Latest / Becker’s Healthcare Podcast / Akhil Narang, MD & Amogh Karnik, MD of Northwestern Medicine Bluhm Cardiovascular Institute

Transcript

Akhil Narang, MD & Amogh Karnik, MD of Northwestern Medicine Bluhm Cardiovascular Institute · Becker’s Healthcare Podcast

• 0:02
  - Just five years ago, a big portion of
• 0:04
  - what's now being done in ambulatory surgery centers
• 0:07
  - wasn't considered outpatient work. The cases have changed,
• 0:10
  - the tech has changed, the patient expectations have
• 0:13
  - changed, and the leaders who are thriving are
• 0:16
  - the ones who saw it coming.
• 0:18
  - Becker's twenty third annual Spine Orthopedic and Pain
• 0:21
  - Management Driven ASC Conference
• 0:23
  - exists for exactly that kind of forward looking
• 0:25
  - practitioner.
• 0:27
  - Here, Becker's hosts more than 1,000 attendees, 275
• 0:31
  - physician and ASC speakers, and three days of
• 0:34
  - sessions that move between the clinical, the operational,
• 0:37
  - and the strategic,
• 0:38
  - and the CME credits to show for
• 0:40
  - it. Join us June eleventh through the thirteenth
• 0:42
  - at the Swiss Hotel in Chicago, and come
• 0:45
  - ready to think about what's next. Register on
• 0:47
  - our events page at beckershospitalreview.com
• 0:50
  - by clicking on the events tab in the
• 0:52
  - upper right.
• 0:55
  - Welcome to the Becker's Healthcare Cardiology and Heart
• 0:58
  - Surgery podcast. I'm Mariah Taylor, assistant editor, and
• 1:01
  - I'm thrilled to be speaking with two leaders
• 1:03
  - from Northwestern Medicine's Bloom Cardiovascular Institute.
• 1:07
  - First, we have Akhil Narang, cardiologist and director
• 1:10
  - of the echocardiograph
• 1:12
  - lab and Amogh Kardnick, cardiologist.
• 1:15
  - Thank you both so much for joining me
• 1:17
  - today. I'm really excited to talk more about
• 1:20
  - how you're using AI in your EC lab.
• 1:23
  - But before we dive into that, I would
• 1:25
  - love for you to introduce yourselves and tell
• 1:27
  - us a little bit about your background. Doctor
• 1:28
  - Narang?
• 1:30
  - Sure. Thank you so much. It's a pleasure
• 1:31
  - to be here. My name is Akhil Narang.
• 1:34
  - I'm an associate professor of medicine here at
• 1:36
  - Northwestern.
• 1:37
  - And I'm a multimodality
• 1:38
  - imager, and I also direct the echocardiography
• 1:41
  - laboratory here at Northwestern.
• 1:43
  - And I have a lot of interest in
• 1:44
  - artificial intelligence and how we use this in
• 1:46
  - cardiac imaging and beyond.
• 1:48
  - Awesome. And thanks for having me again as
• 1:50
  - well. My name is Amogh Karnik.
• 1:52
  - I'm,
• 1:53
  - a clinical assistant professor at Northwestern,
• 1:55
  - and cardiologist here.
• 1:58
  - Also have a very strong interest in artificial
• 2:01
  - intelligence and its application to cardiac imaging,
• 2:04
  - with a specific interest in echocardiography.
• 2:08
  - Excellent. We're really excited to have both of
• 2:10
  - you on the podcast today to dive a
• 2:11
  - little bit more into how you're utilizing and
• 2:14
  - thinking about AI.
• 2:16
  - So let's kick it off with this. What
• 2:18
  - is it about echocardiography
• 2:20
  - specifically that makes it such a compelling entry
• 2:22
  - point for AI compared to other imaging modalities?
• 2:26
  - Absolutely. So I think Echo ends up being
• 2:28
  - the natural training ground for AI
• 2:31
  - and cardiology
• 2:33
  - for a variety of reasons. I think the
• 2:34
  - volume, variability, and and the nature of the
• 2:36
  - data itself. So echo is by far the
• 2:39
  - most frequently performed cardiac imaging tests in the
• 2:42
  - world, probably.
• 2:43
  - It's cheaper than cardiac CT, more accessible than
• 2:46
  - cardiac MRI, and doesn't carry radiation or contrast
• 2:49
  - concerns that some of the other modalities
• 2:51
  - do. And so that ubiquity means there's a
• 2:54
  - massive naturally occurring dataset.
• 2:56
  - Millions of studies throughout the country and the
• 2:59
  - world each year that are often paired with
• 3:01
  - outcomes. And that simply doesn't exist at that
• 3:03
  - same scale for other modalities at present.
• 3:06
  - I think it's also,
• 3:08
  - a degree of variability that exists in echo
• 3:10
  - choreography from interpretation.
• 3:12
  - Because these are dynamic
• 3:14
  - moving images,
• 3:16
  - despite our best efforts with guidelines and recommendations,
• 3:19
  - there can be some differences in how studies
• 3:21
  - are interpreted. Sometimes this is reflected by the
• 3:24
  - expertise of the interpreter as well. And so
• 3:27
  - as a result of the variability,
• 3:29
  - echo
• 3:30
  - choreography also lends itself to being excellent for
• 3:32
  - AI since we can take away some of
• 3:34
  - the variability by really standardizing
• 3:37
  - our outputs.
• 3:38
  - And the third is the nature of the
• 3:39
  - data itself. I would say echo, as I
• 3:41
  - mentioned is real time, it's dynamic, it's moving
• 3:44
  - images.
• 3:45
  - And because of this, there's a lot of
• 3:47
  - a richer signal
• 3:49
  - than we can see from fixed structural images.
• 3:52
  - And this plays directly into what modern deep
• 3:54
  - learning architectures are good at finding patterns across
• 3:57
  - sequences, not within just a single frame. And
• 4:00
  - because of this rich dataset, we're able to
• 4:02
  - take advantage of
• 4:04
  - this in the AI space.
• 4:07
  - Excellent.
• 4:09
  - Now one of the persistent barriers we hear
• 4:11
  - about here at Becker's when it comes to
• 4:13
  - AI adoption is the gap between what's validated
• 4:16
  - in research versus what actually holds up in
• 4:18
  - a live clinical environment.
• 4:20
  - So what has to be in place operationally
• 4:22
  - before an Echolab can deploy AI with confidence?
• 4:27
  - It's a great question. So I think
• 4:29
  - the number one thing that we've come across
• 4:31
  - in our experience here at Northwestern
• 4:33
  - and also having studied,
• 4:35
  - you know, the adoption of clinical AI,
• 4:39
  - kind of
• 4:40
  - system wide and around the country,
• 4:42
  - is that really it's a problem related to
• 4:45
  - infrastructure.
• 4:46
  - It's imperative that systems have appropriate infrastructure in
• 4:49
  - place
• 4:51
  - in order to essentially not only just deploy
• 4:54
  - that artificial intelligence
• 4:55
  - pipeline, but also be able to monitor it
• 4:58
  - for
• 4:59
  - performance
• 5:01
  - in order to adjust to how things evolve
• 5:04
  - over time when we're deploying live
• 5:07
  - with actual real world patients.
• 5:10
  - The other important thing to hear that to
• 5:12
  - sort of take into account here is that
• 5:15
  - it's also important to study these
• 5:18
  - operational tasks and make sure that we're performing
• 5:21
  - at the level that we would expect
• 5:23
  - to perform,
• 5:24
  - by actually generating evidence and data
• 5:27
  - to to show to show that that's the
• 5:29
  - case. And so,
• 5:30
  - obviously, there's been a lot of literature that's
• 5:33
  - been published in
• 5:34
  - major,
• 5:35
  - sort of,
• 5:37
  - scientific journals about research validation of AI pipelines,
• 5:40
  - but less so from a real world implementation
• 5:43
  - standpoint. And so I think being able to
• 5:44
  - generate data and use that data, to justify
• 5:47
  - how tools are being implemented and then supporting
• 5:50
  - the infrastructure from PAC systems to
• 5:53
  - information systems for storage of images, storage of
• 5:56
  - reporting, making sure that's all integrated with the
• 5:58
  - electronic health record.
• 6:00
  - All of those systems need to be in
• 6:01
  - place, before AI can be meaningfully clinically implemented.
• 6:06
  - And I would just add to that that
• 6:09
  - there also needs to be
• 6:11
  - a mechanism to account for discordance. And so
• 6:14
  - if you have an AI algorithm
• 6:16
  - that disagrees
• 6:18
  - perhaps with what the human is seeing, how
• 6:20
  - do we reconcile this? There needs to be
• 6:22
  - a defined pathway, not really an ad hoc
• 6:24
  - conversation
• 6:25
  - about figuring out, you know, how we handle
• 6:28
  - this coordinates between outputs.
• 6:30
  - And then even before that, I think
• 6:33
  - each AI algorithm is often developed
• 6:35
  - in a population that may not be locally
• 6:38
  - relevant to a particular hospital or system. So
• 6:41
  - the idea of trust but verify needs to
• 6:43
  - happen where you almost need a shadow phase
• 6:45
  - where you run an algorithm
• 6:47
  - alongside
• 6:48
  - what a human is typically doing to make
• 6:50
  - sure that the AI
• 6:52
  - works well within the system that you are
• 6:54
  - now deploying it in.
• 6:57
  - Interesting.
• 6:58
  - Can you tell me a little bit more
• 6:59
  - about how AI is changing the workforce needs
• 7:02
  - and team dynamics inside hospitals?
• 7:07
  - So I think there's a a few points,
• 7:10
  - as we as we sort of see AI
• 7:12
  - slowly become more and more integrated into our
• 7:15
  - daily workflows.
• 7:17
  - We know that AI offers the opportunity to
• 7:20
  - enhance efficiency, or at least there's the promise
• 7:23
  - of enhanced efficiency.
• 7:24
  - And with that comes
• 7:26
  - changes in the way that our
• 7:29
  - workforce as a whole is
• 7:31
  - dealing with
• 7:32
  - volumes in terms of seeing patients. So for
• 7:34
  - instance,
• 7:37
  - AI augmented echocardiography
• 7:39
  - offers the chance to automate measurements
• 7:42
  - that would potentially be done by sonographers and
• 7:44
  - echocardiographers
• 7:46
  - in the review of studies. And so
• 7:49
  - with AI assisted imaging technologies,
• 7:51
  - when we're at the point of actually acquiring
• 7:53
  - those images, that means that we can actually
• 7:55
  - speed up the process potentially of acquiring images
• 7:58
  - and then potentially
• 8:00
  - downstream actually do more studies,
• 8:03
  - or sort
• 8:05
  - of deal with increased patient volumes because we
• 8:07
  - can get through,
• 8:09
  - you know, more studies just in a day.
• 8:11
  - Obviously, that has a big impact on the
• 8:13
  - workforce because that means that an individual sonographer,
• 8:17
  - individual reader is gonna be dealing with more
• 8:19
  - images and more studies on a daily basis.
• 8:21
  - And so how that sort of evolves,
• 8:24
  - over time as
• 8:25
  - these tools sort of get implemented and more
• 8:28
  - solidified into daily workflows
• 8:30
  - still remains to be seen. I think there's
• 8:32
  - still
• 8:32
  - some
• 8:34
  - more work and more evidence to be generated
• 8:35
  - about how this actually has an impact on
• 8:37
  - efficiency
• 8:38
  - and workload and
• 8:40
  - also thinking about burnout and and and whether
• 8:42
  - we're
• 8:43
  - putting too much stress or too much tax
• 8:45
  - onto the system.
• 8:47
  - Absolutely.
• 8:48
  - Now we all know AI is very common
• 8:51
  - now amongst all different kinds of imaging, including
• 8:54
  - CTs, MRIs, ECGs, etcetera.
• 8:57
  - How should hospital system leaders be thinking about
• 9:00
  - where echo specific AI investments fit into that
• 9:03
  - broader multi modality imaging strategy?
• 9:07
  - So the way that I would frame this
• 9:08
  - for health systems is not to think of
• 9:10
  - it as one imaging modality competing against another
• 9:13
  - imaging modality.
• 9:15
  - I would say that a lot of what
• 9:16
  - we do in AI, whether it's an echo,
• 9:19
  - cardiac MRI, or CT,
• 9:21
  - is complimentary rather than competing. And so the
• 9:23
  - strategy should meet them from from that perspective.
• 9:27
  - I think first of all, echo and also
• 9:29
  - ECG happened to have a faster cycle time.
• 9:31
  - So they're lower cost per study, higher volumes,
• 9:35
  - and easier to pilot at scale quickly when
• 9:37
  - it comes to AI solutions.
• 9:39
  - Now you can do the same with CT
• 9:41
  - or MRI,
• 9:42
  - it just happens with a different cycle speed.
• 9:44
  - And so I think that's one thing to
• 9:46
  - pay attention to. The second, I think, you
• 9:48
  - know, the prioritization within a hospital system should,
• 9:52
  - allow for a reasonable return on investment. And
• 9:55
  - so often AI solutions are expensive to stand
• 9:59
  - up, and I think you have to be
• 10:00
  - able to make sure that the bills are
• 10:02
  - being paid at the end of the day.
• 10:03
  - And so if you're implementing one solution with
• 10:05
  - echocardiography
• 10:07
  - versus MRI and CT, how does that affect
• 10:09
  - the bottom line with respect to
• 10:11
  - doing what's right for the patient, but also
• 10:13
  - not making sure that the revenue,
• 10:15
  - ends up in in the black? And then
• 10:17
  - the last I think where we we have
• 10:19
  - to think about things is this is really
• 10:21
  - about the connectedness of a patient not really
• 10:24
  - separate wins for separate areas of cardiac imaging.
• 10:27
  - And so a patient for instance with a
• 10:29
  - disease entity like heart failure
• 10:32
  - will have information that might be detected with
• 10:34
  - an echo that might need to be followed
• 10:36
  - with an MRI or CT. And so the
• 10:39
  - real value, I think, brings,
• 10:41
  - comes from bringing together the multiple imaging modalities,
• 10:45
  - not from a single modality.
• 10:47
  - And it's about really improving patient care at
• 10:49
  - the end of the day. And so, you
• 10:51
  - know, I think the the sum of all
• 10:53
  - these parts probably is greater than the individual
• 10:55
  - aspect of them. And so I think really
• 10:57
  - focusing on how does this benefit the the
• 10:59
  - patient
• 11:00
  - and help improve our outcomes for our patients
• 11:03
  - is where is where we need to pay
• 11:04
  - attention to things.
• 11:06
  - Excellent.
• 11:08
  - Well, this has been such a great discussion.
• 11:09
  - Thank you both again for joining me on
• 11:11
  - the podcast today.
• 11:13
  - Thank you. Thank you.