It is my pleasure to welcome you to San Diego and Heart Rhythm Society of 2025, the 46th annual meeting of Heart Rhythm Society. If you have not already done so, please download the mobile app from your preferred app store. This is how you can participate in live Q&A session. We'll accept questions all throughout the talk, but then answer at the very end. Please scan the QR code on the screen to access the session's Q&A. When using the mobile app, log in with your HRS credentials. Please note that visual reproduction of Heart Rhythm 2025, either by video or still photography, is strictly prohibited. It is my great pleasure to chairing this session. I'm Heenan Park from Korea. This is a joint session between HRS and APHRS. It is my great pleasure to introduce first speaker from APHRS side, Professor Yen-Jung Lin, Taichung Veteran General Hospital, Taiwan. His topic will be Advancing Beyond PV Isolation and Persistent PV Ablation, Current Progress and the Role of Artificial Intelligence. Dr. Lin, please. Thank you very much. Good morning, everyone. I'm Mr. President. So I'm Dr. Lin Yen-Jung from Taichung, Taiwan. So I'm honored to share with you with the current landscape and future direction of atrial fibrillation operation as a focus on the persistent atrial fibrillation and imaging role for artificial intelligence. So I have no disclosure regarding to this. So let me begin with the atrial fibrillation. So we know that the atrial fibrillation is the most common arrhythmia in clinical practice. And also, we know that the PV isolation is the goldstone for treatment if a patient is referred to AAD. So according to previous multiple consensus, we know that the PV isolation remains the fundamental therapy for most of the AF cases. However, the efficacy significantly decreased with persistent atrial fibrillation, promine needle adjunctive for the alternative choice. So from the guide consensus document, we know that there are a lot of strategy in addition to primary isolation. But I think most of the strategy is not standardized. So including a lot of strategy, non-PV, also the linear operation or anything. So how effective of this strategy in clinical practice? So we know that the landmark, the STAR-F2, aided either the linear operation or catheter operation to a PV. It did not significantly improve the long-term outcome in persistent atrial fibrillation. We know that after one and half year post-follow-up, freedom for the atrial arrhythmia was below the 50%. So this underscores the limitation for anatomy-based approach alone. And also, some of the trial, like the VENUS study, showed a modest result for ligament muscle operation in patient with persistent atrial fibrillation. But the efficacy is modest. And also, we see some of the trials, say, just describe the efficacy for postural isolation in patient with persistent atrial fibrillation. We show there's no differences between the patient with PVI and the other arm with additional postural isolation. So there's no benefit. So this is an excellent trial from President Park, published five years ago. So we are thinking about in patient with persistent atrial fibrillation, anatomy-guided approach may not be enough in persistent atrial fibrillation. So as an initial physiology, we consider there could be a hot spot in the left atrial. And previously, I think many of the researchers have tried many ways to use the signal analysis to identify the potential driver, such as the domain frequency face mapping, or also some other way to identify this. But I think this is a potential guide to patient-specific operation in real time. So I think this is an approach for identifying the potential driver previously. So like the face mapping, frequency mapping, and also I mentioned the spatial temporal dispersion, and also from our lab is like the prism. So these are all the ways like the ESG approach to identify the potential atrial fibrillation driver in patient with persistent atrial fibrillation. I think this is a later groundwork for incorporation of machine learning into the atrial fibrillation brain. So I think this is, let me introduce one of the important strategy, like by the Julius Seuss, the doctor from French group, introduced the idea for the spatial temporal dispersion. So this merging just mentioned, this is maybe like the synchronized shifting electric pattern, and also this localized dispersion cover the whole cycle of a local region. So this region, they say this is very correlated with the atrial fibrillation driver, and targeting this will have a high yield of the posterior definition site. So the definition is that this is by eyeball detection, so maybe not easy to identify. So I think this method will later be automatic and commercialized by Volta medical, like the VX platform. So we know that the Volta program, they have like automated software, they use the AI to identify the spatial temporal dispersion. And recently they used the multi-center, multi-nation trial, tailored trial, and published, announced in last year's HRSA, and also published recently. They have shown the efficacy from the multi-center, multi-center trial in the retinal control study. So they show that the strategy is that you identify the spatial temporal identification by AI, and also they try to targeting the lesion in addition to pulmonary isolation. In a retinal control study, a multi-nation, they show one arm is only a PBI alone. The other arm also have additional substrate modification guided by AI, detecting by spatial temporal dispersion in that region. And in the tailored trial, they have find, and also the result, show that this is the outcome. The outcome is from the old, from the AF recurrence, the tailored patient-specific approach was 88 percent, and another guy was 67 percent, so we have a better efficacy in AF3 outcome. But total recurrence of HR preparation is the same. So I think this is currently, this is a, currently this is a very large-scale retinal eye control study in patient with persistent atrial fibrillation show the, show the benefit for procedure considering the PBI guided by jocular recording, and the use of AI by machine learning, and to provide the objective and reliable identification of air-pushing target region. I think this is pivotal to achieve the result. So from, from our group, I think in parallel, our team in Taipei, Taiwan, so our team developed another algorithm. This is the PRISM index. Let us use the nonlinear dynamic feature for similarity and periodicity, and try to identify the potential atrial fibrillation driver as an indicator to locate where is the potential, for the potential driver for, of the atrial. So the definition of the, the PRISM indicate that the regional difference in, in jocular recording were assessed by periodicity and similarity. This show the high level for the repetitive, the time evolution of the pattern for the multiple atrial wave, and also the similarity that indicate the waveform resemble each other. And put into the 2D dispersion, this is a sequential similarity between the two electrical and also the overall similarity of that one as an indicator. It could be the marker for, for indicator of the signal characteristic of the specific atrial site. And use that, I think, and from the computer simulation model, we can see that this is almost compatible with the potential driver for atrial fibrillation. And also in the clinical study, we showed that the specific pattern for the this, and also the correlate with the potential atrial fibrillation driver. So previously, we have tried to use this indicator as a guide to, to help us to identify the, the potential atrial fibrillation driver. So in a multi-center renal control study, we find out that in addition to PVI, the, the substrate modification was additional morphology repeating with the, the PRISM mapping. It can help us, and also we can find out the, can be able to have a better long-term outcome. So that is the example that we identified the, the potential driver of here, and also this is the high yield of the PRISM region, and also by substrate modification. So during the long-term follow-up from the RCT, we showed a better efficacy in terms of total AFV recurrence by 50%. Okay. So I think this could be indicator for use to apply into the, the, the artificial intelligence. So we applied the deep learning model, ResNet-50, in another series to, to the, the, the PRISM, the recurrent plot, to find a potential AFV operation, operation site. So our input is the individual signal characteristic of this one, and the output is the procedure termination of each region. And we try to use the AI to predict the, the potential termination region, and it should be the output. So in the preliminary analysis that the, we, we include more than nearly 40,000 local Yehsue Guangfang, more than 100 cases, and differentiate into the, discriminate into the termination site and non-termination site. But termination, the procedure termination site is incidence quite low. So we, we use that signal advance technique to increase the, the prevalence, and to increase to 8%. And we find out termination site is much higher. And use that to augment the imbalanced data and to, to use the training program. So for that, and we can, we can see that use this as a deep learning model, try to use the AI to predict the potential termination site, and use that to guide the operation. So for that one, we, we find out in the total for 4,000 site, and we, for 40,000 interlocal recording, we can find out the, the AOC is quite good. With AOC more than 91%, with a sensitivity of 61 and a safety of 92, to use a single site recording, single catalyst to predict the procedure termination. And how to confirm this from, we confirm this from the computer simulation model and retrospective study and prospective study. So in the retrospective study, we also showed that the same algorithm can use the computer simulation model. So it's almost compatible with the site of the driver, with a high yield of the, the prediction. And also in the prospective study, and we find, we use a, in a prospective study, in a, in a single center, we use, we have 110 persistent atrial fibrillation patient. And we, with that one, and also we have a, and totally we, we have a training model, signal model, and try to use that and, and validate the criteria. And in the prospective study, we have, in the randomized control, we have two arm of the patient. The one arm is only PVI only, the other arm is progeny score match, and the prospective study, and we, in the prospective study, we observe the outcome to compare with the anatomic guide, PV, air proportion. So as in that one, we find out the, the, the prospective study, we can achieve the posterior termination rate about 40% during the operation. And we find that AOC is much better than the previous algorithm, either the similarity or present. So AOC can be the 80, 87%. And also this study also find out the long-term outcome with two years, with two years long-term outcome. The patient with PVI guided by additional AI, guided, identified the posterior termination site as a AF free outcome for 70% as compared to the conventional group is 55% year, 15% difference. Also the recurrence rate in, in terms of the AF recurrence, AD recurrence is quite different. I think this is a good one for, to show the, the result. So to give a one example, this is a patient with a persistent atrial fibrillation with a heart failure. And also you can see this is a, this is a mapping for counter funder, but we use the, the AI to present and to predict the potential AF termination site was in the posterior wall. And very interesting that during the PV isolation, we, we can find out the posterior termination just happen in the location where we predict there is a potential site for posterior termination. So very, very interesting that I think the, this in summary, so in, we use the waveform analysis, use a recurrent plot against, apply to investigate the morphology, repeatness and tempo cluster of the signal. I think the deep learning model can help us to, for a real time automatic waveform analysis during the consideration. So it is a deep reason model. So I think it could be, we did a posterior termination in case of persistent atrial fibrillation enhanced the operation outcome. So ladies and gentlemen, so this is my summary. So we think that the, the first PV isolation remained the gold standard CRP for all type of atrial fibrillation including a persistent atrial fibrillation. However, the additional AF driver in addition to the PV was crucial, especially in patient with persistent atrial fibrillation. Currently, the artificial intelligence may enable the dynamic real time interpretation of intracardiac electrical to help us to identify by, to identify the functional atrial fibrillation substrate with precision. I think combination with electrical mapping, high density electrical mapping, or advanced analysis help us to precise identification of the atrial fibrillation driver. That may be, especially may be beneficial in, especially in patient with the transient persistent atrial fibrillation. I think AI model can enable the data-driven approach that enhance the operation strategy that may predict the favorable outcome by targeting the atrial fibrillation driver beyond the PBI. So thank you very much. So this is our AI research team in Taipei and Taichung VGH. Thank you. Thank you. Thank you very much. Thank you very much, Dr. Lin. We will have discussion session after listening all three speakers talk, and Dr. Nicholson will introduce the second speaker. Thank you. All right. So next up we have Dr. Natalia Trajanova from Johns Hopkins. She will be speaking on integrating AI with clinical practice. Thank you all for being here. It is really my pleasure. I don't know whether you can see me because I'm so short, but yes, I'm here. So I'm really grateful for being invited to speak here. I was asked, or the title that I was given was on ventricular ablation advances. In my team, we do both atrial ablation and ventricular ablation guided by imaging digital twins. So today we will focus on the ventricular part. So digital twin, this word or this phrase has become really buzzword and we need to have a very clear definition of it. The National Academies of Science, Engineering, and Medicine created this document about eight months ago in which they defined what a digital twin is and it is a set of virtual information that mimics the structure, context, and behavior of a natural engineered and social system. So this is a dynamical representation of the processes in the system. It is based on the data from the physical twin and most importantly, it has a predictive capability that informs decisions that realize value. So if you're doing computational model for mechanisms, this is not a digital twin necessarily because it doesn't inform decision that realize value. And as I said, there is a dynamics between digital twins and the physical system. It is not a static entity ever. We have worked on digital twins for a number of years and we have some of the overview was written in a paper in heart rhythm, but we've used patient data, imaging, health records, clinical recordings for two main areas in the ventricles for prognostication, for clinical decision support, mostly prediction risk of sudden cardiac death, and again, interventions that account for the response of the patient and that's ablation. But I want to emphasize the fact that intervention that accounts for the response of the patient, that this is very important in our approaches. So we first published this paper in Nature Biomedical Engineering. For the first time here, we showed what is the strategy of creating a digital twin from contrast enhanced MRI scans, then assigning electrophysiological properties to remodeled zones and normal zones, then predicting where to ablate. And however, the important concept that is here that we don't end with just predicting where to ablate. A very important concept in the digital twins that we use it that once you identify the ablation targets, they are not what gets ablated. What you do is you repeat the pacing protocol that identifies in the digital twin these ablation targets to see what happens post ablation. And this is extremely important because once you have an ablation, this is a scar that's inflicted on the ventricles. So there is a new landscape. You have a original fibrosis or scar, and then you have a lesion. And all that creates a new arrhythmogenic substrate after the ablation. The hope is that it's less arrhythmogenic, but you never know. And so that's what we do in the digital twin. We assess what happens post ablation. And if an emergent arrhythmia is shown, then we ablate that new target. So it is not necessarily only the clinical VT that's in the patient, but also we predict what happens after the ablation. And only after that we construct what we call the optimal set, and we export to CARTA or on any other mapping system. Now I want to say that and state it currently, this is not possible to predict by any other means because you're predicting, as I said, the response of the patient to the treatment and incorporating that in your ablation set. It is completely dependent of any type of ablation, PFA, non-PFA. It is just a tool to do pre-procedure guidance. I want to show you one example from that Nature paper, Nature biomedical engineering paper. So this was a retrospective analysis of ablation in a patient with ischemic cardiomyopathy. So we are blind to the clinical outcome. What we do is we receive only the imaging. And so we predict based on the imaging, incorporate electrophysiological properties, run dynamically the digital twin, and we predicted that these are three locations that need to be ablated in this patient. So when we finished the blinded analysis, this was where the patient was ablated on the first ablation procedure. So it's really good correspondence with our prediction. However, we had two other targets that we predicted that these will occur after this ablation is done. And you can see what happened seven months ago. Seven months later, the patient came, and here is where the patient was ablated. So should our prediction have been used before the procedure, these three targets would have been executed at the initial procedure without the need of the patient to come back and have a redo procedure. So this is the concept. Now, why would you believe me when I say that? In order for me to have the clinicians that I work with to have trust in what we do, you need to do really an extensive clinical validation. So that's what we embarked next. So we conducted the first prospective analysis of heart digital twins in scar-dependent VT. And the paper was just published in circulation at the end of February. So we didn't predict ablation yet there. What we did is, it is the key concept. So if we are predicting locations that are likely to form reentrant driver, there must be some sort of electrophysiological abnormalities there. So we said, if there is, we just want to figure out what is the prediction of the digital twin and electrophysiological abnormalities. This was the first study. And then we also looked at predicting actually the ablation lesions on a second study, which is now in review. So the paper, as I said, was published in circulation. This was in collaboration with the team of Dr. Maggie Saba at St. George's. So in this study, we did 18 consecutive patients requiring VT for scar-related tachycardia at the hospital. And what we did is, again, we are predicting only abnormalities. We are not predicting where to ablate. The patients underwent ablation. We used the images, created digital twin. And here, this indicates the fact that we predict the targets. And then we test again where there is inducibility in the digital twin. And if there is, we would ablate there until the substrate is non-inducible. And then we export our findings. And then they are merged with electroanatomical characteristics, with electroanatomical maps. And we compared voltage, abnormalities, duration, deceleration zones, et cetera. So we also, as I said on the second study, we did a very similar approach. But now we compare the VT circuits and also the ablation locations. And again, we merged to compare that. And so there are three representative case I'm showing here. So here are the voltage maps, the activation maps. You can see abnormal electrograms. And here are, this is the digital twin model. And we are predicting where to ablate. And this is, again, this is not guided by the digital twin. We are just merging them to see where the locations where you can see very good correspondence here. So what is important that the predicted sites, we are talking primary predicted sites and secondary that occur after ablation. All of these locations, because the ones that we predicted secondary, while they are not sustaining any re-entrant drivers, they are latent locations. So that's why we explored abnormalities in these locations as well. And all the predicted sites showed high prevalence of abnormalities and longer durations of the electrograms. We looked at deceleration zones. So you can see here the substrate mapping, the deceleration zones. And then the predicted sites. On top of that, we identified 26 deceleration zones in these 18 patients. And 21 of those were within 5 millimeters of the predicted sites. And the median distance to the predicted site to the deceleration zone was just 1 millimeter. So the hard digital twins showed a really accurately areas of slow conduction. We also, this is the second study which is in review now. We did, predicted the re-entrant circuits. We focused on that. And there were patients also with non ischemic cardiomyopathy. And you can see here the entrainment at the location and our prediction where to ablate here. And they, you can see here is the predicted ablation, ablation site. I want to show you a re-entrant circuit that was seen in the clinic and the predicted circuit in the digital twin. And they had very good correspondence. All right, so thus far what we learned is we did the clinical validation. So this was the first combined prospective study of digital twins and clinical ablation. They displayed those predicted sites had higher prevalence of abnormal and prolonged electrograms. The circuits and ablation targets accurately predicted their clinical counterpart. So there is really a significant potential for these digital twins. Then commissioned an editorial to our paper by Jack Singh, and this was the title, The Future of Ventricular Tachycardia Ablation, and I love the Arthur Clarke, the only way to find the limits of what's possible is going beyond into the impossible. So where are we going with that? The current application is we conducted a VIRT-VT, which is the 10 patient study of direct guidance in patients with ischemic cardiomyopathy was FDA approved, IDE. We completed the study, we are writing the paper, you'll see the results. We also working on something called GenDirect, which is actually a new type of digital twin, which is genotype specific, and we are using it in ARVC, arrhythmogenic right ventricular cardiomyopathy. We have done retrospective analysis indeed, and I'm very proud to tell you that my student Nancy who developed GenDirect to combine imaging with genetic testing in digital twins, she's a finalist in the Young Investigator Award today in the clinical category this afternoon. So you guys, if you can support her, I'm really proud of, she's right there, and we are really very proud of her. Here she is. Thank you. So going forward, what are the challenges to developing and bringing these digital twins in the clinical practice? Personalization of the EP properties. If you don't have patients with genetic testing, you don't have too much information about the cellular tissue properties. So this is a challenge. So for us to overcome this challenge, we went on a very interesting direction, and I want to tell you that we have a poster tomorrow, and this is a machine learning platform for extracting the action potential morphology out of the patient's clinical data. It is a brand new AI approach, and so if you guys are interested, you can attend. Anne is also here in the audience, and she will be presenting that. So this is our pathway. Now we are doing personalization in these ventricular digital twins using AI. Another issue with bringing digital twins in clinical practice is democratization. It takes hours with these simulations. You know, they're high resolution, they run for a long time. What we need is we want to bring it in the hospital, we want to run it on a desktop, and just to take a few seconds. And in doing, our approach to doing so was to combine with AI, not with data science, but actually with scientific AI. And basically, instead of having simulation-based digital twins, we train an algorithm to predict the results of the simulation. And this is a next generation AI-powered digital twins, and we just published a paper in Nature Computational Science in December on a framework to predict solutions of differential equations on any shape. And so this was, we called the algorithm Diamond. So this is a very interesting application. It was developed because we wanted to use it for the digital twins, but it's applicable to everything, anywhere that you need to solve differential equations. You're testing airplanes or doing something, you can use this approach. So this approach is very powerful, and actually in two weeks, we had 16,000 downloads of the paper and algorithm, because it can be used for so many different approaches. And so this is, in this approach, all the cardiac shapes, it works for different cardiac shapes, it works for different pacing sites, the solutions are done on the reference domain. And so when we are doing the press releases for the paper, this was one of the images we suggested. Everything is solved in the reference domain, and then the solution comes to the different hearts for which we want to solve. And here is an example, okay, go back. Here is an example of the prediction. These are different left ventricles, we trained it on over 1,000 patients. And what you see here is the solution for activation, repolarization, and so forth, and some of the, we assessed errors. It was really a theoretical approach, but with a very important clinical application. And one thing I want to show you, execution of one simulation, which took five hours, with the current, we use OpenCARB, now it takes one second, and it takes one GPU on a clinical, it can be in the clinic, any desktop. So the take-home messages is that heart digital twins are demonstrating significant predictive capabilities in clinical outcome prognostication and treatment planning. And together with AI, they post, I believe, to be a major tool in precision cardiology. And here are a lot of collaborations. I particularly acknowledge Magdy, with whom we worked, as well as all the people in my lab that made these amazing advances. Thank you very much. All right. Thank you, Dr. Atreyanova. Now I'd like to welcome Dr. Patrick Boyle to the stage. He's from the University of Washington. He's going to speak on enhancing pre-procedure planning and treatment for arrhythmia ablation, predictive strategies with artificial intelligence. Thank you. It's an honor to be here. I'm a little bit intimidated, presenting right after my post-doc mentor, Natalia, and what's even worse is that one of the former students from my lab, who did a lot of the work I'm going to show you, is sitting at a table back there. So she's here to really call me out. So you already heard the title of my lecture. I want to give you sort of the framing for what I put together. The premise here is that artificial intelligence tools, in particular in our mind, in our hands, those that prioritize explainability of how the models work, are useful to facilitate better mechanistic understanding of arrhythmia, which we think can help inform its treatment. And so there's been a lot said in previous lectures about ML and AI. So what I'm going to highlight is the specific explainability method we use, which is called SHAP. That stands for Shapley Additive Polynomial Analysis. And like I said, Dr. Savannah Bufolco, who's here, who's now at Boston Scientific, was key in bringing this approach to our lab. And Savannah explains SHAP as viewing model features as players in a game, with SHAP showing how much each player contributes to the outcome of the game. And that's a concept that I'll come back to a couple times during the talk. So with the large, rich data sets of image-based computational models of the left atrium that are made possible through our collaboration with the cardiac electrophysiology team at University of Washington, and with careful follow-up data monitoring outcomes like arrhythmia recurrence after catheter ablation of AFib, we now have the foundation for a highly promising application of explainable AI. And I promise we'll get to the clinical applications soon, but first I want to set the stage by discussing a study where we used this technology to better understand the mechanisms of recurrent arrhythmia in simulations, much like the ones that Natalia showed you. So one concept that particularly fascinates us is the idea that proarrhythmic substrate created by ablation lesions might contribute to recurrence. And Natalia talked about that in the context of VT, but we think it's just as applicable in the context of AF. So in Savannah's first paper on this topic, we focused on the potential roles of interaction between ablation-induced scar and the fibrotic substrate left behind after an initial procedure. All the patients I'll discuss in this part of the talk underwent late gadolinium-enhanced MRI scans both before and after their ablations, and that detail is critical because this allows us to combine the post-ablation data showing where durable scar was formed, and that's highlighted in yellow. I hope you can see my cursor, but I don't think you can. I'm going to turn the laser pointer on. There we go. Okay, so that's shown in yellow on these models, and we combined that with pre-ablation data showing the existing disease-related fibrosis highlighted in green, and that's fibrotic tissue putatively that was not targeted and not destroyed by the initial ablation procedure. So in the simulations conducted using these models, we frequently observed episodes of reentrant activity that were anchored to non-conductive regions, most notably ablation scars as shown in these two animations. You can see the reentrant wavefront propagating membrane voltage going around these two islands of ablation-induced scar. Now with this large computational data set, we had the foundation for the first quote-unquote game that we chose to tackle with explainable AI, and that was based on the specific features of these arrhythmogenic ablation lesions in the simulations, can we predict which ones are more or less likely to serve as anchoring points for reentrant activity that might drive future recurrent arrhythmia? And so we trained a random decision forest network for this prediction task, and it performed with fairly reasonable accuracy. But more importantly, we were able to look at the SHAP analysis, and what we see here is that the SHAP gave us an insight of what the model itself was looking at, and I always try and hasten to add that the creators of SHAP, that's not us, this is a group at the University of Washington computer scientist, they are really adamant that this should not be used to impute mechanism necessarily, but rather it's what the model itself is using to drive these decision processes. Nevertheless, it can give us hints to pursue further research in order to try to pinpoint these mechanisms. It can sort of narrow the field, so to speak. And so what we see here is an example that in terms of the perimeter and area of the scar, there's a sort of Goldilocks zone. So scar that's too small tends to not be arrhythmogenic. Scar that's too large tends to not be arrhythmogenic because the wavefront just doesn't have time to meet with the wavetail. There's this intermediate region where, in this case, the non-conductive area of interest was the mitral valve annulus at the bottom of the left atrium, and that was just right in terms of area and perimeter. And then we also looked a lot at residual fibrosis. And so this, of course, was enabled by the post and pre-ablation imaging merged into the common geometry. And in this case, what we saw that generally speaking, having a lot of residual fibrosis surrounding a particular non-conductive obstacle, like this one in the bottom left, tended to be highly arrhythmogenic. But what's really striking to us is that even in cases where there's very little surrounding residual fibrosis, this wasn't always guaranteed. And so this is how the explainable AI has helped us sort of better understand the underlying mechanisms of what drives these reentrant arrhythmias. So the key takeaway here is that SHAP can give us deep insights on which cases in the classification algorithm tend to be borderline because of contradictory factors. OK, so the important caveat here, and these are scatter plots showing the values of those features as they relate to the SHAP values specifically associated with each one. This sort of gives you an impression and an understanding of just how much data we're able to crunch with these relatively simplistic machine learning models. But an important caveat at this point to highlight is that this study used AI to predict simulation features in computational models themselves, not in the context of translational applications. And so we came up with a new game. And here, we decided to try to look at whether we could use these models for patients with both pre- and post-ablation late gadolinium-enhanced MRI to predict who is more likely to recur clinically with a significant arrhythmia. And we used, notably, some of those same features that were derived as part of the original study published in the Journal of the American Heart Association. We also added analysis of simulations in image-based models as a sort of large subset of features. And so there's so many features available here. I won't go through them blow by blow. But needless to say, we had a lot to do. We had not that many models compared to the number of features. And so we began with a lasso feature selection in order to find the most salient features. And then we plugged those most salient features into the random forest model to try to make predictions of recurrence. And then we used the SHAP analysis to try to obtain patient-specific insights on who would recur. And then we were also very pleasantly able to add an additional cohort. And these are 15 patients, also from UW, but for whom the model had never seen those outcomes or those features during the training and validation of the model. So how did we do? The model demonstrated reasonable predictive accuracy as measured by the traditional area under the curve metrics. And that's comparable to prior studies. This one was from Natalia's lab around area under the curve 0.82. Similar problem, but with different types of models. And again, this is from Caroline Roney and Steve Nieder's lab in London. Comparable area under the curve, but completely different types of models, which in and of itself we found striking. Looking at the holdout validation, we thought that that was relatively successful. We got the correct prediction in 80% of the model without any further fine tuning of the model that had been tested and validated on the original cohort of 60 some-odd patients. But more importantly, what the SHAP analysis let us do was receive sort of valuable insights into which features were most information-rich for the model itself. I'm going to highlight just a few. So for instance, a higher left atrial volume index post-ablation tended to steer the model towards predicting recurrence, whereas in patients who had more extensive ablation, and so this is again characterizing one of those features that we can only get from the post-ablation LG MRI scans. So patients who had more of that ablation in a particular region, the atrial floor, the strength of that feature tended to steer the model towards predicting that they would remain free from recurrence. And this is a really interesting one. Again, it's one of these Goldilocks zones. So for durable ablation scar in the right pulmonary vein region, which you would think would be something that would drive ablation success, there seemed to be this sort of intermediate range where, in general, too much scar or too little scar both steered the model toward predicting recurrence. And so it appeared to favor cases where the scar was just right. And here's an example of one of the, I love these plots. These are my favorite plots from the SHAP analysis. It's called a force plot. And so the reason why it's nice is because in this case, we can see that even though the patient had a very low post-ablation LA volume index, which again was our top-ranked feature by far in the overall SHAP analysis, many other features pushed the model in the opposite direction. And that's represented by the yellow chevrons on the right side of the slide. And ultimately, those combined features led the model to the correct prediction that this patient would experience a recurrence. And we believe that it's in these borderline cases where explainable AI truly shines. So this is a short talk, and I'm going to cover a little bit more information about mechanistic understanding of substrate after this. So I can't cover all the details of the SHAP study, but we see many potential applications of this technology. And I think that I happen to be a family member of a stroke survivor who has atrial fibrillation. And that experience really changed the way that I appreciate this because I was, instead of explaining atrial fibrillation mechanisms to people like you in this room, I had to start explaining them to people like my dad. And I'm especially drawn to the allure of using tools like this in the clinic to help patients and to help their loved ones understand the factors underlying their risk of future life threatening events and potentially plan for it accordingly by deciding whether or not they want to have a particular procedure. And again, I want to emphasize SHAP doesn't provide causal information and needs to be interpreted very cautiously. But we think its insights can inspire new mechanistic studies driven by the model's top features. So I have about three minutes left, and so I'm going to give you a little bit of a preview of our next application. We think a lot about substrate because we don't see these reentrant drivers of arrhythmia consistently in all patient-specific models. Natalia alluded to the Optima paper in 2019 where one of the 10 patients was non-inducible because of a relative lack of fibrosis in the model. And in the prior study, also done by myself and Natalia during my time at Hopkins, in that case we had 20 persistent AFib patients. But using the model parameters as we did, we only saw inducible arrhythmias in 13 of the 20 patients. And furthermore, in a study that I conducted when I moved to UW with Savannah, we had persistent and paroxysmal patients combined in a model. And we had only 23 out of 45 patients with any inducible reentrant arrhythmias. And as an engineer, that's always kind of stuck in my craw. The issue is especially bad for paroxysmal AFib patients because they tend to have less fibrosis. But we know that these patients had arrhythmia. We know that potentially, if our hypothesis is correct, they were at least in part driven by reentrant drivers. And so we would think that there should be some better links between model behavior and the clinical phenomena that we know is happening. And so one of the things that we noted is that other exacerbating factors might not be accounted for. So there could be additional functional data that we could use to parameterize the model. We could incorporate other aspects of the substrate. And this is what I'm going to key in on here. And it's a theme that's emerged in all three lectures, is sort of adding complexity to these existing computational models by increasing the level of personalization could be helpful. So specifically, we're talking now about going beyond purely fibrosis-based substrate in these models of the disease-left atrium. We're looking at epicardial fat, which is metabolically active visceral adipose tissue that secretes paracrine factors. It's in direct contact with the atrial myocardium. It shares microcirculation. And we have new, or I don't think I should say new here anymore, but there are imaging protocols like the Dixon MRI sequence that can let us characterize that on a patient-specific basis. And so I don't have time to go through all of the slides in this section, because I've elaborated a little bit too much. But the really interesting thing here is that fat and fibrosis overlap areas in these computational models are rare. When we co-register the areas of fibrotic remodeling with the areas affected by epicardial adipose tissue, what we see is that it's relatively seldom for these things to be correlated. And so what I'm going to do is leave you with a teaser of what we are finding, which is that the tendency in these computational models where we hybridize the substrate, where we have both fat and fibrosis in the same locations, that those tend to be the loci of reentrant drivers. And this is important, because while they are extremely uncommon, these areas of overlap, they're also extremely arrhythmogenic as anchoring sites. So I'm going to attempt to maneuver here by jumping forward to my thank you slide. My conclusion slide is here. So what I hope I convinced you of today is that AF mechanisms are complex, even the complex of these infinitely analyzable simulations. And we think explainable AI can help us predict the need for repeat ablation procedures and more. The patient-specific causes of recurrence following AFib ablation are quite diverse. And we think that in some cases, adding additional complexity by further personalizing the substrate can help us dissect important mechanisms in AF. So thank you to the folks who did the work. And I guess we're now going into the discussion section. Thank you, Dr. Boyle. So it looks like we have one question here. And if anyone has questions, please feel free. There's a microphone here in the center. But also, you can submit questions via the app. OK, let me introduce the social Q&A, the online question, to Professor Tranova. We appreciate that ablation can be pro-arrhythmic. What is prevalence of emergence of novel arrhythmia substrate in digital twin hearts? Oh, come to Nancy's presentation today. And you will see a lot of cases. Well, not a lot. You will see case. I don't know why that's working. Yes, it occurs. In her study, she has patients that recurred once after, let's say, 11 months. And then they recurred after that. And so the digital twin is able to predict all these at once, all these episodes of recurrence. We do have, in our atrial digital twins, we also have quite significant amount of recurrent arrhythmias. So I think this is a very important part of an understanding how to best ablate is what happens to the substrate once you have lesion. We have a new paper in NPG, Nature Publishing Group, Digital Medicine, which further elaborates on the emerging arrhythmias post-ablation in atrial fibrillation. But they occur. If you have, let's say, three drivers that you can identify, often you have one after that. You ablate that one. And there might be another. Yeah, the best illustration of this is that in the context of these simulated episodes of re-entrant arrhythmia, what we often see is one dominant site that is driving re-entrant activity, but there'll be transient, short-lived episodes of sort of pseudo-reentry, incomplete re-entry in the periphery of those drivers, and what we often see happening is that after the virtual ablation of the primary driver sites, the emergent re-entrant drivers are located exactly where that transient re-entrant activity was originally located, and in fact, the most striking case that we had in the Optima study, where we observed during the procedure, during the ablation of the procedure, it was this checkmark-shaped trajectory, and Dr. Calkins was doing the ablation, and lo and behold, arrhythmia terminated as he was ablating that target. That was an emergent re-entrant driver in the computational simulations, and so it goes to show just the critical importance of incorporating information about substrate modification as part of the procedure prior to making your sort of final plan. Dr. Boa, I have a question for you. So as a clinician, so in the PFA era, we clinicians are worrying about too much ablation, overkilling of HR tissue after using the PFA. In your paper, you mentioned about the anchored re-entry after ablation. It's a pro-arrhythmic mechanism. So I wonder, AI can predict those of pro-arrhythmia after extensive ablation. Is the question whether AI could identify those particular, like, what the... Yes. Yeah, so that was the point, and what I would, I guess what I would make sure I want to be crystal clear is that in that study, we didn't look at where the catheter tip was when the ablation was turned on. These were all cryo patients or RF patients in that case. What we looked at was three months later, where was durable scar created? And so I think in the context of that particular study, it would be a tough row to hoe if we wanted to try to predict on the day of the procedure exactly which ones of those lesions are going to be arrhythmogenic, but when we get the three-month post-ablation MRI scan, in our hands, you know, the area under the curve and the sensitivity, specificity, precision accuracy for that assay were very reliable, and it was a relatively, like I said, meat and potatoes application of supervised machine learning just based on the spatial features. So yes. Okay. Yeah. Expanding on that topic and removing ablation from the equation, would you guys care to speculate on the effect of those subdominant arrhythmogenic regions on the deterioration of transiently organized ventricular atrial arrhythmias into fibrillatory rhythms? Do those same areas participate or promote that, or is that a different substrate type of effect? Yeah. So what happens after ablation, there is always organization of the activity. It's always a bit more organized after the ablation, and I wouldn't call it necessarily a different type. There is still a combination of transient re-entries going around, but there is more anchoring to the ablation scar. And so the anchoring at the ablation scar is really prevalent, I would say. We had actually, Eugene Kolmowski is here, who works with a student of mine, and we looked at patients that had ablations and that recurred, and patients who had ablation and didn't recur. And those that recurred, majority of them had scar features that were attracting wavefront locations. So what ends up happening is there is organization by the presence of new scar distribution in the substrate, which becomes attractor to these re-entrant drivers, and the cycle lengths become longer. Which kind of makes sense, and it's understood as well. Nice presentations. Quick question. It seems that most of the atrial work, the modeling is of the left atrium. No. No. All our work is biatrial. Do you model both? Is it biatrial modeling? In my lab, I didn't show that. Everything is biatrial. All of the work we've ever done. Not only that, we have demonstrated the right atrial features of PSAF, of persistent AF. It's exactly the same as in the left atrium. There is no distinguishing features of fibrosis distribution. There are no distinguishing features of the dynamics of re-entry in both. It's just the same characteristics. Do you think that the pre-planning can help guide? Absolutely. Absolutely. In the left atrium, I know what to do. In the right atrium, I can hack it up, and it feels like a bit of a whack-a-mole experience and so I don't do that. Yes. Exactly. Exactly. It's the same for the ARVC study. The right ventricle is very important. It's thin, but you reconstruct. Sure. Very important. Then one follow-up question. Just thinking perhaps years ahead with your digital twin concept, do you envision that the creation of a digital twin and the modeling that you can perform and the potential ablation guidance, can that all be done in a non-invasive manner, or does it require voltage mapping or some type of nips with either non-invasive or invasive mapping? This is a really important question. I think that there are two things. You can say either you have to have imaging to reconstruct some sort of information about remodeling, whatever the imaging is, number one, or you have to have the voltage maps. You can decide which way to go. You have non-invasive, you need to have some imaging, otherwise we don't know anything about the patient. It is possible, though, that with AI, maybe we will be able to reconstruct, using the 12-lead ECG, some distribution of the areas of remodeling. But you don't need activation mapping. No. We don't. You can do this with either some substrate, however you choose to analyze the substrate. Exactly. If you say MRI is expensive or whatever, I don't want to have MRI on atrial for AF, then the approach would be, can we, instead of developing it from MRI, can we use the voltage maps on the fly and reconstruct something similar and then use that to guide ablation? That's the philosophy of the whole approach. Can I ask Dr. Lin a question? Go. Dr. Lin, I was fascinated by the data you presented. I was wondering if you could talk, in your confusion matrix that you showed, which was really impressive, I think there was still a modest number of false negatives. And I'm curious if you've looked at specifically, what does the electrogram activity look like in those areas that end up being false negatives, where it was a termination site, but the algorithm didn't pick that up? Thank you for your question. I think this is important. So because we are targeting each of the region for electrogram, not in the patient base, but in the electrogram base, so patient number will be very big. So we see the characteristic remain to be, I say, the future for prediction determination remain to be rapid, repetitive, and periodic behavior. And false with false negative, false negative part, I think remain to be, that one, we're not clear. We have to show in detail. But I think that that one could be the very rapid fractionation, but not so compatible with periodic activity. So focal as opposed to reentrant mechanisms? Yes. Interesting. So I think a different mechanism of the region may explain. I also have a question for Dr. Lin. You showed that the algorithm predicted very well. Yes, but even we see the computer simulation model we see that the one is not sustained because it's not but algorithm can say Interesting that we use AI as

Heart Rhythm Society 2025

atrial fibrillation treatment

artificial intelligence

pulmonary vein isolation

digital twins

ventricular tachycardia ablation

pre-procedure planning

arrhythmia mechanisms

machine learning

hybrid computational modeling