Hi, I'm Eric Vrstavsky. Welcome to another segment of EPNIP. With me today is Tina Bachner, who is the Assistant Professor of Medicine at Stanford University. Tina, welcome to the show. Thank you, Dr. P. We've known each other for a long time. Ten years, I think. Ten years. We won't go into that story now, right? So you've done some really interesting work on AI and atrial fibrillation. So what I'd like you to go through with us in the audience is, can you walk us through this new interesting area of artificial intelligence and this whole concept of machine learning, how you are looking at it to apply the ablation? Gladly, yeah. I see it as an excellent method that we have underutilized in electrophysiology and now it's getting more and more popular as we knew it would. I see it as a method that can analyze complex things that we weren't able to analyze. It can analyze the signals we can't understand. It can analyze the images we can't understand. So I've utilized it as a method and it is now being used to analyze ECGs, complex signals that we otherwise couldn't do with traditional signal processing methods. So let's go into a little more detail. That's a nice overview. I've read your paper on using this in the lab. Why don't you walk us through what you actually did to try to, I mean, you know, the classic is you go isolate the veins and you go home. But that's really not enough for many people, right? So you're going to apply your methodology to an ablation. So can you walk us through what that means actually? What do you do in your ablation? I think the background is imagine a world that you're the patient with AFib and I have the luxury to tell you before even considering an ablation that what might be the best strategy for you in putting all your information, your biometric data, your imaging studies, your ECG. In an ideal world, I would love to tell you that with 99% certainty, AFib ablation is not going to work for you, but flaconide might be the best drug for you. So that's what a machine learning model can teach us. It can get all sorts of data from different patients, learn which patients did well with what strategy and predict what other similar patients can do either well or not so with similar technologies. So that's a goal, right? That's a goal. So the stuff that I read from your work that you've already done, can you walk us through what you actually did as like first steps? So with this ultimate goal, what we said is why don't we look at all the patients that had ablation? Let's start with ablation as a treatment strategy. They all, at least in our institutions, all had CT scans prior to ablation, so we said why don't we see if a CT scan along with their demographic findings, ejection fraction, all the lab data that are available, et cetera, for a baseline workup that you would normally do for a patient is good enough to predict if someone will do well or not. Can I stop you for a second? When you say do well, is the ablation approach PVI only in your study? Complex. They're by provider's choice, which unfortunately clouds the study a little bit. Yeah, I read that in your paper. I was trying to figure out... About a third of the patients had PVI only. Some had PVI in line. Some had PVI in other targeted areas. Maybe let's take that. I don't know if you were able to. Have you been able to say, well, let's look at the PVI group, because that's kind of the standard, and have you found that looking at the AI that you've been able to retrospectively figure out who would be the winner? We've done the subgroup analysis, but the numbers get smaller and smaller, so my trust in the model's decision, because now in the PVI group, only a few wouldn't do well, for example, and a few would do well, so you're training on a much smaller cohort. But when we applied the final model to only PVI patients or only PVI plus patients, it performed similarly. But you're right. I think it should even start with a cleaner cohort and cleaner strategy. So I think you started with actually a goal, right? And I want to come back to that. It's just like the fibrosis stuff that Nasir and the Utah group were doing, trying to figure out who is the best candidate to be in a lab. I mean, listen, I think their work is nice, but I mean, it's not 100%, right? You're going to look at multiple things, but that is your goal eventually, right? To sort of say, let me look at Eric, hope not, a different Eric, and I know I can get a great result in the lab. Now, that result is going to depend on what you do in the lab, too. So that's why I'm trying to dig into that a little bit. So if you have a longstanding persistent, you've been in the lab for years, I know not more than 10, but you've been in the lab for years, is your goal to look at different phenotypes of AFib, like a longstanding persistent versus a paroxysmal? Because I'm just listening to, wouldn't your model be different for the different types? We've applied it to persistent group and paroxysmal-only group, and it's performed similarly. Oh, it has. So it didn't matter either one. It didn't matter. That's interesting. But, you know, I do understand Nasir's findings, too. If you have too much scar, maybe it's not worth taking you to the lab, but, you know, it's not just scar. Some people without scar just do poorly, too. So I think this is just a more advanced way that potentially incorporates scar or the left atrial geometry into account and some other things to predict why some patients just don't do well. Right. So I'm going to keep pushing you a little bit, because I love the work you're doing and I've enjoyed reading your papers. As a clinician, seeing this patient in my office, you have to make a decision based on symptoms, right? So I think they need an ablation, I hope they do okay. And then you want to apply a model before they're in the lab, right? So if that model is different for persistent or paroxysmal or other factors, I mean, I guess that's what I'm trying to push you to, is your ultimate goal to be able to take an individual and put their phenotype of AFib together with all the other factors before you get in the lab? Because once you're in the lab, you're already in the middle of an ablation, right? I think both are valuable, because if you're in the lab and you know a patient is more likely to have recurrence, you might want to adjust your post-ablation strategy, right? You might want to keep them on entire arrhythmics a little bit longer than you... I think it's still valuable. But you're right. In an ideal world, if we know an ablation won't work for someone, we want to catch them before. So that CT study is, in a way, all pre-ablation. I have a different study that incorporated signals from the procedure, intracardiac electrograms in addition to the surface ECG. I can argue with you that the contribution of the intracardiac electrograms for that predictive model was not any higher than the 12-lead ECG at baseline in sinus rhythm for those who had a sinus ECG. So all of these are available before the ablation. Yeah, you wouldn't have to argue much with me on that, and I'll tell you why. Because I've been fascinated by the AI work done on just a regular ECG and its predictability. I wish I knew why. I mean, I know you can't really just eyeball it and tell it why, but I have great faith that we're going to learn a lot more with this technology. But I'm going to challenge you to come up with something before. Or I'll ask it the other way. You get into the lab, and your standard approach is a PVI. A lot of people feel that way, right? That's your first approach, whether it's persistent or paroxysmal. Based on your approach now with machine learning and AI, you get in there and you add the intracardiac to that, you might change your approach. Is that a fair statement? Exactly. That's a very fair statement. And those, now with the data there is, can be done, right? We can get the ablation locations, we can get their geometry, we can integrate all those into the methods, which was not a feasible thing just a couple of years ago. And how do you scale this for the rest of the world? It's one thing if you guys are doing it, but the hope is you'll scale it, right? So you'll have something that we could use. What is that goal with AI? Are you going to need high processors and things, or is this that you're going to get to a point where anyone in a good solid EP will be able to apply your methodology? I think one problem is, how do you train these models? If you're training these models only for patients in Palo Alto, that may not apply to the patients in France or somewhere else. So I think the model should be more global, which is going to get better with more... Is there something special about Palo Alto? It just might represent a different ethnic, gender of patients. So I think that the model should be trained one broader to test it elsewhere to make sure they're working. Well, I think to work... But I think once the model is there, you don't have to be at a fancy computational center to upload your data for it to spit out, hey, yes, likely, or no, not likely. So give us a timeframe based on where you're at now. When do you think this will be a practical application for us? Oh, gosh. I hope people watching this video in five years won't judge me for that, but I'd say five years. Five years? Yeah. Okay. Well, that's a fair estimate. Tina, it was a great interview on this topic. This is exciting work, and I love your papers. Keep them coming, and appreciate all you're doing for the field. Thanks, Dr. Pistovsky. Thanks.

EPNIP

Dr. Eric Vrstavsky

Dr. Tina Bachner

AI

atrial fibrillation