← Becker’s Healthcare Podcast
Loading summary
Transcript41 lines
- [00:00]
A
Hi everyone, this is Brian Zimmerman with Becker's Healthcare. Thank you so much for tuning into the Becker's Healthcare podcast. Today we're going to explore how AI can humanize healthcare. And joining me for this conversation is Dr. Josh Tamayo Sarver, Vice President of Innovation at Vituity. Josh, thank you so much for being here.
- [00:19]
B
Thank you so much for having me. I'm looking forward to the conversation.
- [00:22]
A
Yeah. And to get us going here, can you just share a bit about yourself, your work in healthcare, any pertinent details that would be helpful for listeners to hear?
- [00:31]
B
Yeah, absolutely. I think what makes me special, or at least makes me feel special is that I have multi dimensional mediocrity. And what I mean by that is.
- [00:45]
A
Never heard that before. Go ahead. I'm curious to hear where this goes.
- [00:51]
B
I have four areas where I have some reasonable competence, but there's someone who's better in that area probably than I am. But it's not very often that you find someone who can go across all four of those areas. And those areas are Soot, Scrub, Geek and quant. Soot. Because I've been in the business and management of healthcare now for almost 20 years, scrub, because I'm a clinically practicing emergency physician and I still do three night shifts every month to make sure that I eat my own cooking. And then geek, because I used to be a software engineer and know what it's like to be in a stand up and do programming and do all the features that the customer wanted and feel it's feature complete and have them come back and say this is nothing like what I wanted. I know what that feels like and how that goes. And then Quant, because my PhD background's in statistics and a lot of the models that later became the foundational models.
- [01:42]
A
For AI Suit, Scrub, Geek and Quant. I think you're, you're in an interesting group. I would be curious to, to see if there's maybe like a Facebook group out there with, with more of your folks in there. Is anything like that, any social connection you have with, with other folks who have this background?
- [02:01]
B
I think that at that point once you, you actually dabble in all those areas, it's more of a group therapy sort of thing.
- [02:06]
A
Okay, got it. Yes. Well, you're, you're the perfect person to come on and have this conversation today. So let's get into it. I think there's, I've seen recent survey findings that identified agentic AI as one of the, the top strategic tech trends for 2025 and beyond. And, and Many health systems are, of course, we know, investing heavily in AI to improve both patient satisfaction and operational efficiency. But I guess, curious, from your perspective, Josh, what makes agentic AI such a pivotal next step for healthcare? We hear a lot about AI in general, but why is this particular kind of AI so important?
- [02:45]
B
I really appreciate that question, and let me see if I can answer it in a way that resonates to answer it. Well, I kind of have to explain a little bit of what generative AI can do really well and can't do very well and why that's so important for understanding why agentic AI is important for us moving forward. So if you think about what AI, especially generative AI, can do, it's really good at summarizing and knowledge retrieval and being kind of that interaction between you and the computer. And when a human can do something like that, we then attribute the human to understanding concepts and being really smart. But the agentic, or rather the generative AI, doesn't actually have any conceptual understanding. So what does that mean? That means if you ask a human what the rules of chess are, they could tell you all the rules. And if you could say, I have this particular scenario, what should I think about? And they could tell you all the things that you should be thinking about for a particular move, you would then assume that that human could probably be a pretty good chess player and make a really good move. And I think he'd probably be right. We then take that same way of assessing intelligence and skills and apply it to generative AI. So when we ask generative AI to tell us what the rules of chess are, it can tell us the rules really well. And then if we say, if we're looking at a particular situation, what would the different good moves be and how should we think about those different moves? It could also summarize that for us. And then if we ask it to play chess, it would suddenly start playing backgammon. And it's because once you ask it to play, it's having to apply a concept, and it doesn't apply a concept very well. Does that make sense so far?
- [04:31]
A
Yeah. I'm tracking. I think even somebody with my rudimentary understanding could pick this up, this idea that AI can execute on things, but once you ask it to play, then there's a different concept it's got to engage with, and it's not. Not great at that.
- [04:47]
B
Yeah. So generative AI does great at this kind of like general knowledge retrieval, but when it comes to specific action taking, it really falls short. And so how does that translate? And what does agentic AI do to fix that? So we were developing a tool. This was actually a few years ago, I'm kind of proud to say, because we were before the bleeding edge. I don't know what that is. That's the suffering edge, maybe. So we had this tool that it was in the ambient space and we were trying to get it to do billing encoding. And it was doing a fabulous job of summarizing what it was hearing and what it could find all over the place to provide the diagnosis textually. So this is a non ST elevation mi. This is a hypertrophic cardiomyopathy. It was doing a great job at summarizing that textually and. And then we would ask it to then execute and do the code of an ICD10. And it was just. It would get it right sometimes, it'd get it wrong sometimes, but it certainly wasn't usable at scale in production. It made a really cool demo, but beyond that it wasn't working. And so we had to switch into what would now be an Agentix solution. And so what the Agentix solution is is you have another AI agent on top of that that identifies the tasks that need to be done and then calls the specific actions that need to happen for those tasks and then takes an action on it. So what that looks like in that billing encoding example is that agentic part of the AI says, okay, what's the textual diagnosis? Figure that out first. And we actually use generative AI to figure out the textual diagnosis. Now call another set of code that is not AI at all, it's just a lookup table. Call that lookup table piece of code to figure out where does that actually map to one to one, which is cheaper, faster, and completely reliable in ways that generative AI aren't. Okay, now we have that diagnosis. Now figure out how that maps to the actual coding that we need to do in this context and now execute by providing that code back. And so now you have this agentic AI is part multi agent, where it can orchestrate what's going on and part identifying what are the best ways to accomplish the different tasks, many of which are not using generative AI and then using that at the end can provide an action. Is that still making sense?
- [07:06]
A
Yes, yes. I'm curious, when you talk about generative AI, I know those are built on, of course, large language models, correct? I guess. How would you describe what the agentic AI is built on? How does the agentic AI, how is it able to understand concepts and call.
- [07:25]
B
Plays yeah, so that's a really good question. It actually doesn't technically understand concepts either, but it can be given the menu of how to go about solving the problem and then follow those steps.
- [07:42]
A
Got it, Got it. So if you give it the proper steps, it will know what to do within the menu you lay out.
- [07:51]
B
Right. And it can even come up with its own menu of steps, which is kind of a weird thing to think of. Right. So it can just like when we were talking about the generative AI for, for chess, it can say, what are the different ways, different moves to look at in this scenario? Right. It can call those different. It can go out to find out what those different moves are, and then it can pass that on to another agent to say, evaluate this particular move and give me a score, and then ask another agent, evaluate this particular move and give me a score, and then.
- [08:24]
A
Bringing that to healthcare. So I think, conceptually speaking, I'm able to play out some scenarios where this could be valuable for healthcare, the agentic AI, but throwing it back to you to paint some of those and tell us what comes to mind for you when you think about integrating this technology into healthcare.
- [08:42]
B
Yeah. So for the one that we just used for the coding and billing, the other area where it happens a lot is in healthcare, you have to put all these pieces together and then come out with an action. Right. It doesn't really. Nothing happens until you have an action step coming out of it. And so, for example, if you wanted to know if a person has sepsis, well, you know, you can use LLMs for pattern recognition and disease, but the variability is still pretty high. And it's not something that you'd want to hang your hat on or your loved one's life on. But you can use that generative AI to call those models that are deterministic. You know, like we have lab tests like Troponin for heart attack. We have an AI tool that we worked on with a company that predicts sepsis 72 hours in advance. It's not generative AI. It's using just machine learning models that have been built up. But you could have a generative AI component on top of that that figures out what diseases does this person have, not by using pattern recognition in an LLM sense, but by calling and orchestrating all those models that are well established to do diagnosis, to go across those diagnoses and then call, formulate that back as an answer, saying, this is what's going on.
- [09:56]
A
Yeah, no, I appreciate that. And I think, you know, we hear a Lot about. Much of the conversation around AI in healthcare has really centered on efficiency and reducing administrative burdens. What you're talking about though, I think could potentially get somewhere else, which is sort of the, the framing for our conversation. Right. The human element of this and something that comes to mind when we think about the human element is that element of trust between patients and clinicians, which is so fundamental, fundamental and foundational and perhaps in these, these times of division and a lot of misinformation sort of put into, you know, put into tension. So curious to hear how you're thinking about this kind of technology, how it could potentially strengthen that trust between patients and clinicians and any, any examples you, you can share with listeners that would kind of help, help bring this to life for them.
- [10:50]
B
Yeah, I, I think there's so many different levels to it in terms of what helps someone trust someone or something. Let me start with, with the whole trust component though, which is healthcare and kind of humans we run on trust, right. Even when we're talking about these systems, you know, there's this, do we regulate more, do we regulate less? And how do you create, you know, go fast with AI without putting things at risk? And really at the end of the day we're, all of the conversation really is trust. And we live in this really bizarre time where healthcare, as doctors, we can do more than we could ever have done before in terms of actual objective effect on your health and life and well being. And yet we're trusted less than we've ever been before. And, and so it's not a competence issue that gets us trust. There's something else going on. And you know, when I took my kids to the doctor and I've been pretty healthy, I'm happy to say, but whenever I took my kids to the doctor or my spouse, I knew the physician, we were friends, but they spent most of their time looking at the computer and that doesn't feel very trustworthy.
- [12:06]
A
Right.
- [12:07]
B
So I think to the extent that we can stop having healthcare take care of healthcare and have healthcare take care of patients, I think that's gonna go a long way to restoring trust. And I think there's a great opportunity for having AI, whether it's agentic AI or other flavors of AI, start doing more and more of the taking care of healthcare so healthcare can take care of patients. If you think about my shift when I'm in the er, I'm making sure we're meeting our metrics and our three hour bundles and, and like, I would hate for you to have to look at the number of screens I have to go through on my ehr to do anything. And 99% of those screens are to meet some quality metric that someone somewhere thought was a good idea at some point. And now that's what I'm caring for, is quality metrics instead of patients. Yeah.
- [12:55]
A
And do you, I guess what is, what are your thoughts about your help with the question is if agentic AI is able to come in and you know, really limit that the amount of screen time and documentation you have to do that administrative burden, how are you thinking about and how do you think about more broadly as an industry that we make sure that that time isn't filled up with new tasks? Am I making, does that make sense? How do we make sure that that time is actually devoted to, hey, let's make sure we're getting face time with patients, interacting with patients and not just what other efficiencies can we grab here now that we have this extra time.
- [13:35]
B
Yeah. And I think there, maybe what you're highlighting is, so let's say that, that before I could see two high acuity patients an hour and now because I don't have any administrative tasks, if I spent the little amount of time I did before with those patients, I could now see five high acuity patients per hour.
- [13:54]
A
Exactly.
- [13:55]
B
Yeah. But that probably doesn't actually increase that trust quotient and the quality of care that I'm providing to those, those patients, even though I'm now providing care for more patients, so it's expanded the access to care and the capacity. So we've addressed that through greater efficiency, but it's not really improving trust or the quality for any one of those patients.
- [14:19]
A
Yeah, truly. Yep. That's sort of where my head was at. So how do you think about, is it just a broad awareness that the industry needs or just more people speaking about, you know, that trust is the crucial element here to pursue?
- [14:34]
B
I think it's probably going to be a really heterogeneous and varied response. Right. So some, some folks are going to say, wow, I have more time to talk to this patient. And it makes my job a lot more fulfilling to really connect and establish trust and alignment and have them leave having accomplished why they came here in the first place. And other folks are going to say, this is great, I can see two more patients and get a nicer car.
- [15:01]
A
Totally. Yep, yep. I imagine too it'll be different for organizations and different physicians will gravitate towards different organizations and how they're deploying this technology. Am I Reading that correctly as well?
- [15:14]
B
Well, I think that's a hypothesis and I would certainly agree with your hypothesis. Right. That I think it's going to end up being, being, you know, to some extent trust is you're connecting with another human as a human. And I think what you're saying, and I agree with is people are going to be people and this opens up capacity and what they do with it is kind of going to depend on who they are as a person.
- [15:36]
A
Right, exactly. Well, then I guess the question I'll ask you as we come to maybe last couple of questions here, but what steps or I guess advice do you have for folks out there who are looking to adopt agentic AI in terms of how should they go about that and what do you recommend folks do out there, healthcare leaders do out there to prepare for the future that's coming alongside this technology and others? It's advancing so fast. Who knows what we could be talking about in a year or two years time. So I guess what advice do you have for folks out there trying to navigate this environment?
- [16:15]
B
Yeah, I think it's a really exciting time to be where we are. And I think there's such rich substrate to work with which, you know, I've been doing technology and kind of on the cutting edge now for 20 years, which feels like, I don't know if you can be on a cutting edge for 20 years, but I guess that's where we're at. I think, you know, when, when we first started to have the Internet and we were trying to figure out what the Internet was for and there was a bubble and there was a bust and you know, eventually we figured out that it was more than just cat videos, but it took a while and we're kind of in the cat video phase largely for AI right now. I think there was an MIT paper a little while ago that showed there was very little ROI that was happening, even though a lot of organizations were trying to put out AI and scale it. I think to me, what really struck me a lot more about that MIT paper was the amount of shadow use, even though very few people were using an AI solution that the organization was putting out. Most people were actually using AI for something in their work, even though it was not sanctioned and it was outside the mainstream. And as a product developing person, I find that incredibly exciting because what it tells me is there are people who are frustrated with getting their job done and they have identified on their own things that they're willing to do to try to solve that frustration. And a lot of it is leveraging AI, maybe in good ways, maybe in bad ways, but regardless, there's a huge signal that there is a frustration. So the first thing that I would say is if you're leading a healthcare system, figure out what's frustrating your users right now, your frontline employees. Right. Like you know what their problems are. But if you're. I've identified what their problem is and that doesn't align with something that frustrates them, then your solution to a problem they're not frustrated by is another frustration and they're not going to use it. So you don't get adoption and you can try and force it, but it's really hard. And at least in my experience, I've found that if I can align with what frustrates them and then find where there's a problem that I identify organizationally, but I can align it with a solution that solves the user's frustration, then I can get really good adoption. And to me, the first thing is that MIT paper shows me that there is a lot of frustration out there that we can attach to and we can solve problems. But I would make sure that it's led by what that user frustration is.
- [18:40]
A
Yeah, you got to start, I mean, hear it often. You got to start with the problem. But I do think we hear it so, so often that, that I. Some folks are probably just really attracted to the technology. How do you recommend, I guess maybe somebody who has a curiosity or doesn't really understand what their people are facing. How do you recommend them getting started to understand what those frustrations really are if they don't have a good understanding of it already?
- [19:07]
B
Yeah, and I would say even if you think you have a good understanding, I would go back because of, you know, it's a little bit of a moving target and it changes and it's just such a wealth of opportunity. So in my experience, if I ask people, you know, what their problems are, they're usually kind of reluctant. And then they, there's like some social desirability of the answer, especially, you know, if they report up to me or, or somehow see me as some authority figure. If I ask them what their pain points are, it's a little bit similar to that they. They. We all interpret when someone says, what's your pain points? It's like, okay, what's your problem? When I ask someone what frustrates them or what, what has frustrated them today or what's frustrated them recently, people are usually pretty transparent and genuine and open about what their frustration is. And so if, and to the, to that extent, we actually, at all of our vituity meetings, we have a frustration station where we invite people to come and just go off on their frustrations. And that has been a goldmine for us to improve some operational things and develop new technologies. So I would encourage any, any leader or anyone who's trying to say, how do I get started figuring out what, what I can solve, ask about people's frustrations. You probably are sitting in a place where you know what the problems are, right? You know, our problems are this inefficiency operationally, it's this collection, it's this market opportunity, it's this competitor. You know what the problems are, but you don't know what the user frustrations are that you can align that problem with to get them to use a solution at least.
- [20:39]
A
That's great. Yeah, that's a great piece of advice, sort of aligning the problems with those frustrations and moving forward from there. Josh, it's been a pleasure having this conversation with you. Thank you so much for coming on the podcast. Is there anything we didn't touch on or something you want to re emphasize? Whatever final thoughts you might have for listeners, I invite you to share those now.
- [20:59]
B
Wow. That might have been a mistake, but happy to do so.
- [21:04]
A
I'll roll the dice. Go ahead.
- [21:07]
B
So I think as we're thinking about this amazing moment we have in healthcare where we really are starting to have technologies that take care of healthcare so that as healthcare, we can take care of patients. I think we need to realign back to what it means to be taking care of patients. And I think fundamentally it gets down to what is a patient trying to accomplish when they seek health care. And I think we've lost that into patient experience of, you know, where you did your doctor dress nice, right? And we've lost that in terms of quality metrics, all well intentioned, trying to get to better quality. But then I think we've lost that focus on why did the patient come here today? What was their emotional need? What was they trying to accomplish? Right. Did they need a note for work? Maybe that's really all they were there for because they didn't want to get fired. And I think we've lost a very sharp focus on why the patient sought healthcare to begin with. Right. What were they really trying to accomplish, like at a more deep, fundamental human level. And so I would like for us to take this cool moment in healthcare to refocus on that golden opportunity to say, how do we engage patients not as okay, you came here with chest pain. You get a chest pain workup. You have diabetes. Here's diabetes management. But but go more human onto what is it they're trying to accomplish? They don't care about their blood sugar level. They care that they don't wake up ten times at night to have to pee, which happens to be because their blood sugar is high, but they don't feel a blood sugar level. They feel having to go pee a lot. How do we get back to focusing on what the patients are really trying to accomplish and then align what we're doing around the patients?
- [22:53]
A
That's a great place to leave it. Josh and I have no regrets at all about asking the question. Truly a pleasure. Thanks for coming on.
- [23:01]
B
It was wonderful to speak with you. Thanks so much for having me and letting me ramble.
- [23:06]
A
Of course. I also want to thank our podcast sponsor, Vituity. You can tune to more podcasts from Becker's Healthcare by visiting our podcast page@beckershospitalreview.com.