Latest / Becker’s Healthcare Digital Health + Health IT / Scaling AI Responsibly While Protecting Clinical Decision Making with Dr. Bryon Frost

Transcript

Scaling AI Responsibly While Protecting Clinical Decision Making with Dr. Bryon Frost · Becker’s Healthcare Digital Health + Health IT

• 0:00
  - What we wanna do with AI is help
• 0:01
  - clinicians go from being reactive
• 0:04
  - to proactive.
• 0:05
  - At Philips, we can recognize patterns in monitoring
• 0:08
  - data across populations
• 0:10
  - and apply those insights to specific patients.
• 0:14
  - Safe and responsible AI can support sharper insights,
• 0:18
  - pave the way for faster diagnosis,
• 0:21
  - and help streamline care.
• 0:23
  - That can benefit clinicians,
• 0:25
  - it can benefit health systems. And most importantly,
• 0:28
  - it can benefit patients and their families.
• 0:33
  - Phillips.
• 0:34
  - Hi, everyone. This is Scott King with Becker's
• 0:36
  - Healthcare. Thank you so much for tuning in
• 0:37
  - to the Becker's Healthcare podcast series. Joining me
• 0:40
  - for today's discussion is Brian Frost, vice president
• 0:43
  - and chief medical informatics officer over at McLeod
• 0:46
  - Health. Brian, how are you doing today? Thanks
• 0:47
  - so much for joining us here. Thanks for
• 0:48
  - having me. Of course. And, you know, before
• 0:50
  - we dive into some some big topics in
• 0:52
  - health care, I'd love to just kinda learn
• 0:54
  - a little bit more about your career journey
• 0:55
  - and your background, if you don't mind sharing.
• 0:57
  - Yeah. So I'm an emergency room physician. I'm
• 0:59
  - still practicing at a level two trauma center
• 1:01
  - at McLeod, and I've been in IT for
• 1:04
  - probably twelve years.
• 1:06
  - Serve as the chief medical informatics officer.
• 1:09
  - I also chair the AI governance committee, which
• 1:12
  - has taken up a lot of my time
• 1:14
  - recently over the last year. So... Yeah. Awesome.
• 1:18
  - Thank you. Appreciate that info. And, you know,
• 1:19
  - doctor Frost, I wanted to first kinda
• 1:22
  - dive into
• 1:23
  - how health systems are receiving
• 1:25
  - a growing stream of patient data from bedside
• 1:27
  - monitoring,
• 1:28
  - ambulatory devices, and care delivered at home. Where's
• 1:31
  - that data creating actionable insights for clinicians,
• 1:35
  - and where does it still create more work
• 1:37
  - than insight? Yeah.
• 1:39
  - So the real problem around all this is
• 1:41
  - the data. And if you think about,
• 1:43
  - you know, the previous revolution we just
• 1:46
  - exited
• 1:47
  - was the
• 1:48
  - data revolution where we had... We we have
• 1:51
  - so much data at our disposal. It's overwhelming.
• 1:54
  - Where AI is helpful... And and there's two
• 1:57
  - two parts of AI that I like to
• 1:58
  - talk about. One is machine learning.
• 2:01
  - So if you do structured machine learning
• 2:03
  - around data on the inpatient side, you can
• 2:06
  - get really good insights. And so the areas
• 2:09
  - that we find value are the deterioration index
• 2:14
  - or the sepsis,
• 2:15
  - score. So whenever you have a problem and
• 2:18
  - you have that data and you're going to
• 2:19
  - answer a very specific problem with a person
• 2:22
  - who's going to receive that information
• 2:24
  - and make a decision on it, it's gonna
• 2:26
  - work really well. The areas that it's not
• 2:28
  - gonna work well are more on the ambulatory
• 2:30
  - side. So
• 2:32
  - if you have a physician who's got remote
• 2:34
  - patient monitoring, patient data coming in, you really
• 2:37
  - have to build a structure around that provider
• 2:40
  - to support them. Otherwise, they're just gonna be
• 2:42
  - overwhelmed.
• 2:43
  - So
• 2:44
  - I I really think it's all about the
• 2:46
  - information overload that's the important point here. So
• 2:49
  - I graduated medical school in 2000,
• 2:52
  - and at the time, they were saying medical
• 2:53
  - knowledge was doubling every five years.
• 2:56
  - Well, today, they say it's doubling every, you
• 2:58
  - know, seventy three days.
• 3:00
  - You can't practice
• 3:02
  - good evidence based medicine with that much data.
• 3:05
  - You really need these AI tools to practice
• 3:08
  - effective medicine.
• 3:10
  - When clinicians are trying to reach a a
• 3:12
  - diagnosis, they often must piece together information from
• 3:15
  - radiology, cardiology, pathology, urology, and other systems. Where
• 3:19
  - does that fragmentation create the most friction in
• 3:21
  - your organization?
• 3:22
  - And how do you bring together together data
• 3:24
  - sources
• 3:25
  - and and address this fragmentation
• 3:27
  - or close the gaps?
• 3:29
  - Yeah. And so what's interesting to think about
• 3:31
  - for me is the scenario we put physicians
• 3:34
  - into. We compress
• 3:36
  - this visit for a patient in seven to
• 3:39
  - fifteen minutes. And in that time, the doctor
• 3:41
  - has to assimilate all this information.
• 3:44
  - And the problem with that is the the
• 3:46
  - big... Biggest risk for patients are really diagnostic
• 3:49
  - errors, not because of a lack of knowledge,
• 3:51
  - but because of either anchoring bias or recency
• 3:54
  - bias. We create the perfect environment by compressing
• 3:58
  - the visit into this small amount of time,
• 4:00
  - and the physician has to assimilate it all.
• 4:03
  - It it... Recency bias,
• 4:05
  - conformity bias,
• 4:07
  - these are the things that I think are
• 4:08
  - gonna be a big risk. And so the
• 4:10
  - value is not just from AI giving more
• 4:13
  - time back to the provider, but it's also
• 4:15
  - giving the provider all the information.
• 4:17
  - And so helping them assimilate,
• 4:20
  - get all the data together so that the
• 4:23
  - patient visit
• 4:25
  - doesn't require so much cognitive work on the
• 4:27
  - provider's part to pull all that data together.
• 4:30
  - Yeah. And speaking of the patient visit, you
• 4:32
  - know, patients are increasingly kinda on the go.
• 4:34
  - They they move among the hospital,
• 4:36
  - outpatient settings, and home. Where does clinical context
• 4:39
  - most often get lost during those transitions,
• 4:42
  - and what would it take to give care
• 4:44
  - teams a more continuous view of the patient?
• 4:47
  - Yeah. The the thing of the area within
• 4:50
  - artificial intelligence that makes me most excited is
• 4:53
  - the agentic workflows
• 4:55
  - that are
• 4:56
  - really going to allow AI agents to pull
• 4:58
  - this data together like we talked about. But
• 5:00
  - specifically for the patient, the value is,
• 5:03
  - right now, for instance, a patient comes into
• 5:05
  - a clinic and tells their story
• 5:08
  - to three different people.
• 5:09
  - They go see a specialist. The specialist has
• 5:12
  - no idea what the primary care physician wants,
• 5:15
  - doesn't understand the plan.
• 5:17
  - Discharge instructions are written for that patient more
• 5:20
  - for the EMR and not necessarily for the
• 5:22
  - patient. So I really think what's exciting
• 5:25
  - is thinking about an AI agent that basically
• 5:28
  - follows the patient across the continuum of care,
• 5:31
  - is able to pull that information for them,
• 5:33
  - and and knows their preferences. So they know...
• 5:36
  - The provider knows the patient wants a telehealth
• 5:38
  - visit, for instance, instead of an in person
• 5:40
  - visit.
• 5:42
  - I think that's where we're going to find
• 5:43
  - the value in the patient and put them
• 5:45
  - back at the center of
• 5:47
  - care. So... Yeah.
• 5:49
  - Yeah. You you just mentioned a great kinda
• 5:51
  - AI use case with the agents there. With
• 5:54
  - so many AI use cases competing for attention,
• 5:57
  - how do you distinguish
• 5:58
  - a compelling pilot from from something ready to
• 6:01
  - become a part of everyday care, and and
• 6:03
  - what must be in place to scale it
• 6:05
  - responsibly?
• 6:06
  - Yeah. So I'm at a nonacademic
• 6:09
  - medical center, which,
• 6:11
  - really by saying that, we we don't have
• 6:14
  - huge margins. We are not funded by a
• 6:16
  - foundation, and so we have to be very
• 6:17
  - judicious and careful with the money that we
• 6:19
  - spend. And so one of the things we've
• 6:21
  - done at McLeod is create a vendor evaluation
• 6:24
  - framework
• 6:25
  - where we can evaluate the vendor space
• 6:28
  - very methodically.
• 6:30
  - One of the biggest risks is someone comes
• 6:31
  - back from a conference like this with a
• 6:33
  - vendor and says, this is the greatest thing
• 6:35
  - ever, and we go to implement it,
• 6:37
  - later realizing it really didn't solve the problem
• 6:40
  - we thought it was. And so we're very
• 6:42
  - careful in our framework to evaluate
• 6:45
  - the vendor space and make sure
• 6:47
  - this vendor is actually gonna produce what we
• 6:49
  - need.
• 6:50
  - Then we've gotta figure out how to govern
• 6:52
  - it, and that's the next piece. And so
• 6:54
  - we might pass on a vendor that is
• 6:56
  - going to be a high risk from a
• 6:58
  - governance perspective, from an AI perspective.
• 7:01
  - So what we've created is an AI risk
• 7:03
  - tiering system at McCloud.
• 7:05
  - So, essentially, I can put our entire AI
• 7:08
  - portfolio
• 7:09
  - into that risk tiering system and focus our
• 7:12
  - energies on those higher risk areas
• 7:15
  - that we know we need to pay attention
• 7:17
  - to. And those high risk are,
• 7:19
  - you know, any AI that produces a score
• 7:22
  - or influences a provider is going to create
• 7:25
  - automation bias. And so we're going to
• 7:28
  - really be careful with those as well as
• 7:31
  - all the agentic
• 7:32
  - workflows.
• 7:34
  - That's where it becomes extremely risky, and we
• 7:35
  - spend the most of our time trying to
• 7:37
  - govern those. And I'll just be honest. Nobody,
• 7:39
  - I think, has figured this out. It's one
• 7:42
  - thing to create a framework. It's one thing
• 7:44
  - to create a
• 7:46
  - risk structure. It's another to find an individual
• 7:50
  - in the organization who's gonna be responsible for
• 7:52
  - monitoring
• 7:53
  - that system. That is right now our biggest
• 7:55
  - challenge.
• 7:56
  - As technology relationships move beyond onetime purchases, what
• 7:59
  - should a partner remain accountable for after go
• 8:02
  - live to make the relationship genuinely valuable over
• 8:05
  - time?
• 8:06
  - Yeah. I think you've gotta pick a vendor
• 8:08
  - that is aligned with your
• 8:11
  - personal road map, your strategy, where you plan
• 8:13
  - on going in the next five years. You
• 8:15
  - gotta look close at the vendor's road map,
• 8:18
  - and you wanna make sure their innovation plan
• 8:21
  - mirrors your own. But, also, I would say,
• 8:23
  - and I think many people have said this,
• 8:25
  - don't have long term contracts.
• 8:28
  - With this new space with
• 8:30
  - AI impacting it so rapidly,
• 8:32
  - the last thing you wanna do is get
• 8:34
  - locked in to a contract with somebody
• 8:36
  - because the horizon of what's coming is is
• 8:39
  - very difficult to see. And so you wanna
• 8:42
  - be very careful with the vendor relationships that
• 8:44
  - you select, pick good vendors,
• 8:46
  - have a good relationship, but also be able
• 8:48
  - to get out of that contract if you
• 8:50
  - need to. If you determine,
• 8:52
  - you know, the EMR, for instance, provides
• 8:55
  - everything we need, you don't wanna be locked
• 8:57
  - into a vendor relationship that you can't get
• 8:59
  - out of.
• 9:00
  - The last question I have for you, doctor
• 9:02
  - Frost, what's the future of integrated
• 9:05
  - diagnostics or just diagnostic services,
• 9:07
  - and where do you see yourself in the
• 9:09
  - next five years?
• 9:11
  - So with diagnostic services, I really think that
• 9:14
  - the EMR is is going to be the
• 9:16
  - one that in five years gets a lot
• 9:19
  - of the information
• 9:20
  - because they've got the longitudinal record. They've got
• 9:23
  - the workflow already there.
• 9:25
  - So
• 9:26
  - I think a lot of vendors are going
• 9:27
  - to have a challenge making the argument that
• 9:29
  - their point solution is going to solve the
• 9:32
  - problem, and I do think the EMRs are
• 9:34
  - probably gonna consolidate a lot of this. But
• 9:37
  - I'll just tell you, when I think about
• 9:38
  - five years into the future, the biggest concern
• 9:41
  - I have
• 9:42
  - is honestly,
• 9:44
  - deterioration
• 9:45
  - of our physicians.
• 9:47
  - With cognitive offloading, this is a significant risk
• 9:50
  - that I don't know anyone has solved for
• 9:52
  - yet.
• 9:53
  - Especially vulnerable are the medical students, the residents,
• 9:57
  - but everybody is at risk
• 9:59
  - with cognitive offloading.
• 10:01
  - You know, I'm an ER physician, and I
• 10:03
  - really enjoy the diagnostic process of seeing a
• 10:06
  - patient, figuring out what problem
• 10:08
  - they have.
• 10:10
  - And there's a lot of studies that came
• 10:11
  - out even last year.
• 10:13
  - I believe it was in Lancet, was looking
• 10:15
  - at
• 10:16
  - colonoscopies
• 10:17
  - and,
• 10:18
  - gastroenterologists
• 10:19
  - who are using
• 10:21
  - AI assisted tools actually had a degradation of
• 10:23
  - their skills.
• 10:25
  - That's a big problem, and it's a risk,
• 10:26
  - and I feel like we need to solve
• 10:28
  - it for the future.
• 10:30
  - Thank you, doctor Frost, for your time and
• 10:32
  - insights today. We also wanna thank our podcast
• 10:34
  - sponsor, Philips. You can tune into more podcasts
• 10:38
  - from Becker's Healthcare by visiting our podcast page
• 10:40
  - at beckershospitalreview.com.