♪ ♪ This is Eric Kristofsky and welcome to another segment of EP on EP. This is a very exciting one and it's an area that's really on the cutting edge and I'm delighted to have Dr. Peter Nomsworthy with us today. He's a professor of medicine at the Mayo Clinic and he's also a well-known electrophysiologist there and elsewhere. And I think director of your electropardiography laboratory. Welcome Peter. Thank you so much. Great to be here. It's good to see you. Thanks Peter. So this is, this is the area we're going to talk about. It's going to be on artificial intelligence and electropardiography. And Peter has done several studies in multiple areas and Peter, I'd like to start with maybe his atrial fibrillation and then we'll get into some of the other areas. When I was reading through your papers, several of them, I realized that I didn't know a lot about this area, no surprise to me. And I looked at some of the terms that we use, this convoluted neural network. Maybe you could start and tell the audience just what is the process? How do you, how do you apply artificial intelligence? Yeah, great. I'd love to. First of all, I'll give you the caveat that I'm an electrophysiologist, as you know, and a cardiologist, and I am also relatively new to this field. So we work with a very large team of engineers, computer scientists, and so forth to get this work done. It's a collaborative effort. I'm speaking on their behalf, but I want to make sure that everybody realizes that this is a team effort at Mayo. The technology is similar to the way computers recognize faces and things like that. And you've seen that there's been a huge number of papers about the application of these technologies to the ECG. And the reason I think that we're seeing that is that the ECG is one of these tests that is very reproducible. We always obtain the ECG in the same position using very similar technology with the leads in the same position. And ECGs look more or less alike. So that physiologic signal is relatively easy for the computer to discern. Essentially, what we do is we feed in a raw signal, meaning the matrix of each of the points in a time series data set to the computer, and then it learns various attributes of that signal and then learns to weight them in a way that it can discern a pattern that we're teaching it to learn. Basically, what we'll do is we'll give it a set of ECGs that have either the presence of a condition of interest or a control, and it will learn to make a distinction between those two. We then reserve a group of data that are untouched and unseen by the network, and we use that to test the network. And we can take that process iteratively to refine it. So let's talk first about the atrial fib. I read your recent paper on that, and I know that it's a follow-up in a sense. Maybe you could talk about both studies. One was an earlier study that I believe was published in Rancid that looked at looking at concomitant atrial fib, and the next was incident atrial fib. And one of the questions I wondered as I went through it, there must be changes of aging that are picked up by this process that maybe we can't see. So would something, if you fed in 30 and 40-year-olds, would that be the same as feeding in 60 and 70-year-olds? Yeah, that's a great question. We try to feed the computer a very diverse and representative sample that will represent the entire breadth of ECGs that it's likely to encounter in practice. Just as an aside, when we developed the first model, which was for low ejection fraction, it performed exceptionally well, and we were interested to see if perhaps feeding it in that additional information about age or sex, whether that would improve the model. And to our surprise, actually, it didn't improve the model at all, and the clinicians in the group were looking around thinking, how could that be possible? Something as fundamental as age and sex don't predict ejection fraction. But the computer scientists in the group told us, well, the reason is the model is probably already able to discern age and sex from the ECG, which is sort of a new concept for us. And after that, we developed models actually that can tell with a remarkable degree of certainty the sex of somebody based on their ECG and even their age. And of course, we're interested about, is this telling us something about physiologic age? Is it telling us something about physiologic sex? And what kind of medical applications that might have? We don't need an ECG to tell if somebody's male or female or how old they are, but it might be telling us something much more informative. So that's something that we've been very interested in, actually. So I think it's fascinating, actually, but we're not going to go there today. So let's talk about then, let's talk about, the thing I'm most interested in is the predictability. I know you did the initial one where you looked at concomitant, but the ability to do an ECG, like you said, a relatively simple test and have a reasonable predictability of who might develop AFib is very important for many reasons as you enumerate in the paper. Walk us a little bit through some of the curves in the paper. I know I don't have that to show today, but one of the things I was struck with was that you could sort of change the bar a little bit, it's like moving something up and down to let you know your sensitivity and specificity. It's not like it's an all or none phenomenon. So if I had a group of people, let me just ask you, let's say people who had a Chad Bass score that was relatively high and I'm worried that they might develop AFib and I want to look closer. I mean, how do I play with that in your system? Right. So when we derive the model, we were looking at concomitant AFib, as you've said, and then essentially the model output is a probability between zero and one. And we pick the probability there that best optimizes sensitivity and specificity for that application. But it's an artificial line in the sand, essentially. If we then take that model and try to apply it to a new population or a new application, for instance, incident atrial fibrillation, as you outlined, we may find that that line that was good for concomitant atrial fibrillation is irrelevant for incident AFib, but we need to make it either more sensitive or more specific. And we do delineate that a little bit in the paper. The other thing is that sometimes we take a model that performs very well for one application and it doesn't perform well for a related or adjacent application. So we've looked at the prediction of postoperative AFib with our model and disappointingly, it didn't work as well, or recurrence of AFib after ablation. So that's a situation where maybe we can't just adjust the cut point, but we actually have to do some retuning and refinement of the model for a new application. So just because we have a biomarker for AFib risk doesn't mean it can be applied indiscriminately for all applications. We think a model like this that is good for concomitant AFib may be particularly good for patients, for instance, with cryptogenic stroke, who, if they've already had a stroke, the hypothesis is if they have AFib, they've already had it. So these people may be very valuable in terms of determining which of those patients might warrant anticoagulation or at least more intensive monitoring for atrial fibrillation. Again, fascinating thoughts. Before we move to a different rhythm, how are you using it now? I know the Mayo Clinic is, like you told me, it's actually in your EPIC system. So how are clinicians applying this? What are you finding out? Well, the uptake has been remarkable, actually. So we created what we call our AI dashboard or ECG AI dashboard, and it launches from within EPIC. So if you're in an EPIC chart, there's a tab under the results tab called AI dashboard. If you click on it, it will pull up all of the patient's ECGs. And in real time, it will run all of those AI models. And right now, we have models for low ejection fraction, hypertrophic cardiomyopathy, atrial fibrillation, and we're about to add aortic stenosis, perhaps cirrhosis, perhaps hyperkalemia, and a number of others that we've published on. And then clinicians have that information at their disposal. And we've seen some clinics use this quite vigorously. It's now been used by over 5,000 clinicians at Mayo Clinic. So we've actually been very surprised at the uptake of this technology internally. And of course, we're using it mostly for research and for interest, and we're trying to learn as we go. But it's been fun to try to put this at the fingertips of our clinicians, and it's been fruitful. One of the caveats, I'll just throw this out, because you would never say this about your fellow clinicians. It's like to a child, with a hammer, every object's a nail. So one of my worries would be somebody says, oh, my God, there's this, whatever, 62% chance of A-fib in the next two years, I better start anticoagulants. Yes, right. So got to be a little careful, right? Yeah, you have to be careful. So of course, we have disclaimers, and we've done a lot of grand rounds in internal communications about how to use this, and we try not to make major clinical decisions, but it's part of the picture, I think, about how we can approach patient care. You know, I'm often stopped in the hallway when somebody says, I looked up my ECG, and it says I'm five years younger than I am, I love your system. But nobody tells me when they're 10 years older than it says they are. You know, there's a bit of a reporter bias in terms of the early adopters. Let me switch disease states. I know you've done some really fascinating work in low ejection fractions, and what I wanted to ask was two things. Does it predict who's going to develop a low ejection fraction or recognize when someone has a low ejection fraction? Yeah, great question. That's very similar to the question we just discussed about the atrial fibrillation. So it was designed to pick up undiagnosed low ejection fraction at the time of the ECG. We then, in the initial study, identified patients who apparently had a false positive of the AI, but because these patients were, it was a retrospective cohort, we could look to see which of those developed low ejection fraction. And actually those patients with an apparent false positive AI flag were three or four times more likely to develop low ejection fraction over time. And in fact, did have low EF over time. So, you know, maybe we're actually detecting a latent signal of impending low ejection fraction, or maybe there are changes that are electrophysiologic that predate, that precede the pump function of the heart. So all kinds of interesting hypothesis generating observations in these kinds of data. The fascinating thing to me is, you're now, you and others are doing this work, are taking this little old ECG, and you are making it the king again. You know, so Pick and Langendorff and all those folks, you know, you know, Sir Thomas Lewis, et cetera, et cetera, would be very proud of all you. It seems like the ECG is regaining its stature. Yes. Well, I think, I actually feel the same way. I think we're breathing some new life into the ECG. And although I know you and others have a very subtle interpretation of the ECG and probably never call an ECG normal, the vast majority of ECGs are either called normal or abnormal, but really, they're much more richly informative. And there are things that we can do as expert ECG readers, but there are also things that the computers and new insights that we can find there. And there's a lot of patterns that are essentially hidden in plain sight. And I think it's very exciting. So what I'm hoping is, and I know this is your hope too, I'm sure, someday you'll figure out what it is that it's looking at, because every time I ask you or any of the other people in this field doing the work, I say, okay, so teach me something. What have you learned? And the answer is usually, you got to put it through the computer. Well, that's actually a very interesting question. What can we learn about physiology from these sorts of networks? And that's maybe a topic for another discussion, but one of our, the lead data scientist has actually been working a lot on that and figuring out ways to sort of peer into the black box, and it's been fascinating. And there are a number of techniques that we can use. You can blind the computer to various parts of the ECG and see what it's sensitive to. You could do invasive experiments where you perturb one parameter or another and see what it is sensitive to. You can do these, you can ask the computer to create essentially a caricature of the archetype of one phenotype or another and allow it to teach us what it's looking at. And there are very, very cool insights that are coming from this. But that may be the next wave of learnings that we're going to get from these kinds of technologies. That sounds like a great part to one of these days. Thank you so much for enlightening the audience on a really interesting topic that's going to be clinically, I think, clinically applicable for us, hopefully in the near future, and stay healthy. Likewise. Pleasure. Thank you.

artificial intelligence

electrophysiology

ECGs

low ejection fraction

clinicians