Good evening everyone, good afternoon everyone. It's a great honor to be here at HRS 2025. Today I will talk about a new AI model using ECG to predict all-cause mortality. Before I begin, here are my disclosures. I received a grant from the Korean government, but I have no personal financial interest, no consulting fee. Let's start with a simple question. Why ECG? Why AI? ECG is everywhere. It's fast, non-invasive, and cheap. We do it in ER, ICU, outpatient, not just for heart rhythm. AI has the potential to identify subtle patterns, including changes in wave intervals, which may be imperceptible to human eyes. As you know, many previous studies using AI have shown the ability to predict diseases such as heart failure, arterial fibrillation, anemia, and even one-year mortality. So, why do we care about mortality prediction? Because many patients look completely fine at first glance. Their vitals are stable, their left looks normal, but that day suddenly worsens. Heart rate and blood pressure open don't change until the patient is going worse. So, by the time we realize something is wrong, it may be too late. Therefore, we need an AI algorithm to predict risk of death from sinus rhythm ECG. We aim to develop and validate an AI model using 12-lead ECGs to predict all-cause mortality at multiple time points. This is our study flowchart. We started with a large ECG database, over 5 million ECGs from 1 million patients. Only sinus rhythm included, and each ECG was matched with mortality information obtained from Statistics Department of Korea. Before the training, we divided the surveillance data into two groups, training data and holdout data. Before we train the model, we need to clean the prepared ECG data. We applied frequency filtering to remove noise from the signal. After that, we applied cropping and padding. So, the final input size was 4,096. And then we labeled the strategy using multiple time points. In this example, the patient died within 75 days after the ECG was taken. So, the labels for 3, 7, 30, 6 are all true. But because the ECGs didn't exist 90 days before the death, we set the earlier labels to zero. We call this a multi-label classification program. We used a deep neural network based on residual network, which is the way for detecting features in the waveform data. The model output from zero to one for eight different time points. We trained the model on Amazon Web Cloud Service. It took about two weeks and cost about $1,500 just for cloud service. Among court, the surveillance holdout had the longest. Otherwise, Mimic-4 and UK Biobank had much shorter follow-up duration. Looking at all-cause mortality, Mimic-4 showed the highest ratio. And then, on the other hand, UK Biobank had a very low mortality rate, 1.2%. Comparing patient age, Mimic-4 and UK Biobank had older age, over 60 years old. And then, in contrast, the holdout surveillance data set was youngest, 48. Finally, sex distribution was balanced across all of the groups. We performed internal validation using ECG holdout data set. ECG AI showed excellent performance, short and mid-term prediction. It's short and mid-term prediction within 90 days. The AUC values was greater than 0.9. In the long-term prediction over the one years to five years, AUC values was exceeded 0.84, indicating strong performance. This table shows the detailed performance. The instance rate increased over the time. It was only 0.5 at three days, but increased over 20% at five years. Now, we will look at the external validation using Mimic-4 data set. In short, mid-term predictions, the AUC values greater than 0.8. However, long-term prediction, one to five years, the performance was poor with the AUC value below 0.7. The mortality rate in Mimic-4 increased rapidly over time. In particular, mortality rate was three times higher in the three years and four times higher five years than Severance Quartet. So while the AI model maintained good performance in short and mid-term, but long-term predictions was still more challenging. Now, we evaluated whether AI ECG matched the real-world outcomes using Mimic-4 data set. In low-risk group Q1, only 0.5% patients died, but Q4, high-risk group, increased more than 20-fold. Compared to the lower-risk groups, Q2 showed 2.4-fold higher, and then Q4 group showed 24-fold higher odds ratio. Most importantly, we confirmed that each 1% increase in predicted risk was associated with 41% increase in mortality. But does this model also work in general population? To confirm these findings, we evaluated our model on the UK Biobank. We conducted survival analysis to assess the performance of short-term and mid-term models. In all three models, we observed clear separation between risk groups. Each model, high-risk group had a higher risk of death compared to the low-risk group. However, when we assessed the model on general population, the model showed poor performance for both short- and long-term time points. So, what's next? We are planning a prospective trial to see if this model truly improves outcomes. We are also preparing for integration into hospital system. In terms of clinical use, this model may be helpful in screening high-risk patients who appear stable before their condition worsens. In summary, first, our AI ECG model shows strong predictive power, especially in hospital court. Second, ECG model generalizes across court, but careful interpretation in healthcare population. And finally, AI is permissive, but as always, it should be applied with caution. Thank you for your attention. So, I was mentioning that we won't be able to get questions from you to the speaker, but if there are questions at the end, feel free to approach the speaker. That would be appropriate. I do have one question for you. Just, it was very striking to see the mortality rate in the very high-risk population for inpatients, and as a clinician, it led me to wonder, you know, whether some percentage of that could be ascribed to something that was, you know, adjudicated on the EKG by a cardiologist or a physician, and whether you had access to the EKG interpretations in your study, and whether any percent could be ascribed to something that was diagnosed, just to get at how much of this is a hidden thing that is unmasked by the AI algorithm. Thank you for your helpful question. Excuse me. One more. Did you have access to the EKG interpretations from a cardiologist or a physician, just to see if it accounted for any of the findings, the high-risk findings? I think ECG is the most important thing. ECG has more, many information, so we apply the ECG data, and then we research a previous study. So some prediction models, we found that, so we applied our model. Thank you very much. Thank you. Next we have John Kundryk, Dr. Kundryk, come on up. Predicting all-cause hospitalizations using machine learning. All right. Hello. My name is John. I'm a third-year resident at UPMC, and I'll be talking today about machine learning and wearable fitness trackers and using machine learning to see if we could predict hospitalizations of any cause. So for a bit of background, I have a personal interest in wearable fitness trackers. I think they're becoming more ubiquitous in our society, and currently there aren't many guidelines on what to do with the extensive amounts of physiological and, you know, heart rate and step data that they produce. So we thought that machine learning might be a tool that could be used to analyze these extensive amounts of data that are produced from wearable fitness trackers, and we tried to predict hospitalizations of any cause using this data. And machine learning models. So specifically we used the All of Us research program, which has hundreds of thousands of participants at this point. However, there are, of those hundreds of thousands, there's only about 14,000 that include both their electronic health record information and Fitbit trackers, which are linked together in this study. We used the heart rate and steps data from the Fitbits and aggregated the heart rate data and steps data. So the heart rate data was the average daily heart rate and also included the mean of that, standard deviation, maximum, minimum, range, first and last values, count of entries, range of days wearing the device, total days, number of interruptions, and trend direction. We included those same features for the steps data, as well as a threshold or a time spent below a threshold of 5,000 and 10,000 steps per day. We then linked that Fitbit data to each participant's electronic health record data and evaluated for hospitalizations of any cause. And there were 414 total hospitalizations or an event rate of about 3% over the monitoring period. We excluded hospitalizations that occurred prior to Fitbit monitoring in order to make it a bit more predictive in nature. So these models, there were five models, which I'll show in a later slide, that were trained on 80% of the data and the results are from the remaining 20% of the data. The models were evaluated with performance metrics including the accuracy, receiver operating characteristic curve, and F1 scores. We also included SHAP plots, which were used to evaluate which features contributed most to model performance, which I'll again show on a later slide. Table 1 shows our baseline characteristics of the study cohort. So again, there were a total number of about 14,000 participants that included both Fitbit and electronic health record information and a total number of hospitalizations of 602. However, again, 200 of those had to be excluded due to them occurring prior to Fitbit data being uploaded by the user. So there were a total number of 414 or an event rate of about 3%. Our population was primarily female, about two-thirds, primarily white, about 81%, and primarily middle class with a median income of 62,000 and median age of 54. Table 2 demonstrates the features that were extracted from the Fitbits. So the column on the left is the average daily heart rate and the column on the right is the total sum of daily steps. And so you can see the average of the average daily heart rates was 77 beats per minute and the average total daily sum of steps for across all participants was about 7.7,000. And the other thing I'd like to point out is, again, these were the same for both heart rate and steps. However, we included a few more for the steps features, which were these thresholds below 5,000 and 10,000 steps per day. Table 3 shows the performance of each of the five machine learning classifiers. So the ones that were included were logistic regression, decision tree, random forest, gradient boosting, and neural network. And I'll point out the middle row, which is random forest, as this was our most robust model and had sort of our most promising results with an accuracy of 0.99, area under the receiver operating characteristic curve of 0.91, and an F1 score of 0.88. The other four models varied in their performance. As you can see, there were some that performed relatively worse and some that fell sort of in between the two and were somewhat close to the overall performance of the random forest model. This slide shows on the left the receiver operating characteristic curves for each of the five machine learning models, and the figure on the right demonstrates the SHAP plots showing the top 15 features of heart rate and steps as well as some of the demographic data that had contributed to model performance to predict all-cause hospitalizations. And as you can see, there's a mix of both heart rate and steps features that contributed with higher numbers meaning relatively more contribution, and the top three being maximum steps, the range of steps, and total days of monitoring. And so to summarize the results, the average time from the first day of wearing the Fitbit tracker to the date of any hospitalization was about 1300 days, and again the random forest classifier had the best overall or most robust performance with an accuracy of 0.99, area under the receiver operating characteristic curve of 0.91, and F1 score of 0.88 in its prediction of all-cause hospitalizations. And sort of to summarize that SHAP plot, there were numerous heart rate and steps features that contributed significantly to model performance. Our study has a few limitations. Overall event rates were low, again about 3% for hospitalizations, which impacts the ability of the machine learning classifier and also F1 scores. Additionally Fitbit come, Fitbits come with their own host of potential issues ranging from differences in use case between participants as well as differences in use case across populations. In addition, they are not as accurate as medical grade devices and that may introduce some issues as well. So in conclusion, our study demonstrates some potential in machine learning and its ability to predict clinical outcomes, specifically in our case all-cause hospitalizations using commercially available wearable fitness trackers. And while promising, there's certainly quite a bit of variability across our five models with some performing relatively much better than others and I think that future studies may help expand upon this. And so this is some of sort of my comments on where future directions might go, but I think in summary having larger databases with more diverse populations would certainly be beneficial for this type of research and also in improving machine learning's ability to adequately and accurately predict clinical outcomes. Thanks. Thank you for your presentation. I had two quick questions for you if I could and first question was just around that point that you made at the end about the data source and when I think about a typical person wearing a Fitbit, it's a young, healthy person who's getting after it with exercise, does that more or less encapsulate the data that you abstracted from the trial just in terms of the initial trial? Yeah, so I think our population was, the median age was 54 years. It was maybe skewed in certain directions. It was primarily white, 80% white and two-thirds female. So I think the All of Us program on the whole is a bit more representative of maybe the U.S. population, but the percent or the portion of it that includes both Fitbit and EHR data may not be just yet. And the other question is just in All of Us, was there any kind of cardiovascular hospitalization data in it? I noticed it's mainly all-cause hospitalization, is that right? If there are, you can actually break down sort of, you know, different diagnoses and such. We are looking at that as well. For this one, we just included any hospitalization. Thank you so much. Thanks. Now we have Dr. Zhang, use of artificial intelligence-assisted MRI images to identify underlying atrial fibrillation after ischemic stroke. Thank you so much. I'll just start your presentation for you. Okay, thanks. Thanks. Hello, everyone. Today I'm very happy, glad to share with you all that we developed artificial intelligence-assisted MRI images to identify the underlying atrial fibrillation after ischemic stroke. So ischemic stroke is one of the most leading cause across the world, and AF-associated strokes are more often fatal and debilitating than strokes from other etiologies. And so that detection of AF represents a significant risk factor of stroke recurrence with serious management implications. For example, the shifting of the secondary prevention strategies from anti-pilot strategies to the anti-coagulation therapies, and the long-term rhythm monitoring is recommended according to the AHA or ASA guidelines. However, the firm guidance regarding to the timing, the method, or the duration of the cardiac monitoring is not provided, and inevitably giving rise to the significant heterogeneity in clinical practice across the world. So a notable proportion of the AF patients is underdiagnosed during the next stroke admission. And also, from several previous studies from New England Journal of Medicine, we can see that the anti-coagulation therapies fail to prevent the stroke recurrence in all comers of ESUS patients, which is the embolic strokes of the undetermined sources. So a more efficient approach to detecting the underlying AF is warranted. So since the DWI images is one of the golden standards for the diagnosis of ischemic strokes, we hypothesized that DWI lesion patterns could be used for the differential diagnosis of the stroke etiologists. So we would like to develop an end-to-end AI model so that we can really efficiently find the underlying atrial fibrillation in these kinds of patients. So we conducted a multi-center proof of concept study to develop this model. Our clinical trial is composed of three cohorts. One of them is the internal retrospective group, and the second one is the external retrospective group from three other participating institutions. And the third one is the prospective testing groups for the test of our AI models. So the model structure is made of two branches. One of them is the CNN branch, and the second one is the radiomic-based approach. And we also developed a combined approach, which combines the CNN network and the radiomics features. So since we used the, you can see here from the flowchart, that we used the five-fold cross-validation approach for the development of our model. So first of all, we evaluated the model performance of these five models we developed. So we can see here that we used the following indicators, including the AUC, the sensitivity, specificity, accuracy, PPV, and NPV, to evaluate the overall model performance. So we can see that the combined classifier outperformed the other two classifiers. So we chose one of the models with the best performance to do the following validation and following testing in the other two cohorts. So we can see that the combined classifier really generated the best performance in most of the indicators. You can see, including the AUC, including the specificity, and the accuracy. In both the external validation group and the prospective testing group. Also, since many people argue that AI models can be a black box, so we generated a heat map to really generate the attention mechanism of our CNN model. We can see from this heat map that our model is really paying more attention, as to looking at the lesion of the ischemic stroke in the TWI images, rather than other irrelevant parts. Also, in the right panel, we can see that we listed the top eight significant radiomics features, which is important for the prediction of the underlying AF. We can see that there are some recurrent features, including the features which can be conceived by human experts, which is the shape, or something that is difficult for the understanding of human experts, including the texture of the TWI images. Next, we conducted a subgroup analysis to see what kind of patients may benefit more from our model. We can see that female patients and patients with high NHIS scores, or patients with low CHA2-VASc scores, have a better AUC performance from our model, which means our model can be really helpful, because patients with high CHA2-VASc scores may draw more attention from the neurologist or the cardiologist. However, the patients with lower CHA2-VASc scores, they did not get enough attention from these physicians. So, our model can really help these subgroups of patients, in terms of secondary prevention strategies. So, here comes to the summary of our study, that we developed a combined classified integrating the CNN and radiomics features, which exhibits satisfactory performance in detecting underlying AF in AIS patients. And our model is robust, with high interpretability and scalability. And also, our model performs better in female patients, patients with high NHIS scores and CHA2-VASc scores. And our model identifies all the newly diagnosed AF patients with how to monitor during the index stroke admission. So, our model has two potential application scenarios, including one of them is to guide the selection of certain AIS patients to undergo more intensive cardiac monitoring, or to serve as an indicator of the atrial cardiomyopathy in AIS patients, to guide the administration of anticoagulation therapies. So, thank you for your time. Thank you so much. I wanted to congratulate you on the study. It's a very difficult problem with cryptogenic stroke, and often the imaging findings are inconclusive as to prediction, whether something is embolic or not. So, it's a very important clinical problem. Particularly, I want to congratulate you on the heat map, which it seems like your algorithm is able to look at features of the DWI, that are particularly a signature for embolic stroke, or something that might be related to atrial fibrillation. Did anything else pan out in terms of the distribution? I mean, a common clinical pearl is multiple vascular territories, or cortical infarcts being a particular signature of embolic strokes, but did your data, maybe that data was not available, but just wanted you to comment on that. Yeah, that's a very good question, and that's one of our future directions, to really see the findings of our AI model, which our attention mechanism, it really matches the pathology findings of the stroke patients, but that's what we are trying to do in the future. Thank you. Okay, so now I'd like to welcome Dr. Kim. Changes in Artificial Intelligence-Based ECG Aging, and the Risk of Stroke in Newly Diagnosed Patients with Atrial Fibrillation. Dr. Kim, thank you so much. I'll just hit start. Okay, thank you for having me. I'm Dr. Dae-Woon Kim from Korea, and I'm presenting the Changes in AI-Based ECG Aging, and the Risk of Stroke in Newly Diagnosed AF Patients. Okay, so now I'd like to welcome Dr. Kim. Changes in Artificial Intelligence-Based ECG Aging, and the Risk of Stroke in Newly Diagnosed AF Patients. Okay, so now I'd like to welcome Dr. Kim. I'm Dr. Dae-Woon Kim from Korea, and I'm presenting the Changes in Artificial Intelligence-Based ECG Aging, and AF and mortality, but the predicted value of chronological age decreased with aging due to the increasing biological heterogeneity among individuals. Actually, in older adults, chronological age is less likely to align with their biological age. And what about sex? There is substantial evidence on sex differences and their impact on AF and arrhythmia. Across various countries, the prevalence of AF in women is typically lower than in men, and women also exhibit different electrical and structural characteristics in the heart. While many AI ECG algorithms have shown remarkable accuracy in disease detection, their generalizability across diverse populations remains a significant concern, largely due to their ethnic and regional difference, and also the variations in diagnostic criteria across courts. To address these limitations, we aim to develop and evaluate AI ECG models predicting age and sex, which are two universally applicable demographic attributes, rather than predicting a disease itself. So we use multi-test learning model to simultaneously predict both age and sex from a single 12-lead ECG. And using this model, we previously published some articles. After AF ablation, there is some association between AI ECG aging and AF recurrence. Also, there is some association between sex classification and AF recurrence after catheter ablation. And there was actually sex differential effects. Actually, the effect was not prominent for the men. But for females, especially females predicted as male, has a higher risk of AF recurrence after ablation, suggesting sex discordance is a novel AI ECG biomarker capable of identifying females with disproportionately elevated risk for arrhythmias. Finally, we analyzed among the general population from the multinational courts. We demonstrated AI ECG aging, which was defined as AI ECG at least seven years older than biological age, was associated with a higher risk of new or early-onset AF. But how about stroke as a consequence of AF? So we aimed to evaluate temporal changes in AI ECG age gap and its effects on stroke risk in patients with newly diagnosed AF. So highlighting the potential novel biomarkers role in advancing strategies for stroke prevention in AF patients. To test this, we developed. This is the developmental stage. So we developed and validate an AI ECG model to predict age and sex using data from three multinational courts. And our model performed well compared to the LIMAS model in predicting age across the two of three validation courts, showing a better correlation between chronological and biological age. And at this time, the model's accuracy and AUC were quite strong. Now we're interested in which regions that our AI model focuses on for its prediction for age. For age prediction, it primarily paid attention to the PQ segments like this. So finally, to evaluate stroke risk following AF diagnosis, we analyzed a court of 2,600 individuals with newly diagnosed AF and no prior history of stroke. All included patients had at least one ECG within two years before diagnosis and one more ECG within two years after AF diagnosis, allowing us to assess temporal changes in the ECG-derived features around the time of diagnosis. So AIECG in this study was defined as age gap at least nine years between ECG age and their chronological age, and we identified four groups, AIECG age trajectory groups. The first one is the previously normal ECG age and after diagnosis still normal ECG age was stratified into consistently normal group, and normally ECG age before diagnosis, but after diagnosis they'll have ECG aging happen that we then we stratified into the accelerated aging. And there was some reversal before diagnosis, there was ECG aging after diagnosis, but after diagnosis there was no ECG aging that we defined them as a reversed aging. Finally before or after both, there were ECG aging, we defined them as a consistently aged. This is the baseline characteristics of a cold. The mean age was about 62 years and 60% were male. And interestingly, the consistently normal group was older and more likely to be a female and have more comorbidities. This is the final result during the median follow-up of seven years. Consistently normal group had 0.7% of the cumulative incidence of stroke at 70 years old, but reversed aging, same, similar to the consistently normal group, still 0.7% of cumulative incidence. How about the accelerated aging group? They had 3% of the incidence of stroke at 70 years old. And finally, consistently aged group had worst outcomes as 4.4% of the cumulative incidence. And we assessed the factors associated with the temporal changes in ECG age, and these two groups with having bad outcomes were more likely to be younger and more likely to have CKD and diabetes. To conclude my talk, longitudinal changes in AI-derived ECG aging related to the incidence of stroke after AF diagnosis and accelerated aging and consistently aged groups will have a higher cumulative stroke risk among patients with newly diagnosed AF, whereas the reversed ECG aging showed no increase in AF risk, suggesting still some kind of ECG reverse remodeling. And our study highlights the potential role of ECG age as a dynamic biomarker for stroke risk predictions. Thank you. Thank you. Very thought-provoking work. I was going to ask you, just as a hypothesis generating, it sounded like your model really paid attention to that PQ segment, and I was wondering whether there was any absolute predictive value of the PQ interval or anything about that PQ segment that came out in your data referencing the stroke risk. Thank you for a really great question and really nice points, but unfortunately we do not have any cutoff for the PQ segments for higher ECG age or something, because it's from the silence map and it's from actually some kinds of black box things, but we can try to evaluate what is the cutoff points for the bad outcome or at higher ECG age using the PQ segment or any other segment in ECG. Thank you. Very nice work. There's so much with blood tests and predicting biological age, so interesting to see that we can get something similar with EKGs. Thank you for the work. We have Dr. Bach here for our final talk of the session entitled Enhancing Thromboembolic and Major Adverse Cardiac Events Risk Assessment in AFib Using AI-Driven EKG in Comparison with the CHADS-BAH score. I'll just start up your presentation for you. Thank you, Chairman. Good afternoon. Good afternoon. It's a great honor to be here to present my topic. My topic is AI ECG is can predict thromboembolic risk and MACE in atrial fibrillation patients. And so we're comparing an RYC with conventional CHADS-BAH score. I'm Professor Yong-Soo Baek from Inha University in South Korea. Yes, the previous presentation known is AI ECG useful, is current AI ECG, is can predict this previous talking is biological age, is low ejection fraction, and AF risk and so on. So we developed this since 2018, is all of the, is all is AI methodology. So now is we focus on the, so is each segmentation is research. Because now is when we or 1 million is ECG low data use AI, we made AI platform like this. Because is we using now is supervised technique, combined unsupervised technique. So sometimes is now is a most AI technology is using the, is just the labelling is true or false is versus, but sometimes is humans is mistake is some, that's the influence is a bad effect is to AI performance. So we using now is using big platform is sometimes is unsupervised is characterized by computerizing itself and then is supervised technique. So nowadays is we improved, we can improve the, is AI performance increased. So yes, is my outpatient clinic is, I used like this, is lively, is travel with ECG, is patient take, and now we simultaneously interpreted to the, all of the cardiovascular risk now. So I used this AI is scoring, right, like this. So is this story is I wonder the, is the left side is, of course is atrial fibrillation ECG. Is the left side is my patient is 61 male, atrial fibrillation, but is child birth score is zero. Right is atrial fibrillation is 66 male, but is this patient is child birth score is four. Is very high risk, but we don't know is whether is ECG is some high risk for is anticoagulation. So we sought to investigate whether AI-driven ECG using large-scale travel with ECG low data can predict thromboembolic risk and further is related in MACE is compared with is conventional chatbot or chatbot scoring. So I mentioned we used big platform AI. So this AI is we used 60, over 60 ECGs atrial fibrillation. We excluded atrial fibrillation with mechanical valve and moderate to severe moderate mitral stenosis. As importantly, we supervised the data is we all data we checked is six medical students and two cardiologists during the five months is we one by one is reviewed ECG and clinical data. So we improved is excellent performance. This is we are baseline characteristics is mean age is 72 years and female sex is 40% and like this. Is first we all test is all chatbot score is zero to nine. This is one example chatbot score zero to one versus two, above than two. You can see there is our AI ECG is performed AUROCs arrived 0.8 and spot precision and recall level is not is satisfied is less than 0.5. So we found the most best performance is non-gender chatbot score is zero versus two above than two. You can see our AI ECG is arrived reach the AUROCs 0.86. You can see there is excellent precision, recall and F1 score. Then we test the another code, atrial fibrillation data. So we is this result showed our AI ECG is a high risk and low risk is hazard ratio 1.98 is compared to the conventional chatbot score. Chatbot score is just 1.82. So we can is confirmed check the superior performance of this AI ECG for predicting is thromboembolic risk in atrial fibrillation patient. When you also check the external validation using another is academic university hospital data. You can see there is AUROC is another university hospital is we used more than is 400,000 patient. You can see there is a very good performance. We check also B capillary microbe is another is hospital. We can also is check the good performance. Let me show there is some example. This patient is 44 is atrial fibrillation. You can see the typical atrial fibrillation is 12 with ECG. Yes it's a chatbot score zero, but our AI is risk is low risk is very good performance 87% check. And GRADCAM is explain the AI. As you can check there is fibrillate wave is blue color is attention to this by our AI ECG and two wave attention. Now example is high risk patient is 92 atrial fibrillation patient and hypertension and diabetes is previous stroke patient. Our AI ECG is can predict well high risk patient is 94% is GRADCAM is a little bit I think this is very interesting point a little bit another point AI attention to fibrillate wave and QRS point. And this patient is 53 years atrial fibrillation patient, but this is chatbot score is zero, but we can check a little bit strange this atrial fibrillation because it's average pattern and deep tube inversion. So this patient our AI ECG is can could predict the AI high risk for thromboembolic risk. So I think this is oh wow it's funny case is GRADCAM is fibrillate wave and two wave attention. So this patient echocardiography is diagnosis as apical hypertrophic cardiomyopathy. So we know the now is guideline, hypertrophic cardiomyopathy patient with atrial fibrillation regardless of chatbot score, so we should get this anticoagulation. So we very in detail sub-analysis in each by each case because it's AI ECG is presenting is different scoring for thromboembolic risk this atrial fibrillation is very interesting case interesting me is chatbot score is including is variable factorosis so can different present can different is another is AI GRADCAM is example heart failure like this and hypertension stroke and all elderly patient is different pattern. So we need for the research in this each by each is variable factors including is chatbot score. So now is we have a plan is future multicenter clinical trial using our AI ECG is prospect study is some I focus on the gray zone chatbot score is one and zero. So ladies and gentlemen I would like conclude is our AI ECG system using like this is AI big platform is can show the sub-central efficacy in stratifying thromboembolic risk and cardiovascular event in atrial fibrillation patient. Sometimes we have we are humor is sometimes is mistaking is chatbot score now is current guideline is chatbot score but is sometimes we missed is hypertension diabetes and so on now guideline is recommend the should dynamic change we check the dynamic change every six months is not same is risk. So I think that this AI ECG more in detail and clinical decision making is very helpful to tell is management is each by atrial fibrillation. Thank you very much. Thank you very interesting presentation also very interesting that interesting case where you're able to pick up on a underlying diagnosis that change that patient's management. Two kind of related questions I mean it did seem from some of those tracings that the fibrillatory wave and potentially how course or the amplitude of the fibrillatory wave or maybe the frequency might have something to do with predicting the stroke risk. I was wondering if you know longer acquisition times in your opinion you know of the AFib over instead of just 10 seconds 30 seconds or a minute might pick up on additional information and sort of the application to wearable technologies where we're getting lots of these EKGs now over 30 seconds and you know whether your algorithm might be able to be applied to wearable technologies. Thank you very nice is good comment is I agree with you but is this I used only 12 read ECG conventional ECG so our studies aims check the only using one is 12 read ECG but now is this ECG AI algorithm. I hope the this AI algorithm expanded the wearable ECG on wearable control. I think the ECG available is can maybe there can in the near future is we improve some performance. Well that concludes our session for today. Thank you to all the speakers and thank you for everyone for your attention. Enjoy your meeting.

AI model

all-cause mortality

electrocardiograms

ECG patterns

deep neural network

mortality prediction

short-term prediction

medium-term prediction

clinical integration