DaVita Algorithm Flags Issues With Home Dialysis Patients

One way AI is being used by health systems is to flag risks clinicians might otherwise miss. In a recent conversation with Healthcare Innovation, Jeffrey Giullian, M.D., M.B.A., chief medical officer for DaVita Kidney Care, described how his company is using predictive machine learning to improve outcomes for those receiving dialysis at home.

In describing this new tool, DaVita offered up the following scenario: Consider a patient whose blood sugar levels had been trending slightly upward for several months. While the levels were not yet high enough to flag an intervention, the new Peritoneal Dialysis Loss Model tool spotted a long-term trend and alerted the care team. This enabled them to intervene earlier, avoiding preventable infection or hospitalization and helping the patient remain on their preferred modality.

Healthcare Innovation: Is there a difference in tracking patients receiving dialysis at home versus in a clinic that makes this innovation important in catching trends earlier to alert care teams?

Giullian: Yes, I would say there are two really big differences between home dialysis and in-center dialysis. The first is that we wrap a lot of support around patients, but obviously in-center we physically lay eyes on those patients three times a week. And at home, while we will still often see those patients a few times in a month, they are taking care of themselves, providing their own care with their loved ones, and doing their own dialysis. Additionally, especially with peritoneal dialysis, there is a relatively high failure rate. About 30% of patients return to in-center dialysis after about two years. 

We believe in the importance of home dialysis. We recognize that for some patients home dialysis can't be their only treatment during their entire journey of being a kidney patient, but we want to make sure that we're broadening that level of support. This just allows us to not have physical eyes on the patient three times a week, like we do in-center, but it allows us to have what I would call technological eyes on the patient, because we're getting data from the patient and from the machine on a regular basis. Even though a nurse can't physically see that patient, AI in the background is able to pick up on trends and flags them for the nurse. Maybe that means let's physically see the patient. Maybe that means let's alter the prescription. Maybe that means let me just call the patient or their loved one and check in on them, but it allows us to have those insights that we, already have in our in-center patients.

HCI: What are some of the things that the AI is flagging to suggest that the nurse check in with this patient?

Giullian: That's not an easy question to answer, because it's actually looking at about 150 data points, not just two or three outliers. It is things like vital signs and laboratory values, but it is also information coming from the machine itself. The machine might have alarms overnight. Is that alarm frequency changing? What were those alarms for? Are we seeing things where patients are going on to the machine later at night and coming off earlier in the morning? Any number of things that suggest something is changing in that patient's life or in that patient's physiology.  Some of it might be perfectly reasonable. They might say it's daylight savings now, I like to stay out later, and I want to get on the machine later and sleep in. Fantastic. But if they say I used to have a caregiver who helped get me set up on the machine at 9 p.m. and that person is in the hospital right now, so I have to do it on my own, that's a risk. If that's a short-term issue, let us support you in the short term. If that's going to be a longer-term issue, let us make sure we're retraining you on setting up the machine yourself.

HCI: After implementing this early warning solution, are you seeing a decrease in the number of patients who need to return to in-center care?

Giullian: We are. It’s relatively early, so I don't want to over-hype this. As with all things in machine learning and AI, we will continue to iterate on this. But this model flags patients who are in the riskiest 10% and we know that if we do nothing for those patients, they are significantly more likely to fail at doing home dialysis and return to in-center hemodialysis. If we don't intervene, we know that this group is at high risk, and we are certainly seeing that those early actions are leading to better outcomes. We’re seeing that that high-risk group is now about 15% less likely to return to in-center hemodialysis If we intervene than if we don't. That is a long way from saying we've solved the problem, but that is a big chunk of patients who we're now able to support and get them over whatever is going on — psychologically, physiologically, or socially — and support them so that they can stay dialyzing at home longer.

HCI: What's the time frame for this project so far?

Giullian: We launched this in a couple of iterations. We launched it originally in late 2024, and we kind of pulled it back and made some tweaks. We've been really doing this now for probably the last eight to 10 months. When we see an issue with these patients and we intervene, we then continue to monitor them closely for a while. And that's why I think we're seeing this 15% lower likelihood of returning to in-center hemodialysis

Speed to action is critical. Let me back up and just say philosophically, I have a view that data is great, but data alone is just noise. You’ve got to be able to take data and interpret it in a way that gives whoever's reviewing it— whether it's a human being or AI — a way to generate some insights. Those insights then have to generate actions. And in a perfect world, those actions have to actually lead to the outcomes you want. That four-step process is critical. In everything that we do with regard to data analysis and ultimately AI, we are constantly going back through and asking: is the data giving us the right insights? Are the insights generating the right actions? And do those actions matter? 

What we find over and over again is that speed in going from data to insights and insights to action  is critical, because for a lot of these patients, time matters. If somebody is having either a physiological issue or a psychological issue and it shows up on Wednesday, but we don't deal with it until Monday, well, that's a long time to go, especially if there's an infection brewing or something cardiac brewing. The ability to ingest this data regularly, and then every single day for this to flag our nurses and say something is going on with this patient, and have that nurse reach out immediately, that’s a critical aspect of all of this.

HCI: Is DaVita doing this data science work internally or in partnership with startups or other vendors or a combination of both?

Giullian: It is absolutely a combination of both. We've got an entire team here of AI and data scientist experts. What we've said is that we want to take a holistic approach to patient care. We have an immense amount of data, so we can build predictive models and action plans. We’ve got a large-scale IT team that can help us with that and help us make sure that we're extracting the data the right way from our electronic health record and from health information exchanges. 

We also recognize there's a lot of caring for people that goes beyond kidney care. When you're a kidney provider, it's easy to get blinders on and say, I'm going to make sure I'm doing the best thing for this patient's kidneys and kidney disease. Yet our patients don't just having kidney disease. They have cardiac disease, they have psychosocial issues, they have gastroenterology issues. So we have realized we need to partner with people that have expertise across the spectrum from a technology standpoint and from a specialty standpoint. For example, we work with a partner called Linea, which helps us with our patients who have congestive heart failure. We're working with partners on personalized dosing and working with others on data visualization.

HCI: We're hearing a lot about agentic AI from other health systems. Is there a way that that could come into play as far as answering patient questions in the middle of the night, or in other ways in the education aspect of this or in any kind of administrative way?

Giullian: Yes, absolutely it can. And we want to be thoughtful through all of this. For us, this is technology with intention…..Is there a world where agents take a first-line phone call at 2 a.m. and can troubleshoot a machine problem for a patient? Absolutely. I think that world is not far off, and we want to make sure every step of the way there's an easy button for a human to be involved, a human in the loop, so that our patients are getting the direct, white glove care that they deserve.