Episode 4: The Future of AI & Kidney Disease: Predicting Care Delivery with Len Usvyat, PhD
Len Usvyat, PhD, Vice President of Applied Advanced Analytics, joins Field Notes to explain how artificial intelligence is being used to provide better treatment for people with kidney disease by predicting potential outcomes and future health care needs.

Brad Puffer: Welcome, everyone, to this episode of Field Notes. I'm Brad Puffer of the Medical Office Communications Team at Fresenius Medical Care North America and your host for this discussion today. Here, we interview the experts, researchers, physicians and caregivers who bring experience, compassion and insight into the work we do every day.

I'm particularly excited because today, we're turning our attention to a fascinating topic, artificial intelligence and its potential to help better guide a patient's care by predicting negative outcomes so we can intervene sooner and, hopefully, improve that outcome. For example, with dialysis patients at Fresenius Medical Care, artificial intelligence is already helping predict when someone is likely to be hospitalized so we can intervene sooner or understand when a patient may be at risk for certain infections. And now, it's even being used to try and see when patients may have COVID-19 even before they show any symptoms.

As an organization, we have already developed more than a dozen predictive models providing insights that might otherwise never have been identified without a computer's ability to continuously examine hundreds of variables and data. This panoramic view of the patient is helping us move from what has become standardized care to highly personalized care. Here to talk to us today more about artificial intelligence and its potential is our own expert on this topic, Len Usvyat, who is the vice president of applied advanced analytics at Fresenius Medical Care. Len, welcome to Field Notes. Thanks for being here.

Len Usvyat: Brad, thanks so much for having me. It's really a pleasure.

Brad Puffer: Well first, I wanted to just set the stage. Because outside of Hollywood's view of robots that are smarter than we are and about to take over the world, what really is artificial intelligence?

Len Usvyat: Well, contrary to the-- contrary to most people's interpretation of AI as being these robots, it's really an umbrella term that encompasses a variety of different fields, both statistics, mathematics, algorithmics. And at the end of the day, it's the ability of computers to mimic human behavior in some way, shape, or form. And some of the main sub-components of AI that most people are probably relying on more frequently these days are things like machine learning, or things like deep learning for example. And the term itself, I should mention, was actually coined in the 1950s at the meeting that occurred actually very close to Waltham at the Dartmouth Summer Research Project collaboration.

Brad Puffer: Oh, wow. Well, how did you get into this area of expertise and what for you is really most exciting about the potential for artificial intelligence in healthcare?

Len Usvyat: Well, I mean, I think I actually have a little bit of an unusual background. I ended up getting a degree in accounting and economics, and I did a lot of work in IT, and I ended up working for one of our subsidiaries, Renal Research Institute in New York City. And it's at that time when I realized the power of data and the power that the data can actually make on patient outcomes, and which is something I have never realized, and so that's really how I got into the whole field of data analytics, data science and artificial intelligence.

Len Usvyat: And over the time that you've been involved in this and working on these issues, how quickly has this technology taken off and what are some of the biggest breakthroughs that you're most excited about?

Len Usvyat: Yeah. So I mean, I think it's definitely been an evolutionary process. In the past few years, there's been a number of different, for example, statistical methodologies that have come out. Things like XGBoost algorithm or certain neural network algorithms, things that have come out or have improved the ability to process data tremendously. And, of course, one of the main things that I think is responsible for why AI has become so powerful is the fact that on the one hand, we've collected a lot of data or we are continuously collecting a lot of data but then on the other hand, it's the ability of computers to process data very, very quickly. In the past this would not have been possible if we were not able to process this data so quickly as we can right now.

Brad Puffer: Well, like a lot of health organizations, Fresenius Medical Care is focusing more and more on using artificial intelligence, machine learning, predictive analytics to better help our patients with kidney disease. Tell us some of the specific work that you've been working on that has been most successful.

Len Usvyat: Yeah. So I think, Brad, you mentioned that in the beginning that I think we have a suite of over a dozen different algorithms. There's also-- in other parts of Fresenius, globally, there's also other work that's being done in addition to these-- to the work that I think you mentioned in North America. There's many, many examples of the work that we're doing where-- and some of it has been in different levels of successes, but I think we've had quite a lot of successes with some of our predictive modeling efforts related to hospitalization and knowing which patients will go to the hospital in the next few days. And now, we're working on predicting cause-specific hospitalizations such as fluid or infectious hospitalizations.

And some of the other work has to do around patients' transitions from in-center modalities to home modalities. Who are the patients, for example, who are best candidates? So we've been working with our home therapies team in helping them identify some of those patients who we think, based on the data, fit the profile of somebody who traditionally was a very, very good candidate for home therapies.

Brad Puffer: Well, Len, of course, right now everyone is looking for insights on how to better predict who will contract COVID-19, and identify people earlier so we can isolate them and protect them from exposure. With all of this data that we monitor for our patients, do you think artificial intelligence can help there? Because I understand that's exactly the problem you're working on.

Len Usvyat: Yeah, Brad. Thanks for asking the question because I think one thing I should-- first of all, I should mention that Fresenius is in a very unique position. I think dialysis providers, in general, are in a very unique position when it comes to COVID. We see patients for long before they even develop any COVID symptoms, or that they are diagnosed as having COVID. And so we're able to track and see what happens to patients for weeks or months before they actually develop these symptoms. And we also collect all the data about them. And so I think this really gives us the ability to be able to build these models that would-- predictive models that would help us identify which patients may have already gotten the virus but don't yet display any symptoms.

And so this is-- I think this becomes extremely important and we do have an effort right now to predict those patients who don't have any symptoms, for example, or clear symptoms, but there are some smoke signals in their data that make them more likely that they have already picked up the virus and have not yet developed any clear symptoms. And so trying to identify that before that happens is really important in making sure that we can isolate those patients or test those patients before they can-- they start shedding virus potentially.

Brad Puffer: And what are some of those one or two most important signals in that data, or do you really have to look at the hundreds across to make any predictive ability?

Len Usvyat: Yeah. And so some of them are very obvious. There's over a hundred different variables that go into the model, but some of them are very obvious. And so, for example, it's things like temperature, body temperature. But what's interesting is not your typical classification of fever, for example, that makes the patient more likely to have COVID. And yes, of course, if the patients do have fever, that is more likely that they have picked up the virus, but even just minor increases in temperature. And I think some of those are very patient specific.

So temperature may be one. I think BUN, maybe some of the other measures, oxygen saturation. And, again, I think some of these very small changes in these patient parameters make it more likely that the patient has actually picked up the virus somewhere.

Brad Puffer: Are you looking to potentially publish this data, or are we looking to implement some changes in our centers based on this data? What's the next step?

Len Usvyat: Yeah. So we do think it's very important to publish the information that we're finding. I think it's important for the medical community at large. And I think also to get it-- really to get in the message out there. And so we are publishing what we're finding. We're also collaborating very closely with our colleagues in the European region who also are using their data in Fresenius, but Fresenius in the European region, to build similar type of models to identify patients, or in some cases, clinics that are at a higher risk of having, let's say, a COVID outbreak. And so we're publishing this information right now.

Brad Puffer: Can artificial intelligence really help patients get just the right treatment they need, which is so important when it comes to dialysis?

Len Usvyat: Yeah. I mean, I think if we think of health in general and medical care, if we think of healthcare 100 years ago, 200 years ago, I mean we always think of this very, very personalized care that the physicians delivered to the patients. Most physicians used to or still do, but I think, especially back in the day, they would know their patients very well personally and in a way, they would try to deliver the care that it's extremely personalized. As our physicians are treating more and more patients, this, of course, becomes more and more difficult to deliver that truly, truly personalized care while going through thousands of data points that we may be collecting on these patients. And so this, to me, is truly where AI can come to the rescue, where it can assist in identifying certain patterns that are very specific to that individual patient, which, in essence, would then allow us to deliver a very, very much precise personalized care.

Brad Puffer: For some people, all of this talk about AI and how we deliver care might raise concerns about whether computers will eventually replace physicians. Are you seeing the clinical community excited about these new technologies or fearful about how they'll be used?

Len Usvyat: I think I'd say it's a process and I think we-- I would say that over the last probably 10 years, as more and more of these different technologies are being introduced, I would say there's more and more acceptance. One of the examples that I always use is when the anemia algorithm, the work that's been done by Renal Research Institute, when that was developed, I think in the beginning, there was a lot of these what we call overrides for the prescription of the actual ESA. Meaning the physicians would not agree with what the algorithm would come out with, or maybe they would make small tweaks to the actual prescription of the ESA. And over the years, we have seen less and less of that while we have also seen better and better outcomes. So I think there is-- like everything else, I think in the beginning, we probably don't trust technology and I think over time, there is more belief that I think it only helps the physicians and it is actually providing that personalized care that we want to provide to our patients.

Brad Puffer: How do we, then, turn all these great insights into this data into actionable intelligence? Because you can have a model like that, but unless you can take action, that really drives an outcome, it's just interesting. It's not effective. So how do we link those two together so we're really driving better outcomes for our patients?

Len Usvyat: Yeah, no. Thanks, Brad. I would say that's actually one of the things that, in Fresenius North America, we have been very successful at, and definitely more successful than many other organizations. I think there's a general tendency to just develop these algorithms and to publish them, and maybe not do anything else with them. And I think we always say that our work does not stop at just building the algorithm. The most important thing to us is to actually incorporate these algorithms into the clinical care and figuring out how it can actually work in real practice, how can it actually work across 2,500 dialysis clinics and thousands of clinicians out there. And this is-- to me, this is what makes me very excited is to actually turn these into operational practice and something that really changes clinical care.

Brad Puffer: So when we look out 10 years from now, what does kidney care look like? Are we able to take these insights and prevent kidney disease or really change outcomes for our patients? What is the potential?

Len Usvyat: So I think it's-- we'll be seeing a lot more technology assisting with a variety of different elements of kidney disease care, be it dietary support, or be it social worker support, or be it preparing people properly for dialysis, and I think, as you mentioned, for example, predicting which CKD patients will actually transition to dialysis. I mean, this is a very important area. If patients do transition to dialysis, what kind of patient will they be? What are some of the things we would want to know about that patient on dialysis if we had this information a couple years before they actually start dialysis? And I think all of these technologies will become more and more proliferate in the care of chronic kidney disease patients.

Brad Puffer: And it seems that more and more we have devices that are collecting this data, that's really helping you. We've recently invested in the company, BioIntelliSense, for example, with remote patient monitoring. We're doing a lot more monitoring of our home patients through the machines and bringing data back to our critical care teams. There's Crit-Line that's been in play for a while, but collecting tons of data on our patients, it doesn't seem like what you do could be even possible without all of those additional tools.

Len Usvyat: Yeah, no. That's exactly right, Brad. The AI-- I should have probably mentioned this in the beginning but AI needs data. Without existing data, you can't have AI because, by definition, it needs history to be able to learn how to do things in a way in the future. And so I think we-- the fact that we're able to collect all those different data, the fact that we're able to potentially get data from these BioIntelliSense devices or the Crit-Line data, that is what makes AI possible. And the more of it we have, the better is the personalized care that we can deliver, so this becomes absolutely essential.

Brad Puffer: Well, we're certainly continuing to evolve everyday. We look forward to more of your algorithms and insights and machine learning tools that will help our patients thrive and do better. It's been an absolute pleasure talking to you, Len. So I really appreciate your time today.

Len Usvyat: Brad, thanks so much. It was great. Thank you.

Brad Puffer: And to our audience, you can find Field Notes on the Apple Store or Google Play, or right here at fmcna.com where you can also find our annual medical report and other feature articles. We hope you'll come back and join us as we discuss more important issues in the weeks ahead. Until next time, I'm Brad Puffer, and you've been listening to Field Notes by Fresenius Medical Care. Take care, everyone.