How One ACO Used Predictive Analytics to Fuel Its VBC Efforts

Shania Kennedy:

Hi, everyone, and welcome to Healthcare Strategies. I'm Shania Kennedy. I'm the assistant editor of HealthIT Analytics. And on today's episode, we are going to explore how predictive analytics can provide accountable care organizations with opportunities to advance their value-based care efforts. I'm actually joined today by Dr. Sheila Magoon, Executive Director of Buena Vida y Salud ACO and the South Texas Physician Alliance, and David Clain, chief product officer at the Health Data Analytics Institute, who will be chatting with us today about their partnership to boost value-based care delivery. Dr. Magoon, David, of course, thank you for coming onto the podcast today. I'm really excited to talk to both of you.

Dr. Sheila Magoon:

Thank you so much for this opportunity.

David Clain:

Thank you.

Shania Kennedy:

So I kind of wanted to get us started, Dr. Magoon, talking about how the last time I spoke to you, we were sort of chatting about how ACOs use their data analytics to reduce their financial risk, improve their outcomes, that kind of thing. And during that conversa,tion you highlighted that your approach at your ACO to value-based care centers on keeping patients healthy at home. So when you guys partnered with the Health Data Analytics Institute earlier this year, I was sort of interested to dig in a little bit more into that. So, can you walk us through what the work you're doing looks like from the inside and the use cases that you're targeting in your work?

Dr. Sheila Magoon:

Sure, absolutely. So thank you again for this opportunity. And so, what we have done is, health data analytics gives us a predictive platform to be able to see our patient population from a different perspective. So when we initially looked at it, there [were] just so many different options of being able to look at anything from our at risk for unplanned admissions, or high risk for having falls, high risk for developing pneumonia, or having their heart failure go bad. So we started pulling a whole slew of different types of reports, as well as they can build cohort. They built cohorts for us for who's in your advanced care group or your chronic care group or your rising risk. So we started taking a whole list, because we are now all so excited about we got all this new ways of looking at data and taking it out to our offices with, "Here's your list, let's look at this, let's look at that."

And quickly realized particularly some of the offices were like, they're just as excited as we are, and they take it, and then I had one doctor say, "This information is really great. You've done a lot of hard work, but you know what, there's only one thing I think I can be able to do out of all of this," because he expressed he was feeling overwhelmed. So then we negotiated it down to, "Okay, let's just look at those patients that are high risk of unplanned admissions, and if we could look at that patient list and be able to see those patients." So he said, "You know what? That's something I can do. I already have a workflow for it and I can put that into action." So, "Okay, that's your one thing." And then getting that to several other offices who then you start realizing, even though they might be excited, they're really not sure where to start.

And so, what we decided to do is then take these as not just [an] overall group approach, because [there are] things we can do internally with our care coordinators that we can work on, but what are we going to have for the physicians? So what we did is then we narrowed down for the physician practices, let's look at those unplanned admissions and, because we have a lot of issues historically with high admission rates and high readmissions, and so attacking that particular cohort of patients made a lot of sense for us.

Then we said, "Okay, what else can we do internally?" So then we took those smart cohorts of those that are in rising risk and chronic risk, and advance and started looking at those with our care coordinators that are embedded in some of the practices because then that way we can learn how to work with those cohorts more effectively, and be less overwhelming and then be able to track and see where that progress is going to be.

And then we've done things like, it's flu season, this is a time we really push pneumonia vaccines, pneumonia is one of those things that's in our top 10 hospital admissions. So we said, "Okay, let's just pull everybody that's at high risk of pneumonia." And we recognize that could be clinical, that could be bacterial or viral, but, "Hey, let's just make sure every one of those patients has got their full pneumococcal vaccines done," because that's something we could do. And if we could prevent some of those patients from going in the hospital for pneumonia, that's the whole thing: how can we keep patients healthy at home? Let's look at this information with that end goal in mind.

We also recognize from looking at that data that our patients that have CKD3, whether it's A, or B or a little higher risk of having ER visits and those kinds of things, so we've identified a nephrologist who's willing to work on it. So that's our next new project. So we've decided what we're going to do is start looking at this, and then breaking them down into bite-sized pieces where we think we can actually have an impact. We are excited so many ways of looking at this, how can we make it effective? And so those are the directions that we're moving into right now and how we're using the platform.

Shania Kennedy:

I'm glad you bring up breaking it down into pieces, because I think a part of the challenge, at least that I see when I'm speaking with ACOs trying to use analytics, it's that, "Okay, I have all this data, or I have these solutions, but I don't know how to apply them. There are so many, many options, I don't know what to choose." And so, I think in your case, having that end goal of keeping patients healthy at home, there are many ways you could go about that, of course, but I think having that broad goal helps you narrow down further, and look at something like unplanned admissions, which is an issue for many ACOs. But again, on a case-by-case basis, you're going to need to decide what data you're pulling, from which populations.

That brings us to the question of data, and I want to turn to David for this part because in the press release announcing the partnership, it mentions that HDAI focuses on supporting value-based care delivery by benchmarking against, "precisely matched comparison populations," or digital twins, essentially that leverage predictive models for care management. And that was really interesting to me, because digital twins and predictive modeling are getting a lot of hype right now. I'm seeing a lot of talk, especially about digital twins because a couple of health systems are using them to make digital twins of patients. This isn't exactly the same, but it's still worth mentioning and I think it's worth talking about. So I was wondering if you could talk a little bit about how these are utilized in the context of the use cases that Dr. Magoon is talking about and what sort of data get pulled into those?

David Clain:

It's a great question, and when we think about digital twins and how we do these analyses, we're thinking about the kinds of conversations that Dr. Magoon is having where she's going to individual physicians and saying, "Let's look at specific patients who might require some support, or some specific intervention, but also broadly speaking, how are you doing managing this population?" And the refrain that we've heard from everyone that I think is very consistent across ACOs, and has been around for a long time is, "My patients are different." And if you don't have a response to that, you're going to walk in and Dr. Magoon would say, "Dr. So-and-so, let's focus on why your patients are in the ED, or why your pnuemo vaccination rates are low." And the doctor's going to say, "Well, I have a different patient population. It's not fair to compare me to others." And it's very difficult to overcome that, I think, if you don't have the data to do it.

And there are other non-data questions here too: do the physicians trust you? Are you having good conversations with them? But having the ability to answer that and walk into the room saying, actually we have thought about that. We've looked at the extent to which your patients are different. And we still see that there's an opportunity. Or conversely, you've got a challenging population, the numbers look bad in absolute terms, but given the population, you're doing really well, it's important to be able to do that. And digital twinning is how we do that.

So if you sort of think about step one, going back decades when we first started having this data from payers and Medicare provides some data to Medicare Shared Savings program ACOs, we started with counting stats, basically; how many ED visits do you have? How many primary care visits do you have? What's your pneumonia vaccination rate? It's an okay starting point, but it's just going to provoke that question of, "I've got a different patient population, how can I know this is representative?" Then we started saying, as an industry, "Let's risk adjust these patients in these panels," which is better than nothing, but that risk adjustment is generally pretty crude. So generally, you pick one metric that's sort of a measure of patient acuity, in Medicare we use typically the Risk Adjustment Factor, the RAF score, which is really not designed for this sort of patient-level assessments of risk. And you say, "Well, good news," to the skeptical doctor, "We've looked at your patients, your outcomes relative to your RAF scores, and maybe you're doing better, maybe not as well as we look at that."

But the fundamental challenge there is their patients are complex, health is complex. You can't take one single metric and say, "We've adjusted for all of these possible things." So when we do digital twinning, it's sort of the third wave of how we can do this kind of analysis. We're finding through our access to data on de-identified data on essentially every Medicare patient in the country for every single patient in Dr. Magoon's ACO, and the other organizations we work with, we find digital twins who look as close to exactly like them as possible. So people who have the same age, sex, Medicare enrollment category, primary care utilizations, are they attributable to an ACO or not? Which of those cohorts that Dr. Magoon mentioned, are they end of life patients? Are they transitional rising risk?

And for every outcome we're interested in, we have a predictive model that says, "What is that patient's baseline risk of an ED visit, of a podiatry visit, of a fall, whatever it might be?" And then we follow Dr. Magoon's patients and we follow all of these other sort of Medicare twins for 30 days, 90 days, a year, whatever window we're interested in. And we say, "Now we've found basically a control group." And now we can see for any of those doctors that Dr. Magoon is talking to, how do they do relative to that very similar group of patients around the country who are matched for everything we can think to match. And someone can still raise an objection, and there are always individual patient circumstances that can be challenging, but that way you can at least say, "We've controlled for all of these different things." And now I think the onus is a little bit on the provider. If the provider wants to say, "My patients are completely different," we can say, "Well, they're not different in terms of age, sex, baseline risk, enrollment category, any of these other things." And it puts the burden on the provider a little bit and hopefully overcomes that objection.

So those are the different things we're looking at. And I think you actually do see there are very different patient populations out there, and instead of being defensive about this, we can say, "You do have a very challenging population with a very high diabetes disease burden. We expect that you're going to have more diabetes related hospitalizations than the average provider, but let's see how you're doing relative to that population." I think that makes for a lot more productive conversations when you can start with that baseline.

Shania Kennedy:

Absolutely, and it's interesting to see because you bring up a good point. There's a really key challenge here that I think sometimes gets kind of lost when we're talking about data analytics. And that's the fact that like you said, you have to get payer data, risk data, and then you can do something like these digital twins like you're talking about. But there are so many steps involved in that data collection analysis pipeline. And of course, that all aside, you have to ensure that the data you're getting to begin with is high quality because otherwise it's useless. A lot of software engineers like to say, "Trash in, trash out." It's kind of the same thing here.

So that's sort of where I wanted to segue into next because when I get emails from providers about the things I cover on my website, it's always, "Okay, well I want to hear more about data quality. I want to hear more about how do I ensure that the data being pulled by either the provider themselves, if they have the capability, or a company that they're working with, how do I ensure that data really is the gold standard that I need?" So to sort of follow up on what you just said, David, how do you at Health Data Analytics ensure that the data being pulled for the predictions is high quality and more importantly, almost, unbiased? Because bias is a key issue here, especially when we're talking about value-based care and health equity.

David Clain:

Yeah. Yeah, there are a few different parts of that. So I guess the first step is how do we make sure we understand the input data and that it's high quality. And we do see, and other people who've done research with Medicare data will see this too, we see claims for patients who are deceased that come in after a date of death. We see cases where there is something that is clearly wrong with the data. And so, part of our job as data scientists is to understand that, clean that up. But then even excluding those very clear-cut cases, it's important just to understand what the data that you're working with, if it's claims data, or lab data, or EHR notes or anything else, what it can do and what it can't do.

And so, with billing data, you have to be very careful about making sure [to] understand the limitations and the uses. It wasn't designed initially for use to build predictive models. We think it's very powerful for that. What we do to make sure we're getting the right input data is, leaning where we can on publications that do chart reviews compared to claims data. And as we're building new outcomes to predict, we actually do that with some of our partners. So if you want to predict advanced heart failure, there is no ICD code for that. Similarly, staging for renal diseases is pretty poor, so we actually build new outcome metrics and then compare that to a chart review to make sure we're pulling the right thing.

And then, the second part of the question I think is as we think about especially predictive models, how we make sure they're high quality and they're unbiased. And I think part of that as data scientists is for us to look at all of the standard things that data scientists do when we build these predictive models. And we look at area under the Receiver Operating Curve, and PPVs, and sensitivity and specificity. I will say, I don't think most clinicians care about those things. I think the test for Dr. Magoon and her providers is when she walks in with that list and says, "These are the transitional patients who were recently in the hospital who we think need more care," the doctor doesn't ask for any of those. The doctors, or the nurse, or whoever Dr. Magoon is talking to looks at the first 10 patients and says, "Yep, you got the right list," and they'll trust it. Or they say, "Nope, these aren't the right patients," and they don't, and there's no amount of data that we can bring them to say, "No, actually it's a really high quality model that's going to overcome that."

So I think it's really working with clinicians to say, "Is this serving the need that you have?" And then the bias question is a really important one too. And I think you mentioned, "Trash in, trash out" with models. I think there's a "bias in, bias out" problem with models as well. And what we have to be very careful to do is not to sort of reinforce underlying biases in the data and in the healthcare system broadly. And so, if we want to build a model predicting who's likely to need a primary care visit, or who's likely to go to the ED, to the extent those things are biased based on sex, on economic means, on demographics or location, we don't want to build models that just reinforce that and say, "This is a low-income population. They've historically had lower access to care, therefore they don't need to come in and see a primary care physician." And if that population is going to the ED more because they don't have that stable primary care relationship, we don't want to say, "We expect a ton of ED visits from them, therefore it's okay."

So what we try to do is pick inputs to these models that are not as sensitive to historical utilization especially, or things that are likely to be biased. We want to say, "Let's look at the patient's inherent health status and risk and predict based on that." And we do look at that bias, and I think it's important to say, look, no model is ever going to be perfect and free of bias. Let's understand what it is, and let's work with clinicians to say, "This is what the model does. This is how you want to use it in a way that is going to ensure that rather than reinforcing some of those inherent inequities, we're actually going to try to overcome it." And we'll say, "Yes, we want you to do more primary care visits than we expect based on the model because this is a population that historically hasn't had that access."

So I think, just being really clear on that is important. It is partially a data science problem, but then I think it's really the interaction between people building these models and then the people using them to say, "How do we make sure that we understand what those issues might be?" And we're using them in a way that's getting us the right outcomes we want and not reinforcing some of these inequities.

Shania Kennedy:

And getting that clinician input, of course, is critical, not just for addressing bias and preventing disparities, but like you mentioned, even getting them to want to use the data that you're providing. No doctor is going to just take something that's handed to them and say, "Okay, I'm going to run with this," because that's just not what doctors do. They want to know how it works, they want to understand it, and then if they think that it's a good system, then they'll consider bringing it into their clinical care for their patients. But you've got to build that trust across the board, and I think that's really critical, and that's something especially that I think partnerships like this can really help do.

And to sort of close out talking about the partnership because that's really what intrigued me in the first place when I read about what work you guys were doing. I know it's still in the early stages, so I know you might not have too much to share, but Dr. Magoon, you emphasized that analytics partnerships can play a major role in helping ACOs meet their value-based care goals the last time I talked to you. So now that you've officially been working with HDAI for a few months, are you starting to see some results? And if so, can you share some of those with us?

Dr. Sheila Magoon:

Sure. Well, first of all, what I'd like to do is just reinforce what David said is, in regards to the lists we're bringing up. When using our standard analytics, I would try to create a list of, "Here [are] the people I think that are at risk." And so we'd go through those lists, and then the doctors would throw some of them and tell them I didn't make these choices right. Well, I will say at least now that I'm bringing the Health Data Analytics list, and saying, "These are the patients that are at high risk of an unplanned admission," when they're going through it, I'm not getting told, "Oh no, I didn't choose my list right." So that's been a huge benefit that they look at them and they go, "Oh yeah, I could see where maybe this person could have a decline or have something else going on." And so that awareness has really been wonderful.

With that alone, what we have seen is, right now our admission rates are stable, which is good because we're starting to see some uptick now with the end of the public health emergency of people using the emergency room more, but I'm not seeing the same percent rise of going into the hospital, so that's good. So I think at least these higher risk patients, we're seeing them, and I think we're going to see some benefit from that on the backside here. And because our numbers are small, I really got to take a whole year of looking at this to be able to see those gains.

And also we've been looking at some of what they call their Smart Cohorts as they put people into categories for us, like who's at the most advanced risk of having adverse events? That Advanced Care List, it was one of those when they went through, I heard stories about every single patient on that list. A lot of them were really at end of life stories. And so it's, okay, that's at least good. We at least have an awareness of what's going on there. And then looking at that chronic list, they're like, "Oh yeah, I could see." And so, then we could pick and choose how we were going to handle those, and how we're going to address the patient care that's associated with it. So I want to say that's been wonderful, just to have those lists accepted and being able to be used for care coordination within their offices and seeing patients has been a tremendous benefit.

So at this point, like I said, we're seeing some stabilization, and that, to me, is goal number one, and we are really hopeful of being able to see however the next year of where this is going to take us.

Shania Kennedy:

Yeah, I'm really excited to hear more about your progress as time goes on, because you're seeing some results already, which is pretty incredible because--[it was] April, you guys announced this partnership? So it has not been very long. It hasn't even been half a year, and you're already seeing those stabilization rates.

So that was everything I had. We have run out of time, so I want to thank you both for coming on to the show to chat, because one of the most exciting parts of my job is getting to hear from organizations that are using these advanced analytics on the ground to make tangible, real-world impacts, and I know that our audience is really interested in that and they really value that as well. So again, thank you two, so much.

David Clain:

Thank you.

Dr. Sheila Magoon:

It's our pleasure. Thank you.

Shania Kennedy:

And to all of our listeners, thanks for tuning in. You can feel free to reach out to me over on HealthIT Analytics. If you have thoughts on this topic or if you have any healthcare related stories that you'd like us to consider for coverage, you can reach out to me at skennedy@xtelligentmedia.com, that's skennedy@xtelligentmedia.com. And if you enjoyed today's conversation, let us know by following the podcast on whatever platform you listen to and leave us a review.

• Tagged
• Accountable Care Organizations
• Predictive Analytics
• Value-based Care