Latest / Becker’s Healthcare Digital Health + Health IT / The Path to Large-Scale Adoption of AI in Healthcare - Customization, Trust, and Safety

Transcript

The Path to Large-Scale Adoption of AI in Healthcare - Customization, Trust, and Safety · Becker’s Healthcare Digital Health + Health IT

• 0:00
  - Hello everyone, this is Jacob Emerson
• 0:02
  with Becker's Healthcare.
• 0:03
  Thanks so much for tuning into the Becker's Healthcare
• 0:05
  podcast series where today I'm very pleased to be joined
• 0:08
  by Akiles Bpu, who is founder and CEO at Deep Scribe.
• 0:13
  He's here with me today to talk a little bit more about why
• 0:16
  AI note customization is key
• 0:18
  for widespread clinical adoption at large
• 0:21
  healthcare organizations.
• 0:22
  So Akiles, thank you so much
• 0:24
  for taking the time to be with us today.
• 0:26
  - Thanks, Jacob. I'm super glad to be here as well.
• 0:29
  - Yeah, we're glad to have you with us.
• 0:30
  And, and to get us started,
• 0:32
  I'm wondering if you could tell us a little bit more about
• 0:34
  yourself, your background in healthcare
• 0:36
  and what it is that you do today at Deep Scribe.
• 0:40
  - Yeah, absolutely. So prior to Deep Scribe,
• 0:43
  I was actually exposed to the problem
• 0:45
  of documentation at a pretty young age.
• 0:47
  You know, my dad was an oncologist
• 0:49
  and as a son of an oncologist,
• 0:51
  I actually would say I was desensitized
• 0:52
  to the problem of documentation.
• 0:54
  You know, he would spend upwards of half his day writing
• 0:58
  notes on everything his patient said to him,
• 1:01
  and a lot of that he would bring home.
• 1:03
  But as a kid, not much you can actually do about it.
• 1:07
  So I admitted defeat and a few years later,
• 1:11
  or roughly decade later,
• 1:14
  I eventually found myself at Bear,
• 1:16
  which is Berkeley's AI Research Lab.
• 1:19
  And I was working on summarization at the time.
• 1:22
  And you know, back in 2017, summarization
• 1:27
  was starting to do amazing things, including being able
• 1:32
  to distill one to two page CNN articles into nice,
• 1:36
  concise abstracts.
• 1:38
  It was pretty amazing. But every time I traveled back home,
• 1:43
  it was like I was going 10 years into the past.
• 1:45
  My dad was using the same tools, the same tech,
• 1:48
  and spending the same amount of time on documentation.
• 1:50
  And so I started to ask the question, you know,
• 1:52
  why not build a tool that could listen
• 1:55
  to my dad's conversations and summarize them
• 1:57
  and fill out his documentation for him?
• 2:00
  And if I were to do so, would
• 2:02
  that actually save him the time he
• 2:03
  was spending on documentation?
• 2:05
  Because obviously it meant a lot to me.
• 2:07
  So I took the same first
• 2:09
  principles approach we did in the lab.
• 2:11
  I tried to collect the data and then curate it
• 2:15
  and then train a model on it.
• 2:16
  But the data step is where I got stuck.
• 2:19
  There was no data set of patient physician conversations,
• 2:22
  there was no data set
• 2:23
  of the transcripts, the annotated notes.
• 2:26
  And so it was virtually impossible to train an AI to,
• 2:29
  to do this work at the bar of accuracy
• 2:32
  that's required in healthcare.
• 2:34
  But I think what was more interesting to us
• 2:36
  as technologists was that
• 2:38
  these conversations
• 2:40
  were actually healthcare source of truth.
• 2:42
  So every patient's story started from one
• 2:44
  of these interactions and no one was actually
• 2:46
  collecting this data.
• 2:48
  So, you know, we thought that if we could collect this data,
• 2:52
  not only could we solve documentation,
• 2:54
  but we could start to solve some
• 2:55
  of the larger problems in medicine.
• 2:57
  And so we started Deep Scribe to do exactly this.
• 3:00
  So we launched the product in 2020
• 3:02
  and we quickly became one
• 3:04
  of the market leaders in documentation.
• 3:06
  So today Deep Scribe is deployed in over a thousand provider
• 3:10
  organizations, everything from a private practice
• 3:13
  to a larger health system.
• 3:15
  And you know, on average we save about three hours a
• 3:18
  day for each of our doctors.
• 3:19
  But I think the most important thing to us is
• 3:22
  that we've been able to do this across a diversity
• 3:25
  of workflows and specialties.
• 3:27
  So when you see digital health products deployed, typically
• 3:30
  a lot of groups of clinicians are left out, you know,
• 3:32
  whether it's the specialists or whether it's the inpatient
• 3:35
  docs, because their workflows are highly nuanced
• 3:37
  and diverse to the point that the product can't fit them.
• 3:41
  However, with Deep Scribe, by deploying the primary care
• 3:45
  or private practices early on
• 3:47
  and seeing thousands of different workflow permutations,
• 3:50
  we were able to build what we call customization studio
• 3:53
  that allows any provider to fine tune deep scribe
• 3:56
  to fit whatever workload they had.
• 3:58
  So we've actually, by doing this, found
• 4:01
  that this is the single biggest key to adoption of any sort
• 4:04
  of AI in healthcare.
• 4:07
  - It sounds like this is a very personal passion for you,
• 4:09
  uh, on top of being obviously a professional one
• 4:12
  as well, Quiles?
• 4:14
  - Yeah, absolutely. I mean, I think one of my favorite parts
• 4:17
  of the whole Deep Scribe journey is
• 4:19
  that my dad actually uses Deep Scribe today,
• 4:22
  and then he's one of the clinicians that actually is able
• 4:25
  to save multiple hours in his
• 4:26
  day now because of Deep Scribe.
• 4:28
  So it came full circle and,
• 4:29
  and you know, I know I built it for my dad initially,
• 4:32
  but it's great to have other doctors be able to have access
• 4:35
  to this technology as well.
• 4:36
  - Sure. Yeah. No, it's very cool to hear.
• 4:38
  And I do want to get deeper into what's,
• 4:41
  what's the latest on clinical note taking, but,
• 4:43
  but I do wanna stay, take a step back for a second
• 4:45
  and ask you a little bit about a trend we've been seeing
• 4:49
  over the last few years, uh, just this rise in hospital
• 4:52
  and health system leaders,
• 4:54
  investments in artificial intelligence technology across the
• 4:57
  board, not just in clinical note taking.
• 4:59
  Yeah. So based on your experience ish,
• 5:02
  how are you seeing organizations deploy AI the most, um, or,
• 5:06
  or to support clinical documentation?
• 5:09
  Uh, are you seeing key challenges that they're looking
• 5:12
  to address ultimately?
• 5:15
  - Yeah, absolutely. So it's been amazing from our vantage
• 5:18
  point because when we started the company in 2018, we got
• 5:22
  to see sort of this adoption curve go from
• 5:26
  no one is really looking
• 5:28
  and talking about clinical documentation
• 5:30
  to now everybody is talking about clinical documentation.
• 5:33
  So in the beginning, you know, we actually made the decision
• 5:36
  to serve deep scribe to only private practices
• 5:39
  because the sales cycles
• 5:40
  and the implementation cycles
• 5:42
  that systems were just unsustainable for us
• 5:44
  as a, as a startup.
• 5:46
  And, you know, the key things there are just accuracy and,
• 5:49
  and proof points that tech would actually, you know,
• 5:51
  solve the challenges that the health systems are looking at.
• 5:54
  But over time, as we've built up a name
• 5:56
  for ourselves in the private practice industry,
• 5:58
  we've started to see that, you know, as the AI kind
• 6:03
  of thought space has evolved, a lot
• 6:05
  of health systems are now finally looking at, uh, AI
• 6:08
  as something that's here to stay.
• 6:10
  And so I think if you look at the numbers, you know,
• 6:12
  earlier this year there's a study done, um, that share
• 6:15
  that about 30%
• 6:17
  of systems are currently looking at documentation.
• 6:20
  And now I think, you know,
• 6:22
  the latest studies show something closer to 60%.
• 6:24
  So it's been amazing to see health systems kind
• 6:27
  of embrace AI.
• 6:29
  And you know, at the center of it, I think is the challenge
• 6:32
  of physician burnout.
• 6:34
  Um, I think that's what a lot of these systems are,
• 6:37
  are looking to solve with,
• 6:38
  with documentation tools like Deep Scribe.
• 6:40
  And I think that sits at the center
• 6:42
  of generative AI in healthcare for a couple reasons.
• 6:45
  You know, the first, I think there have been companies like
• 6:48
  Deep Scribe that have been building solutions
• 6:50
  and building the workflow enhancements
• 6:51
  and integrations, so that come time to implement,
• 6:55
  they're ready to go with all
• 6:57
  of the surrounding functionality.
• 6:58
  And now you, you have Cort Tech that's finally mature,
• 7:01
  so now you can actually deploy these mature solutions.
• 7:04
  But I think the second thing is that, um, it's really easy
• 7:07
  to see the value from these solutions.
• 7:10
  So, you know, with Deep Scribe, for example, as soon
• 7:13
  as it's implemented within a few hours,
• 7:15
  you're seeing clinicians actually start to save time
• 7:18
  and send, you know, little notes to administration that,
• 7:21
  you know, software IE deep scribe has been amazing
• 7:25
  because now they don't have to focus on notes anymore.
• 7:27
  They can now focus on their patient.
• 7:29
  Some physicians are that have been scaled down to just two
• 7:32
  to three days of, of practice from the full four
• 7:35
  or five are now going back to the four or five
• 7:37
  because of these documentation tools.
• 7:39
  So I think, you know, you're able to see instant value
• 7:42
  and, you know, I think the tech is mature enough for,
• 7:45
  for this specific space,
• 7:46
  which is why I think organizations see this as the,
• 7:49
  the most obvious first step in their generative AI journey.
• 7:54
  - And you mentioned a, a jump from 30 to 60%
• 7:57
  or so of systems now looking into AI
• 8:00
  powered clinical documentation.
• 8:03
  What advancements do you attribute that to in terms of
• 8:07
  where clinical note taking has
• 8:08
  improved over the last few years?
• 8:10
  And are, are you finding that customization fits in here?
• 8:13
  Um, why is all this significant, I guess is
• 8:16
  what I'm asking, Ailish?
• 8:18
  - Yeah, I think that's a good question.
• 8:20
  So, you know, since the beginning
• 8:22
  and still now, I think the challenges
• 8:24
  to adopting a technology like this have been the same.
• 8:27
  The two biggest are trust and customization
• 8:31
  and trust in our world is akin to accuracy.
• 8:34
  So if those are the problems, I think in terms
• 8:36
  of advancements, a lot of people see that the advancements
• 8:41
  for accuracy have been primarily tailored around GPD four
• 8:45
  and the advancements in foundation models.
• 8:47
  But we actually see that the biggest advancements were in
• 8:51
  open source lms.
• 8:52
  So, you know, while GPD four is incredible,
• 8:55
  without the right checks
• 8:56
  and balances, that mislabel a medication
• 8:58
  almost every single time.
• 9:00
  So this open source of movement has been amazing
• 9:02
  because as you collect a small amount
• 9:05
  of domain specific data
• 9:07
  and train these alums on it,
• 9:08
  you can quickly become on par with GPD four.
• 9:10
  But with large amounts like we have, you can actually
• 9:14
  solve problems like hallucinations or misinterpretation
• 9:17
  or other common mistakes fairly effectively.
• 9:20
  So I think that's the first big advancement is
• 9:22
  that trust has just come a long way
• 9:24
  and accuracy has come a long way to,
• 9:26
  to finally meet that bar.
• 9:28
  But the second thing that I think you mentioned was
• 9:30
  customizations, and I think that's probably
• 9:33
  the other large advancement.
• 9:35
  And so the product, in order
• 9:37
  to get full widespread adoption, can't just work for
• 9:41
  that one clinician that gets all their, uh,
• 9:43
  information from the conversation.
• 9:45
  They're okay with unstructured notes.
• 9:47
  It also has to work for, you know, the oncologist
• 9:49
  that gets most of their node from previous visits,
• 9:52
  or the highly specific geriatricians that want the node
• 9:56
  to follow exact structure that they've been using
• 9:58
  for the last like 10 years.
• 10:00
  So, um, I think the, the second big advancement
• 10:03
  that we've seen is, you know,
• 10:05
  given the advancements in foundation models
• 10:07
  and open source lms, we can now evolve that
• 10:11
  to customizations.
• 10:12
  And Deep Scribe has embraced that
• 10:14
  and released something called Customization Studio
• 10:17
  that is the culmination of all these little things
• 10:19
  that clinicians have asked us
• 10:20
  for over the last five years wrapped into this,
• 10:23
  I guess, seamless interface.
• 10:24
  And you can kind of think of it as clinicians now being able
• 10:27
  to build their own custom model in a few minutes.
• 10:31
  So when it comes to adoption, now you have trust
• 10:34
  that's relatively solved as well as customizations
• 10:37
  that's also solved that are the two keys
• 10:39
  to just rolling out something like deep scribe
• 10:41
  system wide. I
• 10:43
  - Wanna follow up with you on, on the trust aspect of this
• 10:45
  because I think when we're talking about, uh,
• 10:47
  implementing AI tools within healthcare, the other half
• 10:50
  of this conversation has to be around, uh,
• 10:52
  the responsibilities around using this technology.
• 10:56
  And we know the healthcare industry is working
• 10:58
  to set guidelines right now,
• 10:59
  this year they're being released around the use
• 11:01
  and development of these technologies given the, the,
• 11:05
  the range of concerns around accuracy, privacy, uh,
• 11:09
  potential biases and discriminatory concerns.
• 11:12
  And so I would ask you what considerations,
• 11:15
  what best practices do you recommend for,
• 11:18
  for leaders in charge of these tools as they explore,
• 11:22
  as they begin to implement these tools
• 11:25
  and like what you mentioned as they begin
• 11:26
  to customize these note taking
• 11:28
  tools that we're talking about?
• 11:31
  - Yeah, absolutely. I mean, I think there,
• 11:34
  there are three things that I would highlight
• 11:37
  as very important to, to look into and, and get right
• 11:40
  before adopting generative ai.
• 11:44
  I think the first thing is everybody should do their own
• 11:47
  benchmarks or at least develop their own
• 11:49
  benchmarking methodology.
• 11:51
  And in the, you know, AI documentation space, uh,
• 11:54
  especially, you know,
• 11:55
  you can plug in a conversation into GPD four
• 11:59
  and have a note that comes out
• 12:00
  that looks like it would work.
• 12:02
  But when you examine it closely,
• 12:04
  you'll see hallucinations happen.
• 12:06
  You'll see, uh, things like misinterpretation
• 12:08
  or miscategorizing
• 12:10
  or mislabel that can have, um,
• 12:12
  pretty significant downstream effects.
• 12:14
  So having a di having a benchmarking methodology would help
• 12:17
  a system, you know, prevent against that.
• 12:19
  And, and we have our own a deep describe that we're happy
• 12:21
  to share with anyone that's, that's looking at this.
• 12:23
  But I think that's the, the first thing
• 12:24
  to, to catch mistakes early.
• 12:26
  The second thing is human oversight.
• 12:29
  So, you know, and you can even do, you,
• 12:31
  you can do this even if the solution is fully automated.
• 12:33
  So deep describe, for example,
• 12:35
  our solution's fully automated,
• 12:36
  but we have a team of folks internally
• 12:39
  that review at random different notes to ensure that
• 12:43
  there are no hallucinations that are currently being missed
• 12:46
  by, uh, clinicians or misinterpretations
• 12:49
  or, or other errors.
• 12:50
  So that's, I think the second thing that's important.
• 12:53
  The third thing is automated oversight,
• 12:55
  and we've done some work here as well,
• 12:57
  but you can actually configure LLMs to be able
• 13:00
  to catch their own mistakes by having one
• 13:04
  specific dedicated reviewer, LLM to to review the notes,
• 13:07
  compare it with the transcript,
• 13:08
  and see if there's anything, uh, that's been missed.
• 13:10
  So I think those three things I I would say are,
• 13:13
  are probably the most important as it pertains to,
• 13:16
  to guidelines and, and, and safe AI development.
• 13:19
  - Yeah. Well, I appreciate you taking us through that
• 13:21
  and offering all those details.
• 13:22
  The last thing I wanted to ask you about achiles
• 13:25
  is on the ground.
• 13:27
  I know you, you mentioned your father
• 13:29
  and how he's been using this tool in oncology
• 13:32
  and some other physicians you know, as well.
• 13:34
  But can you share some, some on the ground examples
• 13:37
  or some case studies of where we are seeing hospitals
• 13:40
  and health systems deploy AI for customized medical notes
• 13:45
  and ultimately the outcomes that you, that you're seeing
• 13:47
  and that they are seeing from that?
• 13:50
  - Yeah, absolutely.
• 13:51
  So one of my favorite case studies
• 13:55
  is Covenant Healthcare.
• 13:57
  So Covenant Healthcare has used Deep Scribe
• 14:00
  for actually a little bit less than a year.
• 14:03
  They're on Epic.
• 14:05
  And, um, within six months, they were able
• 14:08
  to not just reduce documentation time
• 14:11
  by roughly 53% across all their providers,
• 14:15
  but they were able to see a 72% reduction in pajama time.
• 14:20
  And that's my favorite stat
• 14:21
  because, um, a lot of the work
• 14:23
  that clinicians do ends up being carried home
• 14:26
  and being able to reduce that by three fourths.
• 14:28
  So, you know, if they're spending an hour
• 14:30
  or a day on it in the evening, now they only have
• 14:33
  to spend 15 minutes, I think is a big, big win.
• 14:35
  Second big win is that, you know, even though we like
• 14:38
  to say we're able to save clinicians' documentation 80, uh,
• 14:41
  reduced documentation by 80% in practice
• 14:44
  because of the diversity of clinicians
• 14:46
  and their workflows, um, we've been able to show
• 14:48
  that we were able to do that by 53%
• 14:50
  consistently across everybody.
• 14:51
  So regardless of if you're, you know, the hyper picky doc
• 14:54
  that wants their documentation a certain way
• 14:57
  or specialist that has a highly configured workflow,
• 14:59
  we're able to make an impact, um, for, for everyone.
• 15:02
  So I think that's awesome to see in,
• 15:04
  in the last great tidbit that I wanted to insert there from
• 15:07
  that study is that they actually just in six months made the
• 15:10
  call to replace their virtual scribe program just months
• 15:13
  after implementing the scribe.
• 15:14
  So I think that's a great one that shows some
• 15:17
  of the value you can get from
• 15:18
  generative AI and documentation.
• 15:21
  - Yeah. And those are some incredible metrics you shared,
• 15:23
  especially that 53% reduction in documentation across
• 15:27
  specialties across all kinds of physicians.
• 15:29
  That's, that's amazing. Um,
• 15:31
  before we go though, Aish, any other final thoughts or,
• 15:35
  or key takeaways you would like to share with all
• 15:38
  of our health system listeners, uh, listening in today?
• 15:42
  - Yeah, absolutely. I mean, I think we're about to roll in
• 15:46
  to the biggest year for Gen AI in healthcare yet,
• 15:50
  and I think everybody is extremely excited, um,
• 15:53
  whether you're a startup
• 15:54
  or, uh, that's, that's building a solution for this
• 15:57
  or health system that that's, that's integrating it.
• 15:59
  But I think the two things to really look out for our trust,
• 16:03
  so, you know, how accurate is the ai,
• 16:05
  what benchmarking methodology does this company use in the
• 16:08
  backend to verify that, you know,
• 16:10
  in production it'll be reliable and and trustworthy.
• 16:12
  And the second thing is customization.
• 16:14
  So how can the AI adapt to all the, the big, I guess,
• 16:19
  diverse amount of workflows that your clinicians have
• 16:22
  and how does the product account for that?
• 16:24
  And I think those are gonna be the two most important things
• 16:28
  to get the adoption that everybody here is looking for.
• 16:30
  So, you know, we've done a lot of work on this and,
• 16:32
  and happy to share notes with anyone interested,
• 16:34
  but I think those are the two things
• 16:36
  that we predict are gonna be very important.
• 16:38
  And so I'm super excited to, to see this year play out and,
• 16:42
  and see just how much AI is gonna really impact and,
• 16:47
  and improve healthcare outcomes.
• 16:50
  - Fantastic. Well, Aish, thank you so much
• 16:52
  for taking the time to be with us
• 16:54
  and for sharing your insights
• 16:55
  with us. We really appreciate it.
• 16:58
  - Yeah, thanks for having me. Jacob.
• 17:00
  - We'd also like to thank our podcast sponsor, deep scribe,
• 17:03
  you continue to more podcasts from Becker's Healthcare
• 17:06
  by visiting our podcast page at becker's hospital
• 17:09
  review.com.