Latest / Becker’s Healthcare Podcast / Building an AI Native Orthopedic Practice with Travis Doering, MD

Transcript

Building an AI Native Orthopedic Practice with Travis Doering, MD · Becker’s Healthcare Podcast

• 0:02
  - There's a version of AI in health care
• 0:03
  - that lives mostly in press releases,
• 0:06
  - pilot programs, partnership announcements,
• 0:08
  - theoretical use cases.
• 0:10
  - And then there's the version that lives in
• 0:12
  - the work, the CIO trying to get a
• 0:15
  - model into production without breaking a dozen workflows,
• 0:18
  - the revenue cycle leader weighing automation against accuracy,
• 0:22
  - the CISO who knows that every new integration
• 0:25
  - is a new attack service.
• 0:27
  - Becker's eleventh annual IT and revenue cycle conference
• 0:31
  - is built for the people doing that work.
• 0:33
  - More than 2,500
• 0:35
  - executive level attendees,
• 0:37
  - 600 plus speakers, and four days of sessions
• 0:40
  - designed around the real operational questions behind AI,
• 0:44
  - cybersecurity,
• 0:45
  - revenue cycle performance, and digital health.
• 0:48
  - Join us September 14 through the seventeenth.
• 0:51
  - For the agenda and event details, visit beckershospitalreview.com
• 0:55
  - and click on the events tab in the
• 0:57
  - upper right.
• 0:59
  - This is Carly Beam with the Becker Spine
• 1:02
  - and Orthopedics podcast. And joining me today is
• 1:04
  - doctor Travis Dohring, orthopedic surgeon based in Austin,
• 1:08
  - Texas.
• 1:09
  - Doctor Dohring, thank you so much for being
• 1:10
  - here today.
• 1:12
  - Yeah. Happy to be here. I, I'm excited.
• 1:15
  - Excited to have you on. And before we
• 1:17
  - dive in some some questions, I know we
• 1:19
  - have some great discussions about AI, coming up
• 1:21
  - in a couple seconds. But could you quickly
• 1:24
  - introduce yourself, share more about your work and
• 1:26
  - what you do?
• 1:27
  - Sure. And so I am a orthopedic hand
• 1:30
  - and upper extremity surgeon.
• 1:32
  - And
• 1:33
  - at the moment of this call, I am
• 1:35
  - opening my own new practice in about five
• 1:38
  - days.
• 1:39
  - So wish me luck with that.
• 1:41
  - But part of the reason why I'm opening
• 1:43
  - my own practice is to be able to
• 1:45
  - control,
• 1:47
  - everything,
• 1:48
  - in no small part,
• 1:50
  - the relationship between the practice and artificial intelligence.
• 1:54
  - I am designing this practice to be the
• 1:57
  - first AI native medical clinic in the country,
• 2:01
  - and excited to do that.
• 2:03
  - Awesome. We'll definitely circle back on that. But,
• 2:06
  - so I'll talk about AI. You know, as
• 2:08
  - AI workflows get increasingly
• 2:10
  - increasingly
• 2:11
  - more complicated,
• 2:13
  - cost of running them,
• 2:14
  - they continue to increase. So I'd love to
• 2:16
  - hear strategies that
• 2:18
  - you're thinking about when it comes to this
• 2:20
  - kind of, you know, strategies to what strategies
• 2:23
  - make sense to limit per task and per
• 2:25
  - token costs?
• 2:27
  - Sure. Part of the reason I started this
• 2:29
  - clinic is that I can tackle some of
• 2:31
  - those questions head on.
• 2:34
  - And so one of the things that I'm
• 2:35
  - most interested in is something called local AI.
• 2:39
  - And so this would be hardware that is
• 2:42
  - in your office, in your clinic, in your
• 2:44
  - hospital,
• 2:45
  - where the processing power and the GPUs have
• 2:49
  - gotten good enough that they can power the
• 2:52
  - majority
• 2:53
  - of,
• 2:54
  - daily workflows in the clinic, patient interactions,
• 2:58
  - billing, coding, revenue cycle management.
• 3:01
  - You know, I think we're really just kinda
• 3:02
  - at the tip of the iceberg
• 3:04
  - for the implementation
• 3:05
  - of AI in most practices.
• 3:08
  - But by
• 3:09
  - starting my own practice and keeping it limited
• 3:12
  - to just myself,
• 3:14
  - the the hardware constraints are much less. And
• 3:17
  - beauty of this is that by keeping all
• 3:20
  - the processing
• 3:21
  - local,
• 3:22
  - none of those tasks are getting sent off
• 3:24
  - to the cloud. And so, really, I just
• 3:28
  - once the hardware costs have been paid for,
• 3:30
  - pay for electricity.
• 3:32
  - And there's no subscriptions. There's no not anything
• 3:34
  - else.
• 3:36
  - And then kinda really the the second big
• 3:38
  - advantage of that is that everything is secure
• 3:40
  - because none of the data is leaving your,
• 3:44
  - local processors, local hard drives.
• 3:46
  - None of it's getting sent off to the
• 3:48
  - cloud. And so needing,
• 3:50
  - any of the agreements between yourself and a
• 3:54
  - AI lab,
• 3:56
  - it's just no longer a problem.
• 3:59
  - That's really interesting. So if I'm understanding correctly,
• 4:01
  - this isn't like using AI the way you'd
• 4:03
  - use, like, Claude or ChatGPT,
• 4:05
  - and that words, you know, being
• 4:07
  - all processed and sent off from those big
• 4:10
  - data centers. But you're describing something
• 4:13
  - more local, more central, even just like to
• 4:15
  - the within the office itself?
• 4:17
  - Yeah. I'm actually looking at my stack of
• 4:20
  - Mac Studios as we speak.
• 4:22
  - And so yeah. And so, you know, there
• 4:25
  - are consumer grade and then professional grade hardware
• 4:29
  - at kind of a variety of price points.
• 4:32
  - But that hardware has gotten good enough,
• 4:35
  - that if you are willing to pay a
• 4:37
  - little bit of a premium right now because
• 4:39
  - of the RAM shortages Mhmm. Mhmm. Instead of,
• 4:42
  - sending off the cloud, off to OpenAI, kinda
• 4:45
  - any of those services that you're used to,
• 4:47
  - you can essentially
• 4:49
  - replicate that exact functionality,
• 4:52
  - whether it's just as a chat interface or
• 4:55
  - whether it's doing more of the administrative and
• 4:57
  - back end tasks behind the scene,
• 5:01
  - entirely locally. And so,
• 5:03
  - again, none of your data is going to
• 5:05
  - one of those cloud providers.
• 5:07
  - You don't need to worry if any of
• 5:09
  - your patient's data is gonna get included in
• 5:12
  - future training sets.
• 5:14
  - And so I think that it's something that
• 5:16
  - is really helpful on my end, but I
• 5:19
  - think it also engenders
• 5:20
  - a higher level of trust with patients
• 5:23
  - because, you know, patients don't know where their
• 5:25
  - data is going. They don't know how it's
• 5:26
  - being used.
• 5:27
  - And so by keeping all that locally, not
• 5:29
  - only am I able to offset some of
• 5:31
  - the per token, per task costs,
• 5:35
  - but also able to provide a much higher
• 5:37
  - level of data security than competitors.
• 5:41
  - That's really cool. So it sounds like, you
• 5:42
  - know, a lot of advantages from you you
• 5:44
  - mentioned,
• 5:45
  - patient data privacy
• 5:47
  - cost, and I mentioned maybe there's even,
• 5:49
  - an energy
• 5:50
  - benefit compared to using one of the bigger
• 5:54
  - the bigger groups.
• 5:55
  - In some ways yeah. The, the energy consumption
• 5:59
  - on a per token basis is gonna be
• 6:01
  - higher with local hardware since it's just not
• 6:04
  - as efficient.
• 6:06
  - But the overall cost compared to spinning up
• 6:09
  - a whole GPU cluster, you know, overall,
• 6:12
  - it it should be less.
• 6:14
  - And I I think that that's something that
• 6:15
  - we're really gonna see as
• 6:18
  - more and more companies
• 6:19
  - shift away from a traditional
• 6:22
  - per user or per seat license
• 6:25
  - to actually having
• 6:27
  - their customers pay
• 6:29
  - per token, per task, per whatever it is.
• 6:32
  - And I think that that shift in how
• 6:34
  - licensing and software in general is sold
• 6:38
  - is probably gonna start taking place over the
• 6:40
  - next six months to year as the subsidies
• 6:43
  - that most of these Frontier
• 6:45
  - AI labs are getting go away. Yeah. And
• 6:48
  - I know you mentioned a bit about,
• 6:50
  - the premium cost of this hardware, especially with,
• 6:53
  - as you mentioned,
• 6:54
  - the RAM prices. And I'm wondering, you know,
• 6:56
  - I'm kind of thinking it about this as
• 6:59
  - well just from a perspective of just any
• 7:01
  - kind of, orthopedic or spine technology acquisition. Like,
• 7:05
  - obviously, there are these robots out there that
• 7:08
  - also cost, you know, into
• 7:10
  - the millions. So I'm wondering how
• 7:13
  - practices should think about these types of investments.
• 7:16
  - Should it be the same? Should they think
• 7:17
  - about the same as they would for, like,
• 7:19
  - a spine or orthopedic robot, or should they
• 7:21
  - approach it differently?
• 7:23
  - Yeah. And I I think that there's a
• 7:24
  - few ways to look at it and a
• 7:26
  - few ways to model it. Right? Because although,
• 7:30
  - yes, hardware right now is expensive,
• 7:32
  - if you're comparing
• 7:33
  - the output
• 7:35
  - to a
• 7:36
  - well configured, well outfitted,
• 7:39
  - local AI solution,
• 7:41
  - you know, you should probably be comparing that
• 7:43
  - to the salaries of the staff that they
• 7:46
  - either are going to be augmenting or replacing.
• 7:50
  - And so, like, for example, if you have
• 7:52
  - an AI receptionist that is able to triage
• 7:55
  - calls, put in medication refills, kinda doing all
• 7:58
  - that stuff such that you need one less
• 8:00
  - MA, two less MAs,
• 8:03
  - you know, that would be kind of the
• 8:04
  - direct comparison.
• 8:06
  - But some of it's also the the synergy
• 8:09
  - that you get from being able to link
• 8:12
  - different aspects of your practice, of your clinic,
• 8:15
  - of your hospital together.
• 8:17
  - Right? Like, behind the scenes, now your front
• 8:19
  - desk is talking to billing, is talking to,
• 8:23
  - scheduling, is talking to everybody.
• 8:25
  - And so, like, if a rejected claim comes
• 8:29
  - in from an insurance company,
• 8:31
  - it automatically,
• 8:32
  - behind the scenes, is able to get
• 8:35
  - processed and fed into a self learning, self
• 8:39
  - evolving,
• 8:41
  - claim submission tool. And so, you know, nothing
• 8:44
  - is,
• 8:46
  - nothing in medicine lives in a silo. Right?
• 8:48
  - You know, patients' data has to go from
• 8:50
  - department to department.
• 8:52
  - And so if you're able to streamline
• 8:54
  - and get rid of some of those silos,
• 8:58
  - you know, then not only do I think
• 9:00
  - that you should be
• 9:01
  - evaluating
• 9:03
  - AI on its,
• 9:04
  - per user cost and potentially the users that
• 9:07
  - it can replace, but also some of the
• 9:09
  - synergies that can allow your practice to just
• 9:11
  - be that much more efficient.
• 9:13
  - And thank you for breaking that down. And
• 9:16
  - then, also, you know, I wanna touch on
• 9:17
  - your AI. It's kind of like this black
• 9:19
  - box. We don't exactly know what's happening behind
• 9:21
  - the scenes. And I know you mentioned some
• 9:23
  - of the advantages that, local
• 9:26
  - AI systems have for,
• 9:28
  - protecting data security and compliance. But, you know,
• 9:32
  - even on that level,
• 9:34
  - what kind of strategies are you baking into
• 9:36
  - this to ensure that, you know, no one
• 9:39
  - breaks in locally and takes data? Yeah. No.
• 9:42
  - Absolutely. You know? And and data security and
• 9:45
  - patient data sovereignty
• 9:47
  - are, you know, two huge issues
• 9:49
  - that I think everybody is gonna be grappling
• 9:52
  - as AI becomes more and more prevalent in
• 9:55
  - all of its forms across health care.
• 9:57
  - And so one of the big things that
• 9:59
  - I've adapted actually from the military
• 10:02
  - is that you need to keep a human
• 10:04
  - in the loop. Right? And so there are
• 10:06
  - certain decisions, whether it's refilling a medication
• 10:10
  - or sending off a test,
• 10:12
  - where that final step, you know, needs to
• 10:15
  - have a human verification.
• 10:17
  - And so the way that I'm designing my
• 10:19
  - my practice and,
• 10:22
  - kind of the the clinical operations here is
• 10:25
  - that AI is doing a great job behind
• 10:29
  - the scenes,
• 10:30
  - but then, ideally, kinda just presents a well
• 10:33
  - ordered task list to, let's say, your medical
• 10:36
  - assistant,
• 10:37
  - you know, where the orders are already put
• 10:39
  - in, the dosages have been verified, everything looks
• 10:42
  - good,
• 10:44
  - but you still need a human in the
• 10:45
  - loop to to do that final verification
• 10:48
  - and actually submit that order.
• 10:51
  - And so the big reason
• 10:53
  - for that is AI is nondeterministic.
• 10:57
  - It is
• 10:58
  - a supercomputer
• 10:59
  - of, you know, probabilities
• 11:01
  - and kinda all that sort of stuff, but
• 11:03
  - you you can never get a deterministic
• 11:07
  - answer from an AI.
• 11:09
  - And so because of that, you know, there's
• 11:11
  - no kinda hard coded rules that that an
• 11:15
  - AI is always going to follow.
• 11:18
  - And so for anything that is mission critical,
• 11:20
  - that is for patient critical,
• 11:23
  - keeping a human in the loop as that
• 11:25
  - final stage,
• 11:27
  - you know, is the the solution that I
• 11:29
  - think is gonna work best for most people.
• 11:31
  - It's very well said too, this idea of
• 11:33
  - keeping the human in the loop. And that
• 11:34
  - kinda leads into just the next question I
• 11:37
  - wanted to ask. And I know you definitely
• 11:38
  - touched on some of this already, but how
• 11:41
  - are you
• 11:42
  - identifying and balancing these tasks that can be
• 11:45
  - handled by AI, but also might be better,
• 11:48
  - handled by a human?
• 11:50
  - Yeah. And and, you know, that is what
• 11:52
  - I think that we're gonna be finding over
• 11:54
  - the next few years is
• 11:56
  - the employees and the practices that are most
• 11:59
  - successful are the ones that are going to
• 12:01
  - be able to integrate
• 12:03
  - AI with their humans,
• 12:05
  - you know, to augment their abilities, to augment
• 12:07
  - their capabilities
• 12:09
  - rather than trying to directly replace them.
• 12:12
  - And so,
• 12:13
  - you know, identifying kind of the low hanging
• 12:16
  - fruit, I think that we've already seen that
• 12:18
  - in the marketplace.
• 12:19
  - Right? AI voice,
• 12:21
  - receptionist
• 12:22
  - kinda doing, you know, simple patient triage,
• 12:25
  - rescheduling visits. You know, that's an excellent use
• 12:28
  - case for AI. Although most consumers don't really
• 12:32
  - like kind of the robotic
• 12:33
  - sounding voices,
• 12:35
  - you know, they're very efficient. They're very good
• 12:37
  - at their job. Some of the other things
• 12:39
  - that are low hanging fruit are billing and
• 12:41
  - coding. Most of the insurance companies have published
• 12:45
  - guidelines about what they're expecting,
• 12:49
  - you know, to get authorization
• 12:50
  - for a shoulder replacement or whatever it is.
• 12:53
  - And so figuring out the necessary information for
• 12:57
  - them,
• 12:58
  - how your notes should be structured
• 13:00
  - and coded appropriately,
• 13:02
  - You know, those are relatively low hanging fruit
• 13:05
  - that, you know, most people's,
• 13:07
  - local AI solutions would be
• 13:10
  - readily able to do.
• 13:12
  - You know? But the the real things that
• 13:14
  - I'm most interested in are what are called
• 13:16
  - the orchestration layers. Right? Because you can have
• 13:19
  - these individual tasks or skills that AI has,
• 13:24
  - but how do you put it all together?
• 13:25
  - Right? Like, you know, you can replace a
• 13:27
  - biller or coder with AI,
• 13:30
  - or they can serve as a receptionist.
• 13:32
  - But how do you get those two different
• 13:35
  - modules to talk to one another?
• 13:37
  - And so I think that the practices that
• 13:39
  - are gonna really be most successful
• 13:41
  - are the ones that are best able to
• 13:44
  - integrate those orchestration layers,
• 13:46
  - but then also using them to augment rather
• 13:49
  - than to replace their current staff.
• 13:52
  - Got it. And, you know, after drawing, I
• 13:54
  - want to then pivot to what you're talking
• 13:56
  - about earlier in this conversation about starting your
• 13:59
  - own practice. We're recording this podcast on Wednesday,
• 14:01
  - May 27, and you said you'd be starting
• 14:04
  - in, I believe, June June 1. First. Yes.
• 14:08
  - It's first first, the beginning to the second
• 14:11
  - half of the year, which
• 14:12
  - is great. And I'd love to hear, besides
• 14:14
  - the
• 14:15
  - AI aspect that you're excited about,
• 14:18
  - what more kind of went into this decision
• 14:21
  - to take a leap and, you know, start
• 14:23
  - your own practice? How long have you been
• 14:24
  - thinking about this?
• 14:26
  - Yeah. So I I've been thinking about this
• 14:28
  - on and off for the last,
• 14:31
  - two years probably and, you know, really started
• 14:34
  - formulating it probably about a year ago.
• 14:38
  - Because, man, especially as a solo practitioner,
• 14:41
  - starting a practice is no joke.
• 14:44
  - Dealing with the insurance companies, figuring out all
• 14:47
  - the little gotchas
• 14:49
  - to get credentialed,
• 14:50
  - contracted,
• 14:52
  - setting up all your payment systems, and then
• 14:55
  - kinda everything else that they don't teach you
• 14:56
  - in medical school. Mhmm. And so it's definitely
• 14:59
  - something that has always appealed to me as,
• 15:02
  - you know, having a bit of a entrepreneurial
• 15:04
  - mindset.
• 15:05
  - But I I really started putting this into
• 15:08
  - motion probably about six months ago.
• 15:11
  - And like I said, opening the doors in
• 15:13
  - about five days from now. You know? But
• 15:15
  - as I kinda get further and further out,
• 15:18
  - right now, I have local AI solutions for
• 15:21
  - some of the things that we've discussed.
• 15:24
  - But as I am getting direct patient and
• 15:27
  - staff feedback and continuing to build out some
• 15:30
  - of this functionality,
• 15:32
  - you know, I'm I'm kinda really excited for
• 15:34
  - the opportunities that it's gonna allow,
• 15:37
  - because, really, kinda when it when it comes
• 15:39
  - down to it, you know, it's it's still
• 15:41
  - all about the patients. Right? They need to
• 15:43
  - feel secure. They gotta feel confident. They gotta
• 15:46
  - know that there is a human at the
• 15:48
  - other end of the line if they really
• 15:50
  - need somebody.
• 15:51
  - And so figuring out the balance between
• 15:55
  - human, AI, kinda all the other tools that
• 15:58
  - are at my availability,
• 16:00
  - to ensure a good experience for the patients
• 16:02
  - where I can deliver
• 16:04
  - high level efficient care. You know, those are
• 16:06
  - all the things that I'm excited about starting
• 16:09
  - a practice.
• 16:10
  - You know, it's really, I think, just being
• 16:13
  - able to deliver a higher higher level of
• 16:15
  - care to my patients.
• 16:16
  - Great. And, you know, doctor Doreen, thank you
• 16:19
  - so much for taking the time to
• 16:21
  - chat, today on the podcast. It's been a
• 16:24
  - great conversation, and I definitely would love to
• 16:26
  - follow-up in the next six months and hear,
• 16:29
  - how things are kicking off there. Yeah. No.
• 16:31
  - Definitely. Check back in six months. I,
• 16:34
  - maybe you'll, be talking to an AI avatar
• 16:37
  - of myself instead of
• 16:39
  - we'll see how that goes.
• 16:42
  - Great. Well, thank you again, and I hope
• 16:44
  - you have a good rest of your day.
• 16:45
  - Thank you. You too. Take care.