Latest / Becker’s Healthcare Podcast / Advancing Rare Disease Diagnosis Through AI and Genomic Innovation

Transcript

Advancing Rare Disease Diagnosis Through AI and Genomic Innovation · Becker’s Healthcare Podcast

• 0:00
  Hi, everyone. This is Lucas Voss with Becker's
• 0:02
  HealthCare. Thanks so much for tuning in to
• 0:03
  the Becker's HealthCare podcast series. It's great to
• 0:05
  have you. And I'm so excited to be
• 0:07
  joined by doctor Shannon Heyman. She's the chief
• 0:09
  of pathology and laboratory medicine at Lurie Children's
• 0:12
  Hospital and vice chair of pediatric pathology at
• 0:15
  Northwestern University Feinberg School of Medicine. And I'm
• 0:18
  also joined by doctor Calum Yakubian. He's the
• 0:21
  director of health data enablement and applied AI
• 0:24
  science at IQVIA.
• 0:25
  Doctor Hammond and doctor Ekubian, thanks so much
• 0:27
  for being here today. It's great to have
• 0:28
  you.
• 0:29
  Thank you. Yeah. Thanks for having us. Absolutely.
• 0:32
  I do wanna start off with introductions here
• 0:34
  for those that might not know you yet
• 0:36
  in-depth
• 0:37
  at least.
• 0:38
  Doctor Hammond, I'll start off with you. If
• 0:39
  you could just introduce yourself and and just
• 0:41
  share a little bit about your work in
• 0:42
  health care. Sure. I'm, doctor Shannon Hammond. I've
• 0:46
  been at Lurie Children's Hospital for almost eighteen
• 0:48
  years now and at Northwestern,
• 0:50
  Feinberg School of Medicine. And I lead a
• 0:53
  department that's really focused on trying to transform
• 0:56
  pediatric diagnostics.
• 0:58
  And often, we're trying to fill in the
• 1:00
  gaps where,
• 1:01
  we need testing and other, tools for children
• 1:04
  that don't exist commercially.
• 1:07
  And so we're really this I'm excited about
• 1:09
  this podcast. We're gonna talk about the stuff
• 1:10
  we've been doing with AI and digital transformation
• 1:13
  in genomic testing, particularly. Yeah. So excited to
• 1:15
  hear about it. Doctor Coogan, over to you.
• 1:17
  Yeah. Thanks very much. Yeah. So, Calum, you
• 1:19
  could be in.
• 1:20
  My focus is on the application of AI
• 1:22
  and different
• 1:23
  technologies to derive value by and large from
• 1:25
  patient generated data. So most of what I
• 1:27
  do is working with EMR data.
• 1:30
  I'm at IQVIA where our mission is to
• 1:32
  improve the lives of patients everywhere, and that
• 1:34
  really applies
• 1:35
  everywhere from bench to bedside. So whether or
• 1:37
  not it's supporting with drug development or some
• 1:40
  of the exciting types of work we're doing
• 1:41
  with applications of AI that we're gonna talk
• 1:43
  to today.
• 1:44
  And I'm a medical doctor by background, and
• 1:46
  if the accent doesn't give away, I practice
• 1:48
  in The UK.
• 1:49
  But lots of the work I do is
• 1:51
  over here, in The US. So great to
• 1:53
  be here and excited to talk about this
• 1:55
  topic. Yeah. It's great to have you. I
• 1:56
  I wanna hop right in, and doctor Hammond
• 1:58
  just mentioned it in her introduction as well.
• 2:00
  We're gonna talk a little bit about genomic
• 2:01
  medicine here too. And as we get into
• 2:03
  it, I do wanna start off with rare
• 2:05
  disease
• 2:06
  diagnostic specifically, and I wanna bring up a
• 2:08
  couple of numbers for our audience here that
• 2:09
  are really crucial. Ninety five percent of rare
• 2:12
  genetic diseases have no approved treatment, and forty
• 2:15
  percent of patients
• 2:17
  don't even have a diagnosis
• 2:19
  there, which is striking. Just let those numbers
• 2:21
  settle there. For a health system leader unfamiliar
• 2:24
  with rare disease diagnostics right now, what does
• 2:26
  that diagnostic journey actually look like specifically when
• 2:29
  you're
• 2:30
  looking at the family in and of itself,
• 2:32
  which is so crucial? And what does it
• 2:33
  mean clinically when the answer takes years to
• 2:36
  find and just it's just not there? It's
• 2:38
  it's so incredible. Doctor Hammond, do you wanna
• 2:40
  start us off? Yeah. Those numbers really get
• 2:42
  at at what drives us, and I'll I'll
• 2:43
  add something by also clarifying that even though
• 2:46
  when we talk about these diseases individually, they're
• 2:49
  rare.
• 2:49
  So, you know, they impact one in two
• 2:52
  hundred thousand,
• 2:53
  individuals,
• 2:54
  but there are thousands of them. So when
• 2:55
  we think about the impact of, this problem,
• 2:58
  it's affecting hundreds of thousands of patients and
• 3:00
  their families.
• 3:01
  And, you know, when we talk about their
• 3:03
  journey, we call it an odyssey,
• 3:05
  because it is so long and so difficult.
• 3:08
  And,
• 3:10
  you know, for many patients that you know,
• 3:12
  for some, the average
• 3:13
  depending on some studies, the average is, like,
• 3:15
  five years to get a diagnosis. But as
• 3:17
  you noted, for a large percentage of families,
• 3:19
  some of them never have yet arrived at
• 3:21
  a diagnosis. And so what that means for
• 3:23
  them is that you have a family, has
• 3:25
  a child. We all have, you know, the
• 3:27
  best intentions and and high expectations for our
• 3:30
  our our children and future generations. But, you
• 3:33
  know, along the way, it becomes clear that
• 3:34
  that's maybe not what's gonna happen with them.
• 3:36
  And so they start seeking answers,
• 3:38
  and bouncing around,
• 3:39
  you know, from one specialist to the other,
• 3:41
  trying to find out what's going on. And
• 3:43
  so that's very long. It's very costly. It's
• 3:45
  very frustrating.
• 3:46
  Meanwhile, the child is not getting treatment, you
• 3:48
  know, that they may need. And so what
• 3:50
  this diagnostic odyssey means is,
• 3:53
  it's really a lot of heartbreak and difficulty
• 3:55
  for families,
• 3:56
  but, really, it means kids aren't getting treatment.
• 3:58
  And it also is impairing our ability to
• 4:00
  identify new treatments when we don't understand the
• 4:02
  underlying signatures of of these diseases. And I
• 4:04
  love that you mentioned the word odyssey because
• 4:06
  there's a difference between a journey and an
• 4:08
  odyssey. Right? There is a plenty more challenges,
• 4:10
  along the way, which is so crucial. Doctor.
• 4:12
  Kuman, just wanted to give you a chance
• 4:13
  here too what that looks like from from
• 4:15
  where you're sitting. Yeah. So this has been
• 4:17
  an area that I've actually worked in since,
• 4:20
  2018,
• 4:21
  trying to apply
• 4:22
  AI to to support with this diagnostic odyssey.
• 4:25
  And I remember the first time I heard
• 4:26
  that term, diagnostic odyssey. I think in health
• 4:28
  care and in general, I work in AI,
• 4:30
  so buzzwords are everywhere. But some of the
• 4:32
  language we can use can be emotive and
• 4:35
  and maybe not particularly accurate, but I think
• 4:36
  it's the perfect description of the challenge that
• 4:38
  these families go through.
• 4:40
  And when you've
• 4:42
  had the opportunity to speak to families who
• 4:45
  are affected by rare disease or have loved
• 4:46
  ones affected by rare disease, you realize the
• 4:49
  enormity of the challenge that faces them as
• 4:51
  individuals. And as as Shannon says, that it's
• 4:53
  you know, in in in those individual cases,
• 4:55
  it's it's rare. But, actually, when you bring
• 4:57
  all the cases together, it affects a huge
• 4:59
  number of people worldwide.
• 5:01
  So, yeah, having the opportunity to try and
• 5:03
  apply some technology in this space to make
• 5:05
  a difference,
• 5:06
  is really exciting.
• 5:08
  And, you know, I have a clinical background,
• 5:10
  but lots of the folks I work with
• 5:11
  are technologists, they're software engineers, etcetera. And for
• 5:15
  them to be able to apply their trade
• 5:16
  in an area like this is, you know,
• 5:18
  it's very meaningful for them, and it's certainly
• 5:19
  different to other types of automation where they
• 5:22
  could take their skills too.
• 5:24
  Doctor Hammond, I just wanted to touch on
• 5:25
  an element of this that is, I think,
• 5:27
  equally as important because, again, it is a
• 5:29
  challenge for the family. Certainly, they go on
• 5:31
  an Odysee. But what does this also mean
• 5:34
  for the care team side of things that
• 5:36
  are in these situations as well often with
• 5:39
  the family, in touch with the family in
• 5:40
  those conversations?
• 5:42
  Yeah. It can it can be really challenging.
• 5:44
  And, you know, I I think we all
• 5:45
  feel there's a lot of missed opportunity. Right?
• 5:47
  So
• 5:48
  today in today's world, many of these kids
• 5:50
  are seeing genetic specialists
• 5:52
  years after they should have seen them. Mhmm.
• 5:54
  And so this is what's
• 5:56
  driving us in, at Lurie Children's to really
• 5:59
  work and partner with our, our genetics division,
• 6:03
  to expand access to testing and genetic specialists
• 6:06
  so that we can get these kids in
• 6:08
  very early in their journey,
• 6:10
  get them answers quicker, and then start working
• 6:13
  on treatments, identify what, you know, treatment protocols
• 6:16
  they may be,
• 6:17
  eligible for, and then also
• 6:19
  build on this research and and knowledge so
• 6:21
  that we can devise new treatments.
• 6:24
  You mentioned something really important here, and that's
• 6:26
  I I wanna hit on that here in
• 6:27
  my next question, which is the fact that,
• 6:29
  fortunately, some of this is changing, right, in
• 6:32
  terms of getting testing done earlier, understanding that
• 6:35
  this needs to happen now when it is
• 6:37
  appropriate,
• 6:39
  especially,
• 6:40
  again, right now. What's driving the shift toward
• 6:43
  using it sort of
• 6:45
  as a first line diagnostic tool, and what
• 6:47
  has that
• 6:49
  be true up what has to be true
• 6:50
  operationally for a health system to make that
• 6:52
  transition to be first line?
• 6:55
  Yeah. So I think one of the things
• 6:56
  we're talking about now is is whole genome
• 6:58
  sequencing. So, right, look at the what traditionally
• 7:01
  happens is that there's a sequential number of
• 7:03
  tests that, you know, once these kids actually
• 7:05
  are recognized of having a rare disease or
• 7:07
  or potentially a rare disease, they go on
• 7:09
  this, you know, sequential journey of lots of
• 7:11
  different testing. It could be this or that.
• 7:13
  But we're really believing in
• 7:15
  moving to where a more comprehensive test that
• 7:18
  we call whole genome sequencing,
• 7:21
  is really the first line test.
• 7:24
  And so there's a lot of capital.
• 7:26
  You know, these are expensive analyzers that we
• 7:29
  need. There's a lot of infrastructure.
• 7:30
  There's a lot of compute. We're gonna talk
• 7:32
  about how large the datasets are, and how
• 7:34
  difficult it is to find the answers,
• 7:37
  but also the necessary experts along the way
• 7:39
  to recognize that this is a rare disease
• 7:41
  very early in someone to have the capacity
• 7:43
  to see them, in a a clinical setting
• 7:47
  and get the results and and interpret them.
• 7:50
  And, you know, we know we're working to
• 7:51
  develop this at Lurie Children specifically because we
• 7:53
  we do know from studies that the diagnostic
• 7:56
  yield or our ability to identify
• 7:58
  a answer,
• 8:00
  increases,
• 8:01
  like, at least fourfold if we're doing it
• 8:03
  in the context and in collaboration with a
• 8:06
  multidisciplinary
• 8:06
  team where the child is seeking care versus
• 8:09
  getting a blood test and sending it out
• 8:10
  to, to another lab.
• 8:12
  Doctor Rykubian, when we're looking at that shift
• 8:14
  then, are health systems ready? How ready are
• 8:17
  they to make that shift?
• 8:19
  Yeah. I think there's not one answer or
• 8:21
  or one way to answer that question. Many
• 8:23
  many health systems are.
• 8:25
  I think what's really interesting about what's happening
• 8:27
  at Lurie is you see this really,
• 8:29
  brilliant collaboration between between clinical operations and research.
• 8:34
  I've been at a few sessions here at
• 8:35
  Becker's, and there's a lot of chat about
• 8:37
  whether or not organizations are ready to adopt
• 8:39
  AI. I think you could take the same
• 8:40
  approach to that question as to the one
• 8:42
  that you pose. And there's a sort of
• 8:44
  varying degree of maturity in the the multidisciplinary
• 8:47
  team that Shannon's talking about across the care
• 8:48
  continuum. I think that applies outside of care.
• 8:50
  It's actually
• 8:51
  within the health system.
• 8:53
  How ready are IT with information governance, with
• 8:57
  potentially clinical operations, all to work together in
• 9:00
  order to realize the benefits that can be
• 9:02
  that can be had when that collaboration,
• 9:05
  comes off. So a long winded answer. Yeah.
• 9:08
  In short, many are. I think
• 9:11
  the the sort of acceleration
• 9:12
  of awareness of the power of technology that
• 9:15
  has come with the kind of generative AI,
• 9:17
  maybe Gen2K AI boom, I think is potentially
• 9:19
  driving a little bit more of that that
• 9:21
  organizational readiness,
• 9:23
  but it's going to take time. And, you
• 9:25
  know, we only have so many people working
• 9:27
  in health systems and shortages and staff are
• 9:29
  are are, you know, are things that are
• 9:31
  are certainly challenging in this area. Yeah.
• 9:33
  I wanna be a little bit more specific
• 9:35
  here on some of the issues that we've
• 9:36
  touched on. And a typical whole genome sequencing
• 9:39
  assay identifies several million variants, but only
• 9:42
  a tiny fraction of those
• 9:45
  are potentially relevant to
• 9:48
  a patient's disease.
• 9:49
  How does your team approach that signal to
• 9:52
  noise problem, so to speak, and where does
• 9:55
  AI
• 9:56
  enter the picture? Doctor. Kumin, I can I'll
• 9:57
  start with you. Yeah. So the the way
• 9:59
  that this was first ever articulated to me
• 10:01
  when I got into this space was, you
• 10:03
  know, doing whole genome sequencing is, like, publishing
• 10:05
  the Internet of you,
• 10:07
  and then using technologies, whether it's what we're
• 10:09
  talking about, say, natural language processing or language
• 10:12
  based AI or others, is your almost your
• 10:14
  your search engine. How do you take the
• 10:15
  Internet of you and narrow down on the
• 10:17
  areas that are most relevant? Mhmm. So in
• 10:19
  this particular
• 10:20
  case, what what we're doing is using natural
• 10:23
  language processing
• 10:25
  in order to comb through the patient's electronic
• 10:28
  medical record and identify
• 10:30
  what we call phenotypes. So phenotypes are the
• 10:33
  physical manifestations or the the clinical manifestations of
• 10:36
  your underlying
• 10:37
  genome.
• 10:38
  And by taking that signature, those pictures, or
• 10:41
  those those clinical descriptions and marrying them to
• 10:44
  the information that's coming from the sequencer,
• 10:46
  that's how you're able to then increase the
• 10:47
  diagnostic yield as as Shannon put it. So
• 10:50
  without technology, that process is done by people.
• 10:53
  They go through the medical record. They they
• 10:55
  try and find what are the clinically meaningful
• 10:58
  phenotypes. They are obviously potentially impacted by their
• 11:00
  own clinical biases,
• 11:02
  and it takes time. You know, as a
• 11:03
  herderian,
• 11:04
  I mean, the sessions here at Becker's are
• 11:05
  great. As I heard in the session I've
• 11:07
  just been in, you know, no one went
• 11:08
  to medical school to become a data entry
• 11:09
  clerk. Yeah.
• 11:11
  So we're trying to we're trying to remove,
• 11:13
  that burden and let people operate at the
• 11:15
  top of their license. And maybe Shannon could
• 11:17
  talk about some of the the results that
• 11:19
  have been seen as, you know, it's really
• 11:20
  exciting. Yeah. And, also, the the change. Right?
• 11:23
  What did it look like before? What did
• 11:25
  it look like after?
• 11:26
  It's obviously changing the way that teams work
• 11:30
  as well.
• 11:31
  Absolutely. And,
• 11:33
  I would I would say that before it
• 11:35
  was, like, as you said, it was very
• 11:37
  manual.
• 11:38
  And this is work I'm gonna describe as
• 11:39
  led by, faculty in my department, doctors Kylie
• 11:42
  Yap and Patrick McMullen.
• 11:43
  And we had this idea of this seemed
• 11:46
  like a well suited
• 11:48
  project for an AI
• 11:50
  augmented,
• 11:51
  process. And so we did a little pilot
• 11:53
  with some open source, tools, and it was
• 11:56
  quite promising, which was very exciting.
• 11:58
  And,
• 11:59
  actually, the the first NLP
• 12:01
  technology that we brought into Lurie Children's was
• 12:03
  not for this specific use case. But when
• 12:06
  they were articulating to me their need, I
• 12:07
  said, I think we could, you know, use
• 12:09
  this existing technology for this. And it it
• 12:12
  made it scale because as, you know, Callum
• 12:14
  was saying, we then had the infrastructure that
• 12:16
  we could build on. And I you know,
• 12:17
  it's great that we're now talking about how
• 12:19
  do we,
• 12:21
  build an infrastructure for things like generative AI
• 12:24
  because I remember when we were struggling with,
• 12:25
  like, getting bandwidth enough to be able to
• 12:27
  actually move data files. Right? And so we're
• 12:30
  well beyond that. So this is really exciting
• 12:32
  that people really quickly just at many, many
• 12:35
  institutions now have arrived. This is in your,
• 12:38
  you know, in your native infrastructure, in your
• 12:40
  hospital, your institution, which means you can use
• 12:42
  it with PHI. You can use it with,
• 12:44
  sensitive information, etcetera.
• 12:46
  And people can tinker around and figure out,
• 12:48
  like, how to best use these tools. So
• 12:49
  that's, I think, a really exciting place to
• 12:51
  be today.
• 12:53
  But, yeah, it was, you know, highly trained
• 12:55
  people looking through medical record for these phenotype
• 12:58
  information then matching to these, what we call
• 13:01
  ontologies, right, that link genes to these disease
• 13:04
  states and the phenotype
• 13:06
  and then
• 13:07
  prioritizing them. Right? So as you said, we've
• 13:09
  got this huge
• 13:10
  signal,
• 13:11
  very low signal with a lot of noise,
• 13:13
  and so we were trying to reduce the
• 13:14
  noise.
• 13:15
  And, you know, combing also known databases of
• 13:18
  how common is this in the population so
• 13:20
  that if it's really common, then that gives
• 13:22
  it very low likelihood that it's actually causing
• 13:23
  the disease. Right? So some of the decisions
• 13:25
  that we make. And,
• 13:28
  what's been reported in the literature? Like, have
• 13:30
  we seen this before? Is this particular variant
• 13:32
  associated with a phenotype like what we're seeing
• 13:34
  in this child?
• 13:35
  So that took a lot of time. And
• 13:37
  then you can imagine, you know, Lucas, if
• 13:39
  you're doing it versus me versus Calum, we
• 13:41
  also may arrive at different prioritizations and different
• 13:44
  answers. And so this is also where having
• 13:46
  an automated AI tool
• 13:49
  can help you standardize things and make them
• 13:51
  more consistent. But also just the sheer amount
• 13:54
  of work and data that needs to be
• 13:56
  processed is really a nice use case for
• 13:58
  having an augmented,
• 14:00
  solution.
• 14:01
  And as far as results, I guess, about
• 14:02
  the spinach,
• 14:03
  so what we found is, obviously, it's faster
• 14:05
  to do it with you when you have
• 14:06
  an AI pipeline helping you do it in
• 14:08
  an automated way. And when we compared
• 14:11
  manually,
• 14:12
  with the AI solution, so which genes did
• 14:15
  the AI identify and how did they rank
• 14:17
  them versus what we did, there was a
• 14:18
  really high level of overlap. So it's now
• 14:21
  to the point where we're actually implementing it
• 14:23
  clinically. Like, we feel it's it's validated,
• 14:25
  and, you know, I think we can talk
• 14:27
  more about how we get to that place
• 14:29
  of being a research tool to translate into
• 14:31
  clinical and getting people's buy in and wanting
• 14:33
  them to use it.
• 14:34
  Yeah. I wanna touch on that a little
• 14:36
  bit because I think it's a crucial factor.
• 14:37
  Obviously, we're seeing results now. But, again, what
• 14:40
  we talk about at Becker's quite a bit
• 14:41
  here at the annual meeting, but also generally
• 14:43
  is we talk about the bright shiny object
• 14:45
  in health care, and a lot of organizations
• 14:48
  like to go for the bright shiny object
• 14:49
  without necessarily considering the business case for it.
• 14:52
  What does it actually do? And and, doctor
• 14:54
  Heyman, I'll start off with you. How did
• 14:56
  your team think about that tool selection, and
• 14:58
  what made and I'll give you the right
• 14:59
  approach again to extract
• 15:02
  and standardizing phenotype information for those clinical nodes?
• 15:05
  Yeah. So as I as I mentioned, you
• 15:07
  know, this wasn't our first use case. I
• 15:09
  recognized the the initial use case that that
• 15:12
  drove the implementation of this technology was that
• 15:14
  we have for all of the patients who
• 15:16
  had already had genetic testing,
• 15:18
  medical records, they're getting better, but they're not
• 15:20
  very good at ingesting these,
• 15:22
  unstructured
• 15:24
  text reports.
• 15:25
  And some of them are mixes of text
• 15:26
  and tables.
• 15:28
  Some of them are 30 pages long. There's
• 15:30
  not a very easy way to put that
• 15:31
  into the medical record and have it accessible
• 15:33
  to people. And so how they end up
• 15:35
  in many institutions
• 15:37
  is either today, we've advanced to having a
• 15:39
  searchable PDF.
• 15:41
  But in the past and still today in
• 15:42
  many places, it's a an image, like a
• 15:44
  scanned image. Right? And so I recognize that
• 15:48
  we had a tremendous amount of patients that
• 15:50
  had testing.
• 15:51
  This really, really important data is locked away,
• 15:53
  unaccessible to people. And so we needed a
• 15:56
  tool to be able to process these reports
• 15:58
  and extract that information
• 16:00
  in whole, but also to extract out key
• 16:02
  pieces that we then load into a database
• 16:04
  that we can use for our own clinical
• 16:05
  and then research purposes. And so that was
• 16:08
  really the first step of that. And so
• 16:09
  when you're talking about extracting,
• 16:11
  you know, from report and text, natural language
• 16:13
  processing at the time, this was probably five,
• 16:15
  six years ago. Right? That was, like, the
• 16:18
  tool to do. Like, we didn't maybe I
• 16:20
  mean, we were starting to play with transformers,
• 16:21
  but I didn't know anything about LLMs. I
• 16:23
  couldn't have called it that at the time.
• 16:24
  But that was the tool. And so,
• 16:26
  you know, we implemented this thing, which was
• 16:30
  a a multiyear project to get it fully
• 16:32
  integrated and and fully set up, within our
• 16:35
  our hospital,
• 16:36
  IT.
• 16:37
  And now we've got something working. We know
• 16:39
  how it works. And so it's like, what
• 16:41
  other use cases are there? And, actually, part
• 16:42
  of our journey in doing that was,
• 16:46
  to identify and work with other groups in
• 16:47
  the institution to say, what do you think
• 16:50
  would be applications where NLP would be useful
• 16:52
  to you? And they were highly variable. And
• 16:54
  so we've knocked a couple of those off,
• 16:55
  but that was really where we started. And
• 16:57
  when we did our request for proposal and
• 16:59
  investigated, we had all those use cases in
• 17:01
  mind as we were thinking about functionality.
• 17:04
  Control f is a powerful tool.
• 17:07
  Doctor Ekubian,
• 17:08
  what questions then do organizations
• 17:10
  need to ask,
• 17:12
  especially when it comes to these AI tools?
• 17:14
  What's important?
• 17:15
  Yeah.
• 17:16
  I think, really, depending on your use case,
• 17:19
  those questions will vary, but there are some
• 17:20
  kind of core questions that are super important
• 17:22
  no matter what tool you're applying.
• 17:24
  I think, you know, both of your own
• 17:26
  organization, you know, of the the tool or
• 17:28
  the vendor perhaps,
• 17:30
  kind of what does the governance look like?
• 17:31
  I think it's really important to understand,
• 17:35
  how the technology will be be deployed. It's
• 17:37
  not in isolation. It's gonna be within a
• 17:38
  workflow. So what does the governance look like,
• 17:40
  and how will that vendor support you?
• 17:43
  To your point, you know, by shiny objects,
• 17:45
  it's like everyone,
• 17:47
  on the vendor side talks about their shiny
• 17:49
  objects and how it's the shiniest and how
• 17:50
  it's the best. But,
• 17:52
  there's lots of great technology. I think you
• 17:54
  need to kind of scratch beneath the surface
• 17:55
  a bit when you're working with with solution
• 17:57
  providers to understand,
• 17:59
  you know, beyond the shine, actually, how how
• 18:01
  is this gonna work in clinical practice? How
• 18:03
  does that van vendor think about the implementation
• 18:05
  within the teams? How is trust going to
• 18:07
  be insured? And I think this use case
• 18:09
  in Shannon has touched on
• 18:10
  NLP, generative AI, agent to k I,
• 18:13
  all the LLMs, etcetera, as Shannon was even
• 18:16
  touching on, whether it's transformers, etcetera, they all
• 18:18
  fall kind of under that banner of natural
• 18:20
  language understanding or natural language processing. And for
• 18:23
  this use case, kind of, trust and transparency
• 18:25
  is hugely important,
• 18:27
  and you don't want things like hallucinations.
• 18:29
  So if you're wanting to get clinical adoption
• 18:31
  and clinical trust,
• 18:33
  how do you make sure that if you
• 18:34
  process the same patient's medical records every time,
• 18:37
  it's only going to improve, not get worse?
• 18:39
  It's not gonna hallucinate clinical findings. And in
• 18:42
  use cases like that, for example, choosing something
• 18:45
  with a more deterministic approach, which is what
• 18:47
  we have going on here,
• 18:49
  you know, make makes an awful lot of
• 18:51
  sense.
• 18:52
  What's really exciting is how do we then
• 18:54
  use those types of technologies in broader agentic
• 18:57
  systems that that we're now seeing. So,
• 19:00
  we're really talking about hybrid AI. How do
• 19:02
  you kind of use the power of generative
• 19:04
  for for certain findings or certain features? Say,
• 19:07
  for example, synthesizing literature, but but use the
• 19:09
  power of deterministic
• 19:11
  to make sure that you're always returning the
• 19:13
  same documents, the same scientific publications to inform
• 19:15
  that decision making.
• 19:18
  So, yeah, I think,
• 19:19
  governance I always come back to governance and
• 19:21
  and privacy is kind of number one, and
• 19:23
  the organization takes that seriously because
• 19:25
  that's where you,
• 19:27
  and the vendor will fall flat,
• 19:30
  if you don't have that right. That's its
• 19:31
  own podcast episode, I feel like. It'll adjust
• 19:34
  the adjust the conversation on governance, which is
• 19:36
  so crucial.
• 19:38
  Doctor Hammond, you've touched on this a little
• 19:40
  bit,
• 19:40
  especially with AI tools. With these new tools,
• 19:43
  one of the crucial
• 19:45
  parts of making this work correctly is adoption
• 19:48
  Mhmm. And making sure that the teams that
• 19:50
  are exposed to it are actually
• 19:53
  doing it and using it. Right?
• 19:55
  You took a different approach because you've built
• 19:57
  this onto an existing workflow in and of
• 20:00
  itself at Lurie.
• 20:01
  What role
• 20:03
  did the clinical team play in all of
• 20:04
  this? And, again, what did that design look
• 20:06
  like for you? Yeah. I think it's extremely
• 20:09
  important.
• 20:10
  We've got a number of, you know, anybody
• 20:12
  who develops
• 20:14
  applications and uses technology certainly has their list
• 20:16
  of this was, like, gonna be a, you
• 20:19
  know, an amazing
• 20:20
  solution, and then it just falls flat and
• 20:22
  never gets used,
• 20:24
  that, you know, if you build it, they
• 20:25
  will come doesn't always work out. Right? And
• 20:27
  so what what I've learned in time is
• 20:29
  that you need a champion,
• 20:31
  and you need people with buy in. And
• 20:32
  and especially for,
• 20:34
  what we find like AI or other advanced
• 20:36
  technologies,
• 20:36
  you know, having people who understand the technical
• 20:39
  piece and can work on on the technical
• 20:41
  piece
• 20:42
  married with people who understand the actual workflow
• 20:45
  and the impact,
• 20:46
  is critical. I mean, I can't tell you
• 20:48
  the number of times I've had some buddy
• 20:50
  come with lab data and tell me about
• 20:52
  a problem they've solved that, you know, was
• 20:54
  not even a problem. Like, okay. Great. You
• 20:56
  identified that we had a downtime. That's not
• 20:58
  anything I care about. I know when we're
• 20:59
  in a downtime.
• 21:00
  So,
• 21:01
  having people who understand what the problems are
• 21:04
  and what the metrics should be, like, what
• 21:05
  does success look like beyond just, like, an
• 21:07
  area under the curve or, you know, a
• 21:09
  a typical accuracy metric
• 21:11
  with people who could actually translate those requirements
• 21:14
  into a functional,
• 21:16
  tool or application is is really important.
• 21:19
  And then, you know, again, I think doing
• 21:21
  this pilot and seeing, like, what are the
• 21:23
  advantages and and continuing to to talk about
• 21:26
  that. Like, how is this impacting things? How
• 21:27
  is this making your job better?
• 21:29
  And, you know, something else that we take
• 21:31
  a little bit more for granted now that
• 21:33
  at the time five years ago wasn't a
• 21:34
  given when we were investigating
• 21:36
  different, vendors was
• 21:38
  having the option for human in the loop,
• 21:40
  especially in health care. There that wasn't always
• 21:43
  the case. And so knowing that, you know,
• 21:45
  people are involved and they have the ability
• 21:47
  to stop things or or understand what's going
• 21:49
  on with with the AI.
• 21:51
  But now our our people, you know, they've
• 21:53
  seen the they've seen the benefit. They are
• 21:55
  now looking at, like, how else can we
• 21:57
  apply this? And so I think that's how
• 22:00
  success begets success, but also having champions and
• 22:03
  people who feel engaged or involved have been,
• 22:05
  important for our success. Doctor. Yacoubian, you mentioned,
• 22:08
  obviously, governance as a crucial piece to what
• 22:10
  sets
• 22:11
  vendors apart, what sets partners apart for health
• 22:13
  systems. What are some of the other factors
• 22:15
  that you think are are are crucial from
• 22:17
  what you've just heard Doctor. Hammitt describe as
• 22:19
  well, but in terms
• 22:21
  of initiatives that work versus those that stall,
• 22:23
  what's crucial?
• 22:25
  Yeah.
• 22:25
  I've heard it put before that AI scales
• 22:27
  at the speed of trust. Yeah. And that
• 22:29
  comes back to that that governance piece.
• 22:31
  I mean, the answer to this one's actually
• 22:33
  a relatively boring one, and it's been banded
• 22:34
  around with technology implementations for a lot longer
• 22:37
  than we've had AI.
• 22:38
  You know, the technology is one piece, but
• 22:41
  then you have the people in the process
• 22:42
  into which that technology is going to be
• 22:44
  deployed. And that's some of what Shannon was
• 22:46
  just was was just talking about. So having
• 22:48
  that clinical workflow, the understanding,
• 22:50
  the multidisciplinary team that we've spoken about before,
• 22:53
  the the kind of the interplay between the
• 22:55
  technical teams that have to deploy it,
• 22:57
  and the clinical teams who have to use
• 22:59
  it.
• 23:00
  That's all incredibly important.
• 23:02
  I think
• 23:03
  what has changed in the past few years
• 23:05
  as we kind of look at the applications
• 23:06
  of AI, of generative AI, in fact, anyone
• 23:08
  can
• 23:10
  potentially upload a synthetic and then we'll say
• 23:12
  synthetic medical record to CHAT GPT, for example,
• 23:14
  and get really rich information back. There's this
• 23:17
  huge saturation of of pilots
• 23:19
  and not just having people processing technology. You
• 23:22
  actually you know, you have to have a
• 23:23
  problem to solve and a well documented problem
• 23:26
  to solve.
• 23:27
  AI shouldn't be a hammer looking for nails.
• 23:30
  You need to have problems which are well
• 23:32
  defined that then enable you to have a
• 23:34
  a bot in team,
• 23:36
  to kind of to bring the solution to
• 23:38
  bear. So,
• 23:40
  yeah, I think there was a Gartner report
• 23:42
  from last year that was, you know, I
• 23:44
  think it was 30% of AI pilots of
• 23:46
  generative AI pilots failed. Yes. That's a huge
• 23:49
  amount of investment. If you're investing money in
• 23:51
  these generative AI pilots and they're you know,
• 23:53
  a third of them are going to fail
• 23:55
  and never go into production, that's a lot
• 23:56
  of money, and we're hearing a lot. You
• 23:58
  know, Becker's publishes what the the operating margins
• 24:00
  are for health systems.
• 24:02
  And I think a lot of that is
• 24:03
  because people just go, oh, I've got AI.
• 24:05
  I'm gonna do something cool without ever thinking,
• 24:07
  is this a problem that needs to be
• 24:10
  solved at the organizational level? Will I will
• 24:12
  I kind of get that return of value,
• 24:14
  that return of, how do I call it,
• 24:16
  return of health, you know, the patient impact,
• 24:18
  as well as potentially, you know, obviously, the
• 24:20
  return on investment. Yeah.
• 24:22
  We've covered so much ground today. We've touched
• 24:24
  on so many different topics. It's so great
• 24:26
  to have you both here. I just wanna
• 24:28
  turn the floor over to you. Anything else
• 24:30
  that you think is important for our audience
• 24:31
  to understand? Anything else that you'd like to
• 24:33
  add to our conversation here today? Doctor Hamid,
• 24:35
  I'll I'll start off with you. Yeah. Thank
• 24:36
  you. Yeah. It's been really fun to to
• 24:38
  talk about this project and and think about
• 24:40
  the future.
• 24:41
  I would say we you know, though we
• 24:43
  talked about that genetic testing is well established,
• 24:45
  this concept of applying
• 24:47
  AI and, you know, expanding, transitioning to broad
• 24:50
  genomic testing as a first line is not
• 24:52
  well established. And so,
• 24:54
  a lot of this work then is,
• 24:56
  research
• 24:57
  and is supported by our generous philanthropic partners.
• 25:00
  So I'd like to acknowledge the Edwards family
• 25:02
  and the Lurie Children's Founders Board for really
• 25:04
  helping us bring technologies like this in and
• 25:07
  supporting,
• 25:08
  our innovation in those areas. We're really grateful
• 25:10
  for their support. And then I hope, you
• 25:12
  know, that we can emphasize that
• 25:15
  the real benefit of of AI in a
• 25:17
  in a situation like this is that, you
• 25:18
  know, when we think about testing for rare
• 25:20
  disease, we talked about the hundreds of thousands
• 25:22
  of families who are impacted. I would love
• 25:24
  to see all of them at Lurie Children's,
• 25:26
  but but we likely won't.
• 25:28
  And, you know, we're we're now at the
• 25:29
  place where we're reliant on getting them in
• 25:31
  the door. But, you know, large scale newborn
• 25:34
  screening throughout this nation and others is moving
• 25:37
  into the way of having being done by
• 25:39
  whole genome sequencing. And so when we think
• 25:41
  about how would we actually practically and feasibly
• 25:43
  scale that, having manual workflows is never gonna
• 25:47
  cut it. And so I think we're looking
• 25:49
  at that as even as we increase our
• 25:50
  volumes and and make more capacity for seeing
• 25:52
  kids at Lurie Children's, we need this tool.
• 25:55
  But in the not too distant future, if
• 25:56
  we think about doing newborn screening with comprehensive
• 25:59
  genetic testing, we're gonna need more AI augmentation
• 26:02
  in this way even more. So, just to
• 26:04
  leave that there for how this could impact
• 26:06
  even more kids in the future. Yeah. Doctor.
• 26:09
  Yeah. There's there's I think we we have
• 26:11
  covered a lot as you say, and,
• 26:13
  probably getting the picture that Shannon and I
• 26:15
  could talk about this for a little longer.
• 26:17
  No. I I probably would like to reemphasize
• 26:19
  that, you know, it's it's a real,
• 26:21
  you know, pleasure and honor getting to work
• 26:23
  in this space. You know?
• 26:26
  The application of technology and AI to make
• 26:29
  a patient
• 26:29
  impact and to make a patient difference is
• 26:31
  what I think, you know, most whether or
• 26:33
  not it's AI, and most people at this
• 26:35
  conference is out. So it kind of gets
• 26:36
  them out of bed in the morning, and,
• 26:38
  you know, it's great to be able to
• 26:39
  work with partners like Shannon and Louie to
• 26:41
  to do this work. And, you know, for
• 26:43
  anyone, I think, you know, most people are
• 26:45
  thinking about AI and and deploying AI just
• 26:47
  now.
• 26:49
  And, yeah, I I just would encourage, you
• 26:50
  know, folks to to get on and do
• 26:52
  it, just kind of take on board some
• 26:54
  of the stuff we've talked about about what
• 26:55
  makes it successful. But,
• 26:57
  yeah, the,
• 26:59
  you know, the future is here. You know,
• 27:00
  I think AI is not gonna replace clinicians,
• 27:04
  but, certainly, clinicians who use AI are gonna
• 27:07
  replace those who don't. And that's not that's
• 27:09
  a quote from from The Lancet in 2019
• 27:11
  Mhmm. Believe it or not. But, yeah, it's
• 27:13
  just a pretty exciting time and, you know,
• 27:15
  very glad to be part of it. Doctor
• 27:16
  Heyman, doctor Ekubian, so great to have you.
• 27:19
  Thanks so much for your insights and time
• 27:20
  today. And we also want to thank our
• 27:22
  podcast sponsor, IQVIA. You could tune in to
• 27:24
  more podcasts from Becker's Healthcare by visiting our
• 27:27
  podcast page at beckershospitalreview.com.