Latest / CareTalk: Healthcare. Unfiltered. / Closing the Healthcare Admin Gap with AI w/ Abdel Mahmoud

Transcript

Closing the Healthcare Admin Gap with AI w/ Abdel Mahmoud · CareTalk: Healthcare. Unfiltered.

• 0:00
  Dr.
• 0:00
  Abdel Mahmoud came to the UK as a refugee as a boy, and then he became the youngest infantry officer in the country and then a physician.
• 0:08
  However, the mountain of paperwork soured him on medical practice.
• 0:13
  So now as CEO of Anterior, he wants to make prior authorization and other administrative hassles invisible.
• 0:19
  Can he do it?
• 0:26
  Hi everyone, I'm David Williams, president of Health Business Group.
• 0:29
  And I'm John Driscoll, the chairman of Waystar.
• 0:31
  Well, before we start talking about eliminating the administrative hassle, let's talk about the month of December.
• 0:37
  Some people think wrapping up in a blanket with a mug of hot chocolate or watching a movie with family is the best way to spend the last month of the year.
• 0:45
  For others like John, it's ice fishing or tobogganing or even riding a luge.
• 0:49
  Well, speaking of comfort, therapy is a great way to bring yourself peace that never goes away even when the season changes.
• 0:55
  BetterHelp offers entirely online therapy that's designed to be convenient, flexible, and tailored to fit your schedule.
• 1:01
  Just fill out a brief questionnaire to get matched with a licensed therapist, and you can switch therapists at any time for no additional charge.
• 1:08
  It's helpful for learning positive coping skills and how to set boundaries, and it empowers you to be the very best version of yourself.
• 1:16
  So whether you're dealing with stress, anxiety, or seeking personal growth, BetterHelp connects you with a licensed therapist who can support you on your mental health journey.
• 1:25
  Make it a great season with BetterHelp.
• 1:27
  Visit betterhelp.com slash caretalk to get 10 % off your first month.
• 1:31
  That's betterhelp, H-E-L-P dot com slash caretalk.
• 1:36
  Dr.
• 1:37
  Mahmoud, welcome to CareTalk.
• 1:39
  No, excited to be here.
• 1:40
  So I want to know how you game the system to be the youngest infantry officer.
• 1:45
  mean, how did that work?
• 1:47
  Yeah, and I was going to say with that generous background that gave, it kind of makes me look like I can't decide what job I actually want to do.
• 1:52
  That's the way my parents put it.
• 1:55
  How did I, I think I just applied before going to university in UK.
• 1:59
  I kind of applied to go to San Perce.
• 2:01
  And just so folks have context, San Perce is the equivalent of West Point for the United States.
• 2:06
  It's you and Winston Churchill.
• 2:08
  Yes, although he wasn't in my class, though, just to be clear for any of the viewers following just by audio.
• 2:13
  Well, that's what that would make you very young.
• 2:16
  it would.
• 2:18
  But yes, yeah.
• 2:20
  Now, I thought for moment you're going to ask me, was it West Point better than San Perce or San Perce better than West Point?
• 2:24
  I'm glad we didn't go there.
• 2:26
  yeah, no, I applied to go to military.
• 2:29
  I kind of thought it will teach you some life skills.
• 2:31
  It definitely did.
• 2:32
  I thought you had to kind of also stay up throughout the night and execute and kind of perform on very low energy reserves and everything else.
• 2:38
  And it's actually been one of the most helpful skills as I kind of started on the founding journey.
• 2:43
  But yeah, it was just more of kind of something to prove.
• 2:46
  Well, I would like you to continue this because David's the only one on this who hasn't been through infantry training.
• 2:50
  And so he's feeling more more insecure, but David, perhaps you'd like to talk about the clinical journey.
• 2:55
  Well, I was going to broaden it a little bit, John, because it's certainly impressive.
• 2:58
  And I was what I was trying to do with that intro really is just to compress a bunch of things that had happened.
• 3:03
  I mean, even from the time you were, you know, a child and my broader question is you sort of your personal journey.
• 3:09
  How did that lead you to this point?
• 3:10
  How did it all come together?
• 3:12
  Yes.
• 3:13
  So I think, I think, you know, After one or two years, I did have to leave the military with an injury.
• 3:17
  So I kind of had to go pick a profession and luckily had some good grades.
• 3:20
  So went to medical school, which in the UK is an undergraduate thing, right?
• 3:23
  So you can start pretty early.
• 3:26
  But I think like many physicians, think you very quickly have this realization of going in, you want to save lives and make an impact.
• 3:33
  Then you very, very realize very quickly the way health gets set up, the impact you could have was less about your clinical skills and actually more about the environment and the tools you had access to.
• 3:42
  I think that realization set in very quickly.
• 3:44
  I've been kind of doing stuff with computers growing up, so I had a bit of a knack for technology and programming, and I felt, wait, you're calculating morphine doses by hand.
• 3:52
  I can build a calculator and we can use that.
• 3:54
  And that got me started down the kind of technology and healthcare route.
• 3:59
  Took a year out to do a computer science master's at the same university, and that's when it really started.
• 4:03
  The avalanche couldn't be stopped, so to speak.
• 4:05
  I was going to disappoint my parents and never actually practice clinical medicine.
• 4:09
  I joined, I spent some time at Facebook, or now it's called Meta.
• 4:13
  applying some healthcare technologies, was chatbots five years ago.
• 4:16
  So you can imagine they were pretty terrible and gimmicky.
• 4:19
  Then joined Google for a stop.
• 4:22
  But what I saw at Google was what led me to what I'm doing now at Anterior, kind of tackling health administration and the burden of paperwork.
• 4:29
  And the main thing was, when I was at Google, I wasn't working on this directly, but there was this corner of Google that was excited about this thing called large language models.
• 4:37
  And I remember there was one thing about them in particular that was fascinating, which was their ability to handle huge volumes of unstructured data.
• 4:45
  I kind didn't understand too much of the other stuff, but that bit felt interesting because if you looked across all of healthcare, it wasn't that healthcare wasn't digitized, it was digitized, and it had a lot of software, but the software didn't speak the same language.
• 4:56
  It was a lot of unstructuredly structured data where we kind of rely on expensive humans as the interoperability glue.
• 5:02
  And it felt like there was going to be a moment where something could change.
• 5:05
  was with this kind of large language model where you could plug that in and to top of all the investments we've made in IT and infrastructure.
• 5:11
  and maybe finally have end-to-end workflows that are fully automatable.
• 5:15
  And that felt like it was a moment.
• 5:16
  And that's kind what led to the founding journey.
• 5:18
  At least the idea of it.
• 5:20
  It turned out there's a lot more discoveries and pivots and iterations.
• 5:22
  So Abdul, for those who aren't kind of versed in computer tech, who work with Greg Corrado, all the other great talent that Google had and don't understand that Google really started with a view with a...
• 5:39
  The founders believed that if they grabbed enough information, you could actually make artificial intelligence or insights at scale and speed that we can't imagine.
• 5:48
  Rather than search was the plan, that this large language model stuff sort of makes sense.
• 5:56
  But can you explain what a large language model is?
• 5:58
  Yeah, there's lot of ways you can slice this question.
• 6:01
  There's probably going to be much better explanations out there.
• 6:03
  But the way I think about it is AI has kind of been thrown around for a while.
• 6:07
  And then large language model is a subset of four.
• 6:09
  Since 1956.
• 6:10
  I heard somewhere that AI is basically just the edge of what computers can do today, right?
• 6:16
  And it will continuously keep shifting.
• 6:19
  But yeah, so like I think AI has been in healthcare all the while for a while, right?
• 6:23
  And I think those kind of applications generally be what's considered narrow now, which what it means is you kind of have a question you want or something you want AI to do, maybe, you know, recognize hot dogs.
• 6:33
  Right, computer vision.
• 6:34
  So you train it on previous examples of data that's been labeled with hot dog and not hot dog, for example.
• 6:39
  That's kind of a classic example from TV.
• 6:41
  And you get these narrow systems.
• 6:43
  But the problem is, like the things that actually humans often encounter is like general problems where you have to kind of pull on a few different things and reason through it and break it down.
• 6:51
  It won't match necessarily always the labeled data, right?
• 6:55
  Everything we do in our stuff, like life that's new will match previously labeled data, but we still are intelligent.
• 7:00
  And I think the thing with large language model was What if we didn't bake in these heuristics or control it and just let it loose on all the data that exists, trillions and trillions and trillions of gigabytes, and get the system to almost learn the heuristics and the rules and the patterns and what we can call intelligence?
• 7:17
  And if you do that, the humongous amounts of compute and data training volumes and length of time, you get something out of there that what is now chat GPT and feels eerily human-like.
• 7:27
  It's not human, but it feels like it's intelligent in...
• 7:30
  you can reframe a question a million times and won't break.
• 7:32
  Whereas narrow AI, for example, it would break if it wasn't in the right format and the right structure and what it seemed before.
• 7:38
  I often kind of use that as the medium between narrow and general and large language models more general.
• 7:43
  So let's talk about what you're doing with Anterior and I get the kind of the general view of it, but let's go beyond that.
• 7:50
  So we talk about this, this $950 billion burden of healthcare administration.
• 7:55
  I guess we'll be rounding it up to a trillion by shortly.
• 7:58
  So we've got that, I think people know it, it's big, there's a lot of frictions involved in it, and you've got this idea that you're gonna make things like prior authorization invisible.
• 8:07
  Now, if I take what we just talking about, the large language models, I know that two years ago, I couldn't have said, hey, write a report about this, that, or the other.
• 8:14
  Now, anybody could go to ChatGPT and get like a reasonable draft from that.
• 8:18
  Like it's completely night and day.
• 8:20
  Are we talking about something similar here when you're dealing with some of these things like...
• 8:24
  prior authorization, is the first use case that you've picked.
• 8:27
  Can we actually see something comparable where it goes to being from this incredibly thing you could never deal with to something that's actually simple?
• 8:35
  Yes, not immediately, but definitely, yes.
• 8:38
  We're very, very optimistic about that.
• 8:40
  you know, the 950- if you weren't, you wouldn't have raised money.
• 8:44
  Yes, yeah, you have to promise optimism, I think.
• 8:46
  have to almost to the point where you believe it.
• 8:48
  But no, I do think that- know, this burden of healthcare administration has been going on for almost too long.
• 8:53
  It has to, it has to be close, you know, the solution must be nearby.
• 8:56
  But I think what it is, if you take a step back, right, this 950 billion, it's not just pure waste, right?
• 9:02
  It's actually people doing work.
• 9:04
  If you look at it, it looks like economic output or activity.
• 9:07
  But I think if you look at the break it down, right, there's many different areas you can slice administration.
• 9:11
  Some of it was on the provider side, some of it was on the payer side.
• 9:14
  Some of it straddles both.
• 9:15
  And I think prior authorization where we started felt like it was a pain point for both sides.
• 9:20
  There was a lot of regulation coming in to try and actually get us out of this.
• 9:23
  So felt like a great place to start.
• 9:25
  So we'll use the example of prior authorization.
• 9:28
  Prior authorization means prior permission from someone.
• 9:32
  So a lot of people here may unfortunately be quite well versed in prior oath, but you want to get a knee surgery or something, a knee arthroscopy.
• 9:39
  It's a $30,000 procedure.
• 9:41
  So your physician, your provider may want to ask your insurer, are they OK with going ahead?
• 9:45
  What the insurer is trying to do is hey, is this medically necessary spend?
• 9:49
  Have you tried conservative therapy that is often successful at much better rates?
• 9:54
  Sorry, like much better outcomes for you and much cheaper.
• 9:57
  So it's actually a good thing, right?
• 9:58
  But the problem I think comes with prioritization is how does the insurance company get to understand if something's medically necessary or not?
• 10:05
  And especially if the medical information they're being sent over is a 200 page fax PDF, right?
• 10:12
  So the only way you can do that today is to hire really expensive jargon understander called a doctor or a nurse or pharmacist to sit down reviewing faxes as a full-time job.
• 10:22
  And I think that's where the expense comes from.
• 10:24
  Because if AI could do it, it wouldn't cost as much.
• 10:26
  It would happen in a fraction of a second.
• 10:28
  And I think the thing we're working on is AI doing it.
• 10:31
  By the way, David's really good at jargon.
• 10:32
  Just if you need a human to kind of be human in the loop for the agent.
• 10:39
  As you think about what you're trying to do, what provides the specific application intelligence that solves the managed care riddle.
• 10:52
  The large language models were all trained on generalized, largely internet or print data.
• 10:59
  Often, you know, the game in managed care is figuring out how do you automate something where even the logic changes on a regular basis?
• 11:07
  Yes.
• 11:07
  I think there are a few components to that, to that question.
• 11:09
  So we'll tackle them kind of bit by bit.
• 11:11
  The first component is like, how does large language models be able to deal with the jargon?
• 11:15
  So mainly, yes, most of the volume of the data on the internet is non-medical, should we say, maybe 1 or 2 % is.
• 11:22
  And the thing is surprisingly, actually, even with that amount, you can get chat GPT passing the US of India than if you saw some of these kind of scary sounding things.
• 11:31
  GPT 4 did a nice job.
• 11:33
  Each generalized model upgrade from chat GPT for an open AI gets materially better.
• 11:41
  The more recent paid version actually passed the...
• 11:44
  the general board, medical board for doctors.
• 11:47
  Yeah, if you look at the USMG questions, they're not easy questions.
• 11:49
  They're actually pretty, pretty hard.
• 11:51
  And whether someone gets a knee arthroscopy or not is actually a bit more of a simpler question.
• 11:56
  So the question is not a can the large language models understand medical jargon?
• 12:01
  They already can.
• 12:02
  Whereas previous attempts at AI, had to kind spend a lot of time just training it on medical records, medical data.
• 12:07
  At this point, you can take a large language model off the shelf.
• 12:10
  And I think some insurance companies have tried to build some of this in-house.
• 12:13
  So that's the first part.
• 12:14
  The second component comes from, like you talked on, John, which is the guidelines change very frequently.
• 12:19
  I mean, there's a scary statistic, I think, Harvard Business Review that medical literature doubles every 72 days.
• 12:25
  And most physicians are out of date, actually.
• 12:27
  average...
• 12:28
  And it takes about 17 years before when the Academy actually agrees on a novel innovation that it actually becomes standard of practice across an entire specialty.
• 12:36
  Exactly.
• 12:37
  But the real-world implication of guidelines changing so frequently is we never digitize them because the minute we digitize them...
• 12:42
  because building rules before was so complex and it involved a lot of expensive labor, you just print it into PDF and give it to a nurse to apply to a PDF record, right?
• 12:51
  And what we've said, actually, you know, the dirty secret law of AI companies is it's not the reasoning part.
• 12:56
  It's actually if you can take that data and make it digitize and structure it to feed it into a system is where you take a lot of the work.
• 13:03
  Because if you look at a nurse, right, you know, answering a question of, you know, is there evidence of the failure of conservative therapy for at least six weeks?
• 13:09
  is not a hard clinical reasoning task, it's an information finding task.
• 13:12
  And if you have 300 pages, a lot of your brain power is not being used in thinking, it's being used in finding.
• 13:18
  But large language models are very good at that.
• 13:20
  They can find those instances in a second, and maybe they may not be able to fully reason yet.
• 13:24
  And I think they're making good breakthroughs there.
• 13:26
  That in itself leads to the results that we see today where a nurse goes from an average of about 10 to 12 reviews a day to now 20 to 30 reviews a day.
• 13:36
  Cycle time reductions of 77%.
• 13:38
  And all that we're doing is just helping them find information in the facts, right?
• 13:42
  It's not even that breakthrough.
• 13:43
  Obviously a lot of engineering goes into how do you structure that facts and make sure there's no hallucinations and things like that.
• 13:48
  But that's how low hanging the fruit is.
• 13:51
  So I noticed that your team seems to be quite clinician heavy, which surprised me a little bit because I was trying to understand what part of the problem you're going at.
• 14:00
  So I'm interested in both.
• 14:01
  your philosophy on the kind of people that you're hiring?
• 14:04
  And then also, what do you add from a technology standpoint, given that the LLMs are out there and continue to advance at good speed?
• 14:12
  Absolutely.
• 14:13
  And I think the main way we think about this is from what are we doing?
• 14:17
  Where do we live in the stack of AI houses, foundation AI companies on one side, and a United Healthcare or a Kaiser Permanente, a completely healthcare company on the other side.
• 14:27
  And we're trying to bridge the gap.
• 14:29
  of the frontier of technology into solving domain, kind of real-world domain.
• 14:33
  And you can't build technology without having the domain expertise, right?
• 14:36
  And usually what happens with lot of companies in last 10 years is they're often in silos, so they may bring clinicians and engineers and put them apart.
• 14:43
  So initially, actually, a first one or two, three hires on the engineering front, some of them had MDs who then retrained as computer engineers and worked as machine learning engineers.
• 14:52
  It's not a scalable hiring strategy.
• 14:54
  Trust me, we've tried.
• 14:55
  There are a few that you hand select, but it meant we set a culture.
• 14:58
  where engineers and clinicians work side by side.
• 15:00
  Today, our engineering team will have nurses, utilization management nurses, in the office next to them, working together, and then each being comfortable with each other's jargon, so to speak.
• 15:10
  And I think that's been really important in us to map the domain, to map the technology to the domain, that leads us to kind of outperform.
• 15:17
  And I think that's our specialty, is to bridge those two worlds.
• 15:21
  When you think about the models, the foundational models, the large mass of trillion that kind of monsters.
• 15:29
  Are you using one of those models and then tuning it or using rag?
• 15:33
  Maybe walk through how you're taking the powerhouse models of a meta, a Google, a Microsoft OpenAI, and then are you building a smaller version of it yourself or are you tuning one of those and applying it just so folks understand we're a little bit more technology forward, unlike David and I, kind of how it all works, how those pieces, those puzzle pieces fit together.
• 15:58
  to solve some pretty quotidian problems like how do you get paid?
• 16:02
  Yes, yes.
• 16:04
  You may not understand the technical jargon, but I didn't know the meaning of the word you just said.
• 16:07
  I'll happily say that.
• 16:10
  It was his pronunciation of it, I think.
• 16:13
  Exactly.
• 16:14
  It's a too few, too few years, a few many years of Latin.
• 16:17
  David, I got it, John.
• 16:19
  Yes.
• 16:19
  So, so yes, from a technical answer, I think those, the words you said there, it's actually a mixture of them.
• 16:26
  That's the answer.
• 16:27
  So the first approach that people will generally do, and sometimes even like innovation labs within an insurance company or a healthcare provider will be just to use a large language model provider off the shelf.
• 16:37
  What you'll find is for like writing marketing copy, website copy, whatever, you get good results, but actually you get pretty terrible results for anything more complicated, a bunch of clinical questions in a row and the errors stack up in a sense.
• 16:49
  So the next iteration on that was, okay, how do we use a mixture in kind of in a constellation?
• 16:56
  in sense of an orchestration, which is, hey, there's a model that specializes in structuring the medical record and extracting dates and extracting metadata.
• 17:03
  Another one generalizes or focuses on how do I structure guidelines to know what question to ask at the right time.
• 17:10
  Another one to just focus on question answering.
• 17:12
  Given a question and some evidence, what is the right clinical answer?
• 17:15
  So you get these abstractions on top of these large language models.
• 17:18
  And then you can do techniques such as, I think, RAG, retrieval of mental generation in particular, for those that don't know, it's It's getting a bit out of date because I think in healthcare you need a determinism, you need to be able to replicate questions.
• 17:30
  So you really don't want to be using that too much.
• 17:33
  I think it's good for MVP or proof of concept.
• 17:36
  But then there's other things like distillation.
• 17:38
  So once you've run a model multiple times, you can actually take all that knowledge in example of data you have to train a smaller model to be a hundred times more efficient, a hundred times faster, and even more accurate actually.
• 17:48
  So these are all the techniques we're employing.
• 17:50
  And to give you an idea.
• 17:52
  And that particular...
• 17:54
  Training the smaller model through distillation, think really is the future for applied specific models in things like healthcare.
• 18:01
  Absolutely.
• 18:02
  Because if you think about it, it's like this GPT-4, it's like it's got the world's knowledge in its head.
• 18:06
  And you need just a corner of that brain that's focused on neothroscopies or MRIs or something.
• 18:11
  And what you want to do is you want get that part out and then just ship away the rest.
• 18:15
  So you can be 1 % of the electricity users versus 100 % of the electricity.
• 18:20
  So who is Florence and what are you expecting of her?
• 18:23
  Yes, Florence.
• 18:24
  So there is a famous Florence called Florence Nightingale.
• 18:27
  And this is her namesake.
• 18:28
  So when we have our AI working, it's really just a constellation of multiple calls, like I said, maybe of the 50 different large language models working together.
• 18:37
  But sometimes when you're working with nurses and Florence is putting results, this AI is putting out results, we actually gave it a name.
• 18:45
  Because we found actually a lot of the users can almost bond with it in a sense of even when giving feedback, Florence made a mistake or the user experience because then there's dashboards and software.
• 18:53
  Sometimes you want to differentiate between the errors that the UX, like a bug of logging in and logging out or the PDF not loading, and actually AI reasoning, almost like a colleague making a mistake.
• 19:03
  So we've kind anthropomorphized it.
• 19:05
  I know there's a lot of opinions about anthropomorphization of AI, but we found that nurses actually love it.
• 19:10
  But there was this one moment where one of the nurses was out on leave for a week.
• 19:14
  And it's like, hey, just take email us the customer kind of helpline be like, Hey, just let Florence know I'm just out for a week and not to be too upset.
• 19:20
  I'm actually on holiday, right?
• 19:21
  It's really funny things like that.
• 19:22
  kind of make it more lovable.
• 19:24
  And to be fair, Abdul, that is actually one of the weirder aspects of robots as well as that have any sort of anthropomorphic or frankly like animal like robots or bots where there's any sort of conversation is they're quickly Comfortably anthropomorphized and and it's it's it's just is it's it's been very consistent across many cultures actually so so when you think about you know solving the Prior authorization problem which may sound small to those who are not involved in trying to solve it But it is a pretty material part of the ecosystem I will I said the bad news is chairman of way star that we do revenue cycle software I will tell you that the health care companies have insurance companies have gone from denying or complicating prior authorizations to now making them easier to get authorized and they're just denying the claims.
• 20:17
  How do you think about the future of, and I speak from like real-time experience with certain payers as we're just reprogramming our own algorithms.
• 20:29
  How do you think about the future of anterior?
• 20:34
  Like where do you want to go?
• 20:35
  That's a great question.
• 20:37
  I think I think even like, know, just to touch on recent sad news, right, with the UHC situation and then the way the public reacted.
• 20:46
  And I think, you know, you could easily say, well, that's, that's evil.
• 20:49
  And people shouldn't be reacting like that way.
• 20:51
  But there is a sentiment there.
• 20:52
  And I want to kind of touch on that sentiment of people are frustrated at end of the day, right?
• 20:55
  They feel like they're paying for care and that they're not getting right.
• 20:58
  And I think the way we see and why we get out of bed in the morning, right, is not to work on prior wealth for prior wealth.
• 21:04
  I think why we do that is how do we reduce the friction?
• 21:07
  I don't think providers are evil or payers are evil.
• 21:09
  I think it's very easy from a shadow level to be like, hey, that group of people there just want to deny care.
• 21:14
  I think it's more incompetence.
• 21:16
  It's more like there is a right answer that's very hard to get to.
• 21:19
  So you have to apply heuristic rules because you also have two options.
• 21:23
  spend, if you could hire a million nurses, payers for free.
• 21:27
  If you could hire a million nurses for free, payers would hire a million nurses to go through every single thing and make sure the right care is paid for.
• 21:32
  But that's just too expensive.
• 21:33
  So you have these kind of like weird rules or AI doing it that's pretty poor.
• 21:37
  What we want to do is, and that's why we believe in agentic AI, AI can have computers that feel like humans.
• 21:42
  It's almost like you've had a human go in detail through that person's medical context and medical necessity and giving a thoughtful answer to that, right?
• 21:50
  As opposed to just having a black box heuristic AI trained on a million examples like you, but not you, right?
• 21:57
  know, people from your zip code, your ethnicity, we're looking for this care and denying it.
• 22:00
  And just, know, folks are tracking adult.
• 22:03
  I mean, the, For those who aren't using the models, agentic AI, the agentic system is really thinking about bots like Florence as co-pilots or supports or leveraging them as sort of sitting beside the clinician, the decision maker, the patient, frankly, and helping them navigate, dominate, or liberate themselves from their healthcare journey.
• 22:33
  agentic world is a world where these models become or are formed or self-formed as partners in leveraging technology to bring intelligence to the point of care or insight.
• 22:46
  Yes, yes.
• 22:47
  I always use the analogy of credit cards.
• 22:49
  You go abroad somewhere nice and sunny and you tap your Apple Pay for a coffee.
• 22:54
  You get it instantly.
• 22:55
  The equivalent in healthcare is your bank saying wait three days, I'm gonna call the other bank and agree what currency and whether you have, know, whether this is forging or not and give you an answer in two weeks saying, you couldn't get the coffee, right?
• 23:06
  It sounds crazy because it is.
• 23:07
  And I think in healthcare, the ultimate goal is when a provider wants to do something, they get an instant answer about the health plans view on whether that's medically necessary or not.
• 23:15
  And 95 % of the cases are approved.
• 23:18
  It's just, it's a pain involves a bunch of humans, right?
• 23:21
  We could all be doing better things.
• 23:22
  So John just assumed away prior authorization, which is a huge problem.
• 23:26
  But you've also put on your roadmap where you're to go beyond that.
• 23:30
  And I noticed things like risk adjustment, care management, payment integrity.
• 23:35
  And I'm wondering, what's the logic behind that path?
• 23:37
  And these are all areas that actually, they have vendors in them, they have solutions.
• 23:43
  And how are you going to differentiate from what's out there?
• 23:46
  Exactly.
• 23:47
  The way we see it is often for a health plan, right?
• 23:51
  Health plan is a business really.
• 23:52
  They're not in the business of providing care.
• 23:54
  They're in the business of understanding risk and understanding their clinical population, right?
• 23:57
  And then can be able to interact that, right?
• 23:59
  And I think as a tearful, clinical data powers a health insurance company, right?
• 24:03
  To be able to do great decisions.
• 24:05
  Now, if that clinical data is unstructured, what happens is that clinical data usually comes in the first port of entry into a health plan is there prior authorization.
• 24:12
  And that gets just stored as facts in PDF.
• 24:14
  Maybe you give a decision to two weeks later about whether a prior order, but it just sits there.
• 24:19
  in a data lake or whatever for like two years before it gets sent off to a risk adjustment firm to do it or sent off to payment integrity.
• 24:25
  What we want to do is like how you turn that data into action immediately, the minute it comes, not just for that prior off so that the member or the patient gets a decision right, but it's like, hey, we also noticed that there's an AF diagnosis there.
• 24:36
  Do you want to go and flag that for care management and case management?
• 24:39
  How do you become kind of more proactive earlier?
• 24:41
  We noticed some missing codes that you haven't captured and that's the payers right, right?
• 24:45
  They've taken on that risk, they should be paid from the government, but not two years later, immediately.
• 24:49
  And then the other component to that is we don't see ourselves competing with existing software companies.
• 24:53
  There's a lot of systems of records and companies that do great work from case management software to core administrative platform.
• 24:58
  We see it as more of like trying to build the employees, right?
• 25:01
  Why are nurses and doctors- David and I may argue whether they're all doing great work or not.
• 25:06
  Yes.
• 25:07
  Some are, Yes.
• 25:09
  But would you wager that 80 % of it is like stuff that even the employees themselves would say, this is like, why am I doing this?
• 25:14
  Why am I logging into 30 different systems and reading all these PDFs, right?
• 25:18
  And I think that's the goal here is we're trying to build the employee that uses the risk adjustment suite, right?
• 25:23
  We're trying to build the agentic future for that.
• 25:26
  know, health insurance companies don't need to be 60,000 employee base, right?
• 25:30
  It doesn't make sense that it should be.
• 25:32
  John, last question to you.
• 25:34
  guess I've done like if you roll the tape five years forward, broaden the aperture beyond your company to perhaps the companies that are posterior to your anterior or next to you.
• 25:47
  What's the promise of AI in healthcare and, and, and if you were to paint a picture for those who don't know AI intelligence, healthcare rags, large language models, Nvidia chips versus us renting all the technical mumbo jumbo, what's the, what's the promise of AI and healthcare going to mean for patients and doctors and taxpayers?
• 26:12
  Yes.
• 26:13
  I think immediately it's what our mission is to reduce the burden by $150 billion of health administration.
• 26:19
  I think that that should go down to at least 200 billion to 300 billion.
• 26:22
  I'm sure there will be some, but it's that kind of 80 % chunk that is really just in humans as interoperability.
• 26:28
  A lot of it is labor.
• 26:30
  And I think a lot of that labor, by the way, we have a shortage of nurses and doctors.
• 26:33
  be doing a lot of this administration work.
• 26:35
  mean, just to put a point on that, we'll be about a million clinicians short over the next three to five years.
• 26:42
  Yeah, and the idea that some of them are just full-time jobs looking at faxes, right?
• 26:46
  Doesn't set right.
• 26:47
  So I think just from a high level, that's where it looks like.
• 26:50
  But I think for the patient, I think it should feel like, and I kind of find it funny, I'm using finance as an example, because often finance is seen as a very slow-moving industry, but it should feel like Apple Pay, right?
• 27:02
  It should feel like, I think care should feel like instant for providers, and just you get on with it.
• 27:07
  You can focus on the stuff happening with your care, right?
• 27:10
  It's just the more important stuff.
• 27:11
  not in paperwork and back and forth, the claim edits and prior authorization and so on.
• 27:15
  should help because you just feel like it works, right?
• 27:19
  That's it.
• 27:20
  Well, speaking of that's it, that's it for another episode of Cure Talk as well.
• 27:24
  Our guest today has been Dr.
• 27:25
  Abdelmahmoud, founder and CEO of Anterior.
• 27:28
  I'm David Williams, president of Health Business Group.
• 27:31
  And I'm John Driscoll, the chairman of the Waystar Corporation.
• 27:33
  If you liked what you heard or you didn't, we'd love you to subscribe on your favorite service.
• 27:38
  And thank you, Dr.
• 27:39
  Abdel.