Latest / Becker’s Healthcare Digital Health + Health IT / Scaling AI in Healthcare: Insights from Johns Hopkins’ Dr. Alvin Liu

Transcript

Scaling AI in Healthcare: Insights from Johns Hopkins’ Dr. Alvin Liu · Becker’s Healthcare Digital Health + Health IT

• 0:00
  Exciting things are happening at Becker's Healthcare. Stay
• 0:04
  ahead of industry trends with the new Becker's
• 0:06
  CFO plus Revenue Cycle podcast,
• 0:08
  your go to source for insights from top
• 0:11
  healthcare finance leaders. Tune in wherever you get
• 0:14
  your podcasts.
• 0:15
  And don't miss the tenth annual health IT
• 0:18
  plus digital health plus RCM conference, happening September
• 0:22
  30 to 10/03/2025
• 0:24
  in Chicago.
• 0:25
  Join thousands of
• 0:27
  executives, engage with industry leaders, and explore the
• 0:30
  future of health care innovation.
• 0:32
  Learn more about our upcoming events at beckershospitalreview.com.
• 0:36
  See you there.
• 0:38
  This is Laura Dirda with the Becker's Healthcare
• 0:40
  podcast.
• 0:41
  I'm thrilled today to be joined by doctor
• 0:43
  Alvin Liu, endowed professor in AI and AI
• 0:46
  oversight team member as well as the inaugural
• 0:48
  director
• 0:49
  at the James p Gills junior MD and
• 0:52
  Heather Gills artificial intelligence innovation center at Johns
• 0:55
  Hopkins Medicine.
• 0:56
  Doctor Lewis, a pleasure to have you on
• 0:58
  the podcast today.
• 0:59
  Good morning, Laura. Thanks for having me. It's
• 1:01
  a pleasure.
• 1:02
  Absolutely. Now I'm really looking forward to this
• 1:04
  conversation because I know the work you're doing
• 1:07
  in AI and especially at Johns Hopkins really
• 1:10
  truly,
• 1:11
  is on the cutting edge of innovation and
• 1:13
  health care technology today. And so many organizations
• 1:16
  are trying to figure out how they can,
• 1:19
  use AI most effectively and what it means
• 1:21
  for their health systems and hospitals.
• 1:23
  So this will be a really fun conversation.
• 1:26
  But before we dig in, I'm wondering, first,
• 1:28
  could you introduce yourself, tell us a little
• 1:30
  bit about your background, and,
• 1:32
  what makes your role at, Johns Hopkins unique?
• 1:36
  Sure. Happy to do so. So, again, my
• 1:38
  name is Alvin Liu.
• 1:40
  I did most of my medical training at
• 1:42
  Johns Hopkins University.
• 1:44
  I'm a practicing retinal surgeon. And after my
• 1:47
  training,
• 1:48
  I stayed on as the medical faculty at
• 1:51
  Johns Hopkins.
• 1:52
  Currently, outside of my clinical work, I wear
• 1:55
  several hats which are all related to artificial
• 1:59
  intelligence.
• 2:00
  First,
• 2:01
  as a clinician scientist,
• 2:03
  I'm the inaugural director of the Wilmer Gilkes
• 2:06
  AI Innovation Center,
• 2:08
  which is the first endowed AI center
• 2:10
  at the Johns Hopkins University School of Medicine,
• 2:14
  which was made possible a few months ago
• 2:17
  by a very generous donation of $10,000,000.
• 2:21
  Second, I have been involved at Johns Hopkins
• 2:24
  in the deployments of AI tools for both
• 2:27
  clinical and operational processes,
• 2:29
  such as autonomous
• 2:31
  AI for diabetic retinopathy screening
• 2:34
  and revenue cycle management.
• 2:36
  And third, I serve on the AI oversight
• 2:39
  team,
• 2:40
  which is a system wide leadership team with
• 2:44
  purview over all things AI related
• 2:47
  in both the clinical and operational domains
• 2:50
  across the entire Johns Hopkins Medicine health system.
• 2:54
  That's great to hear. And, you know, really,
• 2:56
  cool projects that you're working on as well.
• 2:59
  I'd love to dig a little bit deeper
• 3:00
  in there. What are some of the things
• 3:02
  you're doing right now, the opportunities as well
• 3:04
  as the headwinds that you had your eye
• 3:06
  on?
• 3:07
  I would say across
• 3:10
  the different trends that I see in health
• 3:12
  care right now, artificial intelligence
• 3:15
  is currently the biggest opportunity
• 3:17
  and will remain the leading force for change
• 3:20
  in the foreseeable future.
• 3:22
  However, despite the generally rosy sentiments,
• 3:26
  significant headwinds do exist for real world scaling
• 3:30
  of these clinical AI tools.
• 3:32
  If you look at, a recently published report
• 3:35
  by the FDA,
• 3:36
  there are now over 1,000
• 3:39
  AI enabled medical devices that have been approved
• 3:42
  by the FDA.
• 3:44
  Yet if you look at real world deployment
• 3:46
  of these tools,
• 3:48
  we are still in the infancy stage.
• 3:50
  I think the key uncertainty
• 3:52
  is who and how are we go going
• 3:55
  to pay for these clinical AI tools.
• 3:58
  From, let's say, a clinical
• 4:01
  AI startup company's perspective,
• 4:03
  getting a new tool to be approved by
• 4:05
  the FDA is very costly.
• 4:08
  And even if you manage
• 4:10
  to get your products through the FDA,
• 4:12
  you have then make sure that there's a
• 4:14
  way to pay for it. And typically in
• 4:17
  The US,
• 4:18
  that's being done through CPT coding. And oftentimes,
• 4:23
  since these AI technologies are new,
• 4:26
  there isn't
• 4:28
  an applicable CPT code
• 4:30
  that could be used for reimbursement.
• 4:32
  So now you have to go through the
• 4:34
  AMA and apply for a new CPT code.
• 4:38
  And let's say you go through the whole
• 4:39
  process
• 4:40
  and manage to get a new CPT code
• 4:43
  to pay for your AI tool,
• 4:45
  then you have to go to the insurance
• 4:47
  payers and convince them to actually pay for
• 4:50
  it. And last but not least, even if
• 4:53
  you get some commitment from the payers
• 4:55
  to,
• 4:56
  pay for your AI tool, there's no guarantee
• 4:59
  that it will be reimbursed
• 5:01
  at a financially viable level.
• 5:04
  However, on the other hand, especially in the
• 5:06
  past couple years, I've seen significant traction
• 5:10
  in the use of generative AI,
• 5:13
  especially
• 5:14
  large language models in administrative tests,
• 5:17
  such as managing the different steps in revenue
• 5:20
  cycle management.
• 5:21
  I think this makes sense because these administrative
• 5:25
  applications
• 5:26
  do not require
• 5:28
  FDA approval.
• 5:29
  And in many ways,
• 5:31
  these administrative
• 5:32
  applications,
• 5:34
  are easier,
• 5:36
  to demonstrate return on investment.
• 5:38
  For example, previous studies,
• 5:41
  estimated that approximately
• 5:43
  30% of health care spending in this country
• 5:46
  can be considered waste.
• 5:48
  Right now, The US spends about 4,500,000,000,000
• 5:52
  per year in health care. So roughly speaking,
• 5:55
  the total addressable market is close to $1,000,000,000,000,
• 5:59
  meaning there are many opportunities
• 6:01
  to do, to use generative AI
• 6:04
  to address the inefficiency
• 6:06
  and waste in the current health care spending.
• 6:10
  That makes a lot of sense, you know.
• 6:11
  And I appreciate you going through,
• 6:13
  those challenges and headwinds,
• 6:16
  because I I think, you know, there's a
• 6:17
  lot to consider here, and, certainly, trying to
• 6:21
  figure out how, not only to bring the
• 6:23
  technology in effectively, but, also, as you mentioned,
• 6:26
  the payment methodologies and and the funding available,
• 6:29
  for some of these things. And so I
• 6:31
  think that's a really, really great watch out.
• 6:33
  And along those lines, you know, when you
• 6:35
  look at AI and you look at technology
• 6:38
  as it grows and evolves, how are you
• 6:40
  thinking about growth and continuing to add value,
• 6:43
  to the health system overall?
• 6:45
  And speaking from the perspective of an academic
• 6:48
  medical center, I think
• 6:49
  we have to critically evaluate our role in
• 6:53
  the rapidly evolving AI landscape and strategically position
• 6:56
  ourselves
• 6:57
  to maximize
• 6:58
  where we can add value.
• 7:00
  One thing to keep in mind is the
• 7:02
  fact that the majority of care in The
• 7:05
  US is delivered
• 7:06
  by integrated health systems,
• 7:08
  and we will always be the ones performing
• 7:11
  the last mile delivery when it comes to
• 7:13
  patient care.
• 7:14
  It is critical for us to strengthen our
• 7:17
  collaborations,
• 7:18
  with start ups, industry, and venture capital investment
• 7:21
  funds
• 7:22
  to ensure that we have a voice in
• 7:25
  how these AI
• 7:27
  health care products are developed and fine tuned
• 7:29
  from day one.
• 7:31
  We should also invest
• 7:33
  in implementation science and change management
• 7:36
  to transform ourselves into world experts
• 7:39
  and real world implementation of AI tools in
• 7:42
  order to maximize their positive impacts on patients'
• 7:45
  care.
• 7:47
  I think that's a a really great advice
• 7:49
  and certainly makes a lot of sense.
• 7:52
  You know? And with AI still being so
• 7:54
  new and and changing so quickly,
• 7:56
  the capabilities,
• 7:57
  rapidly
• 7:58
  evolving.
• 7:59
  From your perspective, as someone who has spent
• 8:02
  a lot of time looking at the technology
• 8:04
  side as well as as you mentioned, you
• 8:05
  know, the implementation science and change management of
• 8:08
  it. What advice do you have for health
• 8:10
  systems as they're taking some of these pilot
• 8:12
  programs that have worked out well, in in
• 8:15
  small groups,
• 8:16
  as they're trying to scale it out through
• 8:18
  their whole system?
• 8:19
  I think this goes back to
• 8:22
  a frequently asked question,
• 8:25
  when it comes to technology that's not just
• 8:27
  confined to AI,
• 8:30
  which is to build or buy.
• 8:32
  Whenever a new technology comes along, for big
• 8:35
  organizations such as an integrated health system,
• 8:39
  typically, they have to figure out are they
• 8:41
  going to build the new technology themselves or
• 8:44
  buy.
• 8:46
  Given the pace of,
• 8:48
  advancement
• 8:49
  when it comes to
• 8:50
  AI, especially
• 8:52
  generative AI,
• 8:54
  I think
• 8:55
  my personal opinion is that we should
• 8:58
  favor buying.
• 9:00
  Just the nature of the technology and the
• 9:03
  way that things get built,
• 9:05
  I don't think integrated health systems are the
• 9:08
  best
• 9:09
  actors
• 9:10
  in very rapid iteration of these new products.
• 9:13
  And if you look at the flourishing marketplace
• 9:16
  when it comes to health care AI products,
• 9:18
  I really think, you know, either start ups
• 9:21
  or industry
• 9:22
  will be better suited in building a product
• 9:24
  in the first place.
• 9:25
  However,
• 9:27
  what integrated health systems can do is to
• 9:30
  be very actively engaged in in a product
• 9:34
  development process in day one. And oftentimes, as
• 9:37
  I mentioned earlier,
• 9:39
  we are the ones who
• 9:41
  are physically interacting with the patients. We are
• 9:44
  the ones who really understand best the actual
• 9:47
  pain points of health care delivery.
• 9:49
  So we are the ultimate customers for these
• 9:52
  health care AI tools.
• 9:53
  So we should be very involved in the
• 9:55
  product development process from day month day one.
• 9:59
  So that's my advice number one. And my
• 10:01
  advice number two is, in general, if you
• 10:04
  look at academic centers or integrated,
• 10:06
  health systems, we tend to move a little
• 10:09
  bit slower and we tend to be
• 10:13
  more on the conservative
• 10:14
  side.
• 10:15
  I do think that,
• 10:17
  we should change our mindset a little bit
• 10:20
  and be open to rapid iteration.
• 10:24
  And in the sense, you know, we should
• 10:27
  aim
• 10:28
  to iterate fast,
• 10:30
  fail fast if necessary, and pivot even faster.
• 10:34
  And the third advice I would give is
• 10:37
  every integrated health system should have a robust
• 10:40
  AI governance structure.
• 10:43
  What I mean is that, I'm sure there's
• 10:45
  not a unique problem to Johns Hopkins.
• 10:48
  I see
• 10:49
  every single integrated health systems being bombarded by
• 10:52
  AI request by different vendors. It is a
• 10:55
  very chaotic situation right now. So there needs
• 10:58
  to be a system
• 11:00
  that can really govern
• 11:02
  the interaction between AI vendors and integrated health
• 11:05
  system. And the ultimate goal is really,
• 11:07
  so that we can pick AI tools that
• 11:11
  are feasible, that make sense in the real
• 11:13
  world, and ultimately,
• 11:14
  are safe and efficacious for patients.
• 11:17
  So, this is something that Johns Hopkins Medicine
• 11:20
  started doing recently,
• 11:22
  essentially starting about a year ago. The health
• 11:25
  system leadership
• 11:26
  recognizes the need
• 11:28
  to really,
• 11:29
  evaluate and implement AI technology at scale in
• 11:33
  a responsible and safe manner.
• 11:36
  And in order to achieve that goal,
• 11:38
  a decision was made to establish an AI
• 11:41
  governance structure
• 11:43
  to really govern and figure out what is
• 11:45
  the optimal best practices
• 11:48
  when it comes to interacting with AI vendors.
• 11:50
  So for example, now at Johns Hopkins, if
• 11:53
  you were an AI vendor that is interested
• 11:57
  in selling to or working with Johns Hopkins,
• 12:00
  there is a standardized intake process that you
• 12:03
  you have to go through.
• 12:05
  There will be a set of standardized questions
• 12:08
  that cover things like the model card,
• 12:11
  the, data provenance,
• 12:13
  the validation process. And after the initial intake,
• 12:16
  depending on the nature of the application,
• 12:19
  the application will then be sent to specific,
• 12:23
  subcommittees
• 12:24
  that are composed of a variety of stakeholders,
• 12:27
  including physician leaders
• 12:29
  who then do a deeper dive into the
• 12:32
  merit and safety of these applications.
• 12:34
  And so every vendor now is expected to
• 12:37
  go through the same process
• 12:39
  being evaluated the same way, and we will
• 12:41
  also provide adequate feedback to the vendors, at
• 12:45
  the same time. It's not meant to be
• 12:47
  a burdensome,
• 12:49
  process,
• 12:49
  but, really, the ultimate goal is to establish
• 12:52
  something that's standardized that can maximize the safety
• 12:56
  and efficacy of these AI tools for patient
• 12:58
  care.
• 12:59
  I love that. I think that's such a
• 13:01
  a smart way of going about it to
• 13:03
  really organize,
• 13:05
  all the noise. As you mentioned, so many
• 13:06
  different pitches coming through and vendors, you know,
• 13:09
  wanting to connect with members of the organization,
• 13:12
  clinical, administrative,
• 13:14
  across the board. And and so I I
• 13:16
  think that is truly helpful to understand how
• 13:18
  you're dealing with that at Johns Hopkins, and,
• 13:21
  I I can imagine it will continue to
• 13:23
  evolve over time.
• 13:25
  And and speaking of that, where do you
• 13:27
  see some of the best opportunities for growth
• 13:29
  in the future?
• 13:31
  One area that I'm particularly excited about is
• 13:34
  the intersection
• 13:35
  of AI and oculomix.
• 13:38
  So as a some background information,
• 13:41
  oculomix
• 13:42
  is a field of study that connects findings
• 13:44
  or biomarkers
• 13:46
  from the retina, which is the back of
• 13:48
  the eye, with systemic health states.
• 13:51
  For example, if we use deep learning, which
• 13:54
  is the current cutting edge AI technique for
• 13:57
  image analysis
• 13:59
  to analyze a retina photograph,
• 14:02
  we can now predict someone's age,
• 14:05
  biological sex, smoking status, blood pressure,
• 14:08
  cardiovascular risks, presence of dementia, and kidney status.
• 14:13
  This is possible because the retina is the
• 14:15
  only place
• 14:17
  in the entire body
• 14:19
  where we can image both blood vessels and
• 14:21
  brain tissue at the same time
• 14:23
  noninvasively.
• 14:25
  This is no longer science fiction, although it
• 14:27
  sounds like science fiction.
• 14:29
  In fact, there are now multiple startup companies
• 14:33
  that are actively involved in oculomix.
• 14:36
  This approach is very exciting to me because
• 14:38
  it has the potential
• 14:40
  to upend
• 14:41
  medicine as we know it. So imagine in
• 14:44
  the not so distant future,
• 14:46
  Laura, you can get your retina
• 14:49
  image at your local pharmacy or supermarket.
• 14:52
  Within minutes, your retina photographs
• 14:55
  will be analyzed by several different AI models,
• 14:59
  hosted in in the cloud
• 15:00
  that would then give you a readout for
• 15:03
  your systemic health states for multiple organs.
• 15:07
  As you can imagine, this kind of community
• 15:10
  screening is obviously much more convenient
• 15:13
  and cost effective
• 15:15
  than getting a brain MRI scan
• 15:18
  or a CT scan of your heart and
• 15:20
  has the potential to reach millions of patients
• 15:23
  who are not even aware of the systemic
• 15:25
  health problems to begin with.
• 15:27
  And, historically,
• 15:29
  retina specialists like myself
• 15:31
  are considered super specialists, so you only get
• 15:34
  to see a retina specialist after multiple referrals.
• 15:38
  But now we are really flipping this health
• 15:41
  care journey on its head and are using
• 15:43
  the retina literally as a window to your
• 15:46
  body as a form of preprimary care.
• 15:49
  This potentially,
• 15:51
  paradigm shifting delivery model could have huge implications
• 15:55
  for the entire health care industry,
• 15:57
  especially for integrated health systems.
• 16:00
  Wow. That's amazing to hear and, you know,
• 16:02
  really, truly an exciting field of study, as
• 16:05
  you mentioned.
• 16:06
  You know, having that opportunity not only to
• 16:08
  see on the clinical side
• 16:11
  a very, very clear, return on investment for
• 16:14
  patients who who, you know, want to learn
• 16:16
  more about what's going on with their bodies,
• 16:18
  but also just the opportunity
• 16:20
  to democratize the that knowledge and information and
• 16:23
  ability to,
• 16:25
  have access to these types of screenings. So
• 16:28
  I I think that's really, really cool. And,
• 16:30
  again, as you mentioned, just another example of
• 16:32
  how the technology can make a big difference,
• 16:34
  within the health care space and and the
• 16:36
  health of communities overall.
• 16:39
  Yeah. I definitely agree.
• 16:41
  When I started, you know, getting involved in
• 16:44
  the field of AI research
• 16:46
  about eight years ago, you know, things were
• 16:49
  already advancing really fast. And at that
• 16:52
  time, retina and radiology,
• 16:55
  were at the forefront of this revolution because
• 16:58
  the current iteration of, AI, as I mentioned
• 17:01
  earlier, the cutting edge, technique is deep learning.
• 17:05
  It is very good at two things. One
• 17:07
  is image analysis, and the second one
• 17:10
  is in understanding human language. But the ability
• 17:13
  to really analyze images
• 17:16
  at a level
• 17:17
  not possible before
• 17:19
  first came to our attention about eight years
• 17:21
  ago. So if you look at the trajectory
• 17:23
  of medical AI, the medical fields
• 17:27
  that were most,
• 17:29
  reliant on images,
• 17:31
  were the ones who really got into, this
• 17:34
  AI revolution early on, and these fields are
• 17:37
  radiology and ophthalmology.
• 17:39
  So we started doing all this research even
• 17:42
  before
• 17:43
  the relatively recent boom in large language models,
• 17:46
  which started about two years ago.
• 17:49
  It's amazing to hear. And it's just really
• 17:50
  cool to know and see that history, and
• 17:53
  I can imagine a continued opportunity for the
• 17:55
  evolution of the field in really, really meaningful
• 17:58
  ways. Doctor Liu, thank you so much for
• 17:59
  joining us on the podcast today. This has
• 18:01
  been so inspiring and a fun conversation for
• 18:04
  me, and I I look forward to connecting
• 18:06
  with you again soon and, just continuing on,
• 18:09
  the journey of the progress along along these
• 18:11
  lines. So thank you so much for your
• 18:13
  time today. Same here, Laura. It's been a
• 18:15
  pleasure. I look forward to chatting with you
• 18:17
  next time.