Read the Full Video Transcript

Diane Newman: Welcome. I'm Diane Newman, I'm an adult nurse practitioner and I'm at the University of Pennsylvania and I'm Adjunct Professor of Urology and Surgery, and I practice in the area of low urinary tract dysfunction. I'm really excited because I have a guest with me today, Dr. Greg Tasian, who is a urologist at Children's Hospital Philadelphia. So I want him to talk today about machine learning. This is an area that I really know very little about and I heard Greg's talk about this at several meetings and here at the American Urologic Association. So he's going to tell us a little bit about what is machine learning, but also some of the research that he's doing.

Gregory Tasian: So I think when you think about machine learning, you're thinking about whatever podcast you heard or like ChatGPT, whatever exists in that lay media, that space. But machine learning in brief is using algorithms to extract features from complex data. So it's using computers to extract these features that may either not be identifiable by humans or too laborious, challenging, difficult to extract. And then as the machine learns from these data, you can ultimately use that information, at least in medicine and clinical care and urology, to guide decision making. By harnessing those vast amounts of data that you were able to extract those features from.

Diane Newman: But you're in pediatrics, so how did you get interested in this?

Gregory Tasian: Well, I think machine learning is one of those really amazing fields that it brings individuals together into a team, so no one person could do machine learning on their own. So it's really a collaboration between clinicians, data scientists, machine learning scientists, and it's an incredible opportunity to bring in students, mentees, trainees into this team. So the way that we've operationalized this at CHOP and Penn is that I have a very deep partnership with Yong Fan who's a machine learning scientists, and it's an incredible opportunity to bring in students, mentees, trainees into this team. So the way that we've operationalized this at CHOP and Penn is that I have a very deep partnership with Yong Fan who's a machine learning scientist in radiology at Penn, and his area of expertise is image analysis. So what we do is bring in the clinical question, the context, the problem that needs to be solved, and then Yong and his team will develop the solutions, the solutions to that problem. And then ultimately it's about how you implement that new information back into clinical care. So it's that integration and that learning cycle.

Diane Newman: Well, and you're doing this at CHOP and then at University of Pennsylvania, but do you also pull in data from other sites? I mean, how exactly are you doing that?

Gregory Tasian: So I think machine learning is like any other large data science field, in which if you have a very limited data set that arises from one institution or one population of patients, you're inherently going to introduce bias, in that that population may not reflect the larger population to which that information would be applied. It can also overfit a model. So it's incredibly important to be able to integrate diverse streams of data. And that means from other institutions that have different population characteristics. And in the space where we're operating, which is in a lot of image analysis using different, for example, ultrasound machines that may produce different quality and type of images, that ultimately allow you to integrate and learn from this diversity of data and thus be able to produce generalizable information that could guide clinical care far outside of one or two institutions. So we've worked with a number of institutions and partners including Case Western and Cleveland, University of Toronto, Sick Kids in Canada, obviously CHOP and Penn. But that collaboration, that team continues to grow.

Diane Newman: You have several abstracts here at the AUA meeting and I want to talk a little bit about the different areas. One, you're looking at your ureteral stones, correct?

Gregory Tasian: Correct. So that's the space from which I came. So I'm a kidney stone researcher and epidemiologist and trialist. So those questions, where our boundaries of knowledge were very clear. One of those simple questions that has yet to be answered is for any person that comes into the ED, what is the likelihood of passing that stone? Right now you get the clinical sense of that stone is small, it's in the distal ureter, you're likely to pass it. But that is pretty rudimentary and it also relies on actual manual drawing and tracing and measurement of that stone in the ureter. That process can be automated. So what we seek to do is identify that stone, measure it in an automated fashion, and then feed in other information about the patient that may help predict that probability of passage, so that you can have an individualized approach, a personalized approach for every patient where you can determine, yes, it's worth the trial of passage or no, it's very unlikely that that stone will pass. It's better to do stent or ureteroscopy at that time.

Diane Newman: So then you can really kind of identify the size of the stone, where it is and then how long it takes possibly to pass or if it doesn't pass, huh?

Gregory Tasian: Exactly.

Diane Newman: So that's really going to help patients, because stones are a real issue. I mean they affect people of all ages, they can't go to work. I mean it really can deter the daily life of someone. So it would really help to actually understand or predict what's going to happen.

Gregory Tasian: That's absolutely right. So stone disease affects about 11% of the population in the US. It used to be a male dominated disease, that is gone now. It's now driven by early onset disease and a disease that occurs in young women and adolescent girls. So when you're thinking about stones, you need to have a broad population that crosses a lifespan. And that includes in machine learning. So you can both have a representative population but also be open to differences in that probability. So what we've learned in this machine learning model is that there are different probabilities for children and adults. So developing two models in which you can stratify based on age will help personalize that approach. It's not simply a one size fits all.

Diane Newman: Another area we're doing research with machine learning is in urodynamics, which is testing, diagnostic testing for bladder function. Talk to us a little bit about that.

Gregory Tasian: It is, so this was a space in which I don't spend my clinical time. So I was able to learn from experts such as Steve Zderic and Ariana Smith and bring in the problem of urodynamics, which is this incredibly complex and commonly performed diagnostic test in which there is a tremendous amount of information. Some of that information is contained in images with fluroscopic images obtained during the study. Other information is the pressure volume relationships throughout the course of the study. And to be able to extract all of those features by any one individual who may not be present at the time of the urodynamic study is a challenge.

So what we sought to do is to be able to extract those features from the urodynamic studies and first classify, are those features representative of a good, bad or ugly bladder? Ultimately what we hope to do is transform that into prediction of at least in children with spina bifida, the probability of chronic kidney disease progression. So that at the point of care during the urodynamic study, you're able to say, this is a bad bladder, we need to intervene early. Or this is a good bladder, this child is very unlikely to develop chronic kidney disease. We can structure our care and our follow up appropriately.

Diane Newman: And we have urodynamics being done differently, no matter where you were in this country. So really to try to pull all that data together would really help this field.

Gregory Tasian: I think that's exactly right, and this gets back to that question of the diversity of data. Urodynamic studies that are done in one institution are often done somewhat a similar way, but there's some differences by provider. When you bring in other institutions, you not only have other different types of machines but different protocols. The rate of fill of the bladder, the type of machine that's used, the environment in which it's conducted, all of those can impact the types of data that are produced. And unless you incorporate all those data and have a gold standard of ground truth for interpreting it, you're not going to be able to ultimately generate these generalizable knowledge that we need.

Diane Newman: So okay, you're doing ureteral stones, we talked about urodynamics. What else are you doing as far as with machine learning?

Gregory Tasian: So one of the things we're doing is creating a urinary tract atlas. So right now stones can happen anywhere from the kidney all the way down to the bladder. Being able to identify where each of those stones are and be able to automate the measurement of all of those stones and all of their features, would greatly facilitate improved clinical care, but it would also advance research. So if you imagine an individual with seven stones, four in one kidney, three in the other, and then they're in a clinical trial. And this is a trial of X medication designed to slow kidney stone disease progression.

End of study, they have another CT, they now have nine stones, but those stones may be in different locations, they're in different sizes. To be able to do that at scale with the current methods of manual measurement is impossible. You have to be able to automate this process, which would both increase the precision and accuracy, but also increase the efficacy and effectiveness of this work. So bringing in machine learning into both the research space and clinical care can ultimately push our boundaries of knowledge further, in a way that wouldn't be possible without that advanced data science.

Diane Newman: So you talk about stones and I guess I'm just really adult centered. You see them in children and adults, right? I mean...

Gregory Tasian: So when I was in medical school and I graduated in 2005, stones were an incredibly rare disease in childhood. It was a disease of white, middle-aged men. And that male to female gap, that ratio was about three men to every one woman. Over the last 20, 25 years, that gap has disappeared and the prevalence of stones, the number of individuals in the US has over doubled. And what is driving that increased prevalence of stones is early onset disease and disease among young women. So that the face of kidney stones is very different now. What I see on a typical day is a young girl, 14 year old adolescent with new onset kidney stones who is otherwise healthy. And the importance of that changing epidemiology is that stones are so much more than something that brings you to the emergency room, that you may need surgery for.

It's really a disorder of mineral metabolism. And when you think about it that way, you can understand the impact that it has on your health. Stones are associated with a higher risk of cardiovascular disease, chronic kidney disease, low bone mineral density resulting in fracture. And when you think of how the disease has changed, you now have early onset disease in which these individuals are going to be living with these greater risks for 30, 40, 60 years. Very different than when disease starts in your forties or fifties. The other problem and challenge is that because stones were rare in women and rare in childhood, the evidence base to support decision making is-

Diane Newman: It's not there.

Gregory Tasian: Very poor. So what we need to do is strengthen that evidence base so that we're effectively able to care for these individuals who are really vulnerable because of that lack of information.

Diane Newman: Well, we have two kidneys. So is it true that one kidney can be stone forming kind of kidney? I always hear that and that that kidney's going to keep developing stones or is that just old?

Gregory Tasian: I would say that is true, but generally it's a disease that affects both kidneys. But I do have patients, even patients with cystinuria, which is a genetic cause of kidney stones, that only form stones in one kidney. I don't know why that happens, but I think that along with everything else just shows how much more that we need to learn and discover.

Diane Newman: And when we look at populations as far as culture, race and that, who's more prevalent then?

Gregory Tasian: So kidney stones were once a disease of white middle-aged men. It is now equal among women and men, largely in the United States. But the groups and the communities in whom stones are increasing at the fastest rate are adolescents, particularly adolescent girls and particularly Black, adolescent females. So we're seeing that these groups, these communities that once were unaffected by stones are now rapidly increasing. And that suggests that there are some differences in our exposome now, everything to which we're exposed in our environment, that are really driving this. And that's both a problem and an opportunity. If we can identify what those drivers are, then we're able to identify potentially new diagnostic tests and importantly new treatments for stones. So if we know what is being perturbed, we can restore that perturbation to hopefully prevent stone progression throughout a lifetime. And a lot of that is in the microbiome.

Diane Newman: And so the issue really is, machine learning is really going to help the whole stone area. I mean that's what you're saying, I mean that to get more and more data across the country is really going to help, really understand how you move forward with stone.

Gregory Tasian: I think machine learning will help stone disease and it'll help all other urologic disciplines that rely on complex data. And I think this is a call to overcome some of the challenges that exist in urology as well as other medical specialties, in which data are siloed. So data are siloed by institution, but they're also siloed by data type. So it is very difficult to extract features from textural data because of concerns for privacy, confidentiality, PHI. And then imaging data has its own complexity. Those data may exist at an institution, but clearly care takes across multiple health systems. So how do you integrate data that are segregated within an institution and may exist across multiple institutions? That's a broad, both policy and informatics challenge. But I think if we can start to move to integrate some of those data sources, we'll be able to generate new knowledge, not only in stone disease and neurogenic bladder, but across the spectrum of urologic disease, both benign and malignant.

Diane Newman: Well thank you very much, this is really informative. And like I said, I know so little about machine learning, so this is very helpful to learn this, and good luck.

Gregory Tasian: Thank you so much.