So, let me start this session entitled Advances in Antiarrhythmics and ECG Technology. My name is Wataru Shimizu from Tokyo, Japan. And let me introduce the first speaker, Dr. James Lip. His title, Treatment, Satisfaction, and Relief of Symptoms with Self-Administered Erythropermial Nasal Spray Analysis of Phase III Randomized Trials. Dr. Lip, please. Thank you, Dr. Shimizu. On behalf of my co-authors, it is my honor to share with you the results of our study of a treatment, satisfaction, and relief of symptoms using self-administered Erythropermial Nasal Spray Analysis of these Phase III studies. My name is James Lip from Weill Cornell Medicine. As full disclosures, I am an investigator in these trials as well as a steering member for Milestone Pharmaceuticals. About 300,000 new PSVT diagnoses are made each year. About a quarter of them require an emergency room visit or hospitalization. PSVT significantly impacts quality of life, causing symptoms such as weakness, fatigue, lightheadedness, fainting, palpitations, chest discomfort, shortness of breath. And often patients need to seek emergency care, which increases the cost for healthcare utilization. The current treatments, if PSVT continues, requires clinical intervention, often requiring an intravenous dose of adenosine, beta blocker, or calcium channel blocker. This highlights the need for innovation for PSVT treatment. Atropamil is a fast-acting, non-dihydropyridine L-type calcium channel blocker that has been in development for use in a medically unsupervised setting. It has a rapid onset of action after intranasal delivery, achieves a peak plasma concentration within seven minutes, as shown by this red graph. It is inactivated by blood esterases, and its pharmacodynamic profile is shown by the PR prolongation here, more than 10% of baseline, lasting up to 40 minutes. It has been developed to satisfy this unmet need for self-administered treatment to be used outside the healthcare setting, specifically for patients with AV nodal-dependent PSVT episodes, when a vagal maneuver is ineffective. There have been two randomized control studies using atropamil, NODE-301 Part 1, involving 419 patients, and NODE-301 Part 2, or the RAPID study of 692 patients. Both were event-driven, randomized, double-blind, placebo-controlled trials to evaluate the efficacy and safety of atropamil in patients experiencing PSVT in a medically unsupervised setting. Patients were randomized in NODE-301 Part 1, 2 to 1, to atropamil or placebo. At the onset of symptoms, patients applied an ECG monitor, performed a vagal maneuver, and if their symptoms continued, they self-administered atropamil or placebo. In RAPID, 692 patients had the option to repeat a second dose of study drug after 10 minutes if their symptoms persisted. RAPID also included an open-label period, and both of these studies were pooled in this analysis I'll present to you today. Our primary endpoint was conversion to sinus rhythm within 30 minutes of treatment, and we looked at a secondary endpoint of treatment satisfaction among patients who were receiving study drug. Patients were specifically asked, regarding their symptoms, of rapid pulse, palpitation, shortness of breath, chest tightness, pain and pressure, feeling dizziness, lightheadedness, or anxiety, and specifically asked the treatment satisfaction questionnaire for medication, or TSMQ-9 question number 2, which was how satisfied or dissatisfied are you with the way the medication relieves your symptoms on a scale of 1 being extremely dissatisfied to 7 being extremely satisfied. The pooled analysis of the NODE-301 and RAPID were shown here among 370 patients who received a trypamil or placebo, 221 the trypamil, 149 with placebo. Patients were able to terminate their PSVT as confirmed by the ECG monitor by 30 minutes in 60.2% in patients receiving a trypamil compared to 31.9% with placebo, with a hazard ratio of 2.385. The median time to inversion with the trypamil was 20.1 minutes compared to 51.6 minutes with placebo. In regard to the patient satisfaction scores across all 6 symptoms of rapid pulse, palpitation, shortness of breath, chest discomfort, feeling lightheaded or anxiety, it was a statistically significantly higher treatment patient satisfaction scores with a trypamil compared to placebo. Here's a summary of the adverse events across patients who received a trypamil, either single dose or double dose, and about half of patients had some reported adverse event, predominantly were mild to moderate in severity. There were no serious treatment inversion adverse events related to the study drug and very few patients actually discontinued the study because of these events. The most common complaints were nasal discomfort or congestion from the nasal spray, few patients experiencing epistasis or throat discomfort. This high efficacy rate as well as the high treatment satisfaction likely explains the results of the pre-specified pooled analysis of these two trials in regard to emergency room visits. There was a statistically significantly lower risk, a 39% risk reduction in this pooled analysis of node 301 and rapid for emergency room visits. And this was recently published this month in JAMA Cardiology. In conclusion, the symptom prompted treatment of a trypamil for PSVT demonstrated clinical efficacy for episode conversion and significantly higher patient satisfaction with symptom relief compared to placebo using a trypamil. Regardless of the symptoms experienced, patients received reported greater satisfaction with the trypamil compared to placebo and these data support the continued use of a trypamil for self-treatment of PSVT in a medically unsupervised setting. Thank you very much. Thank you. There is no, so far no questions. Are there any questions? This paper is open for discussion now. So I have one question. In 32% of the patient with placebo, their PSVT was terminated. Yes, at 30%. So it was spontaneously terminated within the 30 minutes? Yes. So the difference is 30%? Yes, the absolute difference is 30% relative to 2.3 times. Any question? No question? Okay, and also there is not so much difference between the satisfaction between the true drugs and the placebo. So how do you explain the small difference? I think that it was statistically different. I would think that with a successful termination, satisfaction scores would be much higher. We're going to do a sub-analysis of patients who actually, these are just patients who received the study drug. Whether or not they terminated or not was not necessarily included in the satisfaction score. I'm sure if we looked specifically at whether when patients did terminate, what their satisfaction score, I'm sure they would be much higher, but the bigger difference between those that did not terminate. Is there any adverse event for this study? Adverse events, yes. So I'm going to show that slide regarding the adverse events. Yes, there were adverse events. I mean, these are any adverse events. So if patients complained of nasal irritation after administration, that was an adverse event. So most of these, as I said, were mild or moderate. There were none that were serious, like zero or half percent considered severe, but there were no episodes of AV block as confirmed by an ECG monitor, no syncope. So these are all the key ones that we'd be interested in in a drug that is a calcium adrenaline blocker that terminates PSBT. So only the minor side effects? Only minor. Nasal discomfort or something like that? Yes. Any questions? Any comments? Okay. So thank you so much. Thank you. This way, right? Okay. The next speaker is Ms. Reina Tonegawa-Kuji. The title of her talk is SGLT2 inhibitor and GLP-1 agonist, association with atrial fibrillation incidents and mechanistic insights. So please. Thank you, Dr. Shimizu, for kind introduction. Okay. So the SGLT2 inhibitors and GLP-1 receptor agonists are cardioprotective agents that were originally approved for type 2 diabetes. And the association between SGLT2 inhibitors and the reduced incidence of AF has been demonstrated in both meta-analysis of RCTs and large-scale observation studies, including analysis using Medicare data. However, the underlying molecular mechanisms are not fully elucidated. Given the low cardiac expression with SGLT2, off-target or secondary effects are considered likely. And here I show the data from GTEx, which demonstrates that SGLT2 is specifically expressed in kidney with minimal expression in the heart. So on the other hand, the association between GLP-1 receptor agonist use and AF incidence remains inconclusive. A meta-analysis of RCTs suggested that GLP-1 receptor agonist use is associated with a reduced AF incidence. Some observational studies support this finding. However, a study using Medicare data concluded that GLP-1 receptor agonists may be less effective than SGLT2 inhibitors in preventing AF. This is the GTEx data, and GLP-1 receptor is weakly expressed in atrial appendage and left ventricle. However, direct effects of GLP-1 receptor agonists on cardiomyocytes remain unclear. So based on these findings, our study had two main aims. Aim one is to re-evaluate the association between use of SGLT2 inhibitors and GLP-1 receptor agonists and the incidence of AF using a data warehouse based on electric health records. And aim two is to explore the molecular mechanisms underlying AF reduction by these drugs through gene expression analysis in atrial-like engineered heart tissues. So first, we evaluated the effect of GLP-1 receptor agonists on SGLT2 inhibitors using an electric health record database. We used a Northwestern data warehouse, which includes data from 11 hospitals in Illinois between 2011 and 2022. From the data set, we extracted the patients who were prescribed either SGLT2 inhibitors or GLP-1 receptor agonists or comparative drugs. And then the index date was the first prescription date of the drug. And then the one-year period prior to the index date was used as a baseline period, during which the baseline characteristics were collected. And patients with type 2 diabetes were included, while those with type 1 diabetes, gestational diabetes, and previous AF diagnosis were excluded. And patients were followed up for 180 days for the diagnosis of AF. And propensity scores were calculated using logistic regression based on the probability of receiving the drug of interest. And one-to-one propensity score matching was then performed to balance baseline characteristics. And then after matching, a logistic regression model was used to assess the association between the drug use and the incidence of AF. And then sample size for each cohort are shown here. Because the study period was slightly earlier, the number of patients in the GLP-1 receptor agonist group was relatively small. As a result, in the propensity score matched cohort, only about 600 patients were included in the GLP-1 receptor agonist group. So these are the results of logistic regression analysis in the PS-matched cohort. Compared with the comparator drugs, associated inhibitors were consistently associated with a lower incidence of AF. And GLP-1 receptor agonist use was also associated with a lower incidence of AF compared with TZDs or DPP-4 inhibitors. However, no significant association was observed when compared with esophonia ureus. And no significant interaction was observed between drug treatment and obesity. Next, we used atrial-like engineered heart tissues, EHTs, to investigate the transcriptional changes induced by SJ-82 inhibitors or GLP-1 receptor agonists. Using human iPSC cells derived from a healthy male donor, we generated atrial-like engineered heart tissues. And then the EHTs were treated with either mPAG glyphosate or liraglutide for 24 hours. During the final six hours, periodic pacing was applied as a stressor. We then performed bulk RNA-seq analysis. So these volcano plots illustrate dose-dependent changes in gene expression following treatment with drugs. Treatment with either drug did not significantly change the gene expression pattern. So we repeated the experiments without pacing and observed similar findings. So we assessed the expression levels of SAC582 and GLP-1R in EHTs. So SAC582, the gene encoding SJ-82, was barely expressed, while GLP-1R showed weak expression. So the target protein of SJ-82 inhibitors were minimally expressed, and no significant transcriptional changes suggestive of secondary or target effects were observed. And although GLP-1R was weakly expressed, short-term treatment with liraglutide did not induce significant changes in gene expression. So here we summarize the key findings of the study. An EHR-based study analysis suggests that the use of SJ-82 inhibitors and GLP-1R agonists is associated with a lower incidence of AF. However, no differentially expressed genes were detected in EHTs following treatment with mPAG glyphosate or liraglutide. So there are several possible explanations for the lack of observable transcriptional changes in our model. First, ATR-like EHTs are primarily composed of cardiomyocytes. Therefore, it's possible that the drug effects occur through non-cardiomyocyte cells, such as fibroblasts or adipocytes, which are underrepresented in our cardiomyocyte-enriched model. And second, the mechanisms may be non-transcriptional. For example, modulation of calcium and CAMK2 is a known pathway that wouldn't be captured through transcriptional analysis. And third, systemic effects may predominate. SJ-82 inhibitors and GLP-1R agonists may primarily act through other organs, like the kidneys or the liver, and their direct effects on cardiomyocytes could be limited. And as a next step, we aim to compare responses between healthy and disease-model tissues to better understand context-specific effects, particularly in AF-related pathophysiology. So these are team members. I'd like to especially thank Mitari, who did the experiments, and then Chen-Hsuan, Dr. Luo, who did the EHR analysis, and then our outstanding mentors, Dr. Mina Chang, Dr. John Burnham, Dr. Pei-Hsuan Chen, for their invaluable supervision. Thank you so much for listening. Thank you very much. So this paper is now open for discussion. Any comments or questions? Please. How about the cardiometabolic effect by the GLP-1 for reducing the weight, for reducing the risk of AF? Yes, GLP-1 receptor agonists are known to reduce weight, so it's totally possible that it reduces adipocytes in epicardium. So it would be probably interesting if we take, for example, sequential cardiac MRI, cardiac CT, something like that, to see if there's any association with reduction of adipocytes in epicardium and AF instance, something like that. Thank you so much. Any questions? So in your atrial ligand gene at heart tissue, so you didn't find the difference of gene expression between the acidity tissue and the… Yes. No, we didn't. Yeah. So how long did you exposure the drugs? How? How long? How long? Yeah, yeah. It was 24 hours. 24 hours is enough time? It's usually enough, because if we treat the EHTs with, for example, other drugs, such as dandrurium or metformin, we see a lot of DEGs, differently expressed genes. So it was kind of surprising that we didn't see any DEGs in this experiment. Any comments, questions? Please. Please come to the microphone. Just two brief questions. The first one is, to me, surprising the effect with such a short follow-up, 180 days. So I don't know if you have any explanation to justify these effects in such a short period of time. The other thing you probably already answered, that you did the experiment, the experimental part, you didn't compare with no-drugs cells. You compared one with one of the other two drugs, but placebo cells were not analyzed. So for the first question, the reason that the follow-up period is a little bit short is to ensure that we have enough patient number, because at that time, GLP-1 receptors were, especially GLP-1 receptor agonists, where prescription was a little bit limited. If we try to follow up longer, then the sample size will be pretty small. For this dataset, we decided probably 180 is a good range to make sure that we have enough patient number. For the second question, what was the second question? Oh, okay. So that result is a dose-dependent change. We did 0, 10, 50, 100. So we actually compared with no-drug as well. Yeah, yeah, yeah. Any questions? Okay, thank you very much. So let's move to the third paper. The presenter will be Dr. Akash Seth. The title of his talk is, A Long-Term Safety Profile of Class 1-0 versus Class 3 Antiretroviral Drugs in Elderly Patients with Coronary Artery Disease. Results from a large nationwide database. Please. All right. Good afternoon, everyone. Thank you for being here. I'm Akash Sheth. I'm a third-year cardiology fellow at UPMC Harrisburg, and I'm going to be talking about the long-term safety profile of class 1C versus class 3 antiarrhythmic drugs in patients more than 70 years of age with preexisting coronary artery disease. I have no disclosures. So class 1C antiarrhythmic drugs were approved for treatment of supraventricular tachycardias and refractory ventricular tachycardias. But despite the success of these drugs in suppressing arrhythmias, early safety studies demonstrated now well-known proarrhythmic effects and increased mortality with flecainide in certain patients with severe arrhythmias and structural heart disease. And most of these recommendations come from the cardiac arrhythmia suppression trial, also known as the CAST trial, which raised the concern that class 1C agents could provoke life-threatening arrhythmias in post-MI patients. But unfortunately, the findings of CAST were extrapolated to all elderly patients, especially those with coronary artery disease and structural heart disease, to a point where certain current guidelines also recommend restricting the use of these class 1C agents in presence of structural heart disease, including coronary artery disease. But there have been several small single-center studies that have established the safety of class 1C agents in stable coronary artery disease, and these have come from throughout the entire country, Ochsner Group in Louisiana, Emory Group in Atlanta, Mayo Groups in Arizona and in Rochester, and several other centers in Europe. However, a large nationwide data on outcomes answering this specific question, which is class 1C agents in AFib in elderly patients more than 70 with CAD are lacking. That is the objective of this study, is by using a national database, we sought to evaluate the safety and feasibility of treatment with class 1C agents compared to class 3 agents in patients aged more than 70 with coronary artery disease, but without history of acute coronary syndrome, ventricular tachycardia, and ICDs implantation. And for the purpose of this study, we also excluded patients on Amiodarone and Multac. So this was a retrospective, propensity-matched cohort analysis, we used the Trinatex network database. Trinatex is a large multicenter health research network primarily based out of the U.S. This network aggregates all the information from electronic health records, and it has information of over 250 million patients, involving 120 healthcare organizations. So as I said earlier, we included patients more than 70 years of age, with a history of AFib and coronary artery disease, from January 1, 2010 to December 1, 2024, using appropriate ICD-10 codes. Patients were further subdivided into two groups based on the use of class 1C versus class 3 agents. We excluded patients with ACS, acute coronary syndrome, ventricular tachycardia, ICD implantation, and those patients on Amiodarone and Dronadarone. So cohorts were matched using propensity score matching, using a built-in algorithm in the Trinatex database. So the cohorts were divided into two groups, those on class 1C and those on class 3 antibiotics, and then statistics were performed in the usual manner, where continuous variables were analyzed using t-test, categorical variables were analyzed using chi-square test, survival analysis was performed using the Kaplan-Meier curves, and the entire analysis was done on Trinatex online platform, which uses R for statistical computing. So prior to propensity score matching, we had 16,194 patients in class 1C group and 21,308 patients in the class 3 group. Prior to propensity score matching, each of the groups had 13,354 patients. And we like to believe that the propensity score matching was done as extensively as possible through an online database where we matched the age, sex, race, prior history of AFib ablation, their A1C and LVF, along with 15 other comorbidities. Here are the outcomes. The primary outcome was a composite of all-cause mortality, cardiac arrest, syncope, and heart failure readmission. Secondary outcomes were all-cause mortality, cardiac arrest, syncope, and heart failure readmission looked individually. So this is a busy slide, but we can go through the tables together. Basically after the propensity score matching, we can see that both the groups were quite comparable. The mean age in both the groups was about 75 years. Race and sex were also adequately matched, as were all the other comorbidities. Both the groups had preserved ejection fraction of about 59% to 60%. And after propensity score matching, both the groups were similar and had—were essentially similar and had no residual imbalance. And the standard difference was less than 0.1 for each of the covariates. Coming to the results, so these were the five-year results we looked at. And at five-year, we can see there was a 5.2 percentage point difference in the primary outcome, which was a composite of all-cause mortality, cardiac arrest, syncope, and heart failure readmission. As you can see, it was 12.7 in class 1C group versus 17.4% in the class 3 group. And looking individually, the all-cause mortality, cardiac arrest, and heart failure readmission was also lower in the class 1C group. Syncope was higher here compared to the class 3 group, but overall rates of syncope were low across the board with about 1.4 and 1% in the groups, respectively. Surgical analysis was performed with a cumulative percentage comparison of all-cause mortality, cardiac arrest, syncope, and heart failure readmission amongst patients with class 1C and class 3 agents. And this was compared using Kaplan-Meier method, and log-rank test was used to compare the outcomes. You can see here the curves separate gradually, and there's, you know, a little bit of—there's a survival benefit on patients with class 1C agents compared to class 3. So obviously, you know, our study is not without limitations. This is—the first limitation is presence of unmeasured confounding factors in this observational study. Additionally, Trinitex databases, like most other databases, relies very heavily on documentation of diagnostic codes using ICD-10 codes. As inherent to all the other databases, there is a limitation when it comes to miscoding. However, given the large number of patients we had in each group, we can safely say that the above-mentioned limitations would be equally divided between the two groups. Additionally, the database does not account for patients who follow up with healthcare organizations that did not participate with Trinitex. Again, you know, those differences should be adequately divided between the two groups, but nonetheless, that is a big limitation. We also know that the biggest limitation of our study comes from lack of granular details because the diagnosis of CAD and heart failure were based on diagnosis codes. So we were unable to know the distribution of coronary artery disease and the possible size of ischemia, if there was any, in either of the groups because there was no information on stress imaging. We were also not able to subdivide patients on type of coronary artery disease, such as obstructive or non-obstructive, because there are ICD-10 codes available for these, but it has been well-documented throughout literature that the codes for obstructive versus non-obstructive CAD are not reliable. Therefore, although the missing data are missing completely at random, you know, there is a potential for significant bias. So we would like to say that, you know, our data should be viewed as hypothesis-generating for further prospective studies. In conclusion, this is one of the largest sets of retrospective data in elderly patients. Our data suggests that in elderly patients with AFib aged more than 70 with low-risk coronary artery disease, the use of class 1C agents was not associated with worse outcomes compared to class 3 agents. And this data would serve as an avenue of opportunity for patients who are frequently—in which this drug is—class 1C are frequently restricted. So this study was aimed to answer a practical question that we encounter in day-to-day practice, but because this is a large database retrospective observational study, we—these results should not be used to acutely change your practice, but rather serve as a nidus for more prospective larger clinical trials looking at this topic. And that's it. Thank you so much. Thank you. Thank you very much for a nice presentation. So first of all, I have one question. So as a second endpoint, the syncope is slightly lower in class C compared to class 3. How can I explain that opposite direction? So the syncope was—it is interesting because it was—the difference was not a big difference, but it was nonetheless statistically significant. Now syncope could be because of a lot of other reasons in addition to cardiac syncope. So we don't know if 1% was because of non-cardiac syncope as well, because syncope—as I said, this was—syncope has one code, so maybe this was cardiac syncope, but most likely as with most other syncopes, it's most likely non-cardiac in nature. And there's no way of—with Trinitex and databases, there was no way of accurately distinguishing cardiac from non-cardiac syncope. So any question, any comment for this paper? So your study proposal include only stable chronic coronary disease, but actually it is very difficult to exclude their acute ischemia. Correct. Yeah, so that's a major limitation. So it is a limitation. The only thing was because even though we are not able to exclude ischemia, we are still seeing that patients with mostly—13,000 patients—mostly stable CAD, they still had better outcomes or, you know, not worse outcomes than class 3. So even if there was ischemia, a little bit of ischemia, class 1C agents were also probably okay as long as—because we took mostly patients with preserved EF. So as long as EF is preserved and if there is a little bit of ischemia, it would actually be on the worse side for class 1C, which we didn't see. So, you know, all the more evidence that we should not really restrict class 1C agents and most of the patients we actually, you know, just because they're elderly with high calcium score does not mean they would not benefit from class 1C. That was the whole point. Any question, any comment? Okay. Please come. Why did you decide on excluding patients with acute coronary syndrome? So that was already answered by the CAST trial. So CAST trial was post-M.I. patients—obviously, this is not the same population, but post-M.I. patients, you generally don't want to prescribe class 1C because there has been a history of obstructive CAD, and those patients usually do worse, and that was already answered in prior trials, like CAST trials. So—and in clinical practice, nobody is going—you know, in post-M.I. patients, it's very difficult to prescribe class 1C agents anyways. So that's why. Okay. Thank you very much. Thank you. So the next speaker will be Dr. Jason Andrade. The title of his talk is Artificial Intelligence vs. Electrophysiological Adjustment of Atrial Esmeralda, a Validation Study. Please. Hello. Thank you for coming and listening to the presentation. It's my pleasure to be here on behalf of the rest of the research team to present these results. And so, as you heard, our study is Artificial Intelligence vs. Electrophysiologist Adjudication of Atrial Arrhythmias in a Clinical Trial, and so this is a validation study. So we all know that wearable and implantable ECG monitoring devices are commonly used for the diagnosis of arrhythmia outcomes when we perform trials. Clinically, we use this to guide the management of our patients, and so we often use these ECG tracings to make determinations about how we're going to treat them. In trials, when we use these wearable implantable devices, we know that they're generating lots and lots and lots of data. So for example, if you were to do a one-year follow-up trial with a loop recorder, you could generate anywhere up to 30,000 arrhythmia episodes that then require adjudication. And so if you do standard adjudication with agreement, then you need double adjudication, so you're now at 60,000, and then with discrepancies, you're now adding another 10,000 adjudications. This is obviously a very costly endeavor and very labor-intensive because we know that the ECG monitoring can generate a mix of true positives and false positives, and so we really do need to confirm the endpoints that are recorded by these devices. As we move towards the future, we know that artificial intelligence is a great opportunity to try and democratize how we do trials. There are many algorithms that have been developed through deep learning and machine learning that offer the opportunity to provide more accurate assessment of arrhythmia outcomes. These AI algorithms can annotate and classify this device-detected ECG data and can provide some overread of some of the rhythm episodes to provide improved accuracy based on the information within the devices themselves. One hope is that we would have automated AI-guided analysis that could help bring us forward into the future as a cost-effective way to perform trials and improve outcomes as we move forward. The purpose of the current study or analysis was to look at the BeatLogic algorithm. BeatLogic is a cloud-based deep learning algorithm. It employs multiple deep learning algorithms with multiple layers to it to annotate the individual beats and provide interpretation of the arrhythmia that's seen. This is a software component of the BodyGuardian remote monitoring system. It has 35-beat rhythm and wave classifications through the five deep neural networks in an attempt to try and provide better adjudication of the arrhythmia events that are happening. This has been trained on greater than 10 million beats from more than 44,000 patients, so a large dataset to inform the outcomes that we have. What we did within this study was try and evaluate the performance of this AI-guided algorithm against three expert electrophysiologist adjudicators. The objective was to see if the AI could return an interpretation or a level of agreement or performance that was compatible with what was seen as expert clinicians. We took 316 episodes that were detected by the LUX implantable cardiac monitor. This monitor has an AF detection algorithm, an atrial tachyarrhythmia algorithm, and then also patient-triggered event recordings. We ran it through the AI, but then also had human adjudication. The idea was to compare how the AI performed against these three electrophysiologists from the Duke Clinical Research Institute. What we were looking at was performance against atrial fibrillation determination, so episodes longer than 30 seconds as per the current definition, same thing for atrial flutter and atrial tachycardia, but also rejection of other non-classifiable arrhythmias. The performance goal was basically the hope that the AI algorithm would provide an agreement level that was within 0.2 of these three expert adjudicators. We had three levels of comparison, so of the three clinicians from Duke, they were A, B, and C. We were looking at agreement within those three, as well as consensus agreement across all three, and comparing that against the AI algorithm itself. For review, the kappa statistic is a measure of inter-rater reliability and agreement. An agreement over 0.6 to 0.8 is deemed substantial strength of agreement, with almost perfect agreement when you have a kappa of 0.81 to 1.0. What we showed is that there was very good agreement between these algorithms. The clinicians had an agreement of 0.89, meaning the performance goal was aiming for 0.69. What we actually saw was that the agreement between the AI algorithm and the three clinicians from Duke was actually 0.89. The AI agreed with the consensus of the three clinicians, as well as the three clinicians agreed with themselves. The AI was as good as three expert reviewers from Duke. You can see that visually that they have exactly the same mean agreement of kappa of 0.89, which again is almost perfect agreement. The AI functionally functioned as well as expert electrophysiologists adjudicating these arrhythmia episodes. In this study, to conclude, we showed that the expert physicians had almost perfect agreement with themselves, and that the AI algorithm provided the same almost perfect agreement as the expert clinicians. The AI algorithm therefore met its performance goal of being within 0.2 of the clinician consensus. The reliability amongst the inter-rater of the clinicians and the algorithm were all very good. The hope is that in the future, we may be able to move away from the need to have multiple clinicians adjudicating arrhythmia endpoints, because the AI is performing as well as those clinicians. Hopefully, by doing that, we can speed up data generation within trials, and hopefully bring down the costs of trials, because we don't have to account for all of these clinicians doing the reading. Thank you very much. Thank you very much. So this paper is now open for discussion. Any comments or any questions? Hi, thank you for the great presentation. So I have one question. So what's the step for including AI in clinical trials first? Like probably like substituting one clinician to AI first, and then you kind of like increase the numbers? I was wondering what the steps is. Well, so I think the question is, how do we move forward with this practically? And I think the first thing is we have to ensure that there's acceptance, that we're not, that we have enough trust in the AI algorithm that it's not going to miss something important. And so we don't want to throw out adjudication because that's how it's always been. We use multiple reviewers because that's our way of ensuring that there's accuracy and consensus and agreement that then gives us certainty as a clinical group that the findings are valid. And so generating the data to support moving away from the clinical group is the most important thing. And so once we have trust that the algorithm is giving us a reliable answer, then that allows us to make the substitution. And so this is the first step to generating that trust. As with every application of AI, the systems are not always transposable. So one AI algorithm is not going to perform the same as another. So we very much have to validate it. And so in this case, we have something that's been trained on 44,000 people, over 10 million tracings, and we can say that this is getting a reliable outpoint. And then we can use that to validate moving forward and making the transition over time. Because again, we need to make sure that clinicians believe the result, that regulators believe the result, and that we can use this in the future, hopefully. Thank you so much. Okay, so I was wondering, for the failure cases, did you go in and analyze what the examples of when it failed look like? Are there any commonalities of when it would disagree with clinicians? Are there any features that you could use to highlight? Because as it does get deployed more, it'd be nice if there was any way to flag ones where we think the AI algorithm might perform poorly or disagree with clinicians. Essentially, did you do any of that? So in terms of this abstract, no. You know, there is importance, exactly as you're mentioning, in terms of being able to identify where it may trip up and fail. You know, I think in this subset, we were only looking at a relatively small amount of tracings versus the whole sort of cohort of tracings that were generated. And that was the test against the 3,000 people. And that was the test against the 3 clinicians in terms of the agreement. So figuring out where disagreement lies and why disagreement is there is important both for the 3 clinicians, because they had the same level of disagreement relative to themselves as the algorithm, and figuring out why those differences exist. And so we know, even from the clinician standpoint, that the agreement is, you know, a cap of 0.9 or 0.89. It's never perfect 100% agreement. And so there's always going to be a gray zone within the interpretation. And whether the algorithm is failing the same way as the clinicians fail to agree is, you know, an important thing to figure out as we move forward with this kind of data set and analysis. Okay, thank you. Do you have the data for ventricular arrhythmias, not for atrial arrhythmias? So I don't have the data in terms of this data set because this came out of an atrial fibrillation trial. And so it wasn't focused on that. I know that the algorithm, so the body guardian beat logic, has also been looked at in terms of UT. It has also been looked at in terms of QT analysis and other areas. But for this project specifically, I don't have that available. Any more questions? No? Okay, thank you very much. So the final speaker is Dr. Laura Gonzalez Ruiz. The title of her talk is Imageless Electrocardiographic Imaging for the Location of Premature Ventricular Contractions. Thank you very much. No. So, thank you for the introduction and good afternoon. I'm very pleased to be here to share our recent work on image-less electrocardiographic imaging for the location of premature ventricular contractions. Let me start with a simple question. How many times have you been in the situation where you are analyzing a 12-litre CCG of a PVC and you are confident you know where it's coming from, but then you ask a colleague and he's sure it's coming from somewhere else, and then a third person joins and it's still not agreement. We all know that localized premature ventricular contractions with a 12-litre CCG sometimes can be challenging. There are several published methods on how to localize and interpret these tools, but in clinical practice, it's far from straightforward, and there are several challenges to face. For example, noisy or low-quality signals, small errors on the placement of the electrodes can affect this interpretation and make this process even more high-operator dependent, and all this leads to prolonged procedures and increased procedural risk. In order to overcome all these challenges, there is a new approach that has been increasingly used, and it's the electrocardiographic imaging. For the classical ECGI, the first step is to perform a CT or MRI when the patient is placed with a high density of electrodes, and with this process, we obtain a personalized geometry of the heart. Then body surface potentials are acquired with these electrodes and the ECGI is calculated and it is obtained the activation maps of the epicardial surface. While this technology offers valuable insights, there are still some limitations, as it requires a preprocedural CT, the electrodes must be placed on the patient during the image acquisition, and all this involves high cost and a logistical complexity. So all these constraints are limiting this application, this application, this technology on the real clinical cases, so here is where the imageless comes, ECGI comes in, as it's a technology designed to preserve all the strengths of the ECGI while eliminating the necessity of having a preprocedural CT. So to obtain the imageless ECGI, the first step is to place 128 electrodes on the patient's torso, and this vest is completely safe and compatible with all the other parts needed during the procedure. The next step is to do an automated 3D torso scan around the patient to know where is each electrode located. After this, the system is connected to the acquisition hardware, and with the information of the patient torso, after a feeding process, it is estimated a geometry of the heart, and this geometry is included into the system. So then, during the ablation procedure, the ECGI is calculated, and when the PVC occurs, we select the window of interest, and we obtain the activation map of the PVC, where the red region indicates the origin of the PVC. So the objective of this study is to evaluate the performance of the imageless ECGI for localizing this PVC in comparison to the electronatomical mapping that is established as the gold standard for the study, and then compare this performance with the one obtained analyzing just the 12 leads ECG, and also compare this with the classical ECGI, the one that we need, the CT or the MRI. So for this study, we include 52 patients undergoing PVC ablation at the Hospital Gregorio Marañón in Madrid, with a 55 years old, or a mean of 55 years old, and around 54 of them were male. During the procedure, we perform a simultaneous electronatomical mapping and ECGI, and then also we obtain the 12 leads ECG for all the patients, and for a subset of 26 patients, we perform the classical imageless ECGI. For the standardization of the classification through all the methods, we establish this ventricle regionalization, where the base of the ventricles were divided in these areas. So for all the patients, we have the electronatomical mapping, and visually, the electrophysiologists indicate where it was the origin of them. Also, electrophysiologists, following the published guidelines, classify the localization of the PVC, just looking by the 12 leads ECG, and for the imageless ECGI, an automatic analysis was used to detect the early region. For addressing the third objective, we segmented the geometry of the ventricles, then we do the alignment with the torso and the images to do the identification of the electrodes and the torso, and we calculate the ECGI, the classical ECGI, and we also detect the early region of them as the same with the normal one of the ECGI. So moving for the results, in this case, the PVC was coming from the right ventricular flow track, and the imageless ECGI showed a good concordance with the golden standard indicating the same region. Then in this other example, the PVC was coming from the lateral mitral annulus, and the imageless ECGI also had a good agreement with the electronatomical mapping. And then in this other example, the PVC was coming from the left ventricular flow track, and the ECGI also indicated the same region for the PVC origin. However, not in all the cases, we had a good performance, as for example in this one, the PVC was coming from the anterolateral papillary muscle, and the ECGI was not able to localize it correctly, as the ECGI had the limitation that it's only mapping the epicardial surface of the heart. So in this case, where the PVC was coming from more endocardial region, the ECGI was not able, and also it was challenging to locate this PVC with the ECG. So the electronatomical mapping golden standard for the PVC localization distribution was this one, where most of them were coming from the right ventricular outflow track, septal region, and the left ventricular outflow tracks. And the agreement with electronatomical mapping, we calculated the kappa coefficient, and we obtained a 0.6 for the 12-lead ECG. And for the image-less ECGI, we obtained a kappa of 0.8, indicating a better agreement for the image-less ECGI. For the accuracy, considering only the exact region as the correct one, we obtained a 55% with the ECG, and slightly better for the image-less ECGI with the 60%. Considering also the neighbor region as the correct one, we had 86% with the ECG, and also a bit better for the image-less with a 95%. When comparing the results for the image-less ECGI and the classical ECG, we have this example where the PVC was coming from the mitral valve, and for both, for the image-less and the CT-based geometry, we obtained the same region that the electronatomical mapping was indicating. In this other example, the PVC was coming, the label of the electronatomical mapping was the septal region, and for the image-less, it was correctly located, but for the CT-based, it was slightly, had a slightly displacement, and it was considered, it was at the neighbor region for the CT-based. And in this last example, the electronatomical mapping indicated also the septal region, and for the image-less, it was correctly located, but for the CT-based, it was not in the correct one. So the percentage of agreement with the electronatomical mapping for this subset of 26 patients, where we only have the personalized ECGI, we have a good performance for both, for the exact and the neighbor region, but for the image-less than the CT-based. So as a summary, the image-less ECGI can accurate and fast localize the PVC origin before the ablation procedure. This technology is able to add some value for the identifications of the PVC by solving the interoperator variability, and it eliminates the need of having a previous CT or MRI maintaining or even improving the localization accuracy. So let me thank all the people participating in this project, and thank you for your attention. Thank you very much. So how many electrodes does your ECGI have? It has 128 electrodes, yes, in the front and in the back. Any questions? Thanks for your presentation. So I'm very surprised to see that the image-less is doing better a lot of the times than the traditional. So what do you think the sources of error are that are in the traditional that aren't in the image-less? Or why do you think it's actually able to overcome, because I think the classic expectation is that the full traditional ECGI is the gold standard, and it should perform better. So why is it not? The thing is that when they are doing the CT or the MRI, there can be some movements also for the breathing movement and the heart movement, that this can affect to the place where the real, with the real heart is. So that's why in some cases it has a bit of displacement for the early region in this case. So then have you tried applying the geometric corrections that you use for predicting where the heart is using the image-less, using the real anatomy from the, like essentially fusing the two techniques to make a better traditional version? We haven't done that, but I think that it's a good, it could be like a good approach to improve the general overcome use. Thank you. Okay, thank you very much. Thank you. Thank you very much. So we just close this session. Thank you very much for joining us.

antiarrhythmics

ECG technology

Erythropermial nasal spray

PSVT treatment

Class 1C antiarrhythmic drugs

Class 3 antiarrhythmic drugs

AI algorithms

atrial arrhythmias

imageless ECGI

premature ventricular contractions