Loading summary
Transcript58 lines
- [00:00]
A
Hello, everyone.
- [00:01]
B
This is Erica Spicer Mason with Becker's Healthcare. Thank you so much for tuning into the Becker's Healthcare podcast series. So today we're going to talk about scaling ambient AI across care settings. And joining me for this conversation, we have two very special guests. First, we have Roberta Schwartz, and she is the Executive Vice President and Chief Innovation Officer of Houston Methodist Hospital. So a little bit about Roberta and Houston Methodist. Houston Methodist is one of the founding institutions of the Texas Medical center, and she's the Chief Innovation Officer at the hospital. And as Executive Vice president, Roberta oversees all operations at the 979 bed hospital, which U.S. news and World Report named the number one hospital in Texas for 13 straight years. And we also have with us Nikhil Baduma, who is the co founder and CEO of Ambience Healthcare, where he leads a team dedicated to empowering clinicians and health systems with AI technology. Drawing on his deep background in machine learning, and as author of O'Reilly's Fundamentals of Deep Learning, Nikhil brings a unique perspective on how technology can drive both clinical and operational outcomes across healthcare organizations. So, Nikhil, Roberta, thank you so much again for making time for this today. And welcome.
- [01:14]
C
Thanks for.
- [01:14]
A
Great to be here.
- [01:15]
B
Great to have you both. And I'm excited to dig right into this topic. Nikhil, wondering if you can help set the stage for our listeners today. How has the ambient AI space evolved in the last few years? Can you just give us a sense of what's taken shape?
- [01:30]
C
Yeah, you know, to take a step back, this whole phenomenon of AI has really only taken shape in the last handful of years, right? ChatGPT. I think we forget that it only came out in November of 2022. And I think what's surprising about this wave of technologies is the sheer speed at which it's moved compared to previous technology trends. One of the metrics we think about is how long it takes a new technology to get to 100 million people. Do you know how long it took the Internet?
- [02:01]
B
I don't know, 10 years?
- [02:04]
C
Six years?
- [02:04]
A
Six. Okay.
- [02:05]
C
The Internet.
- [02:06]
A
Okay.
- [02:06]
C
For mobile, 512 for mobile, for chatgpt. Do you have a sense so much shorter.
- [02:15]
A
Months.
- [02:16]
C
Months. Two months.
- [02:17]
B
Wow.
- [02:18]
C
I think that's the crazy part. You kind of think about how many technology innovations do we go from no one really is talking about it to almost everyone has experienced it on their phones at their fingertips. And I think that's probably one of the sort of really interesting macro trends that drives a lot of how we think about AI in healthcare, which is all of A sudden everyone had the opportunity to experience it in their personal lives. And now we're starting to beg the question of what would it feel like to have these technologies with us at work. And I think that's largely what's driven the excitement around this class of technologies. I think what's challenging about bridging the gap between the promise of AI technology and then the reality of it in healthcare and the clinical workflow is that healthcare has this incredible last mile problem, which is you think about an organization like Houston Methodist, you've got dozens and dozens of different specialty and subspecialty areas, some of the most sub specialized clinicians in the entire world and the medicine they practice is so different. How they use the EHR HR is so different. All of the coding and billing workflows are so different. And so I think what we've found is that it takes an incredible amount of work, attention to detail and obsessiveness to go from this high level idea of, hey, let's use AI to make clinicians lives better to actually making it such that clinicians use these technologies every single day that they benefit from them and we see real results and outcomes. And so I think that's sort of been the evolution of the last three years since sort of the inception of ChatGPT.
- [04:02]
B
Yeah, I really appreciate the overview, Nikhil. And just a quick follow up to that. You know, based on everything you just outlined, how did ambience approach the space differently than perhaps other technology companies?
- [04:14]
C
You know, I think we learned a lot from Roberta, which is, you know, and we, we respect the hell out of her because she is one of the most pragmatic leaders in our space. And one of the pieces of feedback she gave to us at the very beginning was, look, we are a really complicated organization. Our clinicians have an extremely high bar for technology and there's been several instances of bringing technology into the clinical workflow and ultimately the promise of the technology doesn't actually translate to outcomes. And so if you can get to a place where clinicians are actually pulling the technology out of your hands, then you've succeeded. And so a lot of what we've done is sort of centered around this idea of what does it take to build technology that becomes the clinician's first choice? And I think to serve an organization like Houston Methodist, well, you can't take a single monolithic product and then try to retrofit it to all of the different service lines. You almost have to from first principles, think about how do you build a bespoke product that solves the problem for the primary care physician, that solves the problem for the hematologist oncologist, that solves the problem for the bariatric surgeon. Because you think about, you know, documentation is this interesting thing where it's not actually a summary of the conversation you're having with the patient. It's a reflection of the medical decision making that's happening in your brain. And most of the time, especially in the specialty areas, that decision making, that thinking, is never verbalized. And so to create this magical experience where it feels like the technology is reading your mind requires you to fine tune these models in such a way that they actually are interpreting and reading between the lines of what's spoken versus what's being thought in the mind of the clinician. So there's all this work that has to happen to make that work by specialty. And then you think about the differences in EHR workflows, you think about the differences in coding and billing workflows, and all of a sudden you're building like 100 different products. You're not building one product. And so I think in many ways that advice from Roberta really shaped how we think about building products for health systems.
- [06:24]
B
Yeah, Nikhil, thank you so much for adding on. And it's a great segue to Roberta. I want to get insights from you next. It sounds like I know you have so much respect in this space and I would love to know more about what your experience has been with AI in health care. It's a broad question, a loaded question, but would love to know from your perspective what some of the key initiatives are that you are focused on now.
- [06:47]
A
AI is so broad. I'm going to stick to the generative AI in this space. Sure. I'm just. Sure, yeah, we, we get a little more focused. But I'll also take a step back in history. I've been in this field going on 33 years now. It's a long time. And for most of my career, if you think about it, most of our doctors were documenting in paper charts the emr. And the explosion of the EMR really happens at the point where you get the incentives to go into the EMR and meaningful use. And sometimes I'm still amazed that I can walk to a health system and they can tell me that they still have like five different instances of EPIC or six different EMRs. And I'm like, I don't understand because at least, thank God, you know, we've been on one instance of an electronic medical record. So the concept that There is something to dictate into the fact that there's something that can line up orders. The fact that there's something that can do anything is not that old same thing, right? It's not. You're not talking about 20 and 30 years, you know, you're talking about in the single and double digits that this has even become an opportunity. Now you then marry that with a technology that can take away. Since we've gotten into there, right? Since we got into the electronic medical record era, you have all of your clinicians were spending probably 50% of their time as a slave to this new electronic medical record, which is supposed to help them. Now they get a whole lot of information. They know what is in their next door neighbor's paper chart, right? Because oh my gosh, all that information is there. You get into, you know, oh, wait, I don't know what happened to them in the hospital too. Oh, wait, I could just read about their last visit. Great, right? But. But they then become the slaves to putting in all of that information. And the amount of information that the government requires is substantial, even if it's cut down. And in addition to that, the nurses who wanted to look at things in work lists and the doctors who wanted to look at things in flow sheets, like, everything looks and feels different. That's why the pragmatic View, it's not one product, it's 100, because each one of those clinicians use it and look and feel and have customized the development of that to something that works for them. The GI physicians want pictures that come out of their endoscopy, you know, and they want to see the inside of my stomach guts. Yuck. Okay, but so as they're doing it, it's really important that they're going to be able to describe pictures and images that's very different than a dementia doctor who's going to be talking about, you know, what's happening. So fast forward to our experience of our clinicians, which is when they were given the opportunity, and again, they went into it very skeptically, as they do with almost any product we bring. When they got a hold of Gen AI and this ability for us with ambience to now just have a conversation like we're having, walk away from the conversation, have their note written, have their orders lined up, and basically have the chart suggesting things they may or may not have documented from former records, it was like nirvana had happened. Like they were given back an ability that they never had before, which is we reduce the amount of time they're Spending, we reduce their cognitive burden. Right. And we gave them back time. And if time is the most important thing that we all have, then we just gave them back the biggest gift. So I get notes from our doctors and I ship them on, which are. You will pry this out of my cold, dead hands to. You have no idea how fundamentally you have changed my work, or my work life balance, or my ability to get back to academics that I loved, or my ability to truly help the patients. Because I have a few moments to think rather than being basically a rote scribe.
- [11:22]
B
Yeah, yeah. It's extremely powerful, everything that you've just said. And it's so interesting, too, Roberta, to hear how you're thinking about application of this tool across specialties as well, and what they need and how there are those nuances as well. And the fact that there is a tool that exists right now that is fundamentally changing the way that physicians feel about their jobs and even how long they stay in their jobs too. You know, in other conversations I've had with leaders, ambient technology has kept people in the field working as clinicians when they may not have otherwise. So it's incredibly powerful. And I appreciate you telling me about your experience.
- [12:01]
A
Well, and I think the hardest part now is the transition as we look towards the next hundred products. Right. Is this transition towards that inpatient world or the rehab world or the home world, which are all different applications. And for physicians, once they do the initial note, the rest of it is heavily copied forward. So the question is, when or where are not these tools going to become applicable to nurses who do things in very different ways? Like where are these tools applicable? Where's chart summarization applicable? So it's going to be. I think we're in. We have just scratched the surface of the changing the way we document. And I. I know I ended up speaking with some folks at EPIC and saying, if you had to rebuild EPIC today, the way that knowing. Knowing these technologies existed, could you get rid of all of the craziness of a workflow and a brain and a worksheet and just allow the nurses to just speak whatever they want to speak?
- [13:04]
B
Yeah.
- [13:05]
A
And you take in all that data and then just AI summarize that in a form that works for the nurses? Would you do things differently? And I'm not sure I got an answer, which may be validation and maybe like, hey, crazy lady, go back and figure out something new. But I think we do it very differently.
- [13:25]
C
I think so, because I think the. What's happened with this class of technologies is the Physics of the world has changed. And so all the premises with which we used to build software and we designed workflows and even designed our operational systems, I think we're going to come to question a lot of these. A lot of these design decisions. And if we were to, from first principles, redesign the software we use, how we distribute work across people, what happens proactively, what happens retroactively, I think we would actually redesign a lot of things in a very different way with these new capabilities.
- [13:59]
A
And it's. I mean, it is fascinating that there's a Harvard Business Review article which was always one of my favorites. And the answer was, how do you kill a bad project? Because you feel like if you just make one more step and one more step and one more step. And sometimes I feel like the better answer would be, if you ever play that game Bananagrams.
- [14:18]
B
Yeah.
- [14:18]
A
Right. And the best answer when it gets so complicated is just to shuffle it all and start again at the beginning. And I don't know if that's the right answer, because, I mean, the fundamental guts of the electronic medical record is everything that we really do need to do a great quality job for our patients. The question then becomes, how does that information get in? And then how does that information get viewed? And I think if we start separating the in from the view, then perhaps that's where you see the next evolution of particularly both the inpatient record as well as the support functions within the physician office. Because really, this window has only been opened up to a small percentage physicians and not the rest of everybody else. And the promise is there for everybody else. But there are two things that have stopped it. One is the development which is coming. I mean, people are. These guys are brilliant. And the second part of it has just been the price tag and the chips and the processing power and the amount of information that these chips are able to absorb. Um, and I think that that's. That that price tag is coming down, thankfully. Yeah. And they had promised by 26, like there's. You're going to see that continue to get less expensive. So we need to make them reasonably cost. The amount of information we are. We are handing over.
- [15:48]
C
Yeah, 100%.
- [15:49]
B
Yeah.
- [15:50]
A
I want to know what I got right and what I got wrong. Go ahead, fix my stuff.
- [15:56]
C
I think probably unpack a couple of pieces, which I agree with. Probably. The first is we are barely scratching the surface of 10% of the opportunity. And a big part of it, as Roberto is saying, is, are the capabilities addressing all parts of the workflow for all the stakeholders inside of the system. And I think today what we've done a good job of is addressing the pre, during and post of how do we support a clinician delivering care directly to a patient in an ambulatory setting. And now we're starting to see that work move to the emergency department where we've got some great work together and only now are we being to see that work also move into the inpatient setting. But then the question is well, what does this capabilities look like for the ma? What does it look like for the nurse? What does this look like for OT pt? What does this look like for my mid rep cycle team? And, and that's one of the places where we're starting to scratch the surface.
- [16:53]
A
Yeah.
- [16:54]
C
I also agree that there is a tendency with new technologies to rush to automate an existing workflow as opposed to do the pragmatic thing which as Roberta is saying, take a step back, ask what are we actually trying to accomplish and given how the physics of the world has changed, what's the most efficient way to accomplish the mission today and then redesign the workflows around that. And so for instance, there's likely going to be a lot, we're already seeing a lot of transformation in rev cycle. What previously used to be the role of mid rev cycle we're now handling at the point of care because you can take a lot of the expertise which was generally a rare center of excellence. Now how do you apply that at the point of care? Get things right proactively instead of chasing clinicians down, which is we don't have enough people to do, nor is that a fulfilling job for a mid rap cycle team.
- [17:45]
B
Right.
- [17:46]
C
To do all these things. The reality is we're going to be processing an insane amount of information. The number of conversations, the amount of past data in the ehr. And so I think we're going to have two tensions. One tension is expecting more of the technology over the next five years. My guess is we're going to be using it for 10 times the number of use cases, processing 100 times the amount of information. And as a result in order to do that we're going to have to see cost of compute go down, which I think the trends are that they're going down aggressively and hopefully the ability to get more out of the technology while also seeing cost of compute go down will work out in the health system's favor. But those are equations that can be very sensitive to assumptions.
- [18:32]
B
Yeah, absolutely.
- [18:33]
A
I think one of the fun parts that I think should be next on your agenda just for fun is as long as people are paranoid and think Alexa's listening to them, you should just listen to them talk about their healthcare and then summarize all that information and then send it when relevant to the doctor. And that would be great.
- [18:53]
C
You know, what's, what is, what is kind of interesting is if you have the ability to spark a conversation at the cost of compute, the way you think about the care model probably changes as well because you could have a continuous conversation with the patient. And that is a whole source of new information for the care team that never existed. That's a whole source of surface area to change behavior that never existed. Obviously there's a lot of work to do before that's even possible. We figure out how that fits into the care model and what's the reimbursement model for that kind of work. But you're absolutely right that that's like very much within the realm of possibility. Over the next 18 months.
- [19:30]
A
This is way more fun than walking the Becker's floor. So sorry, Beckers.
- [19:35]
B
It's a great conversation and I really appreciate how forward looking both of you are while acknowledging that we're still in the midst of a lot of evolution and a lot of development. But Roberta, I want to go back to something that you mentioned. You were kind of hitting on some of the anecdotal outcomes of, you know, physicians saying, you can pry this out of my cold, dead hands. Do not take this away from me. But would love to touch on any other early results that you're seeing at Houston Methodist with AI and Nikhil. Feel free to weigh in as well.
- [20:04]
A
Yeah, we've been able to actually track the minutes that were given back to the practition. What we've seen is a willingness on practitioners parts to adjust their templates. We didn't go the route of like, you have to add one appointment and you have to add one appointment. But what we've seen is the natural evolution of their templates being that we are able to see more patients, which is really wonderful. We're seeing, you know, with a confluence of technologies. I think one of the things you can't say is, oh, everything is because of this or because of that. We are seeing our operating rooms get busier. So again, everything. I had a boss that joked that if you tracked everything back, it would all be back to the primary care doctor because they sent everything to the table. But if you start to look at where you are, each one of these technologies that give back the time, that give back Moments to every individual doctor that gets them does have a downstream impact on what they are able to do, both from a quality and a quantity perspective. So yes, we have our sheet that tracks our lovely little outcomes of the pilot. What we tend not to do is measure any of that on a very long term because the answer is once we've played it out and we know what the outcomes are, then we just spend time and money continuing to reinforce what we already know. And so we've gone into large scale implementation. And I don't think from this work that is happening, there's no putting the genie back in the bottle.
- [21:44]
B
Yeah, Nikhil, how would you respond to that?
- [21:47]
C
I think if there's maybe one metric to highlight that's actually really special that I think both teams are really proud of. It's just the utilization of the technology sold out across 22 different specialties, which obviously is a larger number of subspecialty areas. But the clinicians are using it for 76% of all scheduled visits. And I think it's a really exciting metric because I think our industry is going through a soul searching process to understand what does good actually look like from an implementation perspective. What does good actually look like for ambient listening. And if you think about the numbers that are being published around adoption utilization, folks are barely able to really get 30 to 40% of their clinicians onto ambient listening. And even for those clinicians, the published results out there are like only about 40%. 30% of their visits for those clinicians who are adopting actually flow through the ambient listening system to get to like a 76% metric of all scheduled visits flowing through ambient listening. I think is that's one of the metrics I think our teams are both really proud of.
- [22:56]
A
Yeah, I think one of the other ones that kind of stunned me is the willingness to give up scribes in the emergency room. I was, I kind of like, you know, I sit there with my eyes, you know, with some of our older or slower clinicians and in the emergency room they were both serving as I'm scribe and a little bit of gopher. But it was interesting when looked at the cost aspect of for them in terms of how much cheaper this was than a scribe and a little bit of kind of personal decision that they had to make. It was amazing to me. I think I was floored by the amount of people who were willing to give up their scribes.
- [23:38]
B
Yeah, yeah, it's very telling. It seems.
- [23:41]
C
Unfortunately, I don't know if there's a amount of compute that could Get R scribe to bring coffee into the emergency room. So that's the one thing that I don't know compute's gonna solve, but we can do almost everything else.
- [23:54]
A
You know, I'm gonna give it to them. They were going to get samples or papers or education or supplies or. Or maybe coffee. Okay, well, we'll just leave it at that. But. But it really, it was like again. And part of it is me being fascinated by the human aspects of this of. It's like you threw a pebble in the water and you're just watching that ripple continue to happen. And it really, it's very real. This is, this is one of those areas that is just so real, which is why you can't have a conversation. Whenever you talk about AI, everyone just assumes you're talking about generative AI Voice in the electronic medical record. And I'm like, there are other aspects of AI. They're like, I mean, like what? Like, because. Because this one has had such an enormous impact on the industry. So quickly.
- [24:48]
B
Yeah.
- [24:49]
A
That it's stunning.
- [24:52]
B
Absolutely. Well, it's been such a pleasure learning from you both in this conversation. And I just want to. I'd like to end our conversation on a note of looking ahead. And you both are already great at doing that. You've already given some really interesting projections and where you think this is all headed. But as you continue to build with ambience at Houston Methodist to expand into the inpatient setting, what do you Predict for ambient AI in 2026 and even past that?
- [25:20]
A
I'm not going to say that every bit of it will be with our one partner. I think in this space it is possible that you will end up with multiple partners because I think or technology may come out of machines, or nursing technology may come through people who can more quickly look at the workshee. I can't promise. What I can tell you is that I think we will be primarily voice within three years.
- [25:49]
C
Yeah, I think so. The inpatient workflow is also such a different workflow from the ambulatory setting. And so it's almost like we talk about 100 plus different products. Just what does it take to solve the hospitalist workflow? Which is one of the places where we're starting first is you think about the admission, the rounding, the handoff, the discharge. It's a combination of not only ambient listening, but 80% of the problem is actually complex reasoning over the data inside of the chart that's constantly evolving over every 24 hour cycle.
- [26:20]
A
Every two hour cycle.
- [26:21]
C
Yeah, that's true. Every two hour cycle. And so that's sort of the crux of the problem. But if you can nail it, I think what we're excited about is at baseline it means clinicians are spending less time digging through charts to try to find information, spending less time on documentation. One level above that we're excited about is the amount of work and hand to hand combat that goes into making sure that you pick the right ICD10 codes, you get the right CCs and MCCs that you actually figure out and mark what's the primary diagnosis and what's present on admission. All those things that affect revenue cycle and quality that previously took so much work, you can actually get prospectively right at the point of care. And so I think those are the pieces that we're really excited about in the hospitalist workflow for nurses. I think we've got a million doctors in this country, we have 5 million nurses and you know, there's a lot of excitement around automating flow sheets and I think that's going to be something that we're going to take a look at jointly. But the, we think the bigger opportunity is to rethink what the nursing workflow should look like foundationally. And part of it is, you know, nurses spend an insane amount of time trying to catch up on the state of a patient. An insane amount of time in transitions between floors, transitions from shifts. There's so much of the expertise of a 10 year nurse that you would hope our two year nurses also have access to, to have a good intuition for is the patient going to get better or are they going to decompensate and how should I spend my time on the floor? There's so many opportunity areas and I think Roberta's right, not all these problems are going to be solved by one player, but the opportunity to really make a dent on the nursing workflow. We're also really excited about a lot.
- [28:07]
B
Of exciting developments ahead. And Nikhil, Roberta, I just want to thank you both for spending time and sharing those projections, your media insights with Becker's. It's been a real pleasure talking to you both today. So thank you again.
- [28:20]
A
Thank you. Yeah.
- [28:21]
B
And we'd also like to thank our podcast sponsor for this episode, ambiance healthcare listeners. Be sure to tune into more podcasts from Becker's Healthcare by visiting our podcast page@beckershospitalreview.com.