Latest / Becker’s Healthcare Podcast / Faraz Ahmad and Jonathan Hourmozdi on AI's Future in Heart Failure Care

Transcript

Faraz Ahmad and Jonathan Hourmozdi on AI's Future in Heart Failure Care · Becker’s Healthcare Podcast

• 0:02
  - There's a small circle of people who truly
• 0:04
  - understand what it takes to lead a health
• 0:05
  - system right now. The decisions are enormous.
• 0:08
  - The variables are constantly shifting, and the stakes
• 0:11
  - of getting a call wrong don't stay contained.
• 0:14
  - That's not a complaint. It's a structural reality.
• 0:17
  - And it's why the most effective c suite
• 0:19
  - leaders are deliberate about where they go to
• 0:21
  - pressure test their thinking.
• 0:22
  - Becker's fourteenth annual CEO and CFO roundtable
• 0:26
  - brings together more than 1,500
• 0:28
  - executive level attendees
• 0:30
  - and nearly 400 speakers for four days built
• 0:32
  - around growth, financial sustainability,
• 0:35
  - workforce strategy, and the leadership questions that don't
• 0:38
  - have easy answers.
• 0:39
  - Join us November second through the fifth in
• 0:41
  - Chicago.
• 0:42
  - For the agenda and event details,
• 0:45
  - visit beckershospitalreview.com
• 0:47
  - and click on the events tab in the
• 0:49
  - upper right.
• 0:50
  - Welcome to the Becker's Healthcare Cardiology and Heart
• 0:53
  - Surgery podcast. I'm Mariah Taylor, assistant editor, and
• 0:56
  - I'm thrilled to be speaking with two leaders
• 0:58
  - from Northwestern Medicine's Wilm Cardiovascular
• 1:00
  - Institute.
• 1:01
  - First, we have Faraz Ahmed, medical director for
• 1:04
  - AI strategy for the Center of AI,
• 1:07
  - and Jonathan Wormazdi,
• 1:09
  - core faculty for the Center of AI. Thank
• 1:11
  - you both so much for joining me today.
• 1:13
  - I'm really excited to learn more about your
• 1:15
  - strategy for using AI in heart failure care.
• 1:18
  - But before we dive into that, I'd love
• 1:20
  - for you to introduce yourself and tell us
• 1:22
  - a little bit about your background, doctor Ahmad.
• 1:24
  - Thank you. I'm really excited to be here.
• 1:26
  - So I'm a heart failure cardiologist informaticist here
• 1:29
  - at Northwestern.
• 1:30
  - I've I've been here for over ten years
• 1:31
  - at this point. I spend my a lot
• 1:33
  - of my time seeing patients with heart failure,
• 1:34
  - with all types of heart failure. I also
• 1:36
  - help our cardiovascular institute and our health system
• 1:38
  - think about what AI what AI solutions we
• 1:41
  - should develop
• 1:42
  - and also partner in how we should partner
• 1:44
  - with industry. I also run a digital health
• 1:46
  - and AI research lab.
• 1:48
  - And doctor Ormasty?
• 1:50
  - Yeah.
• 1:51
  - Thanks for having me as well. Yeah. So,
• 1:53
  - I am, also an advanced heart failure cardiologist
• 1:56
  - here at Northwestern's,
• 1:57
  - Bloemburg Gastroenter Institute. I've been here at Northwestern
• 2:00
  - for medical school,
• 2:02
  - as well as some training in both bioethics
• 2:04
  - and fellowship in AI. And then I did
• 2:06
  - my fellowship training in both cardiology and advanced
• 2:08
  - heart failure here. Been in Chicago on and
• 2:10
  - off for the last over a decade and,
• 2:12
  - really happy to be here. So some of
• 2:14
  - my research work is in AI implementation, but
• 2:17
  - also thinking about,
• 2:18
  - questions around the ethics of AI use and
• 2:20
  - clinical adoption. And,
• 2:22
  - that's the research that Faraz and I have
• 2:23
  - shared and and really focusing on how, to
• 2:26
  - make this technology benefit patients the most.
• 2:29
  - Excellent.
• 2:30
  - Now, you know, everybody is all about AI
• 2:33
  - these days, finding the best way to implement
• 2:35
  - it across platforms and services.
• 2:38
  - So I'm curious, what are some of the
• 2:40
  - biggest opportunities to leverage AI in heart failure
• 2:43
  - specifically in order to improve patient outcomes and
• 2:45
  - care delivery?
• 2:47
  - Yeah. That's a great question. There's so many
• 2:48
  - opportunities to leverage AI in heart failure.
• 2:51
  - You know, heart failure is a really prevalent
• 2:53
  - condition, affects over six point seven million US
• 2:55
  - adults, expected to grow to over
• 2:58
  - eleven point four million by 2050.
• 3:00
  - And we're seeing a higher and higher impact
• 3:02
  - on young adults as well. Kids have a
• 3:03
  - high mortality burden burden and affect quality of
• 3:06
  - life.
• 3:07
  - In a lot of ways, heart failure in
• 3:08
  - my view is sort of the prototypical condition
• 3:10
  - that AI can be you can be used
• 3:12
  - to help.
• 3:13
  - It's really heterogeneous.
• 3:14
  - So there's a lot of different causes of
• 3:16
  - heart failure, which often leads to delays in
• 3:18
  - diagnosis.
• 3:19
  - So there's a lot of opportunities for using
• 3:20
  - AI to find to find the cause of
• 3:23
  - heart failure, patients with heart failure earlier and
• 3:25
  - correct misdiagnoses.
• 3:26
  - We also, in heart failure, are fortunate to
• 3:28
  - have a lot of evidence based treatments that
• 3:30
  - really improve quality of life and extend survival.
• 3:33
  - While these treatments, they include medications, devices, procedures,
• 3:36
  - a lot of them are actually underused. So
• 3:38
  - I think AI is really well positioned to
• 3:40
  - help improve the delivery of evidence based care.
• 3:43
  - There's a lot of interesting opportunities in heart
• 3:45
  - failure with wearables and how do you take
• 3:47
  - all this data that we can collect from
• 3:49
  - patients now and merge together and use that
• 3:51
  - to track them over time and get them
• 3:53
  - the help they need in a timely way.
• 3:55
  - You know, our health system is super inefficient
• 3:56
  - is inefficient in so many ways, and how
• 3:58
  - can we make the whole care journey more
• 3:59
  - efficient with AI is a really exciting area.
• 4:01
  - So those are some of the main areas
• 4:02
  - we've been thinking about here at our center.
• 4:05
  - Excellent.
• 4:06
  - Now when you're implementing a use case for
• 4:09
  - AI powered diagnostics and heart care, how are
• 4:12
  - you approaching the validation
• 4:14
  - process so that you know that the AI
• 4:15
  - model is performing equitably across all of your
• 4:18
  - patient populations before you scale
• 4:21
  - it? Yeah. I think this is a really
• 4:22
  - great and important question. So, you know, thinking
• 4:24
  - about especially some of our discrete classifier models,
• 4:28
  - things that help us, whether that's
• 4:31
  - detecting disease,
• 4:33
  - at an early state across a health system
• 4:35
  - or population,
• 4:36
  - thinking about, models that are gonna provide some
• 4:39
  - kind of binary output of this patient's at
• 4:41
  - high risk or has this diagnosis or is
• 4:43
  - at risk of it versus not. You know,
• 4:45
  - one of the things that we've learned, not
• 4:46
  - just from medicine, but across all the different
• 4:49
  - fields that have adopted AI models is that
• 4:50
  - we need to be really cautious. Right?
• 4:53
  - These are models and tools that are trained
• 4:56
  - on,
• 4:57
  - discrete populations and and training datasets and are
• 5:01
  - susceptible to all the same kind of biases
• 5:04
  - that human learners are and in ways that
• 5:06
  - are often obfuscated due to the complexity of
• 5:09
  - the underlying,
• 5:10
  - models and the fact that you can't just
• 5:11
  - probe them and ask them, hey. Why'd you
• 5:13
  - make that decision? And so
• 5:15
  - what we have tried to do is be
• 5:17
  - really thoughtful. And I think as a field,
• 5:19
  - everyone is trying to be increasingly thoughtful about,
• 5:23
  - better understanding those risks and thinking upfront before
• 5:25
  - we're sort of scaling these models about what
• 5:28
  - we might be,
• 5:30
  - unexposing our patients to in terms of, any
• 5:33
  - risk of bias and and or just, you
• 5:36
  - know,
• 5:37
  - inaccurate performance by the model. And so one
• 5:40
  - of the ways that we've tried to tackle
• 5:41
  - that, for instance, looking at a condition,
• 5:43
  - cardiac amyloidosis,
• 5:45
  - which is a condition that causes heart failure
• 5:48
  - and, and has some disparities already in the
• 5:51
  - world in terms of how we diagnose it
• 5:53
  - right now. So we know we're pretty bad
• 5:55
  - at catching cardiac amyloidosis,
• 5:58
  - across the board for cardiology. It's a there's
• 6:00
  - often a delay between the onset of the
• 6:02
  - disease and diagnosis. And we know we have
• 6:05
  - treatments that are really important to start early
• 6:08
  - in the disease course because we think that
• 6:10
  - these treatments work much better if it prevents
• 6:13
  - progression,
• 6:14
  - at an early stage. And so this is
• 6:16
  - a ripe opportunity for AI tools to help
• 6:19
  - us better detect these patients, and we also
• 6:22
  - know already that we're we have some disparities
• 6:24
  - in care and that models may be trained,
• 6:27
  - for instance, on
• 6:28
  - datasets that aren't as diverse as the patients
• 6:31
  - we see in our health system. And so
• 6:33
  - one of the ways that doctor Ahmad and
• 6:35
  - I sort of thought about this was actually
• 6:36
  - to look at a variety of models and
• 6:39
  - create a,
• 6:40
  - representative
• 6:41
  - sort of retrospective
• 6:43
  - cohort where we could look at a population
• 6:46
  - of our patients
• 6:47
  - that is gonna represent the specific populations in
• 6:50
  - terms of, you know, racial and gender identity
• 6:54
  - makeup, for the area we see in Chicago
• 6:57
  - and understand and actually test out, the model's
• 7:01
  - performance as we sort of take something that's
• 7:03
  - been trained in a different city or a
• 7:05
  - different health system with potentially different population makeup.
• 7:09
  - And I think one of the most important
• 7:10
  - things is to really understand the current state
• 7:12
  - of the world and where are we already
• 7:15
  - underperforming. Right? So, you know, one of the
• 7:18
  - things that we understand
• 7:19
  - because of enormous efforts by health,
• 7:23
  - disparity researchers is that we are already missing
• 7:26
  - a lot of diagnoses of amyloid,
• 7:28
  - heart disease in specific populations in The US,
• 7:31
  - in particular, those who self identify as black.
• 7:34
  - And so we wanted to make sure that
• 7:36
  - any model that we adopt at Northwestern isn't
• 7:39
  - gonna perpetuate and exacerbate that existing disparity
• 7:43
  - because we've seen and had, you know, warnings
• 7:46
  - about other areas outside of medicine where models
• 7:48
  - have been adopted that have potentially learned
• 7:51
  - bad biases from the dataset that already exist
• 7:54
  - in human systems and then perpetuated
• 7:56
  - those biases and even laundered them because it's
• 7:59
  - a technology and not a human making that
• 8:01
  - bias decision.
• 8:02
  - And so what we were able to do
• 8:03
  - was really compare model performance
• 8:06
  - at looking at a retrospective data set and
• 8:08
  - and better understanding, you know, historically in the
• 8:11
  - patients that we've seen, is this model working
• 8:13
  - well? How well is it working compared to
• 8:16
  - maybe simpler tool that don't use AI or
• 8:18
  - the clinician judgment? And is it working
• 8:22
  - better or worse for those populations who we
• 8:24
  - know
• 8:25
  - already the health system has problems
• 8:27
  - with due to access to care or perhaps
• 8:30
  - other biases
• 8:32
  - aren't getting diagnosed at the rate that we
• 8:33
  - want them to be and perhaps having worse
• 8:35
  - outcomes related to delays in care.
• 8:38
  - And what we were able to show, at
• 8:40
  - least in some of these cohorts from older
• 8:43
  - datasets, is these models are actually not showing
• 8:45
  - any evidence that they're likely to perpetuate a
• 8:47
  - preexisting biases, that there was very low risk
• 8:50
  - there that we were gonna adopt this model
• 8:52
  - and see inequitable
• 8:54
  - care as a result. And so that kinda
• 8:56
  - gives us confidence as we then move into
• 8:58
  - actually adopting those into a clinical space.
• 9:01
  - Excellent.
• 9:02
  - Now on the other half of it, once
• 9:04
  - you have the AI tool, it's been validated,
• 9:06
  - and it's ready to go live, what's your
• 9:08
  - approach for a model governance?
• 9:11
  - Yeah. This is the million dollar question, I
• 9:13
  - think, and we're all still sort of thinking
• 9:15
  - this one through.
• 9:17
  - So I I do think that, you know,
• 9:18
  - an important piece of this is, like I
• 9:20
  - said, is really understanding
• 9:22
  - what the risks are, trying to be aware
• 9:25
  - upfront of where the current status quo of
• 9:29
  - your human system is, and and understanding what
• 9:32
  - are the key sort of metrics and things
• 9:34
  - that you wanna be monitoring for and measuring.
• 9:36
  - So if you're rolling out a model that's
• 9:38
  - focused on diagnosis and detection,
• 9:41
  - really, hopefully,
• 9:42
  - gathering
• 9:43
  - data
• 9:44
  - that's gonna help you and continuously validating that
• 9:47
  - the model is performing as you expect
• 9:50
  - and is not missing cases
• 9:52
  - at a disproportionate level for some, reason of,
• 9:55
  - say, the model isn't fitting well with a
• 9:58
  - certain population or perhaps there's other reasons why
• 10:00
  - technology developed elsewhere isn't coming in. Now this
• 10:03
  - is easier for us to do, I think,
• 10:05
  - for our the models that we train in
• 10:07
  - house and that we're actually really well aware
• 10:09
  - of the data that it was trained on
• 10:11
  - or or widely published on, for instance, so
• 10:13
  - we can be really,
• 10:14
  - queued into those things that we need to
• 10:16
  - to monitor for. And we have lots of
• 10:19
  - tools that we can implement on the information
• 10:20
  - system side to help us as we're trying
• 10:22
  - to sort of monitor for and measure those
• 10:26
  - specific performance metrics,
• 10:28
  - as in real time we're using the model.
• 10:30
  - It it it can become more challenging when
• 10:32
  - you've got, you know, models that are developed
• 10:35
  - and proprietary and perhaps a little bit more
• 10:37
  - close,
• 10:38
  - and you're trying to sort of work out
• 10:40
  - how between you and potentially the vendor, you're
• 10:44
  - gonna decide on who's gonna monitor those things
• 10:46
  - and who's gonna help drive the sort of
• 10:48
  - ship as far as how we're gonna continue
• 10:51
  - to validate that the model is performing as
• 10:53
  - we expect based on perhaps reviewing
• 10:55
  - prior data of what our health systems patients
• 10:58
  - are,
• 10:59
  - you know, like we mentioned with a cohort
• 11:01
  - of past patients. So I think that's a
• 11:03
  - really important piece is is understanding upfront what
• 11:06
  - you're what you're gonna be watching for, where
• 11:08
  - are the risks, and how you're gonna continue
• 11:10
  - to validate,
• 11:11
  - in an ongoing basis after you've adopted a
• 11:14
  - model in the clinical setting that it's continuing
• 11:16
  - to work as you expect and that you're
• 11:18
  - not exposing any specific,
• 11:20
  - you know, weaknesses in the model over time
• 11:22
  - as perhaps there's model drift or population changes.
• 11:26
  - I think, you know, other pieces of this
• 11:28
  - are just sort of how you think about
• 11:30
  - transparency
• 11:31
  - and how you think about,
• 11:33
  - actually trying to understand what it is that
• 11:35
  - the model is doing. You know, a lot
• 11:38
  - of times in the past, people have focused
• 11:40
  - on explainability
• 11:41
  - and and being able to sort of understand
• 11:44
  - model outputs.
• 11:45
  - And I think that that's still important, and
• 11:47
  - and some models do try to offer,
• 11:50
  - you know, versions of this and certainty figures
• 11:52
  - and things like this.
• 11:53
  - But I I would say that I think
• 11:55
  - the most important piece of that first part,
• 11:56
  - which is just really continuing to look at
• 11:59
  - the outcomes and making sure that, you know,
• 12:01
  - where the computer systems interact with human systems
• 12:04
  - in the actual real world, you're continuing to
• 12:06
  - see the the kind of outcomes and results
• 12:09
  - that you hope for and expect in terms
• 12:11
  - of benefiting the patient.
• 12:13
  - I'm curious, Faraz, if you have other things
• 12:15
  - to add as somebody who really thinks hard
• 12:16
  - about this. Yeah. I think you hit on
• 12:18
  - a lot of really great points. I'll just
• 12:20
  - amplify a couple of them. You know, I
• 12:21
  - think for like you said, for internal models,
• 12:23
  - we have a pretty good governance structure. We
• 12:25
  - have a commercial platform we use that can
• 12:27
  - model
• 12:28
  - inputs over time, outcomes over time, and it
• 12:30
  - it works it works pretty well. I think
• 12:32
  - what we're still really learning in our health
• 12:34
  - system and in other health systems too,
• 12:37
  - is for,
• 12:38
  - commercial models and really who owns governance for
• 12:41
  - that, how much should we be doing, and
• 12:42
  - how much should they be doing.
• 12:44
  - And, honestly, what role should the government ultimately
• 12:46
  - have in sort of monitoring these models over
• 12:48
  - time?
• 12:49
  - The other piece of it I think is
• 12:50
  - really important is that the governance looks different
• 12:53
  - for these sort of deterministic models, like a
• 12:54
  - lot of these disease classification models that Johnny
• 12:57
  - mentioned earlier versus generative AI.
• 12:59
  - And how do you monitor those models, whether
• 13:01
  - they're internal again, internal versus external, and that's
• 13:04
  - still something we're sorting out as a health
• 13:06
  - system and learning with others as we go.
• 13:09
  - It's such a good point you bring up,
• 13:11
  - which is that, like, I I think that
• 13:13
  - in the past, you know, six, seven years,
• 13:15
  - we've really just started to catch up to
• 13:18
  - where models were six, seven years ago, and
• 13:20
  - now we're entering a whole new space of
• 13:22
  - agentic and and generative AI, and and incredible
• 13:26
  - advances are happening at such a rapid pace.
• 13:27
  - And so it it is really you know,
• 13:29
  - it's a challenge. We have to all learn
• 13:31
  - together. We have to all work hard to
• 13:33
  - get ahead of some, some of these, like,
• 13:35
  - advances and thinking about how we deploy these
• 13:37
  - models and and and have effective governance.
• 13:40
  - And then we gotta figure out how to
• 13:41
  - resource this too. That's that's one of the
• 13:43
  - biggest challenges.
• 13:44
  - Absolutely.
• 13:46
  - So I love all of this. And often
• 13:48
  - when we talk about AI, we're talking about
• 13:50
  - it from the point of view of the
• 13:52
  - system and the clinician, but patients are using
• 13:54
  - AI more and more these days. So I
• 13:56
  - would love your take on how are patients
• 13:58
  - using AI today, and what will an AI
• 14:01
  - powered future look like for patients with heart
• 14:03
  - failure and their clinicians in the next five
• 14:05
  - to ten years?
• 14:07
  - Yeah. I mean, I think starting with what
• 14:09
  - are patients using today, I think AI shows
• 14:11
  - up in a lot more ways than, you
• 14:13
  - know, even we expect. And
• 14:16
  - one example is plenty of patients I see
• 14:18
  - in clinic
• 14:19
  - utilize wearable
• 14:21
  - technology, and,
• 14:23
  - they may not even be aware of the
• 14:24
  - fact that there are algorithms involved in the
• 14:26
  - wearable technology that are monitoring them and providing
• 14:29
  - a whole lot of health data,
• 14:31
  - maybe to their phone and an app or
• 14:34
  - being able to provide health alerts about sort
• 14:37
  - of abnormal heart rhythms, for instance, and and
• 14:39
  - and and changes in vital signs. And I
• 14:42
  - think there there's just so much cool,
• 14:45
  - ability from wearable data for patients to be
• 14:48
  - able to,
• 14:49
  - you know, identify earlier changes in their own
• 14:52
  - physiology and bring those up to their providers.
• 14:54
  - And I see increasingly
• 14:57
  - with a with a lower barrier to entry,
• 14:59
  - more and more of my patients with heart
• 15:01
  - failure who are engaging with health data in
• 15:03
  - a meaningful way as a result of, you
• 15:06
  - know, some aspect of their wearable consumer technology,
• 15:09
  - which is equipped with, you know, algorithmic and
• 15:12
  - and sensor data that's able to sort of
• 15:15
  - turn that into something relatively interpretable for a
• 15:18
  - patient and and allow for them to alert
• 15:21
  - me as their health care provider that something
• 15:23
  - may have changed. I think the other way
• 15:25
  - is I I think generative AI and large
• 15:27
  - language models are out there, and,
• 15:30
  - people are engaging with them now on their
• 15:32
  - phones, in their
• 15:34
  - on on any Internet enabled device. They're they're
• 15:36
  - asking questions.
• 15:38
  - They're learning from these large language models, and
• 15:41
  - a lot of what they're using it for
• 15:43
  - is to explore,
• 15:44
  - you know, the care that they're receiving and
• 15:47
  - how to get some additional advice about, you
• 15:50
  - know, taking care of themselves. So patients with
• 15:52
  - heart failure often have a lot of questions
• 15:54
  - about diet and exercise and fluid restriction and
• 15:57
  - salt and these kinds of things. And I'm
• 15:59
  - seeing patients increasingly
• 16:01
  - go to LLMs for the first pass and
• 16:03
  - then bring those answers to me. Hey. I
• 16:05
  - got this printout from, you know, whatever LLM
• 16:08
  - they they they subscribe to, and I wanted
• 16:11
  - to go through this with you and see
• 16:12
  - if this advice was was was the right
• 16:14
  - advice, which is really cool to see as
• 16:16
  - it evolves. And and here are some thoughts
• 16:17
  - that my my LLM friend,
• 16:20
  - told me to ask you some questions about
• 16:22
  - these things because I had this whole conversation
• 16:24
  - with that. And I I think it's a
• 16:25
  - really interesting emergent thing that's just happening
• 16:28
  - by virtue of the fact that these are,
• 16:30
  - you know, available to consumers widely.
• 16:32
  - They're really powerful foundational models, and they do
• 16:35
  - bring a lot of general knowledge. But as
• 16:38
  - you know, we have to be careful, and
• 16:39
  - so we have to be careful in in
• 16:41
  - green lighting the use of these tools and
• 16:43
  - and really encourage patients to engage with us
• 16:45
  - as their health care team as they're learning
• 16:48
  - from these models so that they're not going
• 16:49
  - too far astray of the evidence basis.
• 16:52
  - Yeah. And then I think, you know, future
• 16:54
  - state's really exciting. I might turn this to
• 16:56
  - Faraz to give a little,
• 16:58
  - sample of what he thinks the future is
• 17:00
  - gonna look like and then add a little
• 17:01
  - bit to that. Yeah. No. I think the
• 17:02
  - future of HeartFlare is super exciting. But just
• 17:04
  - to add to the comments about how patients
• 17:06
  - using AI, I completely agree that,
• 17:08
  - patients are using this all the time, whether
• 17:10
  - it's in their devices that the the wearables
• 17:12
  - are purchasing, whether it's in the medical grade
• 17:14
  - devices we're using for HeartFlare already, but also
• 17:16
  - just using things like ChatGPT and clog to
• 17:19
  - answer their questions. And I imagine this is
• 17:21
  - a lot like, what people physicians or clinicians
• 17:24
  - experienced back when Google came out, when patients
• 17:26
  - also, sorry, Google information had some access to
• 17:28
  - information.
• 17:29
  - But this is turbocharged because the barrier to
• 17:31
  - using these LMs is is so low, and
• 17:34
  - then the what you get back is so
• 17:36
  - can be so crisp.
• 17:38
  - And so it's it's an exciting time where
• 17:40
  - patients really educate themselves. I think we as
• 17:41
  - a community are still trying to figure out
• 17:44
  - how to educate patients on how to use
• 17:45
  - these tools.
• 17:46
  - In terms of the next three to five
• 17:48
  - years, you know, I think our hope is
• 17:50
  - that we see AI really in every aspect
• 17:52
  - of heart failure care from
• 17:54
  - initial diagnosis through treatment,
• 17:56
  - from when a patient calls us to the
• 17:58
  - health system, get scheduled in our get scheduled
• 18:00
  - for a visit and through their sort of
• 18:01
  - treatment journey.
• 18:03
  - I think, like many are, I'm really excited
• 18:05
  - about agents. We've already seen agents in Utah
• 18:08
  - helping with medication renewals. We've seen FDA clearance
• 18:11
  - for service scope use case, for an LOM
• 18:14
  - for, insulin titration. So I think we'll see
• 18:17
  - more and more use cases around heart failure,
• 18:18
  - especially starting with low risk aspects of heart
• 18:20
  - failure care using agents, talking to patients,
• 18:23
  - making initial recommendations with appropriate oversight, and really
• 18:26
  - being part of care.
• 18:27
  - My hope is that, ultimately,
• 18:29
  - AI
• 18:30
  - brings back more humanity
• 18:32
  - to
• 18:33
  - bring more, I should say, I guess, patient
• 18:35
  - clinician human connection into into care. You You
• 18:38
  - know, I'm already seeing this with my use
• 18:39
  - of ambient listening technology. You know, I used
• 18:41
  - to sit at my computer and type as
• 18:43
  - I I was talking to patients. Now I
• 18:44
  - can just put my phone down. It listens
• 18:46
  - and generates the note. And so I think
• 18:48
  - as we all become more comfortable with this
• 18:50
  - technology, as clinical social support gets more embedded
• 18:53
  - into ambient listening platforms, when it becomes more
• 18:56
  - context aware, meaning it takes in the record,
• 18:58
  - gets even better. I was seeing a lot
• 19:00
  - more time for this human connection that I
• 19:01
  - think is so important in clinical medicine. Really,
• 19:04
  - what we when really
• 19:05
  - important for patients to understand their health and
• 19:07
  - make changes that improve their long term outcomes.
• 19:10
  - Yeah. I I I totally and completely agree.
• 19:12
  - And I think, like, you know, just to
• 19:14
  - sort of reflect that back, I I totally
• 19:16
  - noticed today in the clinic with a patient
• 19:18
  - how using an ambient dictation software, you know,
• 19:21
  - put my phone down, it completely transformed the
• 19:23
  - way the clinic encounter goes. The screen is
• 19:25
  - no longer for me to look at. It's
• 19:27
  - for us to look at, and it's now
• 19:28
  - we're looking at the images together, and we're
• 19:30
  - having a discussion with more eye contact across,
• 19:33
  - you know, this the the fact that there's
• 19:35
  - not a computer between us. And it it
• 19:36
  - just really enables a more human interaction in
• 19:39
  - that moment. I think this is the sort
• 19:40
  - of future that would be amazing that we
• 19:42
  - would hope is is achievable in a three
• 19:44
  - to five year time horizon, which is that
• 19:46
  - right now, AI, when it's used, is very
• 19:48
  - visible.
• 19:49
  - Right? It's it's in your
• 19:51
  - we you kinda know there's a chatbot in
• 19:53
  - there. You know there's a there's an automated
• 19:55
  - phone call that's happening, and I'd love to
• 19:57
  - see it become invisible. And and and incorporate
• 20:00
  - it in all these aspects that Faraz is
• 20:01
  - mentioning as far as, you know, at every
• 20:04
  - access point to care right now, there's friction.
• 20:06
  - There's a difficulty with scheduling appointments. There's a
• 20:09
  - difficulty with getting follow-up testing. There's a difficulty
• 20:12
  - with understanding your lab results and getting real
• 20:14
  - time, you know, feedback about what's been happening
• 20:17
  - with medication changes. And the ability for agents,
• 20:21
  - agentic AI to sort of be an invisible
• 20:24
  - and very natural sort of interface and reduce
• 20:27
  - friction in each of those steps is is
• 20:29
  - coming and is amazing because it's gonna improve
• 20:31
  - access to care and hopefully
• 20:33
  - allow for
• 20:34
  - less of that friction and more of the
• 20:36
  - time that you want with the clinician, which
• 20:38
  - is not necessarily trying to navigate those barriers
• 20:41
  - to care, but actually just talking about your
• 20:43
  - health, asking the questions that may matter to
• 20:45
  - you, and and and and having somebody who's
• 20:48
  - really listening and able to explain things and
• 20:49
  - teach you very well.
• 20:51
  - And so I I think that's, like, sort
• 20:53
  - of we've already had a sample of that
• 20:54
  - in the now, and I'm really excited to
• 20:56
  - see how that continues to evolve and and
• 20:58
  - how in all of these different friction points,
• 21:00
  - we see, these these newer technologies
• 21:03
  - really help us.
• 21:05
  - Amazing. Well, thank you both so much for
• 21:07
  - being here and sharing all of your thought
• 21:09
  - leadership. This has been such a pleasure. Well,
• 21:11
  - thank you. We appreciate it.