Thank you for pulling me in. I mean, if he shows up, then that's good. If not, then that's good, but, yeah, I'm just attracted. Let's go ahead, I guess. All right. Thank you so much. Good afternoon, everyone, and thank you for being here. I'm not supposed to be the moderator for this session, but we're not sure where the moderator is. So I'll fill in until he shows up. And if he doesn't show up, I'll take you to the end of the session. So I think many people here know me. I am Sana Al-Khatib. I'm an electrophysiologist at Duke University. And it is my distinct pleasure to welcome and thank the panelists that we have for this very interesting session titled Digital Twins and Command Stations, Whether the Twain Shall Meet is the topic. Great. So before we start, why don't we ask our panelists to please introduce yourselves and tell us about what you do in this space, starting with you, please, Natalia. Hi. Good afternoon, everyone. Can you guys hear me? Yes. Yep. All right. My name is Natalia Trejanova. My admin forgot to register me, so when you look in the program, the name appears, but all the credentials are not there. So I'm going to tell you about that. So I am a professor of biomedical engineering at Johns Hopkins University. I hold an endowed chair in our department. I'm also a professor of cardiology. And I also have a courtesy appointment in the Department of Applied Math and Statistics. In addition to that, I direct at Hopkins the Alliance for Cardiovascular Diagnostic and Treatment Innovation, which is a center which I created and direct, which combines interventional cardiologists, engineers, mathematicians, with one common goal, to bring to clinical practice every technology that is being developed in that space. And so this is a center that enables all the digital twins that we develop and AI approaches to be brought directly to the clinic. In addition, I recently stepped in even bigger shoes. So I'm currently the director for AI research and health in medicine at Johns Hopkins University, supposed to direct all the efforts in bringing AI to clinical care in the entire university. And Johns Hopkins is putting huge amounts of money into infrastructure and hiring faculty that can fill that space. So my job is to mediate and provide the vision for these developments. Wonderful, thank you so much. Please. Yeah, I'm Hamilton Baker. I'm a pediatric cardiologist at Medical University of South Carolina in Charleston. And my experience in this area, director of the Clemson MUSC AI Hub, which is an organization that's supported by both institutions to promote biomedical AI research across the two institutions. Specifically to digital twin, our children's hospital at MUSC, the Sean Jenkins Children's Hospital, which was built a few years back, has entered into a project with Siemens Health in Years to create a digital twin of certain parts, starting with certain parts of the hospital. So that's my connection to digital twin. Wonderful, thank you so much. Please, go ahead. Hello, everybody. And first of all, thank you very much for the invitation. My name is Alfonso Bueno. I'm Associate Professor of Computational Medicine at the University of Oxford. And our field of research is basically digital twins in the field of cardiology, in the field of electrophysiology. I have been working in this field for approximately the last 20 years. My biggest interest is in inherited cardiomyopathies, hypertrophic cardiomyopathy, Brugada syndrome, different genetic conditions of the patients, also with a focus on atrial fibrillation. And our group also has additional research interest in myocardial infarction, heart failure, different diseases of the human heart. And our main objective is to get a better understanding of the mechanisms that drive the therapy outcomes in these patients, both from the point of view of therapies that can be either interventions or pharmacological therapy. Great, welcome. Thank you guys for being with us. Just as a reminder to the audience, please submit your Q&A. I was in a session yesterday, similar to this one. People kept submitting questions. I mean, we had like 10 questions. It was a great discussion. So we definitely want to make this interactive and I want you to feel included as well. So let me, first of all, start maybe taking it one step at a time and ask you, how would you define digital twins? And give me examples of digital twins that each of you is working on. So I'll start with that. So the way we define a digital twin, digital twin is basically a digital representation of a system. In this digital representation of the system, everything is represented. In other words, the geometry of the system, the different parts, their interaction, the system is dynamical entity and the digital twin replicates that dynamical entity. An example, so this is a concept that actually came from manufacturing. For instance, it was developed for complex machinery and manufacturing processes. You do not want to stop a manufacturing process or you do not want to stop the use of a very expensive machine because one of its parts is about to fail. You would want to know in advance what is about to fail so you can order that part or you can replace that part of the manufacturing process so there is no loss of income, no loss of people's time and so forth. So this concept has been lifted sort of for application in medicine and it actually now is making inroads in many fields. The digital twin, the wording digital twin emphasizes that it is a virtual model of a personalized entity. If we are doing a digital twin of an organ or the whole human, it is the model of that organ, personalized organ or that person. It is not a generic mechanistic model. It has personalized components to it. So in our hands, we build digital twins of patients' hearts. We have a quite advanced sort of pipelines for building these digital twins of patients' hearts. Our goal is to use these digital twins for two main purposes. One of the big purposes that we are using them is to guide therapy in a very personalized way. We would like to predict in a given patient, particularly in patients who have complex diseases, to predict where exactly, what would be the best therapy for that particular patient. Let me give you an example of one of our applications. I currently have an FDA approved clinical trial to predict what is the optimal ablation target in a patient with atrial fibrillation, particularly in patients with persistent AF, particularly those that are complex cases that have undergone ablation numerous times sometimes. And we are trying to predict outside of the pulmonary vein isolation, what would be the best way to ablate that patient. We focused on minimal lesions. We focused on what will best eliminate the current activity, arrhythmic activity. But here is one point I wanna make, and this to me is extremely important when we talk about digital twins. The use of digital twin allows you to predict the response of the patient to the treatment. So we are not just predicting how to execute the treatment, we use the digital twin to see when you execute the treatment, how would the patient respond? For instance, in my clinical trial to predict where to ablate for patients with persistent AF, we don't only predict where to ablate, we implement that therapy within the digital twin and see what will happen after that. Very often patients who are ablated will go home and then will come back in six months or a year to be re-ablated. We don't want that to happen. We predict where to ablate, not only with minimal lesions for the current activity, but how to prevent that recurrence. And so that's a unique benefit of building a digital twin, the ability to predict the response of the patient to the treatment. So that's a big area and I'm gonna, I could talk a lot more, Sana, of course, about that, but I'm gonna let you, sure. But let me ask you a question. So in this particular, because you said like this could be an organ, it could be the- A whole patient. Yeah, the whole patient. Is it also the operator? Are you also like trying to create a digital twin for the operator or is that not? You can do digital twin of the processes. You can create, if you, that's a completely different, there are many, many replicas, as I said, since it starts from manufacturing. You can do the whole, like for instance, surgery at Hopkins, talk to me about, can we replicate in a digital twin the operating room? Yes, there you go. For instance, you can do that. And you have similar, you know, you can do digital twin of anything. What I'm saying is when you create a digital twin, let's say of the procedure room, you can play scenarios that you normally, you cannot when you are there because they have constraints and that allows you to do things that you cannot do in the patient. I cannot poke and prod in the patient's heart to figure out what is best, but I can do it in the model. Okay, great. Thank you so much. Please add. Yeah, I was just gonna add, that was a fantastic definition of what a digital twin is. And I think to highlight something you were mentioning is the potential scale of digital twins. You can go all the way down to the cellular level, all the way up into an entire institution or even a healthcare system. And so, you know, what we're working on at MUSC is much different than doing organ system type digital twinning. We're looking at our procedural area, very akin to what your surgeons were asking about. And so what that allows you to do is start to run different operational workflow simulations and try and find, you know, in the most simple sense, find where your bottlenecks, where problems going on, but then it can become even much more sophisticated. And I think that's another important thing to understand about digital twin technology is that in order to understand it, you really need to define the inputs and outputs and a digital twin of say an organ system, that can mean a lot of different things. That can mean just the electrophysiology part of it, just the anatomical part of it, really just one chamber of the heart. You know, in order to truly digital twin a patient's cardiovascular system, that is incredibly complex and difficult. And in my experience, I'd be very curious to hear both of your thoughts on that, but how close are we actually getting to that? Because mine is more on an institutional level. Please go ahead. Yeah, of course. So following with the discussions that we've been having, because this is something that we just discussed before the panel. So my vision of the digital twin is a bit broader. So I completely agree with the vision of Natalia of a virtual replica of the patient that is usually based on first principles, laws of nature, the physiology, the physics and the chemistry that we know and we can put into a mathematical formulation and create the digital copy of the patient. These models are multiscales with a lot level of personalization. So that is one part of the digital twin in my opinion, but that is also going to be hand by hand with this new plethora of statistical models and machine learning, artificial intelligence that are also going to have a predictive availability, potential to predict the future outcomes of the patient. So for example, if a patient has an infarction, we can recreate again the anatomy of the ventricles taking into account the personalization of the infarction for the certification of risk and the prediction of the ablation. But at the same time, you can run your machine learning, your artificial intelligence algorithm to get a probability of the patient having a secondary infarction in a time of one year. So for me, it's not only the virtual replica, it's the complete set of tools that allows for a comprehensive characterization of all the information of the patient. Great. Well, thank you so much for sharing that perspective. So if we want to expand on that, talking about the potential applications or the potential benefits of digital twins, you mentioned one which is being like selecting people for certain procedures, trying to predict their response to the procedure. Obviously, by predicting their response to the procedure, you'll be able to make better decisions about selecting them. You'll be able to share this information with patients when you are counseling them about a certain procedure or an intervention. What are other potential applications or benefits from this procedure? We'll start with you and then come down. So for example, we are using this framework for the prediction of pharmacological therapy. So it also depends very much of what you consider as personalization because there are also different levels of personalization that maybe do not need to go completely up to the level of personalizing every single detail in the model. So for example, you could personalize the electrophysiology of the heart to the specific mutation that a patient is carrying. And depending on the mutation that they are carrying, different pharmacological therapies can lead to different outcomes. So that is something that we are already using the models for. There are other applications that Natalia was mentioning before. It's not only the modeling of the electrophysiology, it can be the modeling of the fluid dynamics in the heart. And there are clinical trials and there are products already with this idea of the digital twin that allow you to measure drug pressures in a way better manner than in basic procedures. So actually the digital twin can allow the replacement of invasive measurements, getting the quantities of interest for practice in a faster and non-invasive way. That is another application. Great, thank you very much. Would you like to add something? Yeah, I'll just add a little bit onto that. You know, I think as a clinician, I see the ultimate utility of digital twins being once we are able to do multiple organ systems and their interactions, which is extremely complicated, but I think the true clinical benefit there are things we won't even realize, right? We'll be able to do true digital interventions on a patient's twin and from that make good clinical decisions about how precisely to treat that patient and to bring that back to my own specific work in congenital heart disease. That's been a very big push from the surgeon's perspective is even just simply creating different anatomical settings through the different surgeries that patients like that have to go through, the inability to personalize that for each patient has been frustrating in that one patient might do much better with a slight anatomical difference in the construction of their graft, et cetera, than another, but at this point, that's all just anecdotal and surgeons will do different things different ways because that's what they feel that they're getting. And ultimately, to be able to truly do that, to take a patient at their current anatomy, change that anatomy, and not only see the differences in their cardiovascular system, but all the way down to contractility, et cetera, and the complex interplays, not only within the cardiovascular system, but the other systems as well. That sounds great. Can I also ask you in terms of like response to therapy, it's not just, it doesn't only relate to effectiveness. I suspect that it also relates to safety. Are you able to predict someone's risk of complications or issues with a procedure? I suspect that's true, right? Yeah, well, my answer to that would be it's also the ability to predict the downside of a treatment, right? And to avoid hurting your patient in some way. But again, I see, we already have that in medicine in some ways, it's just not nearly as sophisticated as it would be in the setting of a true digital twin, which again, allows all those very complex interplays for us to actually model those precisely for a particular patient. But I'm sure Natalia probably has some comments on this as well. We actually have a ton of questions from the audience. I promise we'll get to them. So Natalia, go ahead. So I wanna sort of come back to what you asked, Sana, and the way you summarized what we do currently with that. I actually wanted to say we go quite a significant step further because we don't only predict which patient would be good for certain therapy. Actually, our clinical trial is the following. We predict where to ablate before the procedure. It's imported in the procedure room and the catheter is navigated directly to it. So it's the first clinical trial that the FDA approved in which a patient treatment is driven by computer modeling, ultimately. So it is like you're now in the realm of interventional cardiology where you really, the decision no longer rests with the clinician who is making, you know, they don't map at all. So they don't map. They go directly to the target. So that is taken away. And in some way, to me, those developments embody what I see the, you know, the future in medicine. We will be, you know, a doctor will be half engineer, if you will, you know what I mean? So these are technologies in which they, a lot of the clinical decision now relies on that. And so we are testing where the clinicians are comfortable with that. Whether they believe, whether we can demonstrate this approach is better than doing the mapping and making a decision based on that, that the trial is randomized. So we are hoping to have answers to that. But maybe I can come back again, but I see another very different also application of digital twins that I want to talk about also. So before we get to that, because I definitely want to hear that point. Tell us a bit more about the randomized clinical trial. Is this like a proof of concept trial, or is it well-powered to show difference? Like what are the outcomes? It's 160 patients, and it is randomized on patients for standard of care, which is pulmonary vein ablation, or the clinic, our approach is called the Optima approach, which is a personalized target prediction. And so the patient comes in, they receive an MRI scan, they get randomized in which arm they go. If they go, if they're randomized in our arm, where we predict where the ablation target is, we do the prediction, the evening of the procedure, we brief the attending. They know very well what the prediction is in the morning of the procedure. The targets are incorporated in the CAR-2 precision, whatever is the mapping system, they are on the screen, and the catheter directly is navigated to that. Now we make our collaborators to map only, not because it's needed or it's in the clinical data, but I want to compare my predictions also with some electrophysiological data that can be collected. So also our patients receive ZEO patches, so they are followed, you know, we record for weeks and weeks, and then they are followed every three, six, nine, 12 months and so forth. They have a questionnaire of quality of life as well. So far, you know, we have no recurrences, as I said, because the approach that we are doing is, it predicts not only where to ablate, but when I ablate there, would it recur somewhere else? So we are catching that. So these targets, when we do that, basically, so far we have no recurrences, while the other arm, there are recurrence. How many patients have you enrolled so far? So we've enrolled 42 so far, we're done. So this is on the atrial side. We have, this is FDA approved. We have another FDA approved study for ventricular ablation. It's smaller. Who's funding them, by the way? Funding is coming from? You know, I write grants for- NIH. Yeah, that's my job, write grants. There you go, great job. In addition to all the other things. That's amazing. That's impressive. And so the primary endpoint is time to recurrence of atrial fibrillation? Well, it is, yes, it is recurrence or not. Okay. Yeah, okay. Okay, all right, great. Thank you so much for sharing that with us. You had another point that you wanted to share with the audience. Yeah, I wanted to share some, another very kind of interesting use of digital twins that one can do. It is this very, I see a very tight interaction and combination between AI and digital twins. Because AI can be used in the construction of the digital twins. We use that like in our pipeline for building the digital twins. But what is very important also is when you make a prediction, for instance, we developed an algorithm to predict risk of sudden cardiac death from raw clinical images, from MRIs. And so, you know, it's a deep learning neural network and it predicts somewhere, we have a heat map where it predicts the image was most important, importantly correlate with risk of sudden cardiac death. But why? And so here you can construct a digital twin of that patient and figure out what is mechanistically happening? Why this protrusion in the infarct or why this fibrosis distribution led to arrhythmias? So in a way, the combination is I develop AI to make a prediction, you know, for risk of adverse event. But then I can take the predictions and create a personalized model for these patients that you have a adverse effect and see what exactly is there. So this gives you an idea how to work in a modifiable fashion. What can happen? What can you do? So you understand for a given patient, what led for the algorithm to pick this area of the scan or these specific clinical covariates that were not known before to lead to that. Wow, that's fascinating. And for example, following on that, just one line. Another application is, for example, to understand genotype, phenotype correlations where you can actually divide your population in different phenotypes and then use, that can be done exclusively with machine learning, with artificial intelligence, but it's going to tell you that they are different about, as Natalia was saying, it is not telling you why. And then you can use the digital twins of the different groups or specific patients to actually explain what is behind the different phenotypes that you see, for example, in the ECG. So it's not only therapy, it's also, for example, collecting the most valuable piece of information that is needed for the diagnostic. It's enabling new biomarkers of disease. So all these are applications of the digital twins in practice. Wonderful, wonderful, thank you. We have actually a ton of questions from the audience, so thank you so much. Starting with the first one, how would I ever know how accurate the digital twin is to the original? It will never be accurate, period. You have to live with that. So, no, seriously. So, and I can tell you that from a personal perspective, I came from basic science and everything is very precise. I was doing mechanistic modeling, you change 10% of the sodium channel, you know exactly how propagation change, and then you move to working, you know, on clinical translation problem where you work with, you know, with dirty clinical data all the time. The leap of faith. Yes. You know, it is amazing to understand that precision in medicine is very different from precision in biological experiment or something like that, or doing a calculation mechanistic. The model, we are not looking how accurate the model is in representing what exactly is happening. We are looking into how useful the model is, right? So every model ever will be wrong, but some models are useful, and that's our goal. How much inaccuracy are we willing to accept, I guess is the question. No, you test outcome. If it predicts correctly, if I can predict where to ablate in these patients and they do not recur, hey, that's a useful model, right? Okay. It depends on the application, of course. It seems like that's what we should use, is it a useful model or not, rather than how accurate. Are they useful? Exactly, exactly. To me, that's a very important distinction. Sorry for... What do you all think about that? I agree. You know, I think a lot of it depends on the complexity of the particular thing you're trying to model or represent with the digital twin. You know, if you're looking at ablations, that's pretty important stuff, right? I think there are... The risk involved in trusting your digital twin or model has to be seen in the perspective of the clinical situation, right? If it's high stakes, you need a much more accurate representation of what you're doing that you can trust, but the trial that you're running is a great example of how we make that progress towards getting clinician buy-in and actual implementation. Indeed. And I think that's how you do it with digital twin or AI or any other technology. And I also, I find myself that I sometimes think detailed explainability as a need for technology, specifically AI or digital twin, is a little bit overblown because I think what you're saying about results, right? I think that's what clinicians really want to see. And their degree of understanding of how that model digital twin or AI helped them get there, I think is important, but it doesn't have to be in crucial detail. Go ahead. And following on that, people shouldn't believe our models at all. It's our responsibility to build the credibility of the models, and that is something that we are building very cautiously in order to present all the evidence. For example, how much the models are representing the ECG of the patients or the contractility, the PV loops, the comparison with all the clinical metrics and for the patient within the variability that you observe in the population. So it's our responsibility to actually build that credibility of the models. Indeed. And if this is going to get to the clinical practice, we need to do better in terms of providing measures that quantify the uncertainty in our predictions. And that is going to be critical. But it's actually, as Natalia was saying, it doesn't need to be a perfect copy. So why should my goal to replicate exactly the ECG of the patient, a perfect copy of the ECG that I have taken from the patient, when if I ask the patient to come two days after, the ECG is most likely going to be slightly different. But you need to build credibility that at least the QT interval of the patient is going to be replicated, that you have some variability. The electrophysiology is going to vary from the morning to the night. So, but we need to build credibility and uncertainty quantification for sure. That sounds great. So let me, oh, do you have something to add? Okay. Go ahead. So there's an interesting question here. Someone asked, what do you think is the most promising disease or condition or use case of digital twins in cardiovascular medicine? And what is the most difficult or least promising? And I guess the question is like, how do you decide like which conditions to focus on and which conditions perhaps not to spend too much time on? I don't know how to answer that. To me, it is like where, what are you doing? Where is your heart in? So I love electrophysiology. I love electrophysiology. I love the complexity of it. I love, so if somebody tells me, well, that's not very promising. Why don't you move to looking at the gut? Well, I don't want to do that. I just love it. So I'm going to stick with that and do my best. And we try. So it's very hard to answer that. It's just also a personal preference, your academic experience, your who you're working with in the clinical arena. I love the people that I work with. I will continue working with them and do my best, so. Indeed, anything else to add? Following what Hamilton was mentioning before, it's not that it's the least promising, but probably the part that still requires our attention the most is the communication of different types of organs. So we are becoming very good at modeling independent organs. So we have very accurate and digital twins of specific organs of the heart, of the lungs, of the different parts of the body. The communication of the multi-organ, I think, in my opinion, is going to be one of the next frontiers. That, for example, in some cardiac diseases is going to be critical. For example, if you are treating heart failure, maybe the pulmonary hypertension is going to be something that you want to take into account. Or if you want to model new treatments, like neurostimulation, then you need a better understanding of the coupling of the heart with the nervous system. So it's not that it's the least interesting or the least promising, but it's the part of digital twins that still is going to be in the years to come. Yeah, I would add to that. I think probably one of the most promising areas is its relationship with precision medicine, which is probably somewhat obvious. But I think if we can truly harness the power of digital, I sound like an ad, but if we could truly harness the power of digital twins, then I think it'll lead to huge strides in personalized and precision medicine. And just for example, if you can take all of these incredible predictive models from AI that are coming out for predicting heart failure, exacerbation, et cetera, but then you personalize that. So that you, rather than just applying this generalized score or even if it is a good one, you can personalize it to all of the different features of that particular patient, if you have there somewhat of a digital twin of them that already exists. Great, thank you. Another interesting question here. Someone asked in a biological system, how would the digital twin account for important environmental factors that influence response? So I can tell you from my experience at Hopkins, because as I said, now I work with a lot of aspects of medicine and trying to bring AI and technology. So for instance, in cancer, they do that. I know somebody who's a very famous researcher in cancer medicine at Hopkins and her lab is bringing digital twins of the environment of where the tumor develops and accounting for the environment allows them to very accurately predict what is the growth, how the growth will occur. So these things, depending on the specific problem, I think if the problem that we're addressing, let's say in electrophysiology, is such that it has the influence of the environment in some way, then that would be very important to account for. But I have seen with my eyes how they do it in cancer and it's definitely possible, yeah. Do you guys have anything to add to this? I would just say that it's all about the inputs, you know, I mean I think if you're looking at macro environment and you're, you know, whether it's the environment that patients in or the anatomical or physiologic environment of what, you know, whatever disease process you're working with, if you find a way to include those inputs in your digital twin and you do so with fidelity and relatively good frequency, then you will have a good representation. It's just, I think that's one of the complexities of large digital twin products, is where do you stop? Where do you stop with inputs? Because it just becomes so complex. Yes, of course. I have seen some recent papers, for example, that are able to establish a very clear link between pollution and inhibition of certain electrical channels in the heart. So, for example, if we know that information that can be one of the input of the model of changes in the potassium-sodium levels because of changes in the diet or because of having any problem with your dialysis, for example, patients of dialysis that we know and the changes. So definitely there are environmental factors that can be taken into account. Another frontier would be a better understanding, for example, of comorbidities. It could be diabetes in our patients, for example. So I think that is, for me, still one of the promising areas that we are still going to build in terms of environment. Great, great. Thank you. Another question. Creating a digital twin in cardiology, how easily replicated are they for the whole body system and will that put us closer to the precise GPT systems in precision medicine? So the heart is by far, from the point of view of digital twin, the one that is the best developed. We have been creating this type of technologies for the last 60 years. This is the development of the first cellular models of electrical activity. So we are by far the farthest in terms of progress and in terms of delivery of applications to the clinic, as Natalia has been highlighting. Second to that, I would say that it goes the respiratory system. I have seen very beautiful work being done in that scope. Then a translation to other organs, maybe you guys are more familiar with? Not specifically. I'm really only familiar with cardiovascular digital twin projects and I'm sure you all are probably even more familiar. I mean, the only comment I would make, I think, you know, to get at that specific question is a multi-organ system, true digital twin representation of a human is, I would say, relatively far off at this point. However, once we get there, I don't think reproducing them is going to be difficult. I think if you can do it in the setting of one particular patient, then translating that to other patients, that's not where the complexity, I think, will lie. It's bringing the various organ-based digital twins together and trying to get them to interact and give each other input and outputs that I think will be very complicated. That definitely makes sense. There's a fascinating question here about, given the dynamic nature of the human biology, are the predicted outcomes with a given intervention fleeting? Meaning, do they change? Are they dynamic? How minor or major of a change in physiology would impact the predictions made? Well, this is a loaded question because, you know, it's really dependent on the application. I'm going to return to this concept. We are not in the realm yet of having this array of possibilities to build digital twins of everything. We have to constrain what we are doing to a system, to a specific application, particularly like showing, to me the most important is, showing that it makes a difference in clinical practice. So, once you do some of these initial applications, you start to build on and include other responses and some of those may or may not be needed for a particular clinical application. Then you move to another and build it that way. I just really try to avoid thinking that we're going to sit down and build everything of everyone. We are not going to do that. I would like to make a difference in a particular clinical application to approve human health in a certain area, demonstrate utility of digital twins and then expand from there. Other thoughts, please, on this? I have to say, I think that's one of the most realistic versions of digital twin and medicine currently that I've heard because I think that, I agree that I think a full, fully, you know, integrated digital twin of a human being is a lot to ask and, you know, whether or not we'll ever get there. And so that's why it's so important to do projects like the one you're doing where you pick a particular thing, you constrain your inputs and outputs and what you're looking at and make it practical and then really to answer a question of care in the end. Exactly, that's the most important. I agree, I agree with what has been said. Definitely, as I was mentioning before, I also agree with Natalia, it depends on the application. If, for example, the application is to predict therapy, to predict whether there is going to be acute prolongation after the administration of this specific drug, we need to build confidence in the results, confidence intervals. So, for example, we are very good at the moment at the personalization of the conduction system, the personalization of the repolarization of the heart, it's trickier, so what we are doing is to complement our predictions with ranges of uncertainty that can also be of use to the clinician in order to bring additional information in those particular parts of the model where you could have more uncertainty. Can I ask how is that used currently? Do you use it using it in clinical practice? No, we are building up to there, but definitely we are extending, so we have one paper recently that is under review, for example, for the personalization of the conduction system and also the repolarization, and we are moving up to there. There is a question about the potential to integrate data from wearables into the creation of digital twins. Is that being done? It's being discussed a lot. I mean, even today I was discussing it with my people. It is actually very important. So, I do see the future of digital twins in more general, let's say in cardiology, is your digital twin has to be constantly updating with patient data. The way I see it is like you're in the clinic, you can pull from, you know, from Epic, let's say, the digital twin of that patient's heart, and you can sort of see what the status would be. Are they likely to develop this and that? And this digital twin is continuously updated from data from monitoring. That should be a goal. We should be going there, and given that how fast digital, you know, monitoring is developing, I think this is, it's really an important next step. So, for me, the realization of critical twins in the practice is going to happen if two conditions are satisfied. The first one is that the models are used in critical moments of delivering therapy to the patients. For example, the ablation that we have been discussing or tailoring the therapy, the pharmacological therapy, to a patient. And the second one is if the models actually guide the patients through the entire life inside of the health system. And for that, I fully agree with Natalia that we need to keep on updating the models through the journey of these patients through the health system, and that is going to come most likely with the data for the wearables. We have had these discussions, for example, for atrial fibrillation. Still not there, but there is a lot of discussion. I agree with that. That's very interesting, actually, exciting to see how that transpires. So, there is a question here that relates to one of the earlier questions about the accuracy of the digital twins. Someone said, if a digital twin is not truly perfect, what needs to change for us to take a leap of faith to use this in the real world? How much are we willing to accept? And I think we kind of addressed it a little bit in terms of how much inaccuracy, we focus more on the usefulness. Yeah, for example, I mean, without even the need of a digital twin, perfect replicas of the specific patients, our research has already proved that even the actual models, the current models that we have for the heart, they are more predictive than animal models in order to do the prediction of the safety of different drugs. So, we are, with computer models, already doing better for the prediction of risk of cardiotoxicity than using animal models. That is the standard in the industry, and that was the reason of the entire SIPA initiative that was promoted by the FDA. And that is without the models being perfect, because they may not be perfect, but they are much closer to the reality, to representing the target organism that is the human that can actually be a rat, a guinea pig, or a rabbit. So, that is one example, and definitely, in terms of in silico trials, there is quite a lot of scope for the use of these technologies. So, not being perfect doesn't mean that they are not useful. Also, Natalia was mentioning before. But I want to address here a different perspective of that question. It is the trust that you have with the clinical colleagues. When you're bringing something in the clinic, it isn't like, hey, here, use that. It's never that way. You have to have trust and rapport, and developing that is equally important as developing a digital twin, honestly. And, you know, it is also how you approach that. How you, if you know you want to do a clinical application, if you want to, if you're going to want to bring your technology into clinical practice, it's so important to convey the ideas, to be able, look, I, and, you know, my view of, as an engineer, where I meet my clinical colleagues in understanding of a project or understanding of an application, I never meet them at 50%. I'm going to go at the 80%. On my part, I really do. I believe that's how technology, they're the people who do the care. I am the one who wants to make the care better. So, I have to work with them as much as I can. I was, we had a really fun discussion today at lunch, and I said, I treat my lab, I, the way I want my people to be educated is bilingual. They have to speak clinic the same way they can speak technology. No difference. Equally between the two. Only then you have this seamless acceptance of ideas, and ideas can be conveyed in a way that the clinician understands. Just, just the communication, to me, is equally important as an accuracy. You, you have to have the buy-in. You have to have them wanting you to be there. You know, I love it when they cancel clinical meetings because I don't show up, and I'm at the meeting, right? I love that, because they believe that, that technology becomes the center of clinical, clinical application. Wow, that, that's, that's, I completely agree with you. That, that type of collaboration is so important, and that type of trust-building as well. So, speaking of trust, one question had to do with, are people, like, as patients, are they having more trust or more confidence in terms of undergoing procedures based on the results of digital twins? And I suspect we're still not sure. This is not an implementation yet, but it's an important question, because, like, once we have the answers, once we know, yes, you can actually use digital twins to select patients, to predict response, to do this, to do that, how do you sell it to the patient? And, and the medical community is an important question. Your thoughts? I, I have to say, there is a group of patients that are so well-informed. They really, they've taken their healthcare in their hands, particularly, I'm talking about, let's say, AF ablation, OVT. There are patients that follow my papers. They follow what I do. There are patients that have asked me so many times, can I come and be part of your clinical trial? Can I come to Hopkins and be treated? I have one that I did yesterday, so I'm trying to organize that they get scanned in somewhere else and send us the scans, and I'm gonna just do the prediction there, and I've talked to their doctor in California, and the doctor is willing to do it. Now, this is not part of my clinical trial, but people actually, there is a group of patients that are really interested in these novel technologies and want to be, you know, there isn't, there is no necessarily that big of a safety issue. I mean, it is, you know, anyways, what do people ablate? Extra PBI, anything. So, from that perspective, they, they really get that, you know, self-educated, and I think this group of patients is going to drive that. Also, kind of, you know, spread these applications beyond, sort of, the original institution, the original application. I mean, that's actually, at this point, a minority of patients, but I can definitely see it expanding. Oh, without a doubt. Wow, I'm actually impressed that they read your publications. That's, that's very impressive. There is a lot of engagement with different patient organizations, atrial fibrillation, patients with cardiomyopathies, they are really engaged. Still, this is something that is going to be developed even further, but definitely the interest is there in the community. So, that's fantastic. To build on this, there's another very relevant question related to any issues that relate to patient privacy and security, and do patients own any of it? Are they able to access the file, the records? Can you comment on that? Okay, so, first of all, for any patient that's enrolled in our clinical trial, they sign a consent that we, as engineers, can see any of their data. So, we don't own the data, but we can see it. We don't have to even anonymize the data, which typically is not the case, because we are outside of the covered entity. So, I normally cannot see patient data, but of these patients, I can see the data. Okay, so, that's one aspect, and, and if you, when you want to reuse the data, we typically, you know, we can consent the patients again, and so forth. But, there is another very important aspect, and it's kind of tangential to that, as well. It's when you use AI, as well, and that was, it was another session here. I was listening yesterday. It's the whole issue of using patient data for some of the developers of algorithms that are outside of the covered entity, like me. I have figured it out only

digital twins

cardiology

personalized models

atrial fibrillation

ablation targets

wearables data integration

trust between engineers and clinicians

patient engagement

data privacy and ownership