I'm a journalist and I report a lot
on these neural networks, a system that's beter
at reading mammograms than your human radiologists.
Or a system that's better at predicting
which patients in the ICU will get sepsis.
These sort of fascinating academic studies
that I don't really see being translated very quickly
into actual hospitals or clinical practice.
Seems like a lot of the problems are regulatory,
or liability, or human factors,
but how do you get past that
and really make this stuff practical?
>> I think there's a few things that we can do there
and I think the proof points of the technology
are really important to start with
in this specific space.
In other places, sometimes, you can start with other things.
But here, there's a real confidence problem
when it comes to health care,
and for good reason.
We have doctors trained for many, many years.
School and then residencies and other kinds of training.
Because we are really, really conservative with health care.
So we need to make sure that technology's well beyond
just the paper, right?
These papers are proof points.
They get people interested.
They even fuel entire grant cycles sometimes.
And that's what we need to happen.
It's just an inherent problem, its' going to take a while.
To get those things to a point where it's like
well, I really do trust what this is saying.
And I really think it's okay to now start integrating that
into our standard of care.
I think that's where you're seeing it.
It's frustrating for all of us, believe me.
I mean, like I said, I think personally
one of the biggest things, I want to have an impact.
Like when I go to my grave,
is that we used machine learning to improve health care.
We really do feel that way.
But it's just not something we can do very quickly
and as a business person, I don't actually look at
those use cases right away
because I know the cycle is just going to be longer.
>> So to your point, the FDA, for about four years now,
has understood that the process
that has been given to them by their board of directors,
otherwise known as Congress, is broken.
And so they've been very actively seeking
new models of regulation
and what's really forcing their hand is regulation
of devices and software
because, in many cases, there are black box aspects of that
and there's a black box aspect to machine learning.
Historically, Intel and others are making inroads
into providing some sort of traceability
and transparency into what happens in that black box
rather than say, overall we get better results
but once in a while we kill somebody.
Right?
So there is progress being made on that front.
And there's a concept that I like to use.
Everyone knows Ray Kurzweil's book The Singularity Is Near?
Well, I like to think that diadarity is near.
And the diadarity is where you have
human transparency into what goes on in the black box
and so maybe Bob, you want to speak a little bit about...
You mentioned that, in a prior discussion,
that there's some work going on at Intel there.
>> Yeah, absolutely.
So we're working with a number of groups
to really build tools that allow us...
In fact Naveen probably can talk in even more detail
than I can,
but there are tools that allow us
to actually interrogate machine learning
and deep learning systems to understand,
not only how they respond to a wide variety of situations
but also where are there biases?
I mean, one of the things that's shocking is that
if you look at the clinical studies
that our drug safety rules are based on,
50 year old white guys are the peak of that distribution,
which I don't see any problem with that,
but some of you out there might not like that
if you're taking a drug.
So yeah, we want to understand
what are the biases in the data, right?
And so, there's some new technologies.
There's actually some very interesting
data-generative technologies.
And this is something I'm also curious what Naveen
has to say about,
that you can generate from small sets of observed data,
much broader sets of varied data that help probe
and fill in your training for some of these systems
that are very data dependent.
So that takes us to a place where we're going to
start to see deep learning systems
generating data to train other deep learning systems.
And they start to sort of go back and forth
and you start to have some very nice
ways to, at least, expose the weakness
of these underlying technologies.
>> And that feeds back to your question about regulatory
oversight of this.
And there's the fascinating, but little known origin
of why very few women are in clinical studies.
Thalidomide causes birth defects.
So rather than say pregnant women can't be enrolled
in drug trials,
they said any woman who is at risk
of getting pregnant cannot be enrolled.
So there was actually a scientific meritorious argument
back in the day when they really didn't know
what was going to happen post-thalidomide.
So it turns out that the adverse, unintended consequence
of that decision was we don't have data on women
and we know in certain drugs, like Xanax,
that the metabolism is so much slower,
that the typical dosing of Xanax is women
should be less than half of that for men.
And a lot of women have had very serious adverse
effects by virtue of the fact that they weren't studied.
So the point I want to illustrate with that
is that regulatory cycles...
So people have known for a long time
that was like a bad way of doing regulations.
It should be changed.
It's only recently getting changed in any meaningful way.
So regulatory cycles and legislative cycles
are incredibly slow.
The rate of exponential growth in technology is exponential.
And so there's impedance mismatch between the cycle time
for regulation cycle time for innovation.
And what we need to do...
I'm working with the FDA.
I've done four workshops with them on this very issue.
Is that they recognize that they need to
completely revitalize their process.
They're very interested in doing it.
They're not resisting it.
People think, oh, they're bad, the FDA, they're resisting.
Trust me, there's nobody on the planet
who wants to revise these review processes
more than the FDA itself.
And so they're looking at models
and what I recommended is global cloud sourcing
and the FDA could shift from a regulatory role
to one of doing two things,
assuring the people who do their reviews are competent,
and assuring that their conflicts of interest
are managed,
because if you don't have a conflict of interest
in this very interconnected space,
you probably don't know enough to be a reviewer.
So there has to be a way to manage
the conflict of interest
and I think those are some of the keypoints
that the FDA is wrestling with
because there's type one and type two errors.
If you underregulate, you end up with another thalidomide
and people born without fingers.
If you overregulate, you prevent life saving drugs
from coming to market.
So striking that balance across all these
different technologies is extraordinarily difficult.
If it were easy, the FDA would've done it four years ago.
It's very complicated.
>> Jumping on that question,
so all three of you are in some ways entrepreneurs, right?
Within your organization or started companies.
And I think it would be good to talk a little bit
about the business opportunity here,
where there's a huge ecosystem in health care,
different segments, biotech,
pharma, insurance payers, etc.
Where do you see is the ripe opportunity
or industry, ready to really take this on
and to make AI the competitive advantage.
>> Well, the last question also included
why aren't you using the result of the sepsis detection?
We do.
There were six or seven published ways of doing it.
We did our own data, looked at it,
we found a way that was superior
to all the published methods
and we apply that today,
so we are actually using that technology
to change clinical outcomes.
As far as where the opportunities are...
So it's interesting.
Because if you look at what's going to be here in three years,
we're not going to be using those big data
analytics models for sepsis that we are deploying today,
because we're just going to be getting a tiny aliquot of blood,
looking for the DNA or RNA of any potential infection
and we won't have to infer that there's
a bacterial infection from all these other
ancillary, secondary phenomenon.
We'll see if the DNA's in the blood.
So things are changing so fast
that the opportunities that people need to look for
are what are generalizable and sustainable kind of wins
that are going to lead to a revenue cycle that are justified,
a venture capital world investing.
So there's a lot of interesting opportunities in the space.
But I think some of the biggest opportunities
relate to what Bob has talked about
in bringing many different disparate
data sources together and really looking for things
that are not comprehensible in the human brain
or in traditional analytic models.
>> I think we also got to look a little bit beyond
direct care.
We're talking about policy
and how we set up standards, these kinds of things.
That's one area.
That's going to drive innovation forward.
I completely agree with that.
Direct care is one piece.
How do we scale out many of the knowledge kinds of things
that are embedded into one person's head
and get them out to the world, democratize that.
Then there's also development.
The underlying technology's of medicine, right?
Pharmaceuticals.
The traditional way that pharmaceuticals is developed
is actually kind of funny, right?
A lot of it was started just by chance.
Penicillin, a very famous story right?
It's not that different today unfortunately, right?
It's conceptually very similar.
Now we've got more science behind it.
We talk about domains and interactions,
these kinds of things
but fundamentally, the problem is
what we in computer science
called NP hard, it's too difficult to model.
You can't solve it analytically.
And this is true for all these kinds of
natural sorts of problems by the way.
And so there's a whole field around this,
molecular dynamics and modeling these sorts of things,
that are actually being driven forward
by these AI techniques.
Because it turns out, our brain doesn't do magic.
It actually doesn't solve these problems.
It approximates them very well.
And experience allows you to
approximate them better and better.
Actually, it goes a little bit
to what you were saying before.
It's like simulations and forming your own networks
and training off each other.
There are these emerging dynamics.
You can simulate steps of physics.
And you come up with a system that's
much too complicated to ever solve.
Three pool balls on a table is one such system.
It seems pretty simple.
You know how to model that, but it actual turns out
you can't predict where a balls going to be
once you inject some energy into that table.
So something that simple is already too complex.
So neural network techniques actually allow us to start
making those tractable.
These NP hard problems.
And things like molecular dynamics
and actually understanding how different medications
and genetics will interact with each other
is something we're seeing today.
And so I think there's a huge opportunity there.
We've actually worked with customers in this space.
And I'm seeing it.
Like Rosch is acquiring a few different companies in space.
They really want to drive it forward,
using big data to drive drug development.
It's kind of counterintuitive.
I never would've thought it had I not seen it myself.
>> And there's a big related challenge.
Because in personalized medicine,
there's smaller and smaller cohorts of people
who will benefit from a drug that still takes
two billion dollars on average to develop.
That is unsustainable.
So there's an economic imperative
of overcoming the cost and the cycle time
for drug development.
>> I want to take a go at this question
a little bit differently,
thinking about not so much where are the industry segments
that can benefit from AI, but what are the kinds of
applications that I think are most impactful.
So if this is what a skilled surgeon needs to know
at a particular time to care properly for a patient,
this is where most, this area here,
is where most surgeons are.
They are close to the maximum knowledge
and ability to assimilate
as they can be.
So it's possible to build complex AI
that can pick up on that one little thing
and move them up to here.
But it's not a gigantic accelerator,
amplifier of their capability.
But think about other actors in health care.
I mentioned a couple of them earlier.
Who do you think the least trained actor in health care is?
>> John: Patients.
>> Yes, the patients.
The patients are really very poorly trained,
including me.
I'm abysmal at figuring out who to call and where to go.
>> Naveen: You know as much the doctor right?
(laughing)
>> Yeah, that's right.
>> My doctor friends always hate that.
Know your diagnosis, right?
>> Yeah, Dr. Google knows.
So the opportunities that I see that are really, really
exciting are when you take an AI agent,
like sometimes I like to call it contextually
intelligent agent, or a CIA,
and apply it to a problem where
a patient has a complex future ahead of them
that they need help navigating.
And you use the AI to help them work through.
Post operative.
You've got PT.
You've got drugs.
You've got to be looking for side effects.
An agent can actually help you navigate.
It's like your own personal GPS for health care.
So it's giving you the inforamation that you need
about you for your care.
That's my definition of Precision Medicine.
And it can include genomics, of course.
But it's much bigger.
It's that broader picture and I think
that a sort of agent way of thinking about things
and filling in the gaps where there's less training
and more opportunity,
is very exciting.
>> Great start up idea right there by the way.
>> Oh yes, right.
We'll meet you all out back for the next start up.
>> I had a conversation with the head of the
American Association of Medical Specialties
just a couple of days ago.
And what she was saying,
and I'm aware of this phenomenon,
but all of the medical specialists are saying,
you're killing us with these stupid board recertification
trivia tests that you're giving us.
So if you're a cardiologist, you have to remember something
that happens in one in 10 million people, right?
And they're saying that irrelevant anymore,
because we've got advanced decision support coming.
We have these kinds of analytics coming.
Precisely what you're saying.
So it's human augmentation
of decision support that is coming
at blazing speed towards health care.
So in that context,
it's much more important that you have a basic foundation,
you know how to think,
you know how to learn,
and you know where to look.
So we're going to be human-augmented
learning systems much more so than in the past.
And so the whole recertification process is being
revised right now.
(inaudible audience member speaking)
Speak up, yeah.
(person speaking)
>> What makes it fathomable is that
you can--
(audience member interjects inaudibly)
>> Sure.
She was saying that our brain is really
complex and large and even our brains don't know how
our brains work, so...
are there ways to--
>> What hope do we have kind of thing?
(laughter)
>> It's a metaphysical question.
>> It circles all the way down, exactly.
It's a great quote.
I mean basically, you can decompose every system.
Every complicated system can be decomposed
into simpler, emergent properties.
You lose something perhaps with each of those,
but you get enough to actually understand
most of the behavior.
And that's really how we understand the world.
And that's what we've learned in the last few years
what neural network techniques can allow us to do.
And that's why our brain can understand our brain.
(laughing)
>> Yeah, I'd recommend reading Chris Farley's last book
because he addresses that issue in there
very elegantly.
>> Yeah we're seeing some really interesting technologies
emerging right now where neural network systems
are actually connecting other neural network systems
in networks.
You can see some very compelling behavior
because one of the things I like to distinguish
AI versus traditional analytics
is we used to have question-answering systems.
I used to query a database and create a report
to find out how many widgets I sold.
Then I started using regression or machine learning
to classify complex situations
from this is one of these and that's one of those.
And then as we've moved more recently,
we've got these AI-like capabilities
like being able to recognize that
there's a kitty in the photograph.
But if you think about it,
if I were to show you a photograph
that happened to have a cat in it,
and I said, what's the answer,
you'd look at me like, what are you talking about?
I have to know the question.
So where we're cresting with these connected sets
of neural systems, and with AI in general,
is that the systems are starting to be able to,
from the context,
understand what the question is.
Why would I be asking about this picture?
I'm a marketing guy, and I'm curious about
what Legos are in the thing or what kind of cat it is.
So it's being able to ask a question,
and then take these question-answering systems,
and actually apply them so that's this ability
to understand context and ask questions
that we're starting to see emerge from
these more complex hierarchical neural systems.
>> There's a person dying to ask a question.
>> Sorry.
You have hit on several different topics
that all coalesce together.
You mentioned personalized models.
You mentioned AI agents that could help you
as you're going through a transitionary period.
You mentioned data sources,
especially across long time periods.
Who today has access to enough data
to make meaningful progress on that,
not just when you're dealing with an issue,
but day-to-day improvement of your life and your health?
>> Go ahead, great question.
>> That was a great question.
And I don't think we have a good answer to it.
(laughter)
I'm sure John does.
Well, I think every large healthcare organization
and various healthcare consortiums
are working very hard to achieve that goal.
The problem remains
in creating semantic interoperatability.
So I spent a lot of my career working
on semantic interoperatability.
And the problem is
that if you don't have well-defined,
or self-defined data,
and if you don't have well-defined and documented metadata,
and you start operating on it,
it's real easy to reach false conclusions
and I can give you a classic example.
It's well known, with hundreds of studies looking at
when you give an antibiotic before surgery
and how effective it is in preventing a post-op infection.
Simple question, right?
So most of the literature done prosectively
was done in institutions where they had small sample sizes.
So if you pool that,
you get a little bit more noise,
but you get a more confirming answer.
What was done at a very large,
not my own, but a very large institution...
I won't name them for obvious reasons,
but they pooled lots of data from
lots of different hospitals,
where the data definitions and the metadata were different.
Two examples.
When did they indicate the antibiotic was given?
Was it when it was ordered, dispensed from the pharmacy,
delivered to the floor,
brought to the bedside,
put in the IV,
or the IV starts flowing?
Different hospitals used a different metric
of when it started.
When did surgery occur?
When they were wheeled into the OR,
when they were prepped and drapped,
when the first incision occurred?
All different.
And they concluded quite dramatically
that it didn't matter when you gave the pre-op antibiotic
and whether or not you get a post-op infection.
And everybody who was intimate
with the prior studies just completely ignored
and discounted that study.
It was wrong.
And it was wrong because of the lack of commonality
and the normalization of data definitions
and metadata definitions.
So because of that,
this problem is much more challenging than you would think.
If it were so easy as to put all these data together
and operate on it, normalize and operate on it,
we would've done that a long time ago.
It's...
Semantic interoperatability remains a big problem
and we have a lot of heavy lifting ahead of us.
I'm working with the Global Alliance, for example,
of Genomics and Health.
There's like 30 different major ontologies
for how you represent genetic information.
And different institutions are using different ones
in different ways in different versions
over different periods of time.
That's a mess.
>> Our all those issues applicable
when you're talking about a personalized data set
versus a population?
>> Well, so N of 1 studies and single-subject research
is an emerging field of statistics.
So there's some really interesting new models
like step wedge analytics
for doing that on small sample sizes,
recruiting people asynchronously.
There's single-subject research statistics.
You compare yourself with yourself
at a different point in time, in a different context.
So there are emerging statistics to do that
and as long as you use the same sensor,
you won't have a problem.
But people are changing their remote sensors
and you're getting different data.
It's measured in different ways with different sensors
at different normalization and different calibration.
So yes.
It even persists in the N of 1 environment.
>> Yeah, you have to get started with a large N
that you can apply to the N of 1.
I'm actually going to attack your question
from a different perspective.
So who has the data?
The millions of examples to train
a deep learning system from scratch.
It's a very limited set right now.
Technology such as the Collaborative Cancer Cloud
and The Data Exchange are definitely impacting that
and creating larger and larger sets of critical mass.
And again, not withstanding the very challenging
semantic interoperability questions.
But there's another opportunity
Kay asked about what's changed recently.
One of the things that's changed in deep learning
is that we now have modules that have been trained
on massive data sets
that are actually very smart
as certain kinds of problems.
So, for instance, you can go online
and find deep learning systems
that actually can recognize,
better than humans,
whether there's a cat, dog, motorcycle, house,
in a photograph.
>> From Intel, open source.
>> Yes, from Intel, open source.
So here's what happens next.
Because most of that deep learning system
is very expressive.
That combinatorial mixture of features that
Naveen was talking about,
when you have all these layers,
there's a lot of features there.
They're actually very general to images,
not just finding cats, dogs, trees.
So what happens is you can do something called
transfer learning, where you take a small or modest data set
and actually reoptimize it for your specific problem
very, very quickly.
And so we're starting to see
a place where you can...
On one end of the spectrum,
we're getting access to the computing capabilities
and the data to build these
incredibly expressive deep learning systems.
And over here on the right,
we're able to start using those deep learning systems
to solve custom versions of problems.
Just last weekend or two weekends ago,
in 20 minutes, I was able to take one of those
general systems and create one that could
recognize all different kinds of flowers.
Very subtle distinctions, that I would never be able to know
on my own.
But I happen to be able to get the data set
and literally, it took 20 minutes
and I have this vision system that I could now use
for a specific problem.
I think that's incredibly profound
and I think we're going to see this
spectrum of wherever you are
in your ability to get data and to define problems
and to put hardware in place
to see really neat customizations
and a proliferation of applications
of this kind of technology.
>> So one other trend I think, I'm very hopeful about it...
So this is a hard problem clearly, right?
I mean, getting data together, formatting it
from many different sources,
it's one of these things that's
probably never going to happen perfectly.
But one trend I think that is extremely hopeful to me is
the fact that the cost of gathering data
has precipitously dropped.
Building that thing is almost free these days.
I can write software and put it on 100 million cell phones
in an instance.
You couldn't do that five years ago even right?
And so, the amount of information we can gain
from a cell phone today has gone up.
We have more sensors.
We're bringing online more sensors.
People have Apple Watches and they're sending
blood data back to the phone,
so once we can actually start gathering more data
and do it cheaper and cheaper,
it actually doesn't matter where the data is.
I can write my own app.
I can gather that data
and I can start driving the correct inferences
or useful inferences back to you.
So that is a positive trend I think here
and personally, I think that's how we're going to solve it,
is by gathering from that many different sources cheaply.
>> Hi, my name is Pete.
I've very much enjoyed the conversation so far
but I was hoping perhaps to bring a little bit more focus
into Precision Medicine and ask two questions.
Number one, how have you applied
the AI technologies as you're emerging so rapidly
to your natural language processing?
I'm particularly interested in,
if you look at things like Amazon Echo or Siri,
or the other voice recognition systems
that are based on AI,
they've just become incredibly accurate
and I'm interested in specifics
about how I might use technology like that in medicine.
So where would I find a medical nomenclature
and perhaps some reference to
a back end that works that way?
And the second thing is, what specifically is Intel doing,
or making available?
You mentioned some open source stuff
on cats and dogs and stuff
but I'm the doc, so I'm looking at the medical side of that.
What are you guys providing that would allow us
who are kind of geeks on the software side,
as well as being docs,
to experiment a little bit more thoroughly
with AI technology?
Google has a free AI toolkit.
Several other people have come out with
free AI toolkits in order to accelerate that.
There's special hardware now with graphics,
and different processors,
hitting amazing speeds.
And so I was wondering, where do I go in Intel
to find some of those tools and perhaps learn
a bit about the fantastic work
that you guys are already doing at Kaiser?
>> Let me take that first part
and then we'll be able to talk about the MD part.
So in terms of technology,
this is what's extremely exciting now
about what Intel is focusing on.
We're providing those pieces.
So you can actually assemble and build the application.
How you build that application specific for MDs
and the use cases is up to you
or the one who's filling out the application.
But we're going to power that technology
for multiple perspectives.
So Intel is already the main force
behind The Data Center, right?
Cloud computing, all this is already Intel.
We're making that extremely amenable to AI
and setting the standard for AI in the future,
so we can do that from a number of different mechanisms.
For somebody who wants to develop an application quickly,
we have hosted solutions.
Intel Nervana is kind of the brand
for these kinds of things.
Hosted solutions will get you going very quickly.
Once you get to a certain level of scale,
where costs start making more sense,
things can be bought on premise.
We're supplying that.
We're also supplying software that makes
that transition essentially free.
Then taking those solutions that you develop in the cloud,
or develop in The Data Center,
and actually deploying them on device.
You want to write something on your smartphone
or PC or whatever.
We're actually providing those hooks as well,
so we want to make it very easy for developers
to take these pieces and actually
build solutions out of them quickly
so you probably don't even care
what hardware it's running on.
You're like here's my data set,
this is what I want to do.
Train it, make it work.
Go fast.
Make my developers efficient.
That's all you care about, right?
And that's what we're doing.
We're taking it from that point at how do we best do that?
We're going to provide those technologies.
In the next couple of years,
there's going to be a lot of new stuff
coming from Intel.
>> Do you want to talk about AI Academy as well?
>> Yeah, that's a great segway there.
In addition to this, we have an entire set of
tutorials and other online resources
and things we're going to be bringing into the academic world
for people to get going quickly.
So that's not just enabling them on our tools,
but also just general concepts.
What is a neural network?
How does it work?
How does it train?
All of these things are available now
and we've made a nice, digestible class format
that you can actually go and play with.
>> Let me give a couple of quick answers
in addition to the great answers already.
So you're asking why can't we use medical terminology
and do what Alexa does?
Well, no, you may not be aware of this,
but Andrew Ian, who was the AI guy at Google,
who was recruited by Google,
they have a medical chat bot in China today.
I don't speak Chinese.
I haven't been able to use it yet.
There are two similar initiatives in this country
that I know of.
There's probably a dozen more in stealth mode.
But Lumiata and Health Cap are doing chat bots
for health care today, using medical terminology.
You have the compound problem of semantic normalization
within language, compounded by a cross language.
I've done a lot of work with an international organization
called Snowmed, which translates medical terminology.
So you're aware of that.
We can talk offline if you want,
because I'm pretty deep into the semantic space.
>> Go google Intel Nervana and you'll see
all the websites there.
It's intel.com/ai or nervanasys.com.
>> Okay, great.
Well this has been fantastic.
I want to, first of all, thank all the people here
for coming and asking great questions.
I also want to thank our fantastic panelists today.
(applause) >> Thanks, everyone.
>> Thank you.
>> And lastly, I just want to share one bit of information.
We will have more discussions on AI
next Tuesday at 9:30 AM.
Diane Bryant, who is our general manager
of Data Centers Group
will be here to do a keynote.
So I hope you all get to join that.
Thanks for coming.
(applause)
(light electronic music)