Episode 53: Predicting Population Trends in Kidney Health Using Advanced
Mathematical Modeling with Doris Fuertinger

Episode Transcript:

Frank Maddux: Globally, changes in lifestyle have led to a steady increase in obesity and diabetes, major drivers of chronic kidney disease. Advanced epidemiological -type models are valuable tools in assessing the impact novel therapies and demographic changes may have on the size and characteristics of future kidney disease populations. Dr. Doris Fuertinger, Fresenius Medical Care's head of computational medicine, joins us to discuss the use of science -based modeling tools to predict the future characteristics of kidney disease populations. Welcome, Doris.

Doris Fuertinger: Hi, Dr. Maddux. Thanks for having me today.

Frank Maddux: It's good to see you. And let's start by just some description on how do you look at population trends and why is it important that we analyze this as we try to look at what our patient population today and our future population begins to look like.

Doris Fuertinger: There is a continuous change in the landscape of kidney disease. We've seen that over the past decades and we will continue to see that in the future. So, there are multiple things that happen at the same time. On one hand, we have an ever-aging society, so people get older and older. We have a surge in diseases that can cause kidney disease, like for instance, obesity and diabetes. We especially see that in well-developed countries where it's to some extent a consequence of our lifestyle decisions. We saw, especially like in the last decade or last two decades, a lot of advances in medical technology, let it be new dialyzers or new dialysis machines, new approaches to transplantation and other things. We see new medications coming to market which can actually delay the progression of kidney disease but also can impact diseases that cause kidney disease as for instance diabetes. So, there's a multi-of things happening at the same time. And this makes it really difficult to assess where differential effects will be heading.

So, on one hand, we expect that with the aging society, we will take care of older patients because of new advances in medicine, we kind of are healthier for a longer period of time. On the other hand, because of our lifestyle decisions, we get to some extent a certain part of the population gets sicker earlier on. So, I think it's really difficult to predict what the kidney disease population in let's say a decade will look like. And I think it really needs some quantitative and systematic approach to actually synergize the information that we have from various public health trends and kind of make an educated guess on what the future population size and demographic characteristics will be.

Frank Maddux: Let's dive just a little bit deeper and I know most of us won't be able to get to the technical level that you and your team work at. But describe a little bit about the techniques, the mathematical techniques that you use, and how you've approached the problem of this population modeling in general.

Doris Fuertinger: When we started this project a little more than a year ago, we made some very conscious decisions on how we actually want to model and think about the past population trends and future population trends. So, we decided to go for a type of model that we've seen also during the COVID pandemic where epidemiological models were being used to actually predict how certain interventions, COVID measures would impact the number of infected patients that we see. So, we use a similar type of model to predict how the patient population will develop in the renal population landscape. It's very important to also think about with what kind of data such a model is being fed. And again, we decided here very consciously to go for public health data only. So, the population model that we developed is based on publicly available data and this is just simply for the reason that we really wanted to give others the chance to actually look into the underlying assumptions of the model and to be reproducible also for other people out there.

Frank Maddux: A little bit of what we've been dealing with in the last number of months have been the increasing awareness and utilization of some new drugs that are in the marketplace and how they might affect the population. And this effect is this balancing effect of impact on chronic kidney disease progression, impact on mortality and cardiovascular health. Describe a little bit about how the interactions of these things all go into the modeling framework that you have.

Doris Fuertinger: We've seen new drugs coming to the market. That's a little bit far-fetched because the drugs that have gained so much media time recently, so GLP-1 receptor antagonists, they are on the market for two decades now, one has to say that over the years, different types of this molecule were developed and the more recent ones, like for instance, semaglutide, have quite an impact on weight. So, they got a lot of media attention because of the potential to lower the weight without actually going on dietary restrictions. 

That kind of led to an increased acknowledgement of these drugs in the wider population. So, they are currently not only used as antidiabetics for which they were developed in the beginning, but we now see them being used also for other indications such as obesity. Now, that being said, they have an impact on your blood sugar levels, they have an impact on how your body empties your stomach, what your satiety is. And that, downstream leads to effects on your cardiovascular health. And on how well your kidneys function. Now, there are a number of clinical studies that have been published over the years around cardiovascular mortality in different populations. And they showed quite promising effects in these clinical studies. And then very recently, we saw the Flow trial being first of all, interrupted early on, so they finished the study a year earlier than initially designed because they had very good effect in their population. And that was done in patients that had kidney disease where the primary end part really evolved around progression into ESKD, so into end -stage kidney disease and into mortality related to renal health and cardiovascular mortality. 

Frank Maddux: Before we get to talking about Flow, which I'll come back to in a moment, tell me a little bit about how disruptive COVID-19 was and the pandemic to these population modeling efforts. When you begin looking at the impact of COVID-19, we know we had a lot of unexpected deaths that occurred in our population. And I'd be interested in getting your perspective on, through the mathematical modeling tools that you are working on and have developed, how do you look at both the pandemic and the recovery time from the pandemic, and what are your perspectives on that?

Doris Fuertinger: I really have to start a little earlier by maybe explaining how we extract trends from previous data. So, a lot of the work that we do is not writing down a model and implementing them, but to actually understand the data that is out there and to extrapolate trends that we see in the past data. So, when I say past data, we look back 10 to 20 years, depending on the data set that we have available. And we look into different things that drive the structure and the size of the renal disease population, that being incidence into CKD as well as into ESKD, mortality in those two populations, and progression from CKD to ESKD. Now, from, when I say historic data, the data sources that we use, again, are publicly available. So, we base most of our analysis on USRDS data and then some further analysis on CKD, so in the earlier CKD stages using NHANES. The data that we have there is very good, especially for ESKD, because there's a registry basically so, once you enter the ESKD stage, you're very well tracked in the system.

And that also means we have a very good idea what the mortality rate of those patients is. So, at what age did they die, for what reason? And we use that to actually look into historic trends. So, the expectation in general is that you see a growth in the CKD and the ESKD population following something similar to the growth that we see in the general population, right? A certain percentage of the general population at any given point in time develops CKD. Now, the question when we structure by age and look into how patients with different age progress into CKD and into ESKD and how they die in those different stages, then we see very clear trends over, for instance, the last decade with quite a decrease in the mortality of CKD patients and specifically ESKD patients. 

That's one part of the story, right? So, it's kind of understanding what has happened historically. And you mentioned the COVID-19 pandemic. So, there are some rare occurrences where we see disruptions of trends in population growth, in mortality development. And the COVID-19 pandemic was one of the severest ones we've seen in a very, very long time. Now, what does that actually mean? When we look at the CKD and ESKD population, and let's stick to mortality for a second, just to pick one thing, we see that the kind of the survival gain that we had over the last decade was diminished during the COVID pandemic. So, for the CKD population, it kind of threw us back almost 10 years, eight years, nine years. And for the ESKD population, it was even worse. And we see a reverting to trends similar to what we've seen in the early 2010s. 

What does that mean for us when we think about how we recover from the COVID pandemic. From the publicly available data, we can look back until like the end of 2021, the beginning of 2022. That's the data that's right now available. And what we've seen is this really severe disruption in trends in 2020. And then already in 2021, you see kind of recovery in the population trends. Now, the assumption is, and we talked to a lot of different experts in the domain, is that we will come back to previous trends over time. Obviously, the question of when is one that is very speculative. But in our model, we kind of have to give the model some numbers. So that's what we do. We say, for instance, we recover to the pre-COVID trends in 2025. But we know that this is really speculative, that this is an assumption where a lot of guesswork is happening.

So, what we actually do is we say, okay, let's say it's 2025 or anything between 2025 and 2030. And that kind of gives you a wide range of scenarios that we are actually testing. It's what we call sensitivity analysis. So, for all the data and the information that is being fed into the model, we do sensitivity analysis. So basically, there are certain known uncertainties, there are unknown unknowns, and there are, some known unknowns and there are unknown unknowns. And there are some tools to deal and how to incorporate that into your predictions moving forward so that you actually really get the very solid and well-founded prediction for your population trend.

Frank Maddux: Describe a little bit how you utilize simulations and repeated opportunities to run these simulations on a given set of assumptions and how that gives you the yield of your confidence intervals around the modeling.

Doris Fuertinger: Let's start with something that is a data uncertainty first, where we know people always think that data is 100 % factual, right? And the problem even with data is that there are some uncertainties in the data. When we look, for instance, into the information that we have in the CKD population, we face various problems. There's no registration system that kind of logs at what point in time a patient entered CKD3, CKD4, CKD5. It's a fluid thing that's happening.

The majority of patients that have chronic kidney disease do not know that they have chronic kidney disease. So that means that two-thirds are undiagnosed. Even higher percentages in the earlier CKD stages and then in the later CKD stages that becomes a little less. That being said, it's really difficult to actually say very precisely what the number of patients with chronic kidney disease in the US is. You get a number with a spread around it basically that's based on uncertainty, but you don't know the exact state of the system. So, what that means is when we look, at the data from year to year, there's some noise in the data. We try to extrapolate a trend that has some continuity over time. We see in our advanced analytics that there is a certain uncertainty around that. So, for the CKD population when we just only look at the growth rate. So, by how much does the population grow every year of diagnosed and undiagnosed people with CKD? We see anything between 0.5 and 0.9 million. Now that's quite a difference when I start to predict into the future for the next 10 years, right?

So, what we do is we again use sensitivity analysis and we say, okay, we assume that the population growth for CKD is anything between 0.5 and 0.9 million per year. So that's the thing that we know and that comes from data uncertainty. Now, when we talk about effects, of a drug, we enter an even more a uncertain realm because we have to rely on clinical trial data. Now clinical trial data in itself is difficult to extrapolate to the use to the more general population than later on. Very often what you see in clinical studies is a little bit different then what you experience when drugs are really used more generally. That's for a multitude of reasons that maybe at that point in time I'm not going into. But we again see even in the clinical studies a certain effect of the drug, usually that's expressed in hazard ratios. 

Now around these hazard ratios we already see, so the scientists already give us a confidence interval, so how confident they are that the effect is actually within a certain range. Again, we use that kind of approach. We take a range around hazard ratios and say, let's say for GLP-1RAs, we assume that the progression into ESKD the hazard ratio is 0.84, then what we do is we actually test everything between 0.75 and 0.94. Again, this gives us a wide variety of effects. We do that with all other parameters that are in the model. And I just previously mentioned this when we expect a certain return to previous trends.

That's going to be varied. I mean, right now, nobody can honestly claim that they know how many patients or how many people, I should even say, will take GLP-1 in 2030 or in 2035, this is guesswork. Now, again, we account for that by testing a multitude of scenarios. What's very important to us is that we are not looking into the effect of one of those things, but we actually do a few thousand simulations for each scenario that we are interested in, where we vary all these things at the same time, so that even cumulative or some differential effects that these different parameters have are accounted for and what we report out of the model is a prediction around a reference scenario, so where we have very clearly defined parameters and then the sensitivities around these reference scenarios where we vary all these parameters.

Frank Maddux: Just as you looked at the data that came out of Flow, obviously we had a series of assumptions before we went into Flow. How did those assumptions change from the outcome of the Flow trial, and what do you think the impact of that was?

Doris Fuertinger: Before the FLOW study was published, we had some workshops with experts where we already picked their brains. FLOW was not the first study conducted with a GLP-1 receptor agonists. It was just a different population that was studied compared to previous studies and a very specific compound being used.

So, what I'm trying to say is we already had an idea where we think things would fall. And we determined using the knowledge from the experts a certain set of assumptions that we used in model predictions before the Flow data was published. So long story short, we ended up being very close with our assumptions to what we actually saw in the Flow trial. So, we did some really minor adaptions in our hazard ratios to reflect the information that came out of Flow. But overall, that was very similar to what we've already had incorporated in the model predictions. So not very surprising. Also, the outcomes were very similar with this population model prior to the Flow announcement and then after the outcomes were reported from Flow.

Frank Maddux: As you mature this process of doing these mathematical population models, where do you see some of the opportunities in the future for the models to become better models than they've been from what you started with? How do you look at the future maturing of these?

Doris Fuertinger: I think there's a lot of things that you can actually do with these types of models because it's a tool to test hypothesis. So, I have a certain idea of what I think could happen in the future and then I want to understand quantitatively and systematically how that would impact and shape my population. Now really the cool and fun thing with hypothesis is, that you can start off by doing some counterfactual simulations where we say, okay, let's assume all things stay the same as they are today. I mean, we know that's not reasonable, right? But let's assume that's the case. And then I can start to investigate, but what if everybody starts to use GLP-1RA? And what if people additionally use SGLT2, so another anti-diabetic drug. What if we switch our patients to a new technology? So last year during the ERA-EDTA, one of the major nephrology meetings in Europe, the outcomes of the CONVINCE study were published where we saw that HDF has a survival benefit for patients on dialysis, which is a huge thing and a game changer to a certain extent.

Now, as I said, the cool thing is you can look at all these interventions separately, but you can look at all these interventions combined at the same time and see how do they balance each other, what happens actually to the age distribution because, at one end delayed progression means that patient comes a little bit older to dialysis, but then we also have maybe a survival benefit. Is that the same survival benefit for a 60-year-old one and an 80- year-old one? So how much time do you really gain? You can incorporate all these things in these models without having to speculate. There's some very clear systematic way to do that. Even though we're doing something very speculative because we're testing hypothesis that we came up with and by we as the experts in the medical field and that are aware of how trends probably will develop in prescriptions and these things, it gives you a solid understanding of what you might expect if the scenario that you are thinking of comes true.

Frank Maddux: I've been here with Dr. Doris Fuertinger talking about epidemiologic population modeling that we've been highly involved in through our computational medicine team that she leads. Doris, thanks so much for joining me today and helping people understand a little bit more about this particular work that your team's been working on.

Doris Fuertinger: Thank you, Dr. Maddox, for having me and for giving me the opportunity to present the work of the team.