All right, so thank you very much for having me. I'm really humbled by the recognition. It's a privilege to present this on behalf of my colleagues. So this is Graph Neural Network Automation of Anticoagulation Decision Making. Let me try and advance the slide. Okay. Here are my disclosures. So as we all know, atrial fibrillation is the most common sustained arrhythmia with increasing prevalence globally and in the United States, accounting for around 50 million patients in 2020. And this is, of course, multifactorial with an aging population, obesity, and improved detection with smartwatches and personalized waveform acquisition devices. And the guidelines recommend a risk stratification, and if you look here to the right, inpatients with atrial fibrillation should be evaluated for their annual risk of thromboembolic events using a validated clinical risk score such as the CHADS-VASC. Now in addition to that, patients should also be treated with oral anticoagulation in whom they have an annual risk of stroke of perceived risk of thromboembolic events of at least 2%, which is reasonable and has been the standard of care for years. However, the CHADS-VASC shows variable performance across populations. And here you can see the area under the receiver operating characteristic curve numbers as well as the C-statistics, which are often comparable, or often equivalent rather. And here you can see that with an AUROC of .5, which would be chance, the CHADS-VASC score really provides marginally better than chance discrimination as to which patients will actually have a stroke. And so it's an imperfect tool wielded by imperfect practitioners. Now if you think of us as clinicians and how we try and manage patients, we assimilate whatever information we can, and then we make a risk-benefit decision based on that. And I would posit that right now a lot of the artificial intelligence work tries to use these novel tools to accomplish a particular task for an individual prediction. And I would say that with the vast availability of data in the modern electronic health record, that we're really trying to shove a square peg into a round hole. And so the objective for us was to automate anticoagulation decision-making using the vast amount of data present in the EHR to provide individualized patient-level recommendations and net-benefit assessments. So how did we do this? So we took around 1.73 million patients with 19 million visits, 81 million notes, and about 1.2 billion feature instances, and we extracted that data from the Mount Sinai Data Warehouse, including ICD-10 codes and medical concept extractions. We made no exclusions for sex or comorbidities, and the only thing that we essentially excluded were patients under the age of 18. We obtained the whole EHR data, medications, labs, notes, vitals, codes, demographics, diagnoses, et cetera. And we processed this by embedding these medical concepts. So what that does is it takes discrete data and transforms it into continuous vector representations that capture semantic meaning in relationships. And so what we did was we essentially compressed the whole EHR into 128 dimensions. And we take these graphs with embeddings as node features, a node being a data point, and these represent the EHR over the relevant index periods. And then we give this to a graph neural network for processing, which learns distinctive clinical elements and temporal dynamics and learns phenotypic patterns. Then we used conformal prediction. So conformal prediction is a statistical framework that provides reliable uncertainty estimates about the predictions. So it produces predictions with a 95 percent confidence, such that essentially you have at least 95 percent of your predictions are going to correctly reflect reality, another way of saying this is that it guarantees we're only going to be wrong at most 5 percent of the time. And so this is, by the way, not generally guaranteed with probabilistic classification on an individual basis. Then we compute net benefit decision curves. These essentially take the balance of true positive benefits against false positive harms. So in patients with atrial fibrillation who are being anticoagulated, we're talking about stroke and bleeding. And so this practically allows us to determine how a decision-making threshold can actually impact an individual patient outcome. And so, again, I'll just highlight that alone this technique only confers information on a population level and doesn't address individual cases. However, the GNN resolves this issue, and from the distribution of likelihoods we can compute an individualized expected net benefit. So then what do we do with all of this data? We provide it to a meta-learner, which learns to take the patient representation, the EHR graph embeddings, and takes predictions of bleed and stroke risk and expected net benefit. And then it decides whether or not to treat the patient and whether or not to be able to optimize the outcome. Now for patients with atrial fibrillation on anticoagulation, during the continuous level of exposure to anticoagulation, an optimal label was assigned if the patient was treated and they had no stroke and didn't bleed. That's what you would hope to see in someone who's anticoagulated. But an adverse label would be someone who was anticoagulated, they had a stroke, and they bled. Okay? And then there's obviously room in between. So these are the results. The red dots you can see are the treatment recommendations to treat the patient with anticoagulation, and the blue dots represent its recommendation to not treat the patient. On the y-axis, you can see the net benefit for stroke risk prevention, and on the x-axis you can see where the blue is accumulated, the net benefit for bleeding. And so you can see here that the model aligned its treatment recommendations, prioritizing, preventing stroke while avoiding bleeding. And just to zoom in here a little bit, you can see that this isn't clearly delineated. There are some high-risk bleeding patients that the model still recommended for treatment. You can see here these interspersed red dots amongst the sea of blue. Now to try and understand this a little bit, let's take a look at the discriminative performance of the model itself for the individual test. So if we look at the area under the curve for stroke prediction, here you can see that the graph neural network had an area under the curve of .8 for actually discriminating who's going to have a stroke. The CHA2DS2-VASc only had an area under the curve of .576, which is slightly better than a coin flip. Okay? Not ideal. Likewise, the GNN also outperformed bleeding prediction. So here the graph neural network had an area under the curve of .78 compared to the HasBlood of .642. So again, for the individual tasks of predicting the outcomes, the graph neural network is inherently better just at those. Now what does this mean for patients? So if you look at the top left here, let me get the laser pointer back. Okay, never mind. So at the top left in the light blue, you can see that the CHA2DS2-VASc, low CHA2DS2-VASc score patients, the GNN agreed with not treating 15% of all the patients. On the top right, you can see that there were no patients that the CHA2DS2-VASc would not have treated that the GNN would have. So there weren't additional patients that wouldn't have been treated already that we would have treated. On the bottom right, you can see that the GNN agreed with treating about a third of the patients. But what's really interesting is if you look at the bottom left, the GNN would have reclassified 51.8% of the patients to no anticoagulation. So particularly amongst these more moderate risk patients that would have been treated if you were acting on the CHA2DS2-VASc score, we would actually, the GNN would not have treated them to optimize their low risk of, relatively low risk of stroke and avoiding bleeding. Now there are of course limitations to this. It's retrospective. We didn't have external validation, although we had a pretty large cohort in a large, diverse multi-hospital health system. There was no assessment beyond atrial fibrillation. We know that patients with AFib aren't the only ones who have strokes. There's diabetes. There's other patients, for example, with interatrial block who we've demonstrated in patients who don't have a history of atrial fibrillation that that's an independent risk factor for having a stroke, whether or not they were found to have atrial fibrillation. And so there's clearly more work to do here. But I would just leave you all with this and this kind image and how I feel sometimes trying to assimilate all of this data, is that this novel approach automates anticoagulation treatment recommendations, and it takes into account the net benefit to the individual patient across the whole EHR. The GNN has better discrimination than the current risk models, CHADS-VASC and HASBLED, for the individual tasks of discriminating stroke and bleeding risk. And ultimately, I would say that, you know, clinical teams were really not capable of assimilating the vast amounts of data that are present in the modern electronic health record. And I would posit and suggest that potentially this could represent a theoretical but pragmatic break from a paradigm of clinical decision-making with further work. And so with that, I just want to thank my team, my colleagues at Mount Sinai, my great EP colleagues, some of which are in the audience, Dr. Wang over here. And so I really just want to thank the – it's really a team effort to be able to do this kind of work. And with that, if anyone has any other questions, feel free to reach out, and thank you all for your attention. Excellent. Thank you. Any clue about – you probably will look at it – post-ablation, successful ablation predicting? Who can come off and who can't? That's the first question many patients ask us. Yes. When can I come off the anticoag? Half the doctors here will never come off, and half will. Yes. That is an excellent question, and that is going to be one of our subsequent sub-analyses for patients who have undergone ablation. I think it's an incredibly challenging question, and some clinicians vary in how they even approach it. If the CHADS-VASc score is, you know, two or three, they may offer withdrawal. Some of the CHADS-VASCs, four or greater, there are patients never coming off of anticoagulation, but I think what I've just showed is that these tools that we're using are not very good at discriminating that risk, and so hopefully we'll be able to pan that out. When it comes to the patient's risk, when you take a patient for an ablation, will it change if there's no patient in your on-layer? We haven't assessed that yet. It's a great question. It's going to be part of the other analysis that we're going to do. Excellent question. Right, so that is something—so when we—so this gets to another point. I don't want to go too deep into this, but we actually had a first iteration of this where the area under the curve was higher, and the clustering of predictions was much more haphazard, and that was due to the labeling, kind of getting to your point. We actually used a lot of advanced natural language processing pipelines to be able to tease out and get more granular information about the actual patients from all of these notes. And so, you know, we were able to partition for the training and testing to avoid those types of questions—those issues and overfitting. But also, you know, again, we were operating with almost 2 million patients and over a billion data points, and so some of these considerations will come out in the wash. And if we included too many patients who didn't have events, generally the area under the curve would be artificially high because you will include more true negatives, and the model can make competent predictions about that those patients are going to have a negative outcome. So, you do use a training, testing, sort of breakdown of the data set—what does that look like? I mean, it was comparable. Yeah, we did, you know, the general 80-20. We didn't do external validation, which I acknowledge is not a limitation. Yeah. Similar to DOACs? You mean DOACs? Oh yeah, so we included DOACs, Coumadin, all of the medications. So, we included everything—patients who were on triple therapy, who were on Coumadin, who were on, you know, additional NSAIDs and other things that would compound bleeding risk. So, again, this was a very large scale and included all of that EHR data. So, we didn't—one of the benefits of this is that we did not make many exclusions. Other than an age less than 18, everybody was included—all of the medications. By plotting all of these embeddings over in this graph along these 128 dimensions, essentially what the model is doing is that all these have a magnitude and a distance from each other, and that's how it learns those relationships. And so, this all kind of comes out in the wash at the end, that by having all that data available and those interrelationships available for analysis, that's how the model helps understand what these relationships are. I have a question. How does it differentiate between a hemorrhagic, terrible bleed versus a hematuria? Great. And if you get a decision about a high-risk patient— Yes. So, excellent question. So, for this first iteration—we were already on, like, a third iteration— but for this iteration, we used BARC2 bleeding or greater. Essentially, bleeding requiring an intervention. Now, is that a perfect threshold to make? No. Inherently, you are going to have clinicians argue about what that threshold should be. We are never going to get to a point where everyone will unanimously agree on where that should be. In the future, we will also be able to take preferences into consideration to help inform the model's treatment decision. Excellent question. Okay. Thank you. Wonderful questions. What's that? Do you want to take a picture of it? Sure. I'll bring it here. Sure. Okay. Sorry. Okay. Next. David, you receive the award. Again, it's from the Cardiovascular Research Foundation, Amelie Gang, as I told you before, it's through the generosity of our benefactor. And your abstract is extraordinary, and we look forward to hearing it. Congratulations. Yes, absolutely. Let's take a picture first. Yes. Everybody okay? Yes, sure. Thank you. You can introduce yourself and all of the other colleagues, whoever participated in the study. Okay. Okay, great. So, I have nothing to declare. My name is David Chang. I'm actually a new member of the faculty in the Division of Genetics at Brigham and Women's Hospital. But the majority of this work was done during my postdoc with Colin McCrae. So, thank you again for the honor to be able to speak here today. I know this is a new experience for me and for all of you with the headphones, so hopefully we'll get through it. Hopefully no more band playing. So, loss of function SCM5A variants are classically associated with Bugatti syndrome, which leads to a decrease in NAV1.5 function. And this is associated with conduction slowing and reentrant arrhythmias. So, although Bugatti syndrome by itself is a rare genetic disorder, NAV1.5 dysfunction is also associated in more common pathologies, including atrial fibrillation, some cardiomyopathies, and heart failure. Despite decades of research, disease modifying therapy targeting NAV dysfunction has remained elusive. So, with this in mind, we aim to develop novel pharmacotherapy to rescue NAV dysfunction using zebrafish as a powerful drug discovery platform. And, by the way, I did a quick search. This is the only abstract in the whole meeting that has zebrafish. So, you all are here for a treat. So, why zebrafish? So, besides the fact that 70% of human genes have at least one zebrafish homologue, other advantages include its rapid organ development, including the heart, within two to three days it's fully developed, just needs to mature. It is transparent during development, so you can easily observe the phenotype. It has high fecundity, where one mating pair can give hundreds of embryos every week. So, all of these characteristics make it ideally suited for high-throughput phenotypic screens. Importantly, for our purpose, their action potential resembles that of humans because of the similarities of most of the underlying currents, including the sodium current. So, with this in mind, we sought to create a zebrafish model of cardiac sodium channelopathy for drug discovery. So, in fish, there are two orthologs of SCN5A. The major isoform SCN12AB contributes to about 90% of the NAV function, whereas the minor to about 10%, and they share about 75% similarity with human NAV1.5. So, using CRISPR-Cas9, we knocked out the major isoform and obtained germline mutants, which were in the expected Mendelian ratios. Here, I want to note that the homozygous knockouts are both viable and fertile, which is unlike homozygous SCN5A knockout in mouse and humans, and we think that this is because of the residual NAV function from the minor isoform. So, after validating the loss of SCN12AB biochemically, we perform optical mapping on hearts isolated from five days post-fertilization larvae, where the electrical circuit is already fully developed. So, on the top left, that's an isolated heart from a larvae, and on the right, it's just a representation of loading it with a voltage-sensitive dye. And using this, we can extrapolate key EP parameters, including conduction velocity through the entire heart, as well as the maximal upstroke velocity as a surrogate for NAV function and AP duration. So, using this method, we found a significant and gene-dose-dependent decrease in ventricular conduction velocity, as shown in A. We see also a decrease in ventricular Vmax, as well as atrial Vmax, as shown in B. And finally, there's a small increase in ventricular APD in the homozygous, as shown in C. So, together, these findings recapitulate loss of function in sodium channelopathy, at least during development. So, next, we grow these fish to adulthood and perform single-leaf surface ECG on anesthetized fish. In the raw tracing, you can appreciate the similarity to human ECG, with the notable exception of the inverted T waves. And what we found was that with SCN12AB knockout, it leads to both atrial and ventricular conduction slowing, as demonstrated by the prolonged P wave, PR interval, and QRS durations, as well as a slight increase in the QT duration in the homozygous, as shown in G. For both HETs and HOMOs, we also see a decrease in heart rate, as shown in C. Now, if you look closely, you'll notice that the heterozygous do not have a conduction slowing at baseline. So, therefore, we pre-treated the fish with ajmelin, and after that, we were able to see a significant increase in QRS duration compared to their wild-type siblings, which is similar to what we do clinically with the ajmelin challenge. And, finally, we directly measured the sodium current using patch clamp on ventricular myocytes isolated from adult fish. And, again, in A through C, you can appreciate a gene dosage-dependent decrease in the sodium current without significant changes in the activation and inactivation kinetics. And note here that the homozygous still has some residual current, which we suspect is coming from the minor isoform SCN12AA. So, altogether, these data established this mutant as a bona fide fish model of sodium channelopathy. So, next, we sought to explore whether this model could be used for a phenotypic screen, which requires the identification of a screenable phenotype. So, since decreased heart rate has been causally linked to NAD dysfunction, and we did observe that in our fish, adult fish, we wanted to know if we can detect this decrease in heart rate during early development. So, what we did was we took these cardiac videos at different developmental time point at different temperature, and we identified that as early as two days post-fertilization, we can already see a decrease, a small but significant decrease in the heart rate in both the heterozygous and the homozygous. So, that is a screenable phenotype. So, although we could detect this heart rate, this decrease in heart rate in both hets and homos, we ultimately decided to use the homos for the first couple of screens, and the reason is because of their stronger phenotype. So, since they still have the minor isoform from the AA, the screen is assuming that whatever compounds that can rescue the homos knockouts is acting on either transcription, mRNA stabilization, translation, trafficking, or membrane stabilization of the minor isoform, and then once we have those hits, we will go back and validate this also in the hets to see if it can also affect the major isoform. And ultimately, we'll need to validate this in IPSL as well, which I'll talk about later. So, to carry out the screens, we plated homozygous embryos in quadruplicates for each compound and measured their heart rate using cardiac videos at two days post-fertilization. So, we actually did this in two different libraries already with significant improvement in our efficiencies. Just to give you an idea, the first screen was only about 1,200 compounds. It took about four months. The second screen was about 4,000 compounds, which also took about four months. The next iteration, we can even do this five times faster using machine learning, which I will not talk about here. But in total, out of these two screens, we identified 49 that passed the primary screening endpoint, which is just a heart rate increase by 5%. We selected 37 for downstream studies, and 12 of those passed secondary screening in homos and were heterozygotes. And then finally, four passed voltage mapping screens in homos and heterozygotes, and now we're doing more mechanistic studies, including patch clamping. So, in the next few slides, I will primarily be showing data from the most potent candidate, which we're calling compound X. So, this is using compound X and doing voltage mapping. So, on the left, A, C, and E are the homozygous larvae, and on the right, B, D, and F are the heterozygous larvae. So, you can see that X, compound X, is able to rescue the ventricular conduction velocities in the homos, and is able to increase significantly the ventricular Vmax. So, in C, you can see that it goes from about 40 to about 60, and then in the heterozygotes, it starts out about 60, but it can rescue it almost to 80. And just to remind you, the wild type is about 90 per second. And finally, we also see a rescue of the APD duration using compound X. So, to directly test the effect of compound X on sodium current, we also perform patch clamping, this time on isolated ventricular velocities from headfish only, and we treat them first with DMSO versus one micromolar of compound X. So, interestingly, we observed two groups of responses. Some cells did not respond to compound X, whereas others had a very large increase in sodium current, as you can see in panels A through C. And even though these cells were from the same fish and treated in the same wells, and they had no significant difference in activation and inactivation kinetics, we also checked the capacitance, and there was no significant difference. So, this suggests that there are different subpopulations of ventricular myocytes, which responds differently to compound X, at least at one micromolar. Or there could be a threshold effect, where some of them cross the threshold, and they're responding, and the other ones have not responded. I have other preliminary data from a different model, where I used a higher concentration, and in that case, more proportionally, more cardiomyocytes respond to this compound X. So, after we have shown that we can rescue sodium current in the Bugatta fish, we wanted to see if we can rescue this in other cardiomyopathy, where there's also a decrease in sodium function, a decrease in sodium current. So, one of the fish that we have available, the Callum, made many years ago, well, it's more than 10 years ago, it's an ARVC model, where there's an overexpression of the human nexus mutation in Placogobin, and they also have a decrease in heart rate and decrease sodium current. So, we tested the four lead compounds, compounds X, Y, Z, and B, and we showed that they can rescue the heart rate in that model. So, next, beyond heart rate, we also performed voltage mapping, and we showed that compound X can also rescue ventricular Vmax, as shown in B, in the nexus model. And, finally, besides compound X, we also started testing the other compounds, and so you can see this is compound Y in A, B, and C, and compound B in D, E, and F, and it's the same trend where we can rescue the ventricular Vmax in the nexus model, with a trend for rescuing conduction velocity. So, in conclusion, we created a bonafide zebrafish model of loss-of-function cardiac sodium channelopathy. We set up a platform that's useful for high-throughput phenotypic screening for discovery of novel pharmacotherapies, and we identified multiple lead compounds capable of rescuing NAV loss-of-function in multiple disease models. We're actually doing other models at the moment, as well, which I hope to share in the future. The next step is that we want to figure out what is the mechanism of rescue, right? What is compound X doing? Again, we have preliminary data, but that's still a work in progress. We want to validate this in Brugada syndrome iPS-derived cardio myocytes to see how relevant it is for SCM5A, and we want to study the generalizability to more common cardiac diseases. And finally, as I alluded to, we're further improving the screening platform for larger screens. So, with that, I'd like to acknowledge Calum McRae, my postdoc mentor, as well as other members who have contributed to this work and the funding sources. So, thank you very much. Thank you very much. To what do you attribute the slowing of the heart rate? That's not usually a sodium channel dependent thing. So, that's a very, very interesting question. So, the question was, how do we explain the decrease in heart rate? So, it's actually not very well known, but it's thought to be associated with a decrease in the pacemaker cells, that there might be exit block or there might just be a slower reaching of threshold, a slower firing from the pacemakers themselves. So, that is probably not because of the ventricular phenotype. It probably has more to do with the pacemaker phenotype. But we're using that really as a surrogate because, obviously, a lot of things can increase heart rate, and that's why we didn't select all the hits. We only selected ones that are not obvious, like beta-adrenergic agents. We didn't touch those. And then, later, we still need to validate it, you know, using more sophisticated measurements, including PacScan. If I call my pharmacist and ask for a compound X, he won't have it yet. All right. That's fantastic work. Congratulations. And once again, congratulations. We'll have to move on. So, next. Excuse me. All right. Shohei Kataoka. Congratulations. You're going to tell us about long AV intervals leading to worsening atrial function. It's all yours. I think it's a wonderful abstract. Hi everyone. I'm Shohei Kataoka, a research fellow at UCSF. It is a great honor to be here and presenting our research today. So pacing therapy is an established treatment, but frequent RB pacing can lead to ventricular desynchrony and subsequent LB dysfunction, which is known as pacing-induced cardiomyopathy. There are two ways to reduce our risk. One is pacing algorithm to minimize RB pacing by prolonging LB interval to allow intrinsic conduction. Bottom panel shows RB pacing with standard LB delay. If patients have intrinsic conduction, device algorithm prolongs LB delay setting to avoid unnecessary ventricular pacing at the expense of prolonged PR interval. The other is conduction system pacing, which also minimizes our risk by allowing synchronized activation through the intrinsic conduction system. Both are good for ventricular function. However, it's not been determined which is better for atrial mechanics, and previous studies have shown an association between a prolonged PR interval, the risk of AR. Therefore, the aim of this study was to compare the effect of dual chamber left bundle branch pacing with standard and long PR interval on atrial remodeling. We hypothesized that long PR interval will have greater LB enlargements, low conduction and inducible AF than short PR interval. We used 10 minutes wine in this study. At baseline, we performed a NIPI study, pacemaker implantation followed by AB junction ablation and echo study. 10 peaks were randomly assigned to short and long AB delay setting in a one-to-one fashion. AB delay was set at 161 native for short and 350 for long AB delay. Every four weeks, we performed echo study. At 16 weeks, we did terminal study including echo study. We do NIPI study and conduction velocity measurement. In NIPI study, we measured ERP and AF inducibility. ERP was measured at three sites including right side appendage, atrial septum and CSOs. AF was captured at 170, not captured at 160. Also, we performed an AF induction three times from three sites, nine attempts for total. A significant AF was defined as AF lasting more than five seconds. Inducibility was calculated as a percentage of inducible significant AF among the total of nine induction attempts. We created short and long PR model by implanting pacemakers and AB junction ablation. Left panel shows atrial lead at right side appendage and ventricular lead at left bundle branch area. Looking at AB delay on the right panel, short AB delay is showing 160, long AB delay is showing 350, pacemaker duration is 74 and 72 respectively. In addition to standard echo measurement, we evaluated atrial function such as LLA reservoir strain and LLA desynchrony using speckle tracking echo. LLA desynchrony was defined as a time difference between the lightest and earliest segments. Let's move to ECG results. ECG data shows no significant difference in QRS duration pre- and post-pacemaker implantation, suggesting the ventricle is paced through the conduction system. Right panel shows paced QRS duration in both settings. Also, there is no significant difference in paced QRS duration between the two. Looking at the ventricular function, left panel shows LV desynchrony. Median value of LV desynchrony went within 10 millisecond in both settings. LVEF on the right panel, LVEF was preserved in both settings over four months. LVA area, short AB delay shows gradual increase in LVA area, which is not statistically significant. On the other hand, long AB delay setting shows significant enlargement over four months. Atrial function, left panel shows LV desynchrony. Short AB delay is showing no significant change in LV desynchrony from baseline to four months. However, long AB delay setting is showing significant increase in LV desynchrony from baseline to four months. LV strain on the right panel, again short AB delay is showing no significant change in LV strain from baseline to four months. On the other hand, long AB delay setting, we are seeing a clear drop in LV strain over four months. This is the EPS data at 16 weeks. Average ERP shows no significant difference between the two groups. And the long AB delay setting shows significantly slower conduction compared to short AB delay setting. This is also the data at 16 weeks. AF inducibility was 44 percent in long AB delay setting, but interestingly, no significant AF was induced in short AB delay setting. Median AF duration was 18 seconds in long AB delay setting. This is a representative case with long AB delay setting. Red corresponds to lateral, yellow corresponds to septal. So at baseline, LV strain was 36 and LV desynchrony was 12. A follow-up echo study at 16 weeks shows decrease in LV strain and greater LV desynchrony as shown here. When comparing short AB delay and long AB delay, long AB delay shows diastolic MR, which may cause atrial desynchrony and subsequent atrial remodeling. In conclusion, long AB delay leads to worse atrial functions, low conduction, and greater AF inducibility than conduction system pathing with a standard AB delay in this model. This also demonstrates that a prolonged PR interval may be causative and not just associated with adverse atrial remodeling and AF. Clinical implication of this study is that conduction system pathing with a standard AB delay may be better for atrial mechanics and the future risk of AF than allowing intrinsic conduction with a long PR interval. Thank you very much. Thank you. That was excellent. What do you think is the mechanism of the atrial desynchrony with the long AB delay? Thank you for your question. I think there are two possible explanations. One is the diastolic MR, which may cause volume overload and elevated pressure. That's one possibility. Our study shows that long AB delay setting shows significantly slower conduction, which may reflect fibrosis. So far, we have not obtained histology data, but the slower conduction may reflect inhomogeneous fibrosis, which may be associated with elevated desynchrony. The other is the timing of P-wave. In long AB delay setting, the timing of P-wave is overlapping the timing of T-wave. In other words, atrial contraction starts dualling ventricular function, which also leads to elevated pressure. Thank you. Thank you. Thank you very much. Okay, Mario, you're next. Mario Maloof. Okay. It's a pleasure. We are honored to hear your abstract, The Phenotypic Rescue of Genetic Disease, Rescue of Cellular Pathophysiology and Cardiac Function in Arrhythmogenic Cardiomyopathy. We're waiting to hear what you have to say. Thank you very much for this kind introduction. My name is Mario Malouf. I'm a postdoctoral research associate in the Shaw Lab at the University of Utah, and I will be presenting our work today on GJA120K and arrhythmogenic cardiomyopathy. So beginning with a brief introduction, arrhythmogenic cardiomyopathy is a hereditary disease. It arises due to mutations in desmosomal proteins such as placofilin, desmoglion, and desmoplacin. And as you can see in this diagram right here, all these proteins come together to form the desmosomes, which are proteins found at the intercalated discs, and they link adjacent cardiomyocytes together. And disruption in these desmosomes is the triggering event that leads to the cascade, which produces the ACM phenotype, which is the fibrotic phenotype, the pro-arrhythmogenic, and ultimately results in arrhythmias and sudden cardiac death. In order to study the disease, we used a desmoplacin mutant mouse model that was generated to, based on the Carvajal-Huerta syndrome in humans. And these mice have a frameshift mutation in the desmoplacin gene, so it produces a truncated form of the protein. And as is obvious from this image, the mutated mice have a very distinctive, disheveled appearance of their coat, and that gives them their name, Ruffled, R-U-L. In order to treat this disease, our approach was through GJA120K. Now, for the majority of the audience that doesn't know what 20K is yet, it's an internally translated isoform of Connexin 43. It's not generated by cleavage of the full-length protein, rather, the ribosome binds the center of the Connexin 43 mRNA, and it produces this isoform independently of the full-length one. And 20K has been shown to be essential for proper trafficking of the Connexin 43 subunits to the membrane, particularly the intercalated discs, through its ability to modulate the cytoskeleton. So now that we know that 20K is important in proper trafficking of Connexin, it would be interesting to see if we introduce 20K in a disease that has reduction in Connexin 43 trafficking, whether this has any therapeutic effect. And that's exactly what Joe Palatinis, former postdoc in the Shaw Lab, did in his study on a desmoglian form of arrhythmogenic cardiomyopathy. They showed that by introducing 20K to their desmoglian model, they were able to rescue Connexin 43 trafficking and reduce the arrhythmia burden in this disease. So this shows that 20K has efficacy and therapeutic efficiency in treating this form of arrhythmogenic cardiomyopathy. And now back to our study, we used a desmoplacan form of arrhythmogenic cardiomyopathy, this mouse model. We took the mice at four weeks of age, performed retro-orbital injections of AAV9, expressing either GJA120K-GFP or the control. And from this point onwards, every four weeks, we followed up these mice with ECHOs and EKGs up until 20 weeks of age, after which we harvested the hearts and performed subsequent imaging and biochemical assays. Looking at the ECHO data, we can see that the ejection fraction was similar between the GFP-treated desmoplacan mutant mice and wild-type mice at four weeks of age. So they have a similar baseline. But starting at 16 weeks of age, there's a significant decrease in the ejection fraction of the GFP-treated desmoplacan mutants, which persisted up until 20 weeks of age. And this is consistent with ACM disease development. The mice developed heart failure as a result. However, when we treated these mice with 20K, we did not see any decrease in their ejection fraction throughout the experimental period. Rather, their ejection fraction remained constant, and it was similar to that of wild-type mice. And here we can see the ejection fraction at the weeks 16 and 20, to show that we have a significant decrease in the ejection fraction of GFP-treated mutants. However, 20K preserved the ejection fraction. Now, we also performed trichrome staining on the hearts. And as you can see in the image, there's significant and extensive fibrosis in the GFP-treated desmoplacan mutant mouse hearts, which is expected with development and progression of arrhythmogenic cardiomyopathy. However, upon treatment with 20K, the hearts looked very similar to the wild-type mice, and there was no fibrosis. These images were analyzed using ImageJ, and we did color deconvolution to quantify how much fibrosis is present in these hearts. And the data matches what we see on the slides. We have an increase in fibrosis in the GFP-treated desmoplacan mutants. However, 20K was able to prevent the development of fibrosis. Now, at the cellular level, we first wanted to confirm what we saw in the desmoglion model of arrhythmogenic cardiomyopathy that 20K is able to rescue connexin trafficking. So, in these same hearts, we did immunofluorescent imaging with N-cadherin in red, delineating the intercalated disc, connexin in green, and their overlap in yellow indicates proper trafficking of connexin to the intercalated disc. So, comparing the wild-type cardiomyocytes to the GFP-treated desmoplacan mutants, we see a significant reduction in the connexin signal at the intercalated disc, which is expected with other arrhythmogenic cardiomyopathy because it impairs connexin trafficking. However, treatment with 20K was able to rescue connexin trafficking and restore connexin back to the intercalated disc. Now, our data in this desmoplacan model with the previous data in the desmoglion model show that 20K is able to rescue connexin trafficking in a manner that's independent of the underlying genetic mutation, so it gives us a broad therapy for ACM. However, rescuing connexin doesn't really explain why we saw a preservation of ejection fraction or prevention of fibrosis, so I'd like to shift focus now and talk about another important pathway that gets disrupted in arrhythmogenic cardiomyopathy, and that's the beta-catenin signaling pathway. So normally, beta-catenin is present at the intercalated disc bound to E-cadherin, part of the cadherin complex, which similarly to desmosomes links adjacent cardiomyocytes together. This beta-catenin is able to dissociate from the cadherin complex, go to the nucleus where it acts as a transcription factor, and then cardiomyocytes produce a pro-survival and pro-myocyte phenotype, and some of it gets degraded by the beta-catenin destruction complex, and that's really the normal cycle of what happens to beta-catenin. However, when we have disruption of the desmosomes by mutations, in the case of arrhythmogenic cardiomyopathy, there is release of plaquoglobin from the desmosome. And as the Merian group showed around 20 years ago, plaquoglobin is antagonistic to beta-catenin. Plaquoglobin causes increased degradation of beta-catenin and it inhibits the gene expression of beta-catenin-dependent genes. And this is a major contributor of the pro-fibrotic and adipogenic phenotype that we see in arrhythmogenic cardiomyopathy. So knowing that 20K was able to rescue connexin, we wanted to investigate whether it had any effect on beta-catenin. Similarly to what we did before, we did immunofluorescent imaging for beta-catenin at the intercalated disk with N-cadherin in red, beta-catenin in green, and again, their overlap in yellow indicates we have beta-catenin at the intercalated disk. And comparing the wild-type cardiomyocytes to GFP-treated desmoplacan mutants, we see a significant reduction in beta-catenin at the intercalated disk. However, treating these desmoplacan mutant mice with 20K was able to restore beta-catenin at the intercalated disk, showing there's rescue. Now beta-catenin doesn't just exist at the intercalated disk, it goes to the nucleus, so we performed nuclear fractionation on the heart lysates to test if it's able to rescue also beta-catenin localization to the nucleus. First, we have the cytoplasmic fraction. There's no major difference here between all three groups, and that's what we see in the literature. However, looking at the nuclear fraction, we see a significant reduction in the amount of beta-catenin present in the nucleus. However, treating the mice with 20K was able to bring the beta-catenin back up in the nucleus to levels similar to that of the wild-type mice. So we have evidence that beta-catenin is rescued in ACM, both at the intercalated disk and within the nucleus, when we introduced 20K, despite the desmoplacan mutation causing the disease. We also wanted to take this a step further and perform a functional assay to check for whether this improved rescue and localization had any functional consequences, so we performed in a cell model a beta-catenin-dependent luciferase assay. We took these cells, knocked down desmoplacan, and then measured the luminescence as an indirect way to measure beta-catenin activity. So the more luminescence we saw, the more luciferase was present, so it was produced by beta-catenin activity. And the cells were treated with either our control GST-GFP or GJA120K-GFP. Looking at the controls, there's really no difference in either GST or GJA120K, they both have the same level of beta-catenin activity, but when the cells were knocked down, the DSP was knocked down in the cells and they were only given GST-GFP, there was a significant reduction in beta-catenin transcriptional activity. And the introduction of 20K in that system was able to partially rescue beta-catenin transcriptional activity, indicating that it does have a functional consequence. So this increased nuclear translocation is accompanied by an increase in beta-catenin functional activity. So taken together from a mechanistic point of view, what happens in arrhythmogenic cardiomyopathy is disruption of the desmosomes leading to instability of the desmosomes, so we have a reduction in connexin-43 trafficking and antagonism of beta-catenin signaling by placoglobin, which produces this pro-fibrotic, pro-adipogenic phenotype. However, with treatment using 20K, we were able to rescue connexin trafficking to the intercalated disk and we were able to prevent beta-catenin degradation and promote its nuclear translocation and activity. So the major takeaways are that 20K restores connexin trafficking and beta-catenin signaling in this DSP model of arrhythmogenic cardiomyopathy. 20K preserves cardiac systolic function and prevents fibrosis in our model. And taken from a global perspective, combining the desmoglion and desmoplacan model that I've just showed you, we can say that gene therapy is able to correct trafficking disturbances in a manner that's independent of the underlying genetic mutation. And with that, I would like to thank my lab and all the members for their help. It wouldn't have been possible without them and I'm open for any questions. I have a naive question. You may have answered it. Is the increased trafficking also phenotypically seen in function and or reversal of fibrosis or is that a naive question? Could you explain a bit more? Once you give the 20K and you have increased transcription and increased trafficking, can you reverse the damage that you had seen to the... So in our experimental model, we weren't targeting reversal. We were more focused on prevention. We haven't given 20K after we've established the disease has developed, but it's something we're working on. Do we have any more questions? Okay. Thank you. Thank you. Excellent. Thank you. Sure. And finally, let's see. Mustafa, I mean, despite the fact that you work with Miguel Vandora, we're going to let you talk. Thank you. You're an excellent mentor. I've known him well and we look forward to hearing your work. Thank you so much. Congratulations for being part of the presentation. Thank you. Appreciate it. Thank you. Hello, everyone. I'm Mustafa. I'm Mustafa, a second year cardiology fellow at the Houston Methodist. I'll be presenting to you a study about venous ethanol ablation as the sole treatment strategy for intramural ventricular arrhythmias. A brief overview, intramural ventricular arrhythmias, especially those arising from the LV summit, are often inaccessible by RFA. There are a few limitations for RFA, including inability to achieve the necessary depth to target the intramural substrates. And there's also a risk of damaging the nearby coronary arteries or the phrenic nerve. Since the left men usually bifurcate at the LV summit region, applying RFA in the LV summit always comes with a risk of causing damage to the coronary arteries. However, venous ethanol ablation can help achieve more deeper and targeted ablation without the risk of damaging the coronary or the phrenic nerves. Our study objective was to evaluate the safety and the efficacy of venous ethanol-only ablation as a standalone ablation strategy for intramural ventricular arrhythmias. We had two key questions. Can venous ethanol-only ablation eliminate ventricular arrhythmias without the need for RFA? What are the clinical outcomes and complications? So this was a retrospective multi-center study. It included 49 patients at the Houston Methodist Hospital and AZ Xinjiang Hospital in Belgium. So all patients who come to our centers get LV and RV and often CS mapping. If the earliest signal was identified in the endocardium or the epicardium, these patients underwent RFA ablation and were not included in our study. However, if the patient has earliest signal in the CS branches with poor and delayed signals in the endocardium or epicardium, these patients were considered to have intramural substrate. So we applied venous ethanol ablation in all of these patients and we tried to induce the tachycardia after the ethanol infusion. And if the tachycardia or the PVC was not inducible, no RFA was delivered. Regarding the venous delivery technique, so ethanol was infused in one millimeter increment over one to two minutes. We used single balloon technique to target smaller veins with minimal collaterals. Double balloon technique was used for larger veins or when the signal was in the mid vein and we wanted to prevent the ethanol spill over. Sequential delivery was utilized when two adjacent veins had optimal signals. This is the baseline characteristics in our study. So the mean age was 58 years. 73% were males. HFRIF was present in 37% of patients. Structural heart disease in 41%. ICD in 41%. The median left ventricular ejection fraction was 55%. Class 1 or class 3 antiarrhythmics were used in 45%. The indication for the ablation was PVC in 67% and VT in 33%. Prior RFA was performed in almost half the patients and the other half underwent de novo venous ethanol ablation. The arrhythmia mechanism was focal in 73% and substrate in 27%. The arrhythmia origin was from the LV summit in 90% and lateral left ventricle in 10%. Regarding the procedural parameters and lesion assessment we used the contrast venography and the ice ecogenicity to give us an idea about the lesion assessment during and after the venous ethanol ablation. The median ethanol volume that was delivered in our study was 7 milliliters. The median ice ecogenicity volume that was seen after ethanol delivery was 2.5 milliliters. And this is an example of ice ecogenicity as you can see in this image. Targeted veins were AIV branches in 82% and GCV lateral branches in 18%. In PVC patients the median intramural venous signal was minus 40 millisecond preceding QRS compared to only minus 8 millisecond in the endocardium or epicardium. Pace map matching was 97%. Regarding the clinical outcomes in the PVC patients we were able to reduce the PVC burden from a baseline of 22% to only 0.3% after ablation. Only two patients had recurrences. Both were treated with redo ablation. Regarding the VT patients before the ablation 14 out of the 16 patients required ICD therapy and after ablation only 4 patients required ICD therapy. So we had a total of four recurrences. One required redo ablation but as you can see in the bottom right figure even in patients who had recurrence there was a significant reduction in the ICD therapy and the VT episode. On the left side you can see the Kaplan-Meier for the freedom from ventricular arrhythmia. This is combining both PVC and VT patients. On the right you can see the antiarrhythmic drug use that was significantly reduced after ablation. Almost a total of 22 patients before the ablation were on antiarrhythmics and only 8 after ablation remained on antiarrhythmics. Regarding the cardiac MRI data the scar volume at baseline was 5.5 milliliter and after ablation it went up to 11.5 milliliter. All MRIs shown microvascular obstruction in the targeted zones. The median left ventricular ejection fraction did not change before or after ablation and the median was 55% either before or after. No structural complications were seen, no VSD, no pseudoaneurysm. This is a very nice picture on the bottom right. You can see the right picture shows the microvascular occlusion that was seen one day after the venous ethanol delivery and the one on the right shows the lesion on the MRI after 40 days from ablation. The area of the microvascular occlusion has turned into a scar. Regarding the procedural complications we had three patients with postoperative pericarditis. One patient experienced a groin hematoma and one patient or two patients had transient right bundle branch block. This is one of the examples on a patient who underwent venous ethanol ablation for LV summit PBC. On the top left you can see the venous ethanol, the earliest signal between the GCP and the AIV junction on the activation map and we placed the ectopolar catheter in S1 and we found a very nice signal that preceded the QRS by 52 ms. We placed the wire and the balloon and then we delivered alcohol. Bottom right you can see the area of contrast venography under etch and above it you will see the ethanol equigenecity on cartomap. This is another example of a patient who underwent ethanol ablation for large epicardial VT substrate. This patient had sarcoidosis and you can see the baseline on the top left. The patient's scar was mainly epicardial. We placed the decapolar and the octopolar catheter in the lateral vein and we found a very nice fractionated potentials and mid-diastolic signals. We delivered ethanol and on the top right you can see the post-ethanol delivery voltage map. We were able to achieve scar homogenization. In this patient we had to use both single and double balloon technique. On the bottom right you will see the MRI before and after. The baseline is on the top and the post-ethanol ablation is on the bottom. I don't know if you can see the black area on the lateral wall. This represents the microvascular occlusion after the ablation. In summary in our study we were able to show that venous ethanol only ablation was able to significantly reduce the PVC burden, the VT burden and associated ICD therapy without significant complications. In conclusion venous ethanol only ablation is a reasonable first-line strategy to target intramural ventricular arrhythmias. Excellent presentation. Do we have any questions for Dr. Amin? In that case, thank you everyone for attending and thank you everyone. Good luck.

Graph Neural Network

anticoagulation

atrial fibrillation

electronic health records

CHADS-VASC

stroke prediction

bleeding risks

personalized treatment

clinical decision-making

medical AI