I welcome all of you to the Volta Medical Rhythm Theater. I guess you're all confused as to whether or not you're in the right place. Well, I think it's very interesting and very exciting. In the high-impact science session, we actually just heard some new data that was presented, the RESTART trial. We're not gonna see those data here because these slides were prepared before the trial, before the data was presented, but we will also talk about that. So I think there's a lot of fresh data to talk about. So I'm just gonna do a quick introduction, but then we'll have presentations from three experts in the area. So what are we talking about with the Volta Medical Technology? This is a machine-learning-driven innovation for AF ablation procedures. These are my own disclosures. This technology is FDA approved, though there are aspects to it that are not yet FDA approved, and we'll talk about those when we get to them. So this is a technology that's really designed for EPs. I think it's fair to say this is the first machine-learning company that is really focused on EP, and particularly atrial fibrillation. And there's multiple aspects to this technology, the AF Explorer technology. And the idea is to take in data, whether it's electrogram signals, the database of clinical outcomes, et cetera, and try to predict and identify exactly the right locations to ablate. So the fundamental concept behind this technology is the concept of spatial temporal dispersion. Specifically, the idea is if you take a multi-electrode catheter and sweep it along the chamber, so you're looking in a regional fashion, are you gonna see something that is potentially interesting, potentially reminiscent of, let's say, rotational activity is what you see on the left, where you can almost imagine in this localized region, you have potentially the whole AF cycle length, or are you gonna see something that looks somewhat passive, as you see on the right, where the activation is just proceeding along the splines. So this is the basic concept. It was a concept that Julian Seitz and Clement Baars came up with years ago. There are multiple publications. I would argue that it sort of culminated in this nature of medicine manuscript that was published by Isabel Dysenhofer and her colleagues that I think you're all aware of. This is the large randomized trial that demonstrated that you have improved outcomes in precision AF patients. This is the technology itself. You use a multi-electrode catheter here. You can see it's a grid-type catheter. As it's swept along within the chamber, you get sort of immediate information from the catheter. It analyzes signals, and then you'll get an idea as you see those orange dots light up, and you'll see some of the electrodes lighting up, suggested that, yes, there is something very interesting happening at this location. So with that very brief introduction, let's have our presentations. First, Dr. Tina Baikner is gonna present, and then we'll hear from other colleagues. And at the very end, we'll have a discussion. So, Tina? Hi everyone, my task today is to take a step back and talk to you about what AI is, how does VOLTA use AI, how is that technology applicable to the VOLTA system, what their software platform includes and how their Here are my disclosures again. So the terminology, just brief terminology. AI is a term that is any machine capable of performing tasks that typically requires our human intelligence. Machine learning is considered usually as a sub-category of AI that involves algorithms, models that can learn from data without explicit programming. And up top you see an example of the more earlier versions of machine learning. There is an input, but there's a lot of manual feature extraction to put into the model to then learn a label without really teaching the model, without really programming the model how to learn that label. And deep learning came about in the last 10 to 15 years. That is a subfield of ML that are more complex models that require much less feature extraction, feature engineering by humans before they get input into the model. So there machine learning has been used in literature, in research for many years. This is any article, any manuscript that uses the terms applicable to machine learning since 2006. But you see a big uptake since the introduction of deep models to the machine learning world in 2016. And why do we need AI and AF mapping? One, there's increasing amount of ablation data. Here is the the annual increase in ablation volume. The left is from Finland, the right is from NYU. And this does not even include the increase in the last five to ten years of pulse feed ablation and how fast we are now doing AF ablations. There's a clinical need, right? We are not happy in how we are doing AF ablation these days. Paroxysmal patients, persistent patients, with the best, with the newest, latest technology, there's still a lot of recurrence that we are dealing with with any cohort we are dealing with. And lastly, electrograms are tough to work with, right? In AFib, we are dealing with electrograms that look like this and our usual computational tools can gather certain metrics from them. We can gather cycle length. We can gather fancy irregularity indices, fractionation indices, voltage amplitude. But we use these in every clinical trial. None of them really panned out to tell us what else to ablate, how else to help these patients do better. And we have been interested in these complex looking signals for many years. This is the early papers from Naramani from 2004. But there have been positive studies and negative studies since then, many of them, about what happens if we just target complex looking fractionated electrograms until 2017 when the concept of dispersion in addition to fractionation was introduced. And there's some mechanical, there's some mechanistic background to dispersion. These are papers now from 20 years ago from Jerome Khalifa that I see sitting there, as well as the Bordeaux group. Sorry, I'll pull your paper here, Jerome. That if you put electrodes around a stable site in atrial fibrillation, you're likely to see fractionated electrograms. And again, these are very, very early observations. And this is again what Vivek showed a little bit earlier in a video form. If you put electrodes in a rotational stable site in atrial fibrillation, you're going to get signals with dispersion that are going to look like this, that are going to go across the whole cycle length. But if you chose to put your electrodes at the periphery, your signals will be the same time front, at the same timeline, and will not have dispersion. And that is the background, the thought of Volta. So how do neural networks work? It trains, if there's a model that has known labels, such as, these are signals of interest, these are dispersion signals, or you can, I'm going to claim them boring signals. You have a model, and you're going to, you're trying to teach the model certain labels. You give them signals with dispersion, and you teach the model, learn that this, these are signals with dispersion, and, or you give them something that does not look like dispersion, and you teach the model, okay, these are boring signals. And how do you test the model's performance? You give it signals that it hasn't seen before ever. You give it many examples of signals with dispersion, you give it many examples of signal without dispersion, and you test, you figure out its accuracy, precision, and recall. How is it, how is it performing and actually evaluating these signals that it wasn't trained on previously? And that's how all machine learning models work in general. And that's what this group did. They got a model, they gave it, they learned, they taught the model with the labels of dispersion and no dispersion. And this was one of the earlier papers to see what if we do ablation based on manual annotation of dispersion, right? Or in another center, we just do the ablation based on an AI model that is tagging the dispersion sites. And this was the early results from that study. So how does multiplex data collection and annotation ecosystem work? In other words, how do we get those signals that we teach the models? There's a recorder, it's a tablet based software that's in your EP labs, and as your case is going, as you're acquiring the signals, you have the luxury to annotate. Okay, these are dispersion signals. This is when I had AF termination. That's what the recorder does and it helps you. And then there's an uploader, a different software environment that not just only uploads the signals, your annotations, but also the 3D mapping system data. And of course, this is if your center is collaborating with the larger group to help them teach the Volta model. The visualizer, on the other hand, is what you have at the very end. It's a post-procedure software platform where you have all the data from the 3D maps, from the signals, from the annotations. But now you can sit down and have 10 other people, 10 other experts, annotate those signals as dispersion or not. Where were they? Where are the termination sites? How do we visualize them? How do we re-annotate these signals then to teach the model? And then there's, of course, the machine learning model. That's the inside the box. It has clinical features that are derived from the signal. So not just the raw signal, but the things that you can gather from the signal, such as a cycle length, activation, fractionation, voltage, etc. As well as the raw signal that goes into the machine learning, into a classifier, to then learn if that signal was a signal of interest or a boring signal. And these are a little bit more details about the machine learning platform that you have. It uses a rocket, which is short for random convolutional kernel transform. It's actually a relatively simple deep learning. Some would argue it's not even a deep learning. It's a one-dimensional convolutional neural networks inside it. And it's hugely advantageous in using, in processing large amounts of data, large amounts of time, voltage series data, without much back propagation. So it can learn fast. It doesn't take up that much computational effort and space. And then the final classifier is a very simple, in a way, decision tree model that then takes the features that the deep network spits out, in a way, and learns if it's a good signal, if it's a boring signal, or if it's a signal of dispersion. So the model, original model, was trained with over half a million annotated samples by experts. The data was collected internationally, all five channels, because you're trying to teach the model dispersion. So it's not one signal going into the model. All five channels, lasting one and a half seconds, plus some other things that you gather from those, other features that you gather from those signals, are going into the models. And again, the things that I highlighted about the advantages of the machine learning model, they work. And it's interesting because it's fast. It scales well to large data sets, which is what Walter is working with. And it works extremely well with the light GBM model that they use eventually to learn the transformed features. So this is the performance of the model. On the left, you see the area under the ROC curve of the model, that its performance is 0.94, which is considered very, very good, very, very high. And on the right is a more applicable one. On the x-axis, you see the number of cardiologists agreeing that an electrogram is dispersed, right? If five electrophysiologists are agreeing that the electrograms are dispersed, you see the probability that the model is showing on the right three. So if there's more agreement between the clinicians, the model seems to definitely show those sites as sites of dispersion. So this is the final platform. There is live annotation during the cases. There's an uploader that can then upload these cases, not just the signals, but the 3D and atomic maps. All your annotations about the signals. And then other experts can continue to re-annotate those signals and re-analyze what is more important for us. Is it just the signals? Can we tag termination sites? Can we tag, can we highlight patients who have better outcomes? And then you keep training the model, evaluating the model, and update the model every year with hopefully better outcomes. That's what the system is. Thank you. All right, well, you know, thank you so much. You know, it's great to follow that presentation because it's sort of like a look under the hood of how the system works. And she, as you can see, always gives such a clear presentation on helping us understand how AI really works. And I'm gonna really bring it more practically now from a clinical perspective. AI-guided patient selection considerations. And the way I'm gonna frame this is a little bit of discussion about that idea. You know, where is this applicable? Which kind of patients are maybe optimally suited for using a system like this? And then go through a case example that I think sort of highlights that. Here are my disclosures. So we'll start with the case. So I have a case that I recorded actually a couple years ago. 68 year old guy with hypertension, otherwise in good health. First diagnosed with PAF eight years ago. Eventually found to have flutter, underwent a CTI ablation about five years ago somewhere else. And then by the time I met him, he'd been in 18 months of persistent atrial fibrillation, highly symptomatic, failed medications, and cardioversion with ERAF. ECHO was normal LVEF, but severe left atrial enlargement. So I think you would, I think everyone would agree, this is someone we'd consider long-standing persistent AF for first-time ablation. And the idea is, of course, that right now, as Tina also discussed, our current approach to catheter ablation, you know, there have been improvement tweaks. We've all, you know, seen a lot of improvement in outcomes, but the overall concept and strategy is the same. It's an anatomic approach with PVI and routine additional ablations beyond PVI is really not indicated. We've thrown stuff out there to look at various approaches, and there have been hit or misses and mostly misses. And so we're really in a need for improving the procedure. And this is partly why. You can sort of break it down on the left as maybe thinking about how aphid substrate progresses over time. And on the right, then sort of correlating that a little bit, so to speak, with extent of ablation. And at some point, you're going to, you know, be ablating things anatomically, maybe not have a good understanding of what mechanisms are, and your outcomes, as has been shown in many studies, has been inadequate. And this is sort of why. So if you look, if you go from your typical paroxysmal patient where you've done PVI, maybe you have a decent result with that. But then as the substrate progresses, you might do some additional anatomic ablation, like a posterior wall isolation. But then what if you have these really extensive types of lesion sets that you're creating in a very large atrium? So what's the right approach? And that's why I think this is, we're all so excited about this. And everyone cites this, but this is from STAR-AF2. When you at least look at a typical study that's evaluated various anatomic approaches compared to PVI, results have been less than exciting. So you already saw something akin to this slide, but the idea then is if you have a system that can identify potential, not only to identify potential targets that have been shown in studies like what Julian and colleagues have done to look at spatial temporal dispersion, but then to have that augmented machine learning process that then really helps you further refine what you're targeting for ablation. That is a really nice combination. And as shown by one of Vivek's slides, and if you were at the session today with Dr. Hummel, you'll see that the proofs in the pudding, that our data or the data have been quite, I think, promising. So you've seen something like this as well, but this way the system works is that you annotate your signals and you then decide how to do your ablation. And I'll go through a little bit of that practical stuff in the procedure, but then also focusing on how to do the ablation. So, and that's based on the ideas I mentioned a minute ago, that we have data from really a first-time ablation population and persistent AF through the tailored AF data from last year, and then just presented the restart data, which is really looking at repeat procedures. So if you take that context, so how do I think about what patients to select for this? So when I was thinking about this, I came up with this sort of like a Venn diagram approach of thinking about this, and I'll get to that in a second, but you can think of potentially some of the patient characteristics that might influence your decision. Are they in PAF or persistent? Is it a first or repeat procedure? Do you have some measure of substrate characteristics like LA size, fibrosis measures, etc.? Comorbidities. And then there may be some practical procedural considerations. Is there a compatibility with the mapping system you use? What kind of energy for ablation are you thinking about using? Are you someone that is willing to anticipate perhaps a longer procedure or not? And then, of course, being willing to do ablation beyond, say, pulmonary vein isolation. And speaking of that Venn diagram, so this is how I think about it. This is not perfect, but this is just sort of what I came up with. If you imagine this big circle being everyone that might be undergoing an ablation for AF, maybe you can roughly divide it among persistence and paroxysmals, and then among that, there's a subset of, of course, redo procedures. So among this big population, who would you think about ablating with, with additional strategies? And I think the other factors that you're going to think about, of course, in this day and age, of course, with, with PFA being so, so widespread, relatively quick, and performing a simple anatomic approach, I think it's fair to argue that that first-time simple early paroxysmal patient, I probably, honestly, in this forum, I'll tell you, honestly, I wouldn't use advanced mapping. It's probably just not necessary. But those patients are going to have some redos, and then these long-standing persistent AFibs and some subset of persistent AFibs will have very clearly, I think, likely benefit from additional ablation. And then this is where we want to figure out what that additional ablation should be. And so I tried to carefully place this circle, this red circle, of who I would personally think of as a good candidate for Volta mapping, or Volta mapping using this kind of a strategy. Many, if not most, redo procedures, but I also think there's a fair number of persistence, especially the long-standing persistent AFib population, that would be, I think, beneficial. I just, if I may just say anecdotally, I'm sure all of you who do a lot of ablation have come across these patients where you go in, maybe even a first-time ablation, you map the left atrium, their PVs have no signals at all. You can, you can pretty much guarantee that if you just did PVI alone, you're not going to get a good, a good outcome with those patients. So, so there's clearly, as that example I think would show, I think benefit from additional ablation. So then, so in this particular case piece, I think it fits into that long-standing persistent kind of a patient. So I want to show you what we went through with this case. The procedure setup, this one we happen to use the CARDO mapping system. We happen to use an octarray for the mapping process. We happen to use a QDOT for ablation, and I happen to perform all my procedures without fluoroscopy. And our mapping strategy is to, we always actually add in mapping of the right atrium. There are signals that do, or signals that show up in the right atrium. We of course map in the left atrium as well, and then based on that mapping, I'll sort of take a step, a moment to stop and pause and decide what is the ablation strategy I'm going to try to embark on. So I'm gonna, these are very highly truncated, sped up videos to, in the interest of time, but this will show you first is the right atrial mapping, and this is sort of how the real process works in this particular, this is that patient. So we'll go through the whole map, and as you're doing your usual voltage and or anatomic map, it'll automatically annotate points, and the dispersion points will be identified in real time, and we always map the RA, and in the end they're annotated. So with the current version that we use, you'll have a two-tier annotation. So we happen to use red as a higher likelihood of impactful ablation versus yellow, which is what we might call a moderate level of impact, and you can see in this particular case that there are a smattering of points that were noted throughout, maybe clustered in certain areas, but what I personally do is then don't start embarking on ablation on the right atrium yet, but then I'll just move over to the left atrium, and then I'll start thinking about at this point already, you know, is this going to be worthwhile ablating eventually? I'll keep that in mind while I go to the left atrium. So then in the left atrium, we start doing the same thing, and as you can see here, this patient being in persistent atrial fibrillation, we do this map. In the interest of time, I'm going to skip ahead beyond showing the full mapping process for this slide, but then with the end result, as you can see here, this is, now this, I'd like to show this example because you can see here, this left atrium is just covered with potential ablation points, so I'm sure if I pulled, nobody would say I'm just going to ablate all this. I mean, I wouldn't. So then how do we then start thinking about this? So obviously you're going to do pulmonary vein isolation, or maybe not obviously, but I would, and then I start looking at the clustering, and I say, oh, there's a lot in that posterior wall. There's a smattering along the ridge. There's a smattering in the anterior left atrium, so maybe I'll take this one and do PVI and posterior wall isolations, and that's what we did with this one. So then we start ablating, and this is very interesting because, you know, we're, of course, when people show cases, they always show you the best kind of outcome, but in this case, I think with Volta, if I may say, we've really, I've really, and we've all, I think, transitioned to this outlook of, we're not ablating and then expecting the cardioverte and then see how the patient does. We are actually expecting to ablate the patient into sinus rhythm, and that's borne out in all the data that you've seen presented as well, in fact. And so, again, to fast forward, you look at this ablation. We were doing the PV isolation, and then they're not quite organized, but you see some slowing if you look carefully at the electrograms that are shown on the CS catheter, and then as we start ablating the posterior wall that I decided to target, you'll see, if I can have a few extra seconds, eventually, this will now organize into a flutter right there. And so then, aha, that's great. So, again, not a surprise anymore, frankly, and this is sort of what we expect to see. And so now, once you've finished doing your posterior wall stuff, then I'll move on, and then we start mapping the flutter. And just, again, in the interest of time, I'm not going to show the full map, but this turned out to be a mitral flutter. We ablated, ablated, ablated, got the patient back into sinus rhythm. And so, so that's just a case example, and I think, honestly, a very typical case. In this case, we tried to reinduce. If we reinduce, we'll remap with Volta, but if not, then we'll stop there, and in this case, we stopped there. And just for a clinical follow-up with this particular patient, he has remained in sinus rhythm for two years now. So, so not a, I think, a cherry-picked best case example. Honestly, a typical example of what we do these days. So, to summarize my part of things, I would say, number one, clinical evidence really supports improved clinical outcomes using an AI-guided mapping process for AF ablation, as you've seen from some of the data presented. For most of the patients undergoing first-time ablation, the straightforward anatomic approach, I think, is still reasonable, but in the optimal patients that I've discussed, most repeat procedures, especially persistent AF, first-time ablation patients with a lot of substrate, these are ideal candidates, I think, and it helps you sort of guide the right patient to do this for. So I'll stop here. Thank you very much. All right, can you guys hear me? All right. Thanks, Paul. I think this is very unfair that I should be the last one after these wonderful presenters have to give this, you know, but I'll do my very best. And so, all right, and so you can't come to a conference at HRS in 2025 and not hear about PFA, so we just have to talk about it. And, you know, the combination of these two technologies, I am an equal opportunity employer, that is something that we really do need to discuss. And so I'm going to start with a case and then I'm going to show you as we go. So this was an 80-some-year-old gentleman who has persistent AFib and heart failure associated with it, multiple admissions for fib cardioverses. I did an ablation on him back in 2017 and I will say I've done almost every type of procedure to this gentleman. And so I did an ablation, I put a loop in, he had tachybrady, put a pacemaker in, he had bleeding, put a Watchman in, and then he came back to me with more fib. So, you know, as Paul said, we normally show our best cases, but unfortunately, he continued to struggle with fib, very symptomatic with it, multiple hospitalizations. So I needed to do something for him. And so I did bring him back to the lab. And so I'm going to show this map that we did. Again, Paul showed a case with Cardo. This is a case with InsightX. I will say that the auto tagging and the cooperation and collaboration that Abbott and has been with Volta has been wonderful, and I will continue to advocate for collaboration amongst all of the companies that you see down here today to help us actually do a better job of this. And so it is, again, you can see here, this is probably my least favorite thing to see in the lab when I get in. Veins and posterior wall are pretty much isolated. Turns out I did an okay job the first time. I wish I did a crappier job, to be honest, because when you get in there and you see this, you know that we're not in good shape. We've published this at Penn, and then multiple other centers have showed the same thing. So I will also say that many of us sitting up here and people in this room are on advisory boards to help companies, and I have heard time and time again, more recently, anatomical ablation, less mapping, less looking at electrograms, just get in and get out and move on with your day. And that really upsets me as an electrophysiologist. Dr. Marsalinski and Dr. Cowns, who trained me, would drag me off of the stage if I continue to advocate for that. I think all of us believe that this is important. We're electrophysiologists, not just anatomical ablationists. And so again, this is the map that ended up sort of coming to the end. So what do you do based on this? And we don't know the answer to this. And this is why I think that, you know, collaboration with Volta is particularly helpful. The top two reasons that, you know, when patients are coming to me that have had a previous ablation with me or with someone else, what they ask is, well, Ben, what are you going to do differently? You did this procedure. You said you were going to be successful, and I'm coming back with AFib. The top two things are, one, pulse field ablation, right? Everybody's talking about that now, and we're going to be talking about that as well. And you can see this is the dispersion map. So again, and then the other one is AI. So everyone is interested in AI. Everyone's using chat GPT to write their book reports, and patients, when they hear that, their ears perk up. Okay, you're actually going to do something differently than the first time. And that's what I talk to them about. And patients who even have had long-standing persistent fib, ones that, you know, Paul showed that case, I'm not thrilled about putting those patients in my lab, especially as my administrators want me to do more to pay for pulse field ablation, that, you know, time is money. And so all of these things become particularly important. So what was my ablation strategy based on this? You can see that, again, and I will say I probably should have done a better job of completely mapping all of the chamber, but it is important, as Jerome comments on, meticulous. You have to be meticulous about mapping and find these areas of dispersion. So now that we have a catheter that can just deliver scorched earth to the atrium, should we, or maybe we should be a little bit more elegant about how we do this? And so I think that the marriage of pulse field ablation and Volta is a good one, and that we can now, and no pun intended, tailor our ablation strategy to be able to actually deliver energy appropriately and safely. It is kind of incredible that both Restart, which we were a part of at Penn, and Tailored as well, were RF ablation, you know, cases. We did a lot more ablation than we typically will do, and I will say as someone who has had a few fistulas in my career, it's probably one of the worst feelings of my entire life, that I had a patient that I did just a standard PV ion, everything was smooth, high power, short duration, stuff that we've done, you know, pretty typically. He presented with an effusion and had an arrest. We choppered him up to Penn. The surgeon took him right to the to the operating room. Thankfully, he lived. He had spaghetti in his pericardium. It is the, again, it makes me almost nauseous talking about these kinds of cases. So if we can obviate the fear of doing more ablation with pulse field ablation, and also do better, safer, and more efficacious, that's, I think, what we all are advocating for. So you can see here that I, you know, did a anterior, essentially, mitral line with a ferripulse catheter, and was actually able to, I know this is also sped up like Paul's case, terminate an atrial arrhythmia. I will also say that this is novel. We traditionally did not ablate to termination and or non-inducibility because it's never been proved. And so we, you know, would try to do the best that we could in and do it as safe as possible. And so in this case, I ended up terminating fib, and this gentleman was in fib for a good amount of time, with this anterior ablation. Again, I touched up the floor of the posterior wall and anterior, and he was non-inducible. And I was very happy with this result. What would I have done without Volta? I'm not really sure. And, you know, we've published trigger protocols and other things at Penn to be able to help with this. But I think this is a very helpful tool to potentially be, again, more elegant in the way that we do this procedure. And so there's a few things in the lab that make me smile. So this is one of them. So Dr. Gersenfeld, who was in our lab at Penn previously, he would always go, and it was typically a pathway. When you first burn, get a pathway, go right away. PBC, first burn, go away. You know, those are those are moments that make you smile. This was one of them. So this was actually during this case, and I'm smiling, because I terminate fib. It's fun. And so, you know, I'm super excited for the collaboration we're going to continue to do. We need to continue to do more research as we get new catheters. But what I also really like about the system is that it's agnostic. We can use whatever platform we want to use to be able to deliver pulse field or radio frequency ablation and be able to do better for these patients. So I'll stop there and be happy to answer any questions. Okay, okay, great. I think we have about 20 minutes for questions, answers. Both examples you gave, they both had anterior mitral lines added. If you look at the tailored AF data, majority of their diffraction, diffusion, you know, their signals that they picked up were anterior wall. So if I just didn't buy Volta and instead just went ahead and took a pulse field catheter and did a great job homogenizing the anterior wall, doing a great line, I would have essentially got to the same conclusion you both did. Probably faster, I wouldn't have had to do all the intense mapping and had the same result. So is there any data to support a Volta driven approach over an empiric anterior wall, well done, pulse field, whatever technology you think is going to give you the best anterior line? Because star AF was lateral lines, which are notoriously difficult, particularly with RF. So that was my question. Let me just clarify one thing, as my case was actually posterior wall isolation. And I can just speak in general terms, but when you start looking at the tailored data and you look at the distribution of where where a dispersion is located, it's not all in the anterior wall actually. In fact, there's a, I think, a reasonably even distribution between anterior wall, the interatrial septum actually is a quite a big player, the sort of ridge vein of martial area, and the posterior wall actually surprisingly is not as much as you would think. It's a pretty low percentage. It's in the right atrium, too. And so I would predict, I don't know for sure, but it's my guess, is that if you did a study like that, like all the posterior wall studies and just did an empiric anterior line, you probably get the same result that you got from all the posterior wall isolation studies, because you're going to be ablating a lot of unnecessary anterior lines. The other, it's a bit of an obnoxious answer, so I apologize in advance, but the reality is if you ablate the anterior wall and the posterior wall, you'll probably get most triggers or most of these sources, right? And the reality is what we want is not getting sinus rhythm with no atrial function left, right? So I think that, but I think your general question is still right, which is how do you know that the success rate with your targeted ablation is not just simply debulking? And I think that's a good question. The things I would say are a couple of things. One, the Taylor Day F-Trial, which is looking at additional ablation versus not, doesn't answer your question. But there is additional data. There was some data presented at the AF Symposium that, from Clement Bars and Julian, where their data, with the next generation algorithm, the number of sites that are identified are much more, much more contained than the original, original software. And again, this is a non-randomized study, but at least in that study, it appeared to have a very good success rate. Again, I think your question is, so if you end up doing additional ablation, which is in a relatively modest amount of tissue, and you get improved success, I think most of us would agree that's probably specific. So I think this is still an evolving thing. But, you know, part of the issue here is what we, what do we do today, right? I mean, there's going to be the future and we'll hopefully get better technology. But right now we have a tool, and actually, Tina, did you want to make a comment to this before I go into the next question? No, I just wanted to add that if you historically look at all the AF mapping studies, right, dominant frequency mapping, ECGI, etc., anything, there's always, they try to agree. They try to agree in the locations, pretty much what you said. There's always drivers or triggers, whatever you might want to call them in a posterior wall in anterior, especially anterior to the left atrial appendage. But how do you target them? And doing a focal blob of ablation now makes us believe that it's causing a lot of atrial tachycardia, right? So doing a blob of ablation without completing the mitral, anterior mitral line, I think that's what we are doing these days. Instead of just doing a focal ablation, connecting it to non-conducting boundaries is pretty much what we are doing. But I'm imagining a patient who has some scar in the anterior wall, and then you do some sort of a mapping, it shows absolutely no drivers there. And yet most of the dispersion sites or the driver sites in the posterior wall, then you might be stuck ablating an empiric voltage-based line that may serve no purpose. So that's my worry about that approach. I don't know if you agree. And I think it's a slippery slope. I mean, now with PFA, you have folks that are thinking about veins, posterior wall, anterior line, SVC, CRISTA, CTI, FASA, CS. Where do we stop? And, you know, to Vivek's point, it's just that's a lot of ablation that needs to be done. And I think for many of us who, certainly in the RF days, it was sometimes pro-arrhythmic. And these flutters are very hard to treat, and they're traditionally not, they're worse, the patients are worse than when we started. And so I think that this allows me to be just a little bit more careful as to what I'm going to be able to do. And I think, you know, certainly with false field ablation, we can do all of those things, but should we? Is the important question. Still has to be answered, and we still have to do more research in that in that regard, which is why I think those type of clinical trials are very important. So you can send in questions, by the way, for the app, right, through something. Anyway, some people have sent in some questions. Let me just ask one question while you come up to the mic. Where's the mic? There, she'll give you the mic. One of the questions, actually, Tina, maybe this is more appropriate for you, is this question was quite a, they asked the question about the machine learning algorithm, and the question pointed out something that you alluded to, which is that the initial algorithm was really based on expert operators. I mean, all these algorithms are based on some ground truth, and ground truth there was an operator said, this is a dispersion site. So the algorithm got better at identifying dispersion sites. So the question was, why does it work better than eyeballing dispersion? Faster analysis, smarter annotators? That was the question. Well, go ahead. I think it's for scaling, right? In Julian's lab, he's gonna identify dispersion perfectly. You put it in my lab, and I'm not gonna be as smart as, okay, is this dispersion? Is this true dispersion? Should I target this? So it pretty much makes it objective, and you have a tool that will auto-identify the best way of identifying that signal as, yes, this has dispersion versus not. And I think the benefit of the AI is you can take it beyond an expert labeler of dispersion. You can say, okay, now I'm gonna try to teach the model where AFib, the sites, the dispersion signals, that whenever I burn them, AFib terminated. So you can now make the model smarter in a way for acute termination, for long-term success, etc. So I think it's, if every one of us were as perfect as identifying that, it's great. But I think when you scale technologies, mapping technologies like this, you need an objective method to guide us targeting the same sites. Yeah, I mean, I think it made the, it certainly made the trial much more even. People were doing the same thing as opposed to doing random things. I do want to point out, though, but ultimately, look, we're electrophysiologists. In general, electrophysiologists tend not to believe anybody else, right? That's our nature. So yes, there's some experts, but are they really experts? So I think that's a reasonable concern. I, again, will add that the technology is evolving and that there are, again, different constraints that are being put on the algorithm that are, let's say, a physician is somewhat independent, that do seem to get similar outcomes. But again, we'll need to see more data. Paul, let me ask you a question. There's a question here about when you were mapping the right atrium before you had Volta, what did you do? What were you looking for? Yeah, so, good question. So I'll be honest. I think we, I, like probably a lot of you here, used to think of the right, or still do actually, think the right atrium is this big black box. You know, we, you know, almost all the research on AF mapping ablation has been the left atrium. The right atrium has been this sort of forgotten stepchild. There's an era, you know, whatever it was, ten years ago, where we were doing a lot of SVC isolation and empiric CTI lines for a while, but overall we don't know what's really going on in there. So frankly, before that, I really wouldn't know what to do about that other than, you know, once in a while we'll put in an empiric CTI line. And now, you know, observationally, you know, you do see a fair number of, and actually beyond observationally in the data sets that they, that we, that Volta has presented, there's a fair number of dispersion sites that show up in the right atrium. I forgot, sorry, go ahead please. Thank you for the talk. I have two questions. First is interpretability. So for the future plan of the algorithms, what kind of interpretability strategies are you going to use? And also, I'm asking this to the clinicians, to what kind of level of interpretability will you accept to believe that this is somewhere you want to ablate? And also the second question is, similar to the previous one, I have a question about the ground truth. So is there a golden standard for the ground truth, or is it more based on the subjective point of clinicians' experience, too? So that's, if it's not enough ground truth labels, have you tried synthetic data or simulated data to do the deep learning algorithms? And how does the performance look like? And how do you tackle them? Yeah, let me just start there. There is no ground truth. Unfortunately, the only ground truth we have is a relatively blunt approach, which is a prospective randomized trial with one-year follow-up. So that's the difficulty, and that's why our whole field, I believe, has been moving so slowly. So let's start there. Having said that, there are certain markers you can use that'll move us closer. So for example, if doing something causes an effect in the atrium, like prolonging cycle length globally, or certainly terminating AFib, that suggests that that may be an important area. So I'll start there, and go ahead, please, Tina, if you want to continue. That's exactly right. There is no ground truth quite yet, at least in vivo, in humans, right? We don't really know how AF propagates, because we can't really see action potential propagation, at least in the humans, at least in vivo. And even in vitro, there's so many debates, right? Is it true propagation? But let's assume that in vitro, action potential propagation via optical mapping is your ground truth. As far as I know, your question is, is there correlation of this kind of mapping with in vitro optical mapping? To my knowledge, there is none, but I showed you some early, early studies showing fractionation at the site of mechanistic relevance sites to atrial fibrillation mechanism. If you burn there, AF terminates, and you do see fractionated electrograms, because of the animations that I shared with you. So for now, it is subjective ground truth of the experts analyzing the signals and labeling that. Now, combined with acute endpoints, right? If I'm burning here, and something happens to the cycle length, it likely is relevant to the mechanism of AFib. So I'm gonna alter or focus my ground truth towards those for the model to learn those sites, rather. So it's more like a termination versus non-termination. For slowing, or anything that is making a mechanistic impact on the the AF dynamics, right? Just to add some additional complexity, additional problem is, an AF terminates when you're doing something. We don't know if it's because you did that thing, or you did the thing three minutes ago. That adds additional complexity. Let me ask Ben a question. One of the questions was, why do you think there are so many ATs, atrial tachycardia, seen in tailored AF, with a limited ablation, and without doing a fully validated line? Is that the reason? Also, you may want to talk about pulse fit ablation. It's a good question. I remember years ago, when I first started, I heard a colleague talking to a referring. He said, well, we don't really see a lot of recurrent AFib, but we tend to see flutterers after, that we need to touch up. Well, that colleague does a lot of empiric lines. And so, I think that with tailored, like in anything, we, you know, the question earlier about interpretation, it is our job, based on what Volta tells us, to still decide an ablation strategy, right? And so, you can see dispersion in the right atrium, in the left atrium, but what are you going to connect it to? What are you going to do? And that's our job, to make sure, as an electrophysiologist, that we understand whether we're actually making the patient pro-arrhythmic, versus not. I think we all know, and there's the reason many of us don't do more ablation, beyond pulmonary veins, or even the posterior wall, is that it can be pro-arrhythmic, if not done, you know, properly. Now, will pulse-to-field ablation obviate that? We don't know yet, but I think that that's where clinical trial stuff is going to be very interesting, to see if we did tailored, but if we did it with PFA, would the risk of atrial flutterers, or atrial tachycardias, be less? We hope so, but we don't know that yet. Yeah, and if I could add a quick point to that, Vivek, is that one limitation of all the data here is, we really have no guidance how much ablation to do at each site, so if you specifically focus on RF, it's not, there's no guidance on, you have to use an index rally, or you have to ablate to this amount of time, or this end point, and impedance drop, or whatever it is, so if you're in a thick part of the atrium, you could be under ablating, so these are potentially reasons for recurrent atrial tachycardias. There are a number of additional questions, but I also want to talk about the restart trial, so it was just presented earlier today, John Hummel presented it, and can I ask, maybe Ben, I'll ask you again, since you're involved with the trial, do you want to just talk briefly about the data, and what are your takeaways from it? Yeah, I will say I was very impressed with, you know, just to be clear, this was a single arm trial, and patients were undergoing redo ablation, it was a prospective trial. So again, the restart trial showed, I think, about an 88% success rate from atrial arrhythmias at follow-up, which is, I believe, the best that I have seen in terms of a redo clinical trial. So I participate in the trial, and I will say that, again, it was an RF trial, so there was a lot of, I remember working with our colleagues in Volta saying, Ben, you have to ablate this, you got to induce again, you got to map again, and I'm like, do I really have to do this? I have a four and seven year old that I would love to go home to, but it showed me, and I think the trial shows this too, that if you spend more time doing these for these redo cases, you will have a better success rate. And so I have pivoted from the going in and then just maybe doing more anatomical ablation to a strategy where I map, ablate, map, ablate, map, ablate, especially with pulse field ablation, you know, adding to that. So we have about five minutes left, and four or five questions here, so this will be the lightning round. Right. We'll try to try to get through them if we can. I just got another question, thank you. So first question, are there any surprises from the Volta AI data in terms of the proportion dispersion sites, where you see them compared to what you expected? Tina, maybe I'll throw that one to you. I think it's in line with every other mapping study I've ever seen. About a third of the sources has always been in the right atrium, the rest has been in the left atrium, very in the left atrium, the base of the appendage, the posterior wall, the septum. I think I am happy that the mapping studies are in agreement, actually, that they're looking at the same underlying mechanism, whichever algorithm that they're trying to use. Paul, let me ask you this question. The question is about non-PV sites, about how many sites. So it's a hard question because it depends on the substrate, but in a redo persistent patients, approximately how much area or how many sites do you typically find yourself targeting, and if you don't get sinus rhythm, what do you do? Those are two different questions, obviously, but for the first question, to be very brief about it, it's quite variable. I mean, you'll see some patients where there's a very clustered focal site and your job's easy, you know, you have a very clean anatomic target, but like the one I showed you, there can be quite a broad range in it and you see all of it, honestly, and I think I'll leave it to some sub-analyses that will come out to look at that kind of information. Now, in terms of the second one, yes, so it's not a hundred percent termination in our cases. If you look at the data, it's, you know, in that 60 plus percent range that, and we see that in our cases as well, and so when that doesn't happen, like any other procedure, I mean, there's a point where you just say, let's see how the patient does, and the data bears it out that they still end up generally doing quite well. It's interesting, in the restart trial, it was a number of different operators. Some were using RF, some were using PF, and it was interesting that one of the sites that used, I think the only site that used PF, happened to do a lot of cases, had about a 90% termination rate, so you do wonder what these data are going to look as they continue to evolve with different energies. Ben, let me ask you a question. Here's another one. How consistent do you find the sites to be? It's a great question. If you look over time, you map, and then you remap later, do you see the same results? It's very important, right, because you want to be able to see that it continues, and I think a couple of things to that. One, it has been reliable. We go back and we look again and make sure that it's, as to your point, we need to make sure that it is reproducible, and we're all skeptics. I also think it's interesting, and a concept that I never thought of, that you do ablation and you map again, and there are new sites that were not there previously. To me, that's interesting and novel, not something we normally do. I mean, obviously, we look at voltage, but anything beyond that, we don't necessarily, so I've seen, at least anecdotally, reproducibility, but you do have to continue to do it and continue to go back, and in regards to the restart trial, and something that I've learned as well is that, you know, termination and non-inducibility is important, like that is an important endpoint, and I have to start moving towards that, especially for persistence, long-staying persistence, and definitely redo procedures, because anecdotally, my patients that recurred, I didn't get to that. Restart didn't push me to keep going until you got to non-inducibility, like Taylor did, and I did it, and I stopped after a certain period of time. I'm acknowledging that I need to continue to do more for those patients. And then, let me ask you this question, Tina. The question is, ablation itself can create more substrate for AFib to recur. Can you comment on how Volta mitigates that, or maybe the question is, how do you mitigate the atrial tachycardias after? I think there's some elegant, elegant basic science studies on that. Imagine you have a very small focus of scar, and imagine you have something anchored to that that's spinning around. If you just do one ablation lesion there, it becomes a slower cycle length local circuit, and hence, in my head, that's the mechanism of the atrial tachycardias that can happen, that are relatively micro reentrance still, in a slower, slower cycle length, and at least in the basic, in the animal world of AF mapping, if you extend your AF ablation lesions to a non-conducting border, which is what we do with our common sense, right, in clinical practice. Let me continue that to a posterior wall isolation or to the annulus of mitral valve or the tricuspid valve, but I think that is definitely the right thing to do. I think we have good evidence from animal studies with optical mapping of what happens with limited ablation versus if you extend that ablation to non-conducting borders, and I think that would be my strategy. I think that was a strategy in the tailored AF as well. Okay, fantastic. I think it's going to be the last word. Thanks all of you for staying to the bitter end, and enjoy HRS.

Volta Medical

machine learning

AF ablation

RESTART trial

FDA approved

atrial fibrillation

spatial temporal dispersion

neural network

pulse field ablation