Latest / Becker’s Payer Issues Podcast / Why Domain-Specific AI Models Are Transforming Payment Integrity in Healthcare

Transcript

Why Domain-Specific AI Models Are Transforming Payment Integrity in Healthcare · Becker’s Payer Issues Podcast

• 0:00
  Hello, everyone. This is Erica Spicer Mason with
• 0:02
  Becker's Healthcare. Thank you so much for tuning
• 0:05
  into the Becker's Healthcare podcast series.
• 0:08
  So today, we're going to talk about how
• 0:09
  small domain specific language models are transforming healthcare,
• 0:14
  unlocking new opportunities for innovation,
• 0:16
  operational efficiency, and data driven applications
• 0:19
  across the payer and provider ecosystem.
• 0:21
  And joining me for this conversation, we have
• 0:23
  with us Gene German, the chief technology officer
• 0:26
  at Lyric.
• 0:28
  Gene, welcome to the podcast. It's so great
• 0:29
  to have you with us today.
• 0:31
  Hi. Thanks for having me. I'm Gene. I'm
• 0:34
  the chief technology officer for Lyric. I've been
• 0:37
  with Lyric for approximately three years.
• 0:41
  Prior to,
• 0:42
  joining the company, it's been in health care
• 0:44
  for another, seven years focused on the consumer
• 0:47
  tech.
• 0:48
  When I joined Lyric, the company already
• 0:52
  had decades
• 0:53
  of service to our clients and has built
• 0:56
  trust,
• 0:57
  over these decades through delivering of value,
• 1:00
  to those clients,
• 1:02
  you know, through our products.
• 1:04
  And so, we already had, like, a really
• 1:06
  good place to start at,
• 1:09
  But there was a challenge, that we faced,
• 1:11
  and that challenge was the bulk of our
• 1:14
  software
• 1:15
  was hosted on prem,
• 1:18
  in the customer's,
• 1:19
  data centers.
• 1:20
  And not only is it difficult to deliver
• 1:23
  AI solutions,
• 1:24
  when your software
• 1:26
  is hosted on prem, but also any other
• 1:29
  capabilities,
• 1:31
  that we wanted to give to our customers,
• 1:33
  because, typically, upgrade cycles,
• 1:36
  happen every two to three years, for our
• 1:38
  software in the past.
• 1:40
  And so we've built our platform, Lyric 42,
• 1:44
  that gave us kind of the pipes and
• 1:46
  the capability,
• 1:47
  to deliver,
• 1:49
  multiple products,
• 1:50
  across a whole spectrum of payment integrity,
• 1:53
  solutions, not just, in primary,
• 1:56
  but also our secondary content.
• 1:59
  We've acquired a company called,
• 2:02
  ClaimShark,
• 2:03
  with Replay and,
• 2:05
  Virtusa,
• 2:06
  replay platform and Virtusa products on top of
• 2:09
  it
• 2:09
  as well,
• 2:10
  which allowed us to deliver manual reviews. None
• 2:13
  of that will be possible with our without
• 2:15
  our Lyric 42 platform.
• 2:19
  Well, Jean, that's great to learn more about
• 2:20
  you and also Lyric's platform.
• 2:23
  Great context setting for this conversation because we
• 2:26
  are gonna go,
• 2:27
  a little bit deeper on some of the
• 2:28
  trends that we're seeing here with small language
• 2:31
  models across the industry.
• 2:33
  We know that SLMs are enabling
• 2:35
  really this new wave of purpose built domain
• 2:38
  specific applications
• 2:40
  across industries and including health care. So from
• 2:43
  your perspective, what makes health care uniquely positioned
• 2:45
  to apply these tools effectively right now?
• 2:49
  Well,
• 2:49
  AI generally
• 2:51
  can be very useful in working
• 2:53
  with,
• 2:55
  very large datasets.
• 2:57
  And AI can certainly drive innovation. And in
• 3:01
  health care,
• 3:02
  there's plenty of that. There's,
• 3:04
  datasets
• 3:05
  of claims,
• 3:06
  member data, provider data, eligibility,
• 3:09
  and so on and so forth. And that
• 3:11
  data comes from various different sources.
• 3:14
  It varies in formats. It varies in levels
• 3:17
  of completion
• 3:18
  and many other attributes.
• 3:20
  And so
• 3:21
  we see that LLMs are, very good at
• 3:24
  dealing with this,
• 3:25
  high textual
• 3:27
  complexity.
• 3:28
  We also see that the small language models,
• 3:30
  which Lyric tends to use,
• 3:33
  are at least as good or better than
• 3:35
  large,
• 3:36
  language models
• 3:37
  in,
• 3:38
  solving problems when those,
• 3:41
  small language models are grounded in specific domains.
• 3:45
  We see them being cheaper,
• 3:47
  which is as important as being faster or
• 3:49
  more accurate when grounding them in the specific
• 3:52
  domain knowledge.
• 3:53
  Jean, thanks so much. It's been really helpful
• 3:55
  getting your perspective on
• 3:57
  what makes health care kind of unique in
• 3:59
  this space as they're apply as leaders are
• 4:02
  applying tools.
• 4:03
  At the same time, what unique challenges do
• 4:06
  these technologies introduce for health IT systems?
• 4:10
  Sure.
• 4:11
  Well, being in the health care industry for
• 4:13
  the past,
• 4:14
  decade or so,
• 4:15
  we typically see,
• 4:17
  health care IT comp
• 4:19
  groups,
• 4:21
  focus on deterministic,
• 4:23
  type of algorithms. So typically, for a given
• 4:25
  set of inputs,
• 4:27
  there's a very predictable way to see a
• 4:29
  given set of outputs.
• 4:31
  And LLMs,
• 4:32
  by definition,
• 4:34
  they're much more nondeterministic.
• 4:37
  And so
• 4:38
  it is very important as you think of
• 4:40
  the use cases to which you're gonna be
• 4:42
  applying AI that those use cases kind of
• 4:44
  combine this concept of determinism and, nondeterminism,
• 4:48
  and they deal with a very high cost
• 4:50
  in health care of failure.
• 4:52
  And, typically, that involves having humans in the
• 4:55
  middle,
• 4:56
  having well defined workflows in which the models
• 4:58
  are trained and given feedback as well as
• 5:01
  validated,
• 5:02
  and so on and so forth.
• 5:04
  Mhmm.
• 5:05
  Yeah. And I know another important piece of
• 5:08
  driving innovation
• 5:09
  or at least driving it successfully is data.
• 5:13
  So how should organizations think about leveraging their
• 5:15
  own data to build differentiated applications on top
• 5:19
  of emerging SLM platforms?
• 5:22
  That's a great question. Data is indeed very
• 5:24
  critical in driving innovation, and so it is
• 5:27
  very important, first and foremost,
• 5:30
  the the design tracking and usage of data
• 5:33
  is not an afterthought.
• 5:34
  There's many different kinds of data in health
• 5:37
  care as I mentioned,
• 5:38
  but also, you know, some data, like you
• 5:40
  said, is generated
• 5:42
  by the health health care organization
• 5:44
  itself,
• 5:45
  but some data is received from,
• 5:48
  members or from providers.
• 5:51
  And so all of that data can be
• 5:53
  used to drive,
• 5:55
  improve areas such as efficiency, quality,
• 5:58
  outcomes, and speed.
• 6:00
  The key here is to make sure that
• 6:02
  you focus on the goals,
• 6:04
  upfront, and you define the use cases
• 6:07
  around those goals to make sure that everything
• 6:09
  from data rights
• 6:10
  to data pipelines,
• 6:12
  is, thought through clearly in the beginning. And
• 6:15
  you bring in your, you know, clients,
• 6:18
  along with you around the value that is
• 6:20
  being driven and the usage,
• 6:22
  of data.
• 6:24
  And it's true for your own data as
• 6:26
  well.
• 6:27
  When you combine AI technology and people together,
• 6:30
  you can achieve very many different goals leveraging
• 6:33
  this new tech.
• 6:35
  Yeah. And, Gina, I've heard you emphasize a
• 6:36
  few times now the importance of getting really
• 6:39
  clear on those use cases. So I appreciate
• 6:41
  you,
• 6:42
  underscoring that as a consideration for our listeners.
• 6:46
  And I know at Lyric, you focused on
• 6:48
  payment integrity and claims analytics, which you mentioned
• 6:51
  in some of your introductory remarks.
• 6:53
  Where are you seeing the most meaningful efficiency
• 6:56
  gains today across the payer provider ecosystem in
• 6:59
  this realm?
• 7:01
  Yes.
• 7:01
  For us,
• 7:03
  because we've
• 7:04
  deployed Lyric 42, and there's a very high
• 7:07
  adoption from our client base,
• 7:10
  across the client base of our platform,
• 7:13
  it opens up a whole new set of
• 7:15
  avenues. So we could have done a lot
• 7:16
  of different things, but we've decided
• 7:19
  to focus on four key areas,
• 7:22
  for us. So one is the
• 7:24
  ideation building and validate validating the payment integrity,
• 7:29
  content. So this is our bread and butter,
• 7:31
  both primary and secondary content.
• 7:33
  Second is the adoption of that content. Once
• 7:35
  that content is built and we've seen significant
• 7:38
  increases in how quickly and accurately,
• 7:40
  we can deliver that content leveraging,
• 7:43
  AI technology,
• 7:44
  it's the adoption and deciding how to prioritize
• 7:47
  it
• 7:49
  with our clients and figuring out which content
• 7:52
  will cause more abrasion, less abrasion, and, you
• 7:54
  know, focusing on the holistic,
• 7:56
  health care system.
• 7:58
  Three, it is, like I mentioned,
• 8:01
  a replay product,
• 8:02
  which is a manual review product integrated with
• 8:05
  Lyric 42, our platform.
• 8:08
  Now that we're able to,
• 8:10
  roll it out much quicker to our customers,
• 8:13
  this new product,
• 8:14
  We are leveraging AI,
• 8:17
  for prioritization automation
• 8:19
  and
• 8:20
  the summary
• 8:21
  and,
• 8:23
  search of very complex information that's required,
• 8:26
  for those reviews, whether it's
• 8:28
  information related to eligibility, like, for coordination of
• 8:30
  benefits or itemized bill review or pulling a
• 8:34
  chart.
• 8:35
  AI is definitely helpful in putting all those
• 8:37
  tools into the fingerprint
• 8:39
  fingertips
• 8:41
  of people that are leveraging our software.
• 8:44
  And then fourth,
• 8:46
  last but not least is,
• 8:48
  leveraging our Lyric 42 pipes, and the learnings
• 8:51
  that we get, to
• 8:53
  create new products and test those products for
• 8:55
  our customers because our our clients are constantly
• 8:57
  asking us,
• 8:59
  for new, products that improved,
• 9:02
  their claims flow.
• 9:05
  Yeah, Jean. Again, thank you for for that
• 9:07
  helpful overview. And I'm curious as you're helping
• 9:10
  organizations
• 9:11
  roll out some of these tools
• 9:13
  in the payment integrity and claims space, have
• 9:16
  there been any outcomes that have surprised you
• 9:18
  or maybe even key lessons that you think
• 9:20
  leaders should take away from that work?
• 9:23
  Yes. You know, there are certain areas we
• 9:26
  expected,
• 9:27
  to improve such as
• 9:29
  speed, accuracy, and the outcomes and the value
• 9:32
  prop our technology is driving.
• 9:35
  What one area that we're kind of surprised
• 9:37
  is how much how much quality,
• 9:40
  we can drive with AI,
• 9:42
  even on top of already, like, high quality
• 9:45
  products. So not only are we accelerating the
• 9:47
  throughput,
• 9:48
  but we're also
• 9:49
  accelerating the way that we deliver quality at
• 9:52
  a level that would be just not possible
• 9:55
  without this technology.
• 9:57
  Yeah. So good to know, Jean. Thank you.
• 10:00
  And as we kind of close our conversations
• 10:02
  today, something I wanted to make sure we
• 10:04
  touched on was sustainability
• 10:07
  with some of these tools. So at Becker
• 10:09
  as we hear about leaders
• 10:11
  implementing
• 10:12
  new tools
• 10:13
  such as what we've touched on today with
• 10:15
  SLMs,
• 10:16
  but I think the struggle is in maintaining
• 10:19
  those those tools, making sure that they're driving
• 10:22
  continuous improvements and continuous outcomes that they're aiming
• 10:25
  for and really ensuring that they're going from
• 10:28
  pilot to real scale.
• 10:30
  So as more developer tools come to market,
• 10:33
  for you, what does a practical road map
• 10:35
  look like for for leaders who are aiming
• 10:37
  to balance innovation, trust, and measurable outcomes?
• 10:41
  Yes. That's a great question,
• 10:43
  and I I think it's a challenge for
• 10:45
  everyone.
• 10:46
  It's really important not to get distracted on
• 10:49
  every shiny thing that comes to the market.
• 10:50
  As engineers,
• 10:52
  We tend to do that,
• 10:54
  more often than not. So, it is also
• 10:56
  important to understand that AI is not necessarily
• 10:59
  an answer for everything. It is a very
• 11:01
  powerful set of tooling.
• 11:03
  But whether it's new models or a new
• 11:06
  set of tools, the key here is to
• 11:08
  focus on the problem you're solving versus the
• 11:11
  tool itself.
• 11:12
  So start by cleanly outlining the objectives and
• 11:16
  important KPIs. You know, for us, those were
• 11:18
  the four areas,
• 11:20
  that I mentioned. But for a different company,
• 11:22
  maybe a different set of priorities. And,
• 11:25
  these KPIs will help you measure along the
• 11:27
  way and not, you know, get you distracted
• 11:30
  from what success looks like. And a good
• 11:33
  road map would look like,
• 11:35
  especially in AI,
• 11:37
  where you laser focus on your key use
• 11:39
  cases, where you believe that AI is the
• 11:42
  right solution based on the attributes of those,
• 11:45
  use cases. And road maps we typically see
• 11:48
  is that are successful
• 11:49
  is where you start small.
• 11:51
  You prove it out. We started,
• 11:54
  in kind of a single space state Medicaid
• 11:56
  that was extremely difficult,
• 11:58
  to run-in quality,
• 12:00
  without these tools, and we were extremely successful
• 12:03
  in that. And we've expanded those results,
• 12:06
  to other areas. And then sometimes, you know,
• 12:09
  you may fail. You wanna abandon it and
• 12:11
  not kind of continue
• 12:13
  doing the same thing,
• 12:14
  if you're not seeing a progress there. And
• 12:17
  if it works, then the idea is you
• 12:20
  scale it by kind of building the foundations
• 12:23
  that are needed, then you expand either to
• 12:25
  other populations or to larger sizes of data
• 12:27
  or to kind of adjacent,
• 12:30
  use cases.
• 12:31
  But if you're
• 12:32
  kind of responsible looking at the roadmap that
• 12:34
  way, then more often than not, you're gonna
• 12:37
  be successful.
• 12:39
  Such sound advice, Jean. Thank you. And I
• 12:42
  know we've covered a lot of ground today,
• 12:43
  but wanted to just check-in to see if
• 12:45
  there was anything that we haven't covered that
• 12:47
  you think is important for health IT leaders
• 12:49
  to understand right now.
• 12:51
  Yes. We talk a lot about the value
• 12:55
  that AI can drive in certain circumstances.
• 13:00
  We haven't talked,
• 13:02
  much around the foundation
• 13:04
  that's needed,
• 13:05
  to be able to build
• 13:07
  AI successfully.
• 13:16
  Unfortunately,
• 13:17
  AI is expensive, and it's kind of very
• 13:19
  hard to do it at the side of
• 13:20
  your desk to do it at a scale,
• 13:22
  especially in health care. You need to make
• 13:24
  investments into your governance programs.
• 13:26
  You need to be able to make investments
• 13:28
  into technology,
• 13:30
  tracking both data and model drifts,
• 13:33
  automatically
• 13:34
  and be able to monitor and alert on
• 13:36
  those, when they happen to make sure that
• 13:39
  your models
• 13:40
  change successfully over time.
• 13:42
  And, setting the AI controls to prevent things
• 13:45
  like data poisoning, dealing with edge cases,
• 13:48
  building, like, the right
• 13:49
  user experience to be able to provide,
• 13:52
  feedback and train your AI algorithms and then
• 13:55
  have kind of human in the middle.
• 13:58
  And, you know, that differentiates going from a
• 14:01
  POC to production ready
• 14:04
  solutions. It's somewhat maybe kind of some of
• 14:07
  the boring stuff, not as exciting as the
• 14:09
  ROI. But without
• 14:11
  that boring stuff, you're not gonna be able
• 14:13
  to scale.
• 14:15
  Yeah, Gina. It's a really great point. I've
• 14:17
  I've noticed this through line throughout our conversation
• 14:20
  that going back to the drawing board, really
• 14:22
  establishing the use cases you're looking for, getting
• 14:25
  clear on those, getting clear on KPIs, but
• 14:27
  also going back to the foundational systems that
• 14:29
  are in place.
• 14:30
  Those steps have to be done before success
• 14:33
  can be expected.
• 14:34
  Absolutely.
• 14:35
  Well, Jean, it's been great having you on
• 14:37
  the podcast today. I wanna thank you again
• 14:39
  for your time and your insights,
• 14:41
  and we'd also like to thank our podcast
• 14:43
  sponsor for today, Lyric.
• 14:45
  Listeners, be sure to tune in to more
• 14:47
  podcasts from Becker's Healthcare by visiting our podcast
• 14:50
  page at beckershospitalreview.com.