Breaking Down Machine Learning, Predictive AI, and More
ASN Podcast
Published March 23, 2026 * 00:18:59
Hosts Karandeep Singh, MD, and Navdeep Tangri, MD, explore how machine learning regression and predictive AI impact patient care by examining importance factors, support vector machines, data training, and ways to improve prediction accuracy.
Please note that this transcript was automatically generated and may contain inaccuracies. It is intended for informational purposes only. Refer to the audio for full context.
----Speaker 1: Welcome to Rewired, Navigating AI's Role in Kidney Health. This series will dive into challenges and triumphs as artificial intelligence continues to evolve in kidney care. Each episode will feature leaders in this new intersecting space. Welcome to our podcast series where Nav and I host a conversation on artificial and augmented intelligence and its relevance and application to nephrology. I'm Karandeep Singh. Associate Professor and Chief Health AI Officer at UC San Diego Health and a member of the ASN Partnership for Responsible AI in Kidney Health.
Speaker 2: And I'm Navdeep Tangri, Professor of Medicine at the University of Manitoba in Canada and Chair of the ASN Partnership for Responsible AI in Kidney Health.
Speaker 1: So today, following up on our last conversation, I wanted to talk a little bit about machine learning and predictive AI, which is often that forgotten stepchild that gets left behind because all the focus has shifted to folks' interest in generative AI. I wanted to rewind back to that and start with you, Nav. What is machine learning or predictive AI and really why should we care about it?
Speaker 2: Thanks. Terrific question. First, we should care because it's everywhere. When you turn on Netflix and it tells you, you'll really like this show or you'll really like this movie, that's predictive AI. When you're listening to Spotify and the next song is being picked for you, it's probably based on a predictive AI algorithm. And when you're seeing the ads that pop up on your browser, on your browsing history, there's predictive AI running in the background. So it's really everywhere. And it's often been the basis, in fact, of some of the generative AI work, which we discussed so eloquently on the last podcast. Specifically in nephrology, we should care because predictive AI applications have been around for a long time, and they've actually influenced how we look after our patients. I'm going to throw the KFRA in there, but they're also great models for AKI in CKD, in dialysis and transplant. And we'll cover some of those dialysis and transplant applications and AKI applications in future episodes. But certainly there's been use, real use, development, validation, implementation of these models in nephrology.
Speaker 1: So Navdeep, I think you know this, but actually when your original paper on the kidney failure risk equation came out, that was really the first paper I think I'd ever read on prediction. And actually it was definitely the first paper I'd ever read on the use of Cox regression for prediction, where you have a time to event outcome and you fit a survival model to learn the probability of that outcome happening at different time points in the future. So that was something that I would say really shaped my career, because I still point to the actual paper, the supplement and the Excel calculator that came with it as some of the kind of most forward-thinking tools for kind of education. But what I will say is that when your paper came out and I look at the title of your paper, the title of your paper does not say that you leveraged Cox regression to be able to do this thing. It really focuses on the thing that you were predicting, which is the onset of end-stage kidney disease. But when you see papers come out on machine learning, I would say a disproportionate number of them will say something like, we leveraged machine learning in a way that you would never have said we leverage regression. So tell me a little bit more about machine learning. What's different than regression? And for our listeners, when they read papers that say leverage machine learning, what should they be thinking in their minds?
Speaker 2: No, I think it's a great point. I think first, from my own perspective, regression is a type of machine learning. And if you talk to machine learning people, they'll sometimes use regression. But where I think regression is really great is that it's regression is very explainable. There are these beta coefficients, there's a baseline hazard. I can intuitively sit down with you and I can tell you what's the weight of each individual variable and its relationship to the outcome. With the machine learning models, you start to approximate that. There are these things called Shapley's and there are these things called importance factors or VIMPs, but they are only some approximation of the relationship between the variable and the outcome. It's far from a real, like linear, easy to understand relationships. Now, These models exist, these machine learning models exist. In fact, the subject of this podcast is because they can be incredibly useful. Support vector machines came out earlier on and are thought to be a great classifier, dividing things into two categories. There's random forests and random survival forests, which have been shown to improve outcome prediction. There's the very simplest of machine learning models are the easiest to apply. Probably not the simplest in architecture, but the easiest to apply, XGBoost and CATBoost. And all of these things are forms largely of supervised learning, where like regression, the outcome, you have a set of predictors, you're training on data to help predict. Now, no, In a statistician or an epidemiologist or a machine learning expert, you should probably have a toolbox of things that you apply to make the best prediction, right? And where I think machine learning really shines is when the data is not numbers or binary or ICD codes, when the data is imaging and pathology slides, because regression is just not built for that. You can't do a logistic regression model or a Cox regression model on pixels. because there's just like millions of pixels in a pathology slide or in a radiology emit. So machine learning really and deep learning really shines for that. In other places, I don't know.
Speaker 1: Yeah, and I think you made an important distinction there. We often talk about machine learning and deep learning interchangeably. You'll hear people say, oh, it was machine learning, it was deep learning. Those are two different things. Machine learning is a much broader class, right, that includes things like decision trees, random forests, the different types of boosting, and deep learning. But deep learning is a, like random forest, it's one way to formulate mathematically like that relationship between predictors and outcomes. But I would say the majority of models that I've seen published in clinical journals are actually not looking at images, and they're not looking at pathology slides. They're actually looking at things like age, structured variables, in some cases, billing codes, or kind of comorbidities and conditions. So given that we've got these quote unquote powerful tools that are machine learning tools that are not regression, do those tools always perform better than relying simply on regression?
Speaker 2: I think this is an important thing, not just for us as the audience, but also the editors and the associate editors and the reviewers of journals, right? I think we always should be asking this question. We should be asking, can you convince me? in your own dataset that you're proposing to me? And can you convince me from the existing literature that regression is not the best answer here? Because as we said earlier, regression is explainable, right? It's easily explainable, it's easily implementable. In our JAMA 2011 paper, we built an Excel spreadsheet and we built an online calculator. That was much easier to do with regression. It would've been much harder to do at the time, not much harder now, but at the time, it would've been very difficult to do if we'd used a neural net, for example. So I think practical and empirical evidence, if you look across areas, so kidney disease, non-kidney disease, is that regression often performs similarly well or better. And really the use of a machine learning model like a boosting model or a forest, random forest really depends on the data set size, sample size, the number of predictors. And then this important thing about like, are there non-linear interactions and a ton of non-linear interactions in your data? And it's very hard to know that a priori. Like it's not that you, I mean, theoretically, physiology is very complex. Almost everything we think should be nonlinear. But if you think everything is nonlinear, then you'll probably overfit everything. And maybe you just learn the nuances of your data set and you don't actually generalize. So to say a lot of different things, what I'd want to say is if the data underlying data is very complex and there's a lot of nonlinearity, perhaps the machine learning model is better, but you cannot start with that as your null hypothesis. Your null hypothesis should be that regression is going to be equally good and I'm going to show you why this is better or that's better.
Speaker 1: Yeah, and building on that, there is empirical evidence and evidence from simulation that shows that Modern machine learning models are data hungry. They really need orders of magnitude, more training data. And so, you know, when you say I've got thousands of predictors, that to me is a huge red flag, unless you say we've got millions of samples. So I think just recognize that, you know, just because you can train a model, when there's thousands of predictors and 100 samples doesn't mean that model's any good. And in fact, there are actually sample size calculators out there for prediction. And it's really counterintuitive. When you think of sample size in the context of other epidemiology work, you think about things like effect size. How much effect am I going to find for someone getting this treatment versus not getting it? That's completely different for this space of prediction, where what you're actually thinking is, How solvable is this problem? What is the level of predictive performance I expect to get based on what else is published? And that actually is a key driver of how much sample you need. If you're trying to differentiate something where it's, you know, is someone going to die? That's usually a pretty extreme prediction, and you often don't need as many variables to do that, and you can often get better predictions because it's an extreme outcome. But if you're predicting something more upstream, like development of CKD stage 2, stage 3, that is a much more subtle thing. And that actually usually requires A substantially larger sample size. So I would just say that, you know, I agree with you 100%. Your null hypothesis should largely be that you can get away with regression. And you really need a good justification to, I think, veer off and away from that. And the other thing I'll just say is a little piece about non-linearity. Regression models can model non-linear data, period. I think the linearity we talk about is in a specific dimension or a specific space. And because we often transform variables, either the outcome or the predictors in a variety of different ways, it turns out even if your data when you plot it is nonlinear, there's usually a regression model way to model it. So this is not to say that you should always use regression or only use regression, but just recognize that machine learning is not a free pass. And that if you're going to use machine learning, don't say leverage machine learning. focus on what was the type of machine learning. A decision tree is machine learning, an XGBoost model is machine learning. That doesn't mean they're the same. They're actually quite different in terms of their capability and their strengths and weaknesses. So with that, you know, we alluded to the fact that there are just so many papers coming out focused on development and validation of prediction models to the point where I think clinical journals are really starting to look at those with skepticism and say, okay, you need to move some actual clinical outcome, not simply do development and prediction. However, it's still really important to develop and validate prediction models. So when you read papers that do that development and validation nav, like what do you look for?
Speaker 2: So we often have these conversations, like either when reviewing for journals or with trainees. And I think in some ways, the machine learning models do get a free pass on first glance. People don't think about those same stats and epi principles sometimes with the machine learning model that they would think with the regression model. And I would challenge that we absolutely should. And there's been incredible papers. I think you highlighted the sample size, some of Richard Riley's work. There's tripod reporting guidelines and tripod AI extension. I think Gary Collins to call out another name has really done sort of, they've done great work showing that we should be critical, just as critical of these models, just as critical of study design. I've even seen simple things like leakage of outcome in the predictor data. That's like a model that predicts sepsis, but having receipt of antibiotics in the model. Like you've already made that decision. If you fix you're predicting sepsis, you shouldn't have receipt of antibiotics as in a model trying to predict sepsis. So pay be just as critical. External validation, really, really important because there's a, if you don't have, the models are data hungry, like you said. So if you don't have enough data, you're likely to overfit and external validation becomes even more important. And then there's performance drifts over time. Like, you know, medicine's evolving. Look at kidney disease. 2015 to 2025, we've had a complete change in treatment. Is the same model in IGAN, let's say, which is the most rapidly evolving part of nephrology, right? Like is the same model that predicted outcomes that was well calibrated in IGAN in 2015, is that going to be relevant in 2030, right? Because there's going to be like 20 different therapies that are available and that have a meaningful impact on progression. And finally, I think, All models and cardiology, our cardiology friends have like 10X the models that we have, but very few implemented, right? Probably just as many implemented in cardiology as in nephrology, but with 10X the available models. So it's because most models fail at integration.
Speaker 1: So tell me, Nav, when you talk about integration or implementing AI models, why don't most models get there? What are some of those challenges in your experience?
Speaker 2: So I think the very first challenge is we need to reframe our thinking when we're building a prediction model. So when I look at all the listeners today and the people who are out there working on prediction models, I would challenge them first to say that unless you're willing to do the post-publication hard work of translating this model into clinical workflow, don't even develop it. Now that seems harsh, but really that should be the goal, right? I think both of us would tell our trainees that if you only plan to present a poster, don't worth, not worth to work it on a project. Make sure you carry it forward to a paper. For models, it's really one step forward ahead than that. And the table stakes are just so much higher now, right? Like it used to be when the KFRE was done that internal and external validation was really great. You had really surpassed the mark if you had done an independent external validation. So the next step was to put in a medical calculator, a spreadsheet or online calculator. And that was really new. That was the thing you had to do. But now that's normal. That's like all expected. That's the, that's. That's the price of entry. So what you need to do today, I think, to get your model integrated is you have to have a strategy on how that model is going to integrate into an electronic health record or lab information system or some other part of clinical workflow, right?
Speaker 1: So my take on it is, I think, I agree with that. I think that research scientists have historically viewed themselves as outsiders. That's why the language of internal and external validation exists. It's like I internally validated it here, I externally validated it there. Whereas I think what research scientists need to do is become learning health scientists, where there's not a one group of people who builds a model, validates it, and then hands it over the fence to someone on the receiving end who implements it. You really need the person building and validating to actually be part of the implementation side of that. So I think that, what learning health systems science or learning health science is about is really about you can't just train the model and build the model and evaluate it. You actually need to carry that to integration and implementation. Because if you don't, what have you really achieved? What you've achieved is something that's maybe generalizable, maybe others can look at it, but simply putting a model out there which actually many papers don't do. Many papers describe a model and never actually put the model out there, especially if it's a machine learning model where the way you share it is not a table, but actually is a file that's a supplement or a open source repository. So I think that actually the focus needs to shift towards how do we do better learning health system science where the first question we ask when we build that model is what clinical or patient or experience problem are we trying to solve, work backwards from there such that when that model is quote unquote delivered, there's actually now a burning need to actually get it implemented so that you can do some further validation and then use it to drive actual clinical care, which hopefully produces the better outcomes that we're looking for. So Nath, thank you.
Speaker 2: I don't think I could have said that any better.
Speaker 1: So Nath, thank you so much for joining today. Just as a reminder, ASN members can join us for regular updates and discussion in the ASN Community, AI-Powered Kidney Care Network. To join, log into the ASN website, select ASN Communities in the top navigation, and search for AI-Powered Kidney Care Network. We look forward to seeing you there. Thank you.
Speaker 3: Thank you for listening to this podcast by the American Society of Nephrology. I'm Todd Ibrahim, Executive Vice President of ASN. If you're a nephrologist, fellow, or kidney health professional looking to advance your career, stay informed, and build connections, I encourage you to become an ASN member. Membership, which is free to fellows and other trainees, offers exclusive access to groundbreaking research through our journals, discounts to live and virtual educational programs, including Kidney Week, and support of a global network of nearly 22,000 kidney professionals. To learn more or join ASN, visit www.asn-online.org. This podcast is copyrighted by the American Society of Nephrology, all rights reserved. All content in this podcast is for informational purposes only and is not intended to be medical advice. This podcast should not be used in a medical emergency or for the diagnosis or treatment of any medical condition. Please consult your doctor or another qualified healthcare professional. if you have any questions about any medical conditions or before taking any drug, changing your diet, or commencing or discontinuing any course of treatment. Thank you for listening to this podcast by the American Society of Nephrology.
----
Episode URL: https://www.asn-online.org/media/podcast.aspx?ID=644