00:00:04
my name is Jessica Mega I'm a
00:00:07
cardiologist at Stanford I'm on the
00:00:10
board of directors at danaher
00:00:11
Corporation and one of the co-founders
00:00:12
of verily which is the branch of
00:00:15
alphabet really focused on life science
00:00:17
and Healthcare
00:00:18
and personally my passion has always
00:00:21
been thinking about the intersection of
00:00:23
Technology life science and the
00:00:25
applications to improve patient outcomes
00:00:27
and I think we're in a moment that
00:00:28
Harkens back in my mind to when the
00:00:31
human genome was sequenced and if you
00:00:33
think about all of the Decades of
00:00:35
research and all of the deep insights
00:00:38
that it took to sequence the human
00:00:40
genome there were many many people many
00:00:43
tools but it was then the next several
00:00:45
decades of applications the
00:00:47
understanding of what genes do we
00:00:49
actually have to measure when we think
00:00:50
about oncology and Diagnostics what does
00:00:53
it translate now we're applying direct
00:00:55
crispr therapies but we're in that same
00:00:57
moment of excitement where there have
00:00:59
been Decades of technical advancements
00:01:01
and we have some of the world's experts
00:01:03
with us today to talk about that but now
00:01:05
we're in that critical moment of
00:01:07
application and we heard from Nigam that
00:01:09
there's a Chasm and I think collectively
00:01:11
this group together can help bridge that
00:01:13
Chasm and I think take all of the
00:01:16
opportunity but also all the reality of
00:01:18
what it takes to truly create a business
00:01:20
in this world
00:01:22
as I am going to turn to our group we're
00:01:24
going to have each individual introduce
00:01:26
themselves the company that they're
00:01:27
working with and any high level thoughts
00:01:29
and then what we're going to do is I
00:01:31
have a few questions prepared but I know
00:01:33
that your questions are on your mind
00:01:35
because people have already grabbed me
00:01:37
in the hallway and started questions so
00:01:38
uh start to keep those in your mind as
00:01:41
we get the introductions because I
00:01:42
promise they'll be time for uh for all
00:01:44
of your questions all right if uh if you
00:01:46
don't mind started
00:01:48
hi everyone my name is Rohit takar I
00:01:51
head the technology at carpool so
00:01:53
carpool is a platform that allows you to
00:01:55
explore validate and deploy Medical
00:01:57
Imaging AI Solutions in your clinical
00:01:59
workflow I think we all believe that AI
00:02:04
how can improve the patient outcomes
00:02:06
stream streamline the processes and can
00:02:09
help reduce the cost but what are the
00:02:11
challenges that we are facing right now
00:02:12
so I believe that there are two
00:02:14
challenges uh that I feel ah that are
00:02:17
actually handling the deployment of AI
00:02:20
uh in clinic clinical workflow one is
00:02:23
the implementation so how do you present
00:02:26
these solutions to the end users which
00:02:28
in our case are Radiologists so ah there
00:02:31
could be multiple solution because one
00:02:33
fits for all ah it doesn't happen the
00:02:36
other is once you have deployed a
00:02:38
solution so how do we monitor it over
00:02:40
time so there could be model drift data
00:02:42
drift and there would be some issue with
00:02:45
the pipeline so these are the two
00:02:47
challenges that I feel uh which ah we
00:02:50
need to figure out together as a
00:02:52
collaborative effort and see how can we
00:02:55
build Solutions create proof points and
00:02:58
then deploy them in uh on a wider ah at
00:03:02
a clinical level so that's what I feel
00:03:06
could be something that we can do in
00:03:08
2023
00:03:10
thank you
00:03:12
hi I'm ten viewers either Mahmoud I'm an
00:03:14
IBM fellow at IBM research IBM research
00:03:17
as you know is the academic wing of IBM
00:03:20
where we our Charter is to explore the
00:03:22
Next Generation Technologies in various
00:03:24
fields in healthcare in particular as
00:03:27
you know we were
00:03:28
you know ahead of the curve in trying to
00:03:31
realize the importance of AI and
00:03:32
radiology and some of you have been to
00:03:35
RSA 2016 or 2015 onwards you've seen the
00:03:39
introduction of AI in Radiology it
00:03:42
transitioned into business so we've gone
00:03:43
through the cycle of you know the hype
00:03:47
and practicing putting it in
00:03:50
products and then coming back to the
00:03:52
realization of what is the next things
00:03:54
to do and so I will have some
00:03:56
perspectives on now what works what
00:03:58
doesn't work and where are the
00:04:00
challenges particularly when it comes to
00:04:02
looking at capturing fine-grained
00:04:05
nuances which is where the AI needs to
00:04:08
get to in Next Generation models
00:04:12
okay my name is cu I'm a healthcare
00:04:15
investor at F Prime which is a venture
00:04:18
capital group mostly based in Boston SF
00:04:20
and London but before that I'm a
00:04:23
Stanford Medical School alumni and also
00:04:25
did research at the Stanford machine
00:04:27
learning grouping AI for healthcare as
00:04:28
well and I think right now we're looking
00:04:30
at a lot of early stage opportunities
00:04:32
with people with ideas on twofolds one
00:04:36
is actually similar to the previous
00:04:38
panel talk about speeding up a lot of
00:04:41
the processes that we encounter in the
00:04:43
clinic you know managing the back office
00:04:45
for clinicians and whatnot and the other
00:04:47
side is actually a step forward to to do
00:04:51
things that clinicians cannot do right
00:04:53
now and that's an area that we also look
00:04:56
at currently too
00:04:59
so good morning
00:05:01
good morning
00:05:03
good morning
00:05:04
you know what I said that to weigh me up
00:05:06
because I just fly from the other side
00:05:08
of the Earth from Vietnam to here so so
00:05:12
you guys just remember a Vietnamese guy
00:05:14
on the stage so that's me so
00:05:16
so I'm Stephen Drew I'm the guy for big
00:05:21
brain
00:05:22
so Wing brain is a company around four
00:05:25
years and you know one day I told my
00:05:29
boss a secretive in Microsoft say I'm
00:05:32
not going to do any more Microsoft stuff
00:05:35
I want to do something
00:05:37
a little bit more Humanity so I decided
00:05:41
to get into this field to build AI
00:05:44
to assist grade watches
00:05:47
on the workflow
00:05:49
and improving the efficiency like most
00:05:52
people said and also the by through
00:05:55
spending and also through the
00:05:58
misdiagnosis in cancer area
00:06:02
and then also we do the social impact we
00:06:06
just like dealing with tuberculosis so I
00:06:09
went to MetLife who you know a few
00:06:12
months ago in Geneva tried to tackle
00:06:15
that problem in other side of the world
00:06:18
so basically Wing brain is focusing on
00:06:22
religious workflow we have 52 you know
00:06:26
comprehensive detection
00:06:29
and at the same time we realized that
00:06:32
not just AI we need to focusing on the
00:06:36
Enterprise side which is where we
00:06:39
connect all of the data together because
00:06:41
you know that hospital have risk lists
00:06:43
his love of law so they vary this very
00:06:48
so we bring the connector to bring the
00:06:50
data together and we can train
00:06:53
Ai and we also one of the guys that
00:06:57
interray the activity and all of that
00:07:00
into the workflow so we can improve the
00:07:02
reporting as well and the data
00:07:05
extraction from
00:07:07
La language model for the medical field
00:07:10
so with this I'm really looking for a
00:07:14
lot of questions about like you know on
00:07:17
how we really deployed to a hundreds
00:07:19
Hospital you know the way we built is
00:07:22
two year we never do anything else
00:07:24
except just helped out to be a decent
00:07:27
product and the third year we go from 10
00:07:29
to 20 to 100 hospital and today we
00:07:33
affect around 2 million people live
00:07:35
through our Pipeline and we have a FDA
00:07:40
you know one of the only one in the
00:07:42
region of Southeast got that and as well
00:07:45
we assign agreement with Microsoft
00:07:48
Nvidia and also stand for so great to
00:07:53
work with curl every two weeks we see
00:07:55
his face on life so thank you very much
00:07:59
and great to see all of you again
00:08:02
hi everyone my name is Nish I'm the CEO
00:08:05
of Bunker Hill previously the starting
00:08:07
Bunker Hill I was a student at Stanford
00:08:09
I was a graduate researcher at the
00:08:11
Stanford Amy group actually and Dr
00:08:14
nigamsha mentioned that there were many
00:08:15
algorithms that are being built I'm
00:08:16
pretty sure my algorithm was one of
00:08:18
those in the tally there and you know
00:08:21
after we built those algorithms we
00:08:23
really stumbled stumbled hard on the
00:08:25
question what now
00:08:27
okay you know what the steps are you
00:08:29
know you need to externally validate
00:08:31
your algorithm you know you need to show
00:08:32
that it has some kind of clinical Roi
00:08:34
you know you need FDA clearance you know
00:08:36
you need to deploy but for all of those
00:08:38
phases for all of those questions you
00:08:41
had two pressing problems one I don't
00:08:43
even know what it takes to get there too
00:08:46
I don't even know where to start and we
00:08:48
see this as a pressing problem you know
00:08:50
the only way algorithms are hopefully
00:08:53
going to get out into the clinical
00:08:54
practice if you're able to take
00:08:56
researchers algorithms and actually have
00:08:58
a way to get them into practice the
00:09:01
current solution of starting a company
00:09:03
around a single algorithm and hoping to
00:09:05
get Venture funding is just not working
00:09:07
and so what we do at Bunker Hill is that
00:09:11
we are creating a way to get algorithms
00:09:14
built by the broader research Community
00:09:16
out into the clinical world we have
00:09:18
built a Consortium of academic medical
00:09:20
centers which is effectively a legal
00:09:22
bubble there are 17 academic medical
00:09:24
centers a part of this today and what
00:09:26
this allows them to do is researchers at
00:09:28
each other researchers can collaborate
00:09:30
across institutions share data share
00:09:32
algorithms validate each other's
00:09:34
algorithms and when they are validated
00:09:36
Bunker Hill files for FDA clearance
00:09:38
commercializes it on a platform and then
00:09:40
shares the revenue back with the
00:09:41
researcher hopefully that creates a way
00:09:43
to actually get all of the one wonderful
00:09:46
research work that is being done by such
00:09:48
daring innovators and actually getting
00:09:50
them into clinical practice and making
00:09:52
AI actually a reality
00:09:56
hi my name is tanishq Abraham and I'm a
00:09:59
researcher at stability Ai and
00:10:02
previously I was also a PhD student at
00:10:06
UC Davis and just recently completed and
00:10:10
yeah at stability AI I work on medical
00:10:13
AI research in fact we founded an
00:10:15
organization known as medarc to help
00:10:17
accelerate some of this medical AI
00:10:20
research especially looking at the
00:10:22
intersection of generative Ai and
00:10:24
medicine so of course you know we've
00:10:26
heard a lot about that already today but
00:10:28
yeah we're really excited about about
00:10:30
what we can how we can apply a
00:10:32
generative AI to Medicine especially at
00:10:34
a company like stability which is known
00:10:35
for their efforts in that space with you
00:10:38
know the famous stable diffusion image
00:10:40
generator and you know there are very
00:10:42
surprising and interesting applications
00:10:44
of of generative Ai and you know you
00:10:46
never think like something like an image
00:10:48
generator you know what kind of
00:10:49
applications it could have in medicine
00:10:51
but it turns out you know there's some
00:10:53
really interesting applications and I
00:10:56
think also one of the really exciting
00:10:58
things about working in medical AI
00:10:59
research is also the exchange just yeah
00:11:01
generally the exchange of ideas uh in
00:11:04
just the general AI field but also and
00:11:07
with you know medical AI I I think um
00:11:09
you know of course there's lots of
00:11:10
applications of various AI Technologies
00:11:14
in medicine but also lots of interesting
00:11:16
applications of methods that were
00:11:19
originally developed in medical Ai and
00:11:21
bringing them to kind of the general AI
00:11:23
uh uh Fields so I mean we've seen some
00:11:26
of these examples in the past for
00:11:27
example the unit was originally
00:11:29
developed for you know biomedical
00:11:31
segmentation even clip which was you
00:11:34
know which was developed by open AI it
00:11:36
was kind of a scaled up version of a
00:11:38
paper that was originally published by
00:11:40
Stanford Amy actually so there's lots of
00:11:42
interesting applications in general AI
00:11:44
that kind of start out in some of the
00:11:48
medical AI research so I think that sort
00:11:50
of collaboration that sort of exchange
00:11:52
of ideas is really exciting and I think
00:11:55
yeah the sort of development in parallel
00:11:57
of ideas in general AI generative AI
00:12:00
medical AI That's what I'm most excited
00:12:02
about
00:12:03
excellent and congratulations
00:12:05
so we were talking about the business
00:12:07
models how do we actually create
00:12:10
sustainable businesses and if you think
00:12:12
about outside of the healthcare space a
00:12:15
lot of the AI is embedded in tools where
00:12:17
there's already an ad model or there's a
00:12:20
freemium model where you start but you
00:12:22
pay for additional services or to direct
00:12:25
to Consumer so you if we can start with
00:12:27
you what are some of the business models
00:12:29
that you're seeing that are effective in
00:12:31
healthcare because in many cases it's a
00:12:33
B2B right so you're an industry selling
00:12:35
to someone else who is paying for health
00:12:37
care who are ultimately serving the
00:12:40
customer our constituents patients and
00:12:42
clinicians so it'd be good to just hear
00:12:44
a little bit about how you evaluated
00:12:46
businesses and then we'll hear from a
00:12:48
few of the companies what business
00:12:49
models are working and potentially if
00:12:51
any of them haven't worked yeah
00:12:53
definitely I think that's an era that
00:12:55
I'm also learning as well so I would
00:12:57
love to hear other people's ideas too so
00:12:59
I think we see a couple of different
00:13:01
ways people try to get paid and I was
00:13:03
previously mentioned by other people as
00:13:05
well on the provider side either as
00:13:07
hospitals or clinicians the margin is
00:13:10
quite thin right like there's not as
00:13:12
much of Liberty to just pay for hundreds
00:13:15
of thousands of dollars of software as
00:13:17
what other companies can pay for even
00:13:20
just one single user so one way we've
00:13:23
seen people trying to get around that is
00:13:25
to not just create clinical value from
00:13:27
their product but also direct monetary
00:13:31
value from their AI product to for
00:13:34
example to get it reimbursed from the
00:13:37
payer side for providers so that
00:13:39
provider is not just using it to get out
00:13:42
of their own pocket but also actually
00:13:44
create hopefully Financial value
00:13:46
directly for them as well another way
00:13:49
that we see this being monetized is
00:13:51
actually to go beyond the direction to
00:13:54
patient with back to clinicia angle but
00:13:57
also to tap into the broader Healthcare
00:13:59
ecosystem people who tap into the
00:14:02
healthcare data market which is huge and
00:14:04
you know companies pay hundreds of
00:14:06
Millions for for that in some ways and
00:14:08
also other sides of the healthcare
00:14:11
ecosystem as well which sometimes can
00:14:13
have more Liberty in trying out and
00:14:16
paying for these type of AI Solutions
00:14:18
and from some of the company's
00:14:20
perspectives
00:14:21
what are the models that you're using I
00:14:23
can take a stab at it
00:14:25
um I think their business model it
00:14:28
implies money and I think that's where
00:14:30
AI just reduces down to any other
00:14:32
medical intervention you just have to
00:14:34
follow the money so if you look at who
00:14:36
the healthcare stakeholders are okay so
00:14:38
you have providers the health systems
00:14:40
you have Pharma and then you have payers
00:14:42
on the provider side what can you do to
00:14:44
make a financial case for them okay
00:14:46
could you find them more patients could
00:14:47
you reduce or automate away mundane or
00:14:50
boring or expensive tasks okay so if you
00:14:52
can do that there's a business incentive
00:14:55
there's a financial incentive for the
00:14:57
providers to adopt it and actually pay
00:14:58
for it on the on the Pharma side can you
00:15:01
accelerate their clinical trials by
00:15:03
finding patients faster if you can do
00:15:05
that there's a business case for them
00:15:06
there's money to be saved by them can
00:15:08
you find more patients who would be
00:15:10
eligible for drugs that they just came
00:15:12
out with okay now that's again you're
00:15:13
increasing the market that's a business
00:15:15
model on the insurance side on the peer
00:15:18
side can you reduce their cost of care
00:15:20
within two years okay that's another
00:15:21
incentive there's a there's a financial
00:15:23
uh play for them so again in this case
00:15:26
AI just reduces down to any other
00:15:28
medical intervention it does not really
00:15:29
matter what technology is supporting
00:15:32
that intervention ultimately it's who is
00:15:34
being who financially benefits from the
00:15:37
widespread deployment of that tool so
00:15:39
Stephen yes so I think uh I go back to
00:15:43
the paying the pain point of hospital
00:15:45
and the operation of the hospital so as
00:15:50
you know that in my point of view the
00:15:52
hospital uh you know I.T is uh probably
00:15:57
not as strong at the order vertical
00:16:00
market like you know e-commerce supply
00:16:03
chain manufacturing so we come down to
00:16:06
not only using the AI to increasing the
00:16:10
efficiency
00:16:12
misdiagnosis you know showing the TB
00:16:15
problem of the world we also look
00:16:17
looking at how to really solving the
00:16:19
pain of data display and data platform
00:16:23
and combined with the AI we can connect
00:16:26
on the data and generate the analytic
00:16:29
and forecasting for the operational as
00:16:32
well as for the doctor in efficiency so
00:16:35
you can think of the the model business
00:16:37
we do a software as a service because we
00:16:40
like to be more like TurnKey solution
00:16:43
comprehensive solution rather than
00:16:45
single POI solution and we tackle from
00:16:48
two Dimension One is the Enterprise that
00:16:52
we can connect all of the display data
00:16:54
into the center line with the identifier
00:16:57
model the second is the AI to bring the
00:17:00
value of efficiency and the third
00:17:03
dimension we do is the monetization of
00:17:05
AI so we view a platform to do the
00:17:08
labeling annotation for all of the
00:17:11
company Wai because we know Wing brain
00:17:14
cannot do everything so that's why we
00:17:16
bring you know the monetization AI not
00:17:20
only from labeling and speak up with
00:17:22
hundreds of pre-latches it's very hard
00:17:24
in the U.S to get 100 languages to do
00:17:27
the labeling you know of course we have
00:17:29
to define the standard of how to do the
00:17:32
label correctly and then we're doing
00:17:34
that so we build those platforms not
00:17:37
only for solving the doctor and the
00:17:40
workflow also we're doing with the
00:17:42
labeling and then the third dimension we
00:17:45
also want to make sure that like you
00:17:47
know sharing the data to the communities
00:17:50
in the public matter so that we can
00:17:53
solve the problem together and
00:17:55
Marketplace our platform any other
00:17:58
thoughts maybe I can give the other end
00:18:00
of the spectrum perspective from a large
00:18:02
company and why we got out of the
00:18:04
healthcare business
00:18:05
so um
00:18:07
you know the the kind of Revenue that
00:18:09
makes sense uh for IBM a billion dollar
00:18:12
revenue is still not good enough so so
00:18:14
this field and and just by AI alone uh
00:18:19
surviving is more for the smaller
00:18:22
businesses that's why you saw so many
00:18:24
startups in this space but as the
00:18:25
starters have also discovered just
00:18:28
having an AI model alone is not enough
00:18:30
it is only a small piece of the big
00:18:32
puzzle that you have to get into so if
00:18:35
you think of what are some good models
00:18:36
for Success that where AI will still
00:18:38
succeed and where we see a future for it
00:18:41
will be in a few different deployments
00:18:43
situations one and I remember there's
00:18:45
Intel Inside there's like I think terms
00:18:48
that you see on your laptops uh it
00:18:51
should be something like that where it's
00:18:53
the other thing that in which AI is
00:18:56
built into that is giving the value
00:18:58
proposition whether that is Hardware in
00:19:01
which you have ai doing this naturally
00:19:02
or is in the packs where it's already
00:19:04
built in and it's not being charged
00:19:06
separately so the way you win in this
00:19:09
might be through those vehicles and then
00:19:12
having some attribution back to aisl but
00:19:15
AI by itself
00:19:16
is not something that people will be
00:19:19
willing to pay big bucks for and that
00:19:21
was one of the learnings as such the
00:19:24
other thing I could say is and and I
00:19:26
mean if you've heard of the think and
00:19:28
what's the next announcement recently
00:19:29
and all we are big into AI models still
00:19:31
but the way we're looking off is the
00:19:34
value proposition is using those for
00:19:36
some use cases and so the AI is again
00:19:39
built inside and it's not the focus but
00:19:42
it's the end user use case that is the
00:19:44
focus and there there is opportunities
00:19:46
in Consulting in you know in those sort
00:19:49
of methods of deployment so bundle
00:19:52
inside is one way and I think we can
00:19:55
still make a business case for AI
00:19:59
so I think ah
00:20:02
the biggest question that we get is who
00:20:04
will pay for the implementation of AI
00:20:07
right so whether the radiology
00:20:09
department the hospital or will you cut
00:20:11
it from the radiologist pocket so that
00:20:13
these are some of the questions that we
00:20:14
get
00:20:15
uh I think if you can do an Implement a
00:20:18
implementation where we can really
00:20:20
augment the Radiologists so if let's say
00:20:22
they are doing 100 scans and they are
00:20:24
paid X dollars for that
00:20:26
at some point they can do more than 100
00:20:28
scans and by implementation of AI let us
00:20:33
say so AI Solutions you can get for 30
00:20:35
cents 20 cents for 40 cents per study so
00:20:39
if you can uh give them a value that is
00:20:41
X Plus Delta so I think that would be
00:20:44
something that can be that can come from
00:20:46
the radiologist pocket as well so those
00:20:48
are the business cases that we are
00:20:49
actually trying right now the other is
00:20:51
how can you get more patients so ah from
00:20:54
your x-rays to CD scans can you convert
00:20:57
more patients which need
00:20:59
CT scans to be done so I mean that could
00:21:03
be the other way
00:21:05
businesses wants to ah see the value
00:21:09
coming in yeah so it sounds like there
00:21:11
are two major themes that are coming
00:21:13
together one is this idea if you think
00:21:15
about life science and Healthcare there
00:21:16
are already models out there so if you
00:21:18
can accelerate drug Discovery if you can
00:21:20
increase efficiency and value leverage
00:21:24
those models and the second one that can
00:21:26
be really emphasized is that the AI is
00:21:29
the one piece surrounded by all of the
00:21:31
services and just think holistically I
00:21:33
think about the offering so I think
00:21:35
those are just good reminders I mean I
00:21:37
did want to add that it is possible if
00:21:40
AI could simply be that a reason for a
00:21:43
business if AI had that capability right
00:21:45
now the current models can only do touch
00:21:49
some portion of this like think about
00:21:50
chest x-ray work there's still like 4 14
00:21:52
14 findings I mean what about the rest
00:21:55
and you know so if I have to remove so
00:21:58
so you have to raise the capability of
00:22:00
AI and make it comprehensive to replace
00:22:03
or to augment in a way that is uh
00:22:06
TurnKey so I think until that time
00:22:09
you'll still have that and people don't
00:22:11
have the patience to do that you know
00:22:13
just as we ran the medical Civ Grand
00:22:15
Challenge for six years we took us a lot
00:22:18
of effort to cover all the findings in
00:22:20
chest x-rays lots of clinician
00:22:22
involvement lots of data sets so such
00:22:25
resources should be there you know order
00:22:27
to get to those levels and and that's
00:22:29
just for one modality and you know so
00:22:32
you think about that so that's the level
00:22:34
at which AI should be working if we want
00:22:37
to have that kind of impact but if
00:22:39
you're only doing a little bit of this
00:22:40
or a little bit of that then you'll just
00:22:43
be a tool that people may want to turn
00:22:45
you off so then you don't have a way to
00:22:48
get Revenue yeah it's interesting right
00:22:49
because the Health Care system is
00:22:51
already so fragmented that it's a
00:22:53
reminder to us that if we come in with
00:22:55
one point solution when I put my
00:22:57
clinical hat on I'm thinking how does
00:22:59
that integrate into the Haiku app that
00:23:01
sits in my pocket so you're right it's
00:23:04
very hard to think in fact the best
00:23:06
thing we could do for patients and
00:23:08
clinicians is to bring down the
00:23:09
fragmentation there's another debate as
00:23:12
we were just talking about business
00:23:13
models and how open or closed do we stay
00:23:16
with these businesses open source
00:23:19
becomes there's just a lot you can't
00:23:21
open up the New York Times The Wall
00:23:23
Street Journal without seeing major
00:23:24
companies taking different stances and
00:23:26
and certainly a lot of the the up and
00:23:28
coming companies
00:23:30
um do you mind giving us a little bit of
00:23:31
your perspective on the role of Open
00:23:33
Source in the public-private
00:23:35
Partnerships and actual deployment yeah
00:23:38
I think open source is extremely
00:23:39
powerful for you know accelerating
00:23:41
research and
00:23:43
um yeah there's a lot of opportunities
00:23:45
in that in that sector but also
00:23:48
um the kind of the approach that a lot
00:23:50
of these open source companies including
00:23:52
stability uh tend to take is developing
00:23:55
a kind of General open source models
00:23:58
that can be publicly released but then
00:24:00
utilizing their expertise to fine-tune
00:24:02
these models for specific use cases so I
00:24:05
think this may also have a potential
00:24:07
um this is also potentially something
00:24:09
that can be done in the healthcare
00:24:10
sector where you can have research
00:24:12
models that are developed but also
00:24:14
fine-tuned models for specific clinical
00:24:16
data sets from hospitals you know
00:24:18
specific hospitals situations like that
00:24:20
and I think that will be a really
00:24:21
interesting model as well
00:24:23
um for applying AI to healthcare yeah
00:24:26
excellent any other comments on open
00:24:29
source so I think that besides what the
00:24:33
gentleman here said I think the open saw
00:24:35
is the way to promote
00:24:37
the demoretization of AI and technology
00:24:40
so I think as a company like you know
00:24:44
doing Health Tech we believe in that it
00:24:47
beside that we also have to look at
00:24:50
responsible AI responsible software when
00:24:53
you utilize open source so so like
00:24:57
example we we have to have a board
00:24:59
director with the policy make you know
00:25:02
to make sure that we follow it and we
00:25:04
make sure they identify data and all the
00:25:07
validation of the data point that have
00:25:09
no data leak for you know like personal
00:25:12
information so that is the fundamental
00:25:15
when we view our pipeline for for data
00:25:18
thank you so I'm going to ask one last
00:25:21
question then we're opening it up and
00:25:23
then you are the team will keep us on
00:25:25
time so we've heard a lot this morning
00:25:28
about what it takes to have a algorithm
00:25:31
or product but what it really means to
00:25:33
engage with the regulatory bodies
00:25:34
whether it's the FDA emea
00:25:36
it would be helpful to hear some
00:25:38
practical examples how often are you
00:25:40
refreshing your algorithms for example
00:25:42
how have you navigated the quality
00:25:45
management systems and Regulatory
00:25:47
environment if we could just have maybe
00:25:48
one or two examples with that I think it
00:25:50
brings it home and in fact I think Nishi
00:25:54
said that that was one of your passions
00:25:55
the reason why uh you you moved over to
00:25:58
create the company do you mind just
00:26:00
getting us started on some of that yeah
00:26:02
I think when we started the company we
00:26:04
wanted you know coming from the valley
00:26:06
the underlying notion was the FDA is the
00:26:08
reason why AI was not getting into
00:26:10
clinical practice it was back in 2017
00:26:12
2018 so my co-founder and Ivy flew to DC
00:26:15
we met with the FDA and they're just
00:26:18
they're the nicest people they're really
00:26:20
good at their job it's like and it's
00:26:24
their job is difficult and they've done
00:26:26
a tremendous amount of work in making it
00:26:28
very clear what the expectations are so
00:26:30
we actually just got FDA clearance for
00:26:32
the one of Stanford's algorithms this
00:26:35
was about a month ago and the process
00:26:38
was was 83 days end to end from
00:26:41
submission to clearance and what's
00:26:43
exciting is that we worked with them to
00:26:45
even come up with okay we are we know
00:26:47
that we're going to be deploying this
00:26:48
algorithm it's now live at 200 hospitals
00:26:50
already and as we deploy this we're
00:26:52
bound to find cases where like oh yeah I
00:26:55
think we could come up with improvements
00:26:56
for the algorithm how do we relay that
00:26:58
back to the researcher how does the
00:27:00
researcher work towards you know
00:27:02
defining the model with that additional
00:27:04
information and what does that look like
00:27:06
currently the pros they've actually come
00:27:08
up with something called pccp which is a
00:27:10
predetermined Change Control product
00:27:11
plan and they're really thinking about
00:27:13
this in the sense of we know that these
00:27:15
algorithms will need to be trained and
00:27:16
fine-tuned in the future and how do we
00:27:18
make it such that it's not too laborious
00:27:20
to to to to get them cleared again
00:27:26
right now there seems to be also debate
00:27:28
on how that standard may change and even
00:27:31
the boundary of what will be regulated I
00:27:33
think that's an interesting topic that
00:27:35
we keep following as well for example
00:27:37
clinical decision support like the
00:27:39
sepsis algorithm that was mentioned
00:27:41
although that's killed right now but can
00:27:43
that be regulated should that be even
00:27:45
regulated and I think that's also an
00:27:47
interesting direction to follow as well
00:27:50
so one point I want to share is we
00:27:52
cannot just rely on FDA to really doing
00:27:55
ethical and responsible because it takes
00:27:59
a lot of time and a lot of resource so
00:28:02
the way we view is we you have to
00:28:04
basically build into the culture and the
00:28:07
principle of the company where you want
00:28:10
to ensure that all the data is protected
00:28:13
by privacy and all of that because
00:28:16
that's the way to build sustainable
00:28:18
model for a company you know to go for
00:28:21
the long term and I think that that's a
00:28:24
fundamental I mean like FDA you probably
00:28:27
like I mean for us we have like 52
00:28:29
comprehensive on just x-ray that people
00:28:31
say but we cannot like apply like 52
00:28:35
probably take 10 years for us to or
00:28:37
maybe more to get on of that so I think
00:28:40
that we cannot like just rely only on
00:28:43
that but we and then beside you also
00:28:45
need to get like ISO and all of those
00:28:48
equivalent and CE things like that to
00:28:51
get further but the fundamental is a
00:28:54
principle and culture it's it's
00:28:56
interesting a lot of times quality
00:28:58
Engineers really need to be partnering
00:29:01
it has to be we had many signs about the
00:29:04
quality is everyone's uh problem or
00:29:07
everyone's purpose and we turned it into
00:29:09
this is everyone's purpose and if you
00:29:11
create really efficient Quality
00:29:12
Management Systems regulatory schema it
00:29:14
actually can be a business accelerant
00:29:16
and so one of my personal when I hear oh
00:29:18
they're all these barriers I say as you
00:29:20
said there's a reason that there are
00:29:22
systems in place to protect patients we
00:29:24
just need to help evolve them and I
00:29:26
think if anyone's looking for
00:29:27
interesting jobs I think at the
00:29:29
interface of the private public and
00:29:32
Regulatory boundary whether it's with AI
00:29:35
whether it's with now cell and
00:29:37
gene-based therapies just a little plug
00:29:39
for I think that's an interesting career
00:29:41
to choose all right so we're opening it
00:29:44
up on the regulatory aspects I think you
00:29:47
hinted on this summer as well but one of
00:29:50
the biggest barriers to scaling AI for
00:29:52
broader list of findings is you know
00:29:54
having to do this verification for one
00:29:57
finding at a time and the conducting the
00:29:59
clinical studies for those so that's why
00:30:01
most people go for a few findings and so
00:30:04
on but if you want to make the change
00:30:06
and get greater adoption and and scale
00:30:09
out we need to change the process of
00:30:12
evaluations the for such large when you
00:30:15
have large number of findings to be
00:30:17
evaluated that was one of the barriers
00:30:19
we found when we had the technology but
00:30:21
the company would not go through the
00:30:24
regulatory approvals for all of these at
00:30:27
a time because they're just too much
00:30:31
paperwork and stuff to go through even
00:30:34
if you had the capability right so I
00:30:36
think that there needs to be some
00:30:38
improvements I know FDA is working on it
00:30:39
obviously the other aspect that I think
00:30:42
people have emphasized more and where we
00:30:43
have made a bit more headways in the
00:30:45
generalization capabilities when you go
00:30:47
to new deployments and how to update
00:30:49
your models and so on there you have I
00:30:51
think a bunch of Radiologists helped
00:30:53
shape that thinking there but when it
00:30:56
comes to large scale findings and
00:30:58
coverage of an entire Anatomy or
00:31:02
modality and so on we don't have yet you
00:31:06
know policies that would
00:31:08
you know ratify such software
00:31:11
this one all right any live questions
00:31:17
the panels have been exceedingly clear
00:31:19
and we do realize that we're standing
00:31:20
between us and lunch but uh any
00:31:23
um any questions here
00:31:31
oh yeah
00:31:32
um so I would love to hear your take on
00:31:35
when you might want to categorize
00:31:39
clinical decision support versus uh
00:31:41
clinical operations support and
00:31:44
what might be the toilet challenges like
00:31:47
the difference between those
00:31:51
so I'll take this answer so I think that
00:31:53
we one we target for
00:31:56
diagnosis assistant is clinical have
00:32:00
decision support so we even care about
00:32:03
like naming the product like we call Dr
00:32:05
Ray the aid so we intend to say we
00:32:09
assist we not to replace right and we
00:32:12
believe in land losses the experiment
00:32:14
religious use AI will replace there is
00:32:18
just hold on so I guess that's uh one
00:32:21
and then the second dimension we're
00:32:23
talking about like operational where we
00:32:25
build the Enterprise dimension of how to
00:32:29
Gathering all the data together on the
00:32:32
on-prem on private cloud or public Cloud
00:32:35
depending on the environment and where
00:32:37
we build analytic tool on top of that so
00:32:41
that we can prove the operational
00:32:42
efficiency so that's the way we
00:32:45
differentiate ourselves from the rest we
00:32:48
have two dimensions of capability to
00:32:51
solve the problem and comprehensive is
00:32:53
the fundamental like I mean if you do
00:32:55
like say the liver cancer we cannot just
00:32:58
look at only liver we have to look at
00:33:01
abdomen the whole area of a city and
00:33:04
find out what's wrong with that same for
00:33:07
just actually we cannot look at
00:33:08
pneumatic we have to look at 52 or
00:33:11
something like that that's you can do
00:33:13
with big data
00:33:16
all right so I'm going to close out with
00:33:18
a question for Dinesh um
00:33:20
technically what are you most excited
00:33:22
about as you look in the next five to
00:33:24
ten years it's obviously been a huge
00:33:26
amount of change you've focused your
00:33:28
career really being on the Forefront of
00:33:31
the latest Technologies what excites you
00:33:33
the most
00:33:35
um yeah I think uh the I'm really
00:33:38
excited by I think yeah what's most
00:33:41
exciting is the development of
00:33:42
multimodal systems now and applying
00:33:45
multimodal systems to Medical AI I think
00:33:47
what we've seen so far are very unimodal
00:33:50
applications where you know you're just
00:33:51
looking at images you're just looking at
00:33:54
um you know clinical text or time series
00:33:56
or whatever so being able I mean of
00:33:58
course in the actual clinical workflow
00:34:00
you're trying to integrate all of that
00:34:01
information together in order to make
00:34:03
clinical decisions so I'm really
00:34:05
interested in seeing the development of
00:34:07
multimodal systems that are able to now
00:34:10
take in you know various different
00:34:11
sources of data and you know be able to
00:34:14
provide
00:34:15
um support to clinicians by
00:34:16
incorporating all that information so
00:34:18
I'm really excited by uh developments in
00:34:21
that field and also yeah I'm just also
00:34:23
excited by uh not only I think there's a
00:34:25
lot of opportunities for medical AI in
00:34:29
various other settings like developing
00:34:31
countries and third world countries
00:34:32
where you know the opportunity duties
00:34:35
and kind of I think yeah there's a lot
00:34:37
of opportunities in that that in those
00:34:38
sectors that are not being explored as
00:34:40
much and I think that's also a really
00:34:41
exciting uh area for applying AI to uh
00:34:45
you know help support the the doctors
00:34:48
that and you know kind of ease their
00:34:49
burden in in those areas so yeah
00:34:52
excellent well thank you to the speakers
00:34:53
for giving us technical Vision but also
00:34:56
walking us through the realities on the
00:34:57
ground whether it's with regulatory or
00:34:59
the business models and people will be
00:35:01
around if you want to ask anyone any
00:35:03
further questions so a big round of
00:35:04
applause to our speakers here
