Counting Sand

HIMSS24: Value-Based Care, Health Equity, and AI

Episode Summary

In this episode of our HIMSS 24 coverage, host Angelo Kastrouis is joined by Nick Stepro, Chief Product and Technology Officer at Arcadia, to explore the future of healthcare. They dive into the crucial topics of value-based care, the impact of social determinants of health, and the transformative role of AI in healthcare. The conversation highlights the need for a shift from transactional patient histories to a more comprehensive, holistic approach. By integrating diverse data sets, including social and economic factors, healthcare providers can develop a longitudinal view of a patient's health, enhancing care for individuals and communities alike. Join us as we unpack these complex topics and envision a more integrated future for healthcare. Links: Nicholas Stepro Angelo Kastroulis

Episode Notes


Sound Bites


00:00 Introduction and Overview

06:31 The Role of Data and Analytics in Healthcare

13:59 The Importance of Trust and Transparency in AI in Healthcare


Episode Transcription

HIMSS24 Episode with Nick Stepro of Arcadia

Angelo: I'm Angelo Kastroulis, host of the Counting Sand podcast, and we're continuing our coverage of HIMSS '24. With me today is my friend and colleague, Nick Stepro, Chief Product and Technology Officer of Arcadia. And we're going to talk about value-based care, social determinants of health and other ambient data and artificial intelligence, especially in the concept of healthcare. Nick, thanks for joining us today and appreciate you taking the time to chat.

Nick: Yeah. Thanks for having me. Good to see you again.

Angelo: So let's talk about value-based care. So, just for the listeners who might not know what that is, that is the evolution away from the fee for service model where providers are getting paid for things they do, like you read an x-ray, you get paid, you bill that. If you do other tasks, do a write-up and other things you get billed for those tasks. And that doesn't take into account regardless of whether people get better or not, you may or you may not. The idea is hopefully you will, but in a value-based model that all changes, it shifts so that if we do better, you get paid more, you make [00:01:00] more money. There's more profit. If we don't do so well, readmissions, going back to the hospital over and over again for the same condition, it will cost. So that kind of sets the stage for why I think data and analytics are really important here, because we've got to know how we're performing to be able to improve it. That's your space, data, and analytics. So what do you think about that? How do you think we can enable that kind of thing with technology?

Nick: Yeah. So the shift from fee for service, for fee for value has been underway in the US for honestly depending on how you count, like 40 years or more, or, you know, really in its modern incarnation, the past decade or so. It was catalyzed by the, very reality that U.S. healthcare expenditures go up, up, up, up, up and out clip basically every developed country on the planet. And our outcomes were relatively stagnant and underperforming relative to the spend. And so, one of the theses behind that is that the incentives were a little bit out of whack. Physicians and hospitals are effectively incentivized to render more and more and more services, which costs more and more and more money, [00:02:00] regardless of whether or not they're improving the lives or the outcomes of patients, right? And so these fee for value models have effectively said, hey, instead of us paying you every time you, you know, give Angelo an MRI, we're gonna give you a fixed amount of money to manage Angelo, and if you can keep Angelo healthy and you can do it for under the targeted top line, then you get to recoup the savings. That does two things that are actually really interesting. One of them is it fixes this incentive problem, right? And now both providers and payers are heavily incentivized on doing cost effective services that improve patient outcomes. The other thing it does is enable providers or health systems to provide care or services that otherwise might not be advantageous or reimbursable. So a lot of the drivers, and I think we'll talk about this in a second, a lot of the drivers to negative health outcomes are things like lack of access to transportation, lack of access to housing, or food instability. It's really, really hard to manage your health when you have that much instability in sort of a foundational [00:03:00] functional area, and those were previously like off limits for a health system to take care of because they just have to wait for someone to show up and get sick and then render a service. And now with these incentives, you can deploy those savings and more creative ways to improve the health outcomes of your population, even if that is not a billable asset in a traditional fee for service.

Angelo: That's a great way to say that, but that also illustrates why data is so important because if you weren't aligned, you wouldn't be looking at the data to try to find better ways to do this right outside of the box.

Nick: Well, right. If I'm just waiting for Angelo to show up and tell me he needs an MRI, then I don't have to think a heck of a lot about Angelo outside of the four walls of my health system, right? If now all of a sudden I care about, Angelo as a person, right? In three full dimensions, a 360 degree view of Angelo. Then I have to wonder why isn't Angelo showing up? Is he filling his medications? Is he not filling his medications? Does he potentially have any of these functional barriers to care that might impact his health outcomes? And that inverts the sort of [00:04:00] ITIS infrastructure needs of a health system away from just being basically like a POS transactional mechanism to a holistic population analytics play. And that's where Arcadia comes in. We've been around, you know, for the past 10, 15 years. And the thesis that we had is that you have all of this digitization of healthcare based on the onset of these EHRs, but all of that was just done in a transactional manner. And these fee for value type models require that you harvest that transactional data, blend it and commingle it with other data, like social factors and provide analytics and a holistic person view so you can start caring about the population as a whole and people across the entire longitudinal lifespan. Rather than just when they present in your hospital.

Angelo: Yeah, exactly. And so let's talk about these other factors because they're not captured in an electronic health record, right? They're not necessarily in there unless the physician has the wherewithal to have heard you say something like, yeah, I lost my job, lost my [00:05:00] coverage, you know, I won't be able to continue taking this blood thinner. Okay, red flag, you can't get off the blood thinner once you've started. So how do I make sure now that I know you can't afford it, I have to figure out how to get your samples or something or into a program. Unless they caught that when you said it, or maybe you didn't even say it, I need to know these things. So typically one way we do it is through questionnaires to try to get it into the record, but I like the way you said this to me once that there's, "ambient signal" out there that's happening around the visit. Can you talk a little bit more about that? Like, how do you capture this ambient signal that can signal some of these determinants?

Nick: Yeah. So probably the most exciting thing about the past few years in artificial intelligence, at least for me is, converting what used to be perceived as non-analytics grade data into analytics grade assets, right. And so like the simple representation of that is a paper chart, a paper medical chart, right. That was not considered an analytics grade asset. Because it's unstructured data t's really, really hard to mine. Even if you can apply OCR, which has been around for ages, can you [00:06:00] really derive analyzable tokens from that in a reliable fashion. And it turns out these large language models are able to eat that for breakfast, right. And compute and derive sort of like analytical tokens that can inform these various models. And so when I talk about signal, there's a lot of signal potentially in the visit itself, right. Where there's, a lot of folks are out there with ambient listening devices, that's sort of the most near-term wave that we're seeing in the application of AI. And to point that you mentioned, in the course of a conversation folks might allude to the fact that they have housing instability or you know, maybe they don't have transportation to their follow-up visit or something like that but unless that's caught and coded in the EHR it just gets lost. And I think a real opportunity is to start harvesting that information and use those signals to improve some of the social services around patients. But outside of the visit, there's a heck of a lot of data too. Like everyone uses social networks there's all sorts of consumer data available. There's proxy data. I mean, the sad reality is, is you can tell a heck of a lot about someone based on the block that they live on. [00:07:00] Not everything, but you can get some signal from that and start titrating your population and understanding who should I reach out to. To just proactively see if they need any of these services. So these are all really powerful pieces of data, but in order to crunch them in a real-time fashion that's meaningful for a hospital system that doesn't have a heck of a lot of big data services or capabilities under their roof, you need sort of a proven big data platform.

Angelo: Yeah. And this data is really important. It isn't nefarious always to say, you know, oh, where do you live? Living, say, within walking distance of your provider is super important to know. Or if you live all the way on the other side of town you might have some issue with transportation. If you can walk there, you know, that's one set of problems I can help you with. If you can't get there, that's a whole other set of problems. But you won't know that necessarily in the visit, but I can start to derive it when I bring in all this other data. It's important for people to know. It's not just these nefarious causes of data. I mean, you don't know your credit score, therefore we're labeling you a kind of person. [00:08:00] It isn't about that. It is about, can this give us hints to other things that we can use to help drive healthcare in a better direction.

Nick: Yeah, I think that's right. But also, I fully acknowledge that there's a fair amount of healthy skepticism from the US health consumer as to how their information is being managed and frankly, just trust in the health system generally. We're just coming out of this pandemic and the amount of distrust in proven medical miracles that were invented a hundred years ago, like vaccines, like it's all out the window. People don't necessarily trust their PCPs to be doing the right things. And so now you layer on top of this, the fact that we can harvest basically every footstep that a patient makes and you know, amass that into this massive AI brain. It's a little bit of hyperbole, but that creates a real anxiety and stigma of the healthcare consumer. And I think part of the battle that we have to fight is to build back that trust while deploying some of these bleeding edge tech assets that can actually, really materially improve the life of patients. And the last thing I want to say on that front is that [00:09:00] you can think of like the double edged sword of this AI, right. Some of it can be seen as a real powerful weapon or adversarial because you know, it's a very big brother feeling to a lot of folks. But there's a lot of aspects of it that can actually, really quickly improve health equity and improve access to information. Because if you think about a lot of healthcare information is consumed as really, really dense arcana that is basically impenetrable to anyone that's not in this field. You know, my wife just got a new job. And we were evaluating a benefit summary of two different payers and it's like, how the heck are you supposed to interpret that? That's wild. And so, the first thing I did was I shoved that into an LLM and I said, model out a few things. Say I'm of this age, you know, we're this family structure, we have this health history, what's a few next year could go for me. And, didn't break a sweat, right? So now I'm empowered to make healthcare decisions on my own in a way that even someone that's well over a decade in this space has a hard time understanding these benefit packages wouldn't otherwise be able to do.

Angelo: That's a really great point because yes, we [00:10:00] should be afraid of some of this technology because there's a lot of data being captured, pooled, who gets it? What happens to it after you've trained, you know, there's a lot of those questions that are really, really important, but I think that in the right hands, in your hands, for example, your model helped you. It wasn't that somebody else tried to use it necessarily to manipulate you. You had the control of the model. So I think in the right hands, the technology is really good. LLMs certainly have taken center stage in what they can do just because they're getting really sophisticated. And they're moving. The one thing I want to say about AI, especially when you look at HIMSS, you were inundated. So I had a little bit of PTSD with all the AI there because anybody and their brothers doing AI, you know, OCR. You have a scanning solution now that turns it into text. We're doing AI and okay. Okay. Okay. But there's many ways I think to approach this problem from many directions. And one of the ways that research and the stuff that I've always done [00:11:00] is to go the cutting edge route and try to discover new things. But that takes half a decade of research to get into mainstream. And actually that usually takes the form of a startup and then the startup will introduce some idea and then others will follow, but that isn't, I don't think the best way to use this technology. There's better ways to look at today problems. Like you said, LLMs could easily do this and you have a today problem. And I know that's a space that you live in. You told me about ways of using it to make people more effective, especially in the fact that providers are just overworked. And, you know, we're gonna make mistakes. Can you talk about that a little bit?

Nick: I can, you forgot about the part where that startup then just gets gobbled up by one of the three big tech firms.

Angelo: Definitely. Yeah.

Nick: The other side, just on the HIMSS floor, because I have like the same, you know, whiplash, is one of the small burdens of this AI age is effectively just trying to figure out what the heck any of these companies do. It's like every company I click on, the first, you know, five view ports as you scroll through their webpage is just [00:12:00] AI, AI, AI it's like, well, for what? What's the actual company under this? And then you get to the point, you're just like, okay, you're a rev cycle company that has like a little process automation bot. So now you're in AI company, right.

Angelo: Or my favorite one. We are AI based facts.

Nick: Yeah. It's amazing. Why not, right? Sprinkle some magic on there and get that valuation boost.

Angelo: Yeah.

Nick: That's too funny.

Angelo: Yeah.

Nick: Yeah, I think like the advent of AI is also coming at a down capital market because of interest rates or whatever. So I think like you also have a bunch of people just desperately chasing valuations and trying to like make the claim to be an AI company to justify, you know, the round of raises that they did three years ago at an ungodly valuation. We digress a little bit. But anyways, the question was about how pragmatic to be with AI, right? And I think there's a lot of really, really exciting things that you can do here. I think some of the most innovative and exciting applications of this probably won't be realized for five years because the sophistication has to really, really, really increase to do that. And that's like real-time [00:13:00] live autonomous clinical decision making of you know, a hospitalization or something, right? That's a crazy high-risk scenario to operate in that requires real-time streaming data of all types of modalities, as well as like health history. And a significant amount of literature. You need to have video. You need to have audio all streaming in real-time space. Really, really compelling. It will absolutely happen, right. But Arcadia is just transparently like, we're not spending a bunch of time chasing that. What we're focused on is that the U.S. healthcare system is facing an access cliff, right. I don't know about you, but, you know, I've got really darn good coverage, and when I try to make a primary care appointment, I have to wait 30, 60 days to get a schedule for primary care. And we have an aging population. Medicare is booming. We're not getting more physicians. So the only way you solve the problem of the supply demand shortage is by improving the productivity of the provider. I'm going to say provider because I'm referring to the broader like provider care team. Not necessarily the physician. We have to do it. We don't really have a choice. Panel sizes [00:14:00] need to expand. We can't sustain the growth and the health needs of this population without doing that. And so what I am very encouraged about is that the convergence of that access, that like really imminent access problem that we have with generative AI coming together to improve the productivity of care teams. And that doesn't have to be the most glamorous application. If you look at a lot of care coordinators or clinicians, they spend, you know, 10, 20 percent of their day just gathering context. Just gathering context. When they see a new patient, particularly if it's a Medicare patient or a complex Medicaid patient, you've got tens of pages of health history and to understand what the heck's going on with that patient, you have to trawl through that. And so what do these models do really well? They do summarization. Just applying a simple summarization model against a really, really complex health history to provide a consumable contextual summary of that to a provider at the point of care can do a heck of a lot for that. The other obvious one is transcription, billing, and coding. I mean, for better or worse, even in the fee for value models, people still [00:15:00] have to push codes out the door, and in order to do that, they need to have thorough documentation of the medical record largely for the purposes of liability and that's really taxing. You know, the amount of time providers spend documenting is massive. And why the heck can't we automate 95 percent of that? These are the ways that we're applying this today and a lot of will actually cause today's benefit to today's sicknesses in healthcare and then we can focus on some of the fancier stuff.

Angelo: Exactly! I think the other thing that's kind of interesting here is that you and I have lived in this space a really long time. And we both are trying to tackle exactly the same problem. How do you make a provider effective? And what's cool is where you and I are tackling it from different ways, both using AI, but, the ways I'm doing it from the, how do I help you operate at the top of your license by getting you knowledge that you wouldn't normally have? In other words, it's not fair that if you happen to live next to the best cancer center in the country, you have the best knowledge because it's pooled there. And yeah, they have also the best techniques and equipment, [00:16:00] but the knowledge, at least we can fix. Knowledge we can disperse. I like that you're approaching it also from well, another way to make people effective is to make them efficient. And I can get you information and not make you do the things you shouldn't have to do, so that you can focus on the things you're good at. That's another really valid approach. So I think you need all of it. You need everyone working on this problem from different angles. But the one thing I like about what you're doing is, it's also the best commercial approach. It's viable.

Nick: Right, right. Yes, it's viable today. And it drives real ROI today. But I couldn't agree with you more. I mean, it's kind of an all hands on deck moment for healthcare. And what's really encouraging to me actually about AI is, unfortunately, some of those powerful models are closed models, right. I know OpenAI and the name is open, but it's a closed model, and it's opaque for reasons. Largely being if you're, justifying a trillion dollar valuation, you kind of need to have some proprietary technology. But largely the successes over the past 12 to 18 months have been [00:17:00] open source. And I think that, type of open science driven community of tech advancement is something that can lift all boats and enable all of us to row in the same direction, though, coming at it from slightly different angles, like you said.

Angelo: Exactly. Well, Nick, thank you for joining us today. Thanks for your time. And I really appreciate this conversation. We both, I think, love the same thing. And be sure to follow us both. We'll have our handles for our social media accounts in the notes. And Again, thanks for your time.

Nick: Yeah, thanks. Always a pleasure.

Angelo: And as always, thank you to our listeners. We couldn't do this without you. Please subscribe to the podcast. It's available on your favorite platform. You can follow us on LinkedIn at AngeloK1 or on X and that's AngeloKastr. I'm your host Angelo Kastroulis and this is Counting Sand.