Hello, I'm David Haynes, Director of the Heart Rhythm Center at Beaumont Health in Royal Oaks, Michigan. I'm pleased to meet with you today, and I'm very pleased to introduce Theo Mour-Dordes, who is the co-founder and CEO of Volta Medical, exciting new technology, and we're looking forward to hearing what you have to say to us today, Theo. Thank you, Dr. Haynes. So hi, everyone. Today, I'm going to present artificial intelligence approaches we are implementing at Volta Medical to improve the outcome of the ablation of cardiac arrhythmias. So my only disclosure is quite obvious, as I am the CEO of Volta Medical. So Volta was founded in 2016. We are currently 21 with headquarters in Marseille and a subsidiary in the US, and our objective is to leverage data-driven approaches to assist cardiac electrophysiologists in the EP lab. Here what we mean is that we are developing software solutions to, A, consistently reproduce and provide state-of-the-art clinical expertise in the EP lab, and B, try to learn from past procedural outcomes to improve and go beyond human knowledge. So A, we want to mimic clinical expertise, and B, we want to try and go beyond human knowledge. So from the beginning, we have chosen data-driven approaches to solve these types of problems. Why? The reason why we think these type of approaches particularly fit to EP are the following. First, EP is a highly digitalized medical field, and procedurals generate a lot of numerical data that are both rich and complex in terms of content. Second, human analysis and clinical expertise often goes beyond the description provided in papers, and this is particularly true for electrogram-based approaches in cardiac electrophysiology. And finally, procedural data such as electrograms harbor complex features that are hardly captured by conventional posse-parametric algorithms. While more conventional analysis approaches are based on academic descriptions that are often restrictive, machine learning algorithms, if well-designed, will be able to capture more complex or more subtle features and dependencies within the data. This translates into the use of multiple complex features and statistically determined thresholds that allow for more accuracy. All in all, such approaches are meant to be more representative of real-life situations. Our first challenge at Volta was to mimic a clinical expertise aiming at identifying multipolar electrograms exhibiting a specific pattern known as spatial temporal dispersion. Our goal was to be able to provide physicians with an assistance to identify this electrogram abnormality in real time. While the broad principles for its identification had been laid out in the JAC in 2017 by Seitz et al., it remained a challenge for most operators that tried it in their own lab. In fact, this human ability goes beyond the academic paper description and requires a lot of training. As you can see and appreciate on these examples of multipolar electrograms recorded during AF procedures, several features such as morphology, voltage, activation sequences, fractionation, time-wise evolution, noise, and far-field have to be taken into account. To reliably mimic this expertise in the epilab in real time, we have developed a first software solution, which is now CMART, and the name of this solution is Vx1. To develop Vx1, we have collected a large amount of multipolar electrograms that have been annotated for the presence or absence of spatial temporal dispersion. This database comprises various configurations and has a rich content which makes it highly representative of all situation an operator may face during an AF procedure. We have then trained a machine and deep learning models that have been integrated into a user-friendly product, and this product has been tested extensively in 300 patients in multiple centers already. The results are under analysis, and today I will only focus on a study dedicated to comparing the performances of Vx1 versus the ones of trained electrophysiologists. In most machine learning tasks, unambiguous ground-truth labels can easily be acquired. However, this is a luxury, and this luxury is not afforded to many high-stakes real-world scenarios such as medical image interpretation, where even expert human annotators typically exhibit very high levels of disagreement with one another. In our case, we needed to understand how Vx1 compares with trained human operators and how operators compare to one another. For this, we have conducted what we call the reader study on a very small dataset. It consists of a dataset of 14,000 electrograms recorded in 9 patients with 3 catheters. Please remember, this is not the training database, but a specific small testing subset, so it's like testing a car in very specific weather conditions. These electrograms have been annotated by Vx1 and also by 3 trained electrophysiologists with unlimited analysis time independently. Annotations have then been compared one to another. We have defined composite adjudications for the physician's annotation. Electrograms adjudicated as non-dispersed by each of the 3 readers are considered non-dispersed. Electrograms adjudicated as dispersed by one reader only are considered little dispersed by 2 readers, likely dispersed, and by each of the 3 readers, they are considered very likely dispersed. Now, what you can see on this slide are 3 histograms showing electrograms annotated by Vx1 as normal in blue, dispersed in orange, and highly dispersed in red. These show that Vx1 adjudications strongly correlate with the composite physician's adjudication. The vast majority of blue electrograms are considered non-dispersed, the majority of orange are considered likely to very likely dispersed, and the majority of red electrograms are considered very likely dispersed. Now, if you look at this graph that was already presented by Dr. Khalifa at HRS this year, you can see that operator Vx1 agreement levels are much higher with unlimited analysis time than agreement levels if operators are given a limited time for analysis in blue. It shows that unlimited times really is an added value, but this is not offered in real time in the operating room. Also, what we have understood is that operator Vx1 agreement levels are on par with inter-operator agreement levels. So these are our conclusions for the reader study. First, Vx1 adjudications highly correlate with the likelihood of electrograms to be classified as dispersed by a group of trained operators with unlimited time. Second, agreement levels are much higher when operators are given unlimited analysis time. And third, Vx1 cannot be distinguished from a trained operator with unlimited analysis time, but works in a real time configuration. So finally, what we can say is that Vx1 is able to mimic a very specific clinical expertise at its best. But data-driven doesn't mean magic, and this approach raises a lot of challenges on which our data science team is constantly and intensively working. How to correctly translate a clinical need into a theoretical problem? How to choose relevant performance metrics? How to define the gold standard and ensure data representativity? How to replicate or mimic existing medical knowledge, and should we? How to make algorithm robust? How to efficiently annotate data? And how to collect data efficiently and safely? And this is probably the most critical question. So what's next? What's next in terms of science? We are developing efficient and safe detection tools to grow databases. We are developing new machine learning modules for AT and VT mapping assistance. And these modules will then be integrated in hands-on EP lab software solutions that we intend to commercialize. I would like to conclude by saying that I am extremely grateful for having a brilliant team of passionate inventors and innovators. I am very confident that we will be able to assist electrophysiologists in the EP lab, and ultimately, by improving the procedural workflow, that we will ultimately be able to improve outcomes for patients with cardiac arrhythmias. Thanks to the HRS for hosting this virtual session. Theo, that was a great presentation. Really exciting area that you're developing. Now, my understanding of the spatiotemporal dispersion is that this is the modern evolution of what we used to call CAFEs, basically trying to identify driver sites by areas of fibrillatory conduction wave break from a driving focus. Is this your understanding of the mechanism as well? And how is your system going to enhance our ability to interpret these images? Thank you, Dr. Haynes. Regarding spatiotemporal dispersion, I think that indeed, spatiotemporal dispersion is a refinement of the CAFE definition as laid out by Wina Demamy in 2004 in The Jack. Now, there are a lot of mechanistic explanations. There are also some theoretical diagrams and explanations in academic papers. But, and this is what I wanted also to insist upon in my presentation, at Volta, what we are trying to do is not to say that we are understanding something about the mechanisms, but rather to say there is a clinical expertise that works in some hands, somehow it's very difficult to reproduce for other operators. And what we are going to do is to bring this expertise in everywhere in the world by being able to mimic visual expertise at identifying specific electrograms. So obviously, there are a lot of mechanistic explanations that are linked with spatiotemporal dispersion and why spatiotemporal dispersion may be a signature for regions of interest and regions that may perpetuate the arrhythmia. For this, I think the best is to look at the work of Jerome Califa, but also of Natalia Trajanova, who has done a lot of work on this very specific topic and on the link that exists between mechanistic understanding and mechanistic explanations and clinical situations. So even if there are a lot of mechanistic explanations for spatiotemporal dispersion, we want to mimic a human expertise and a human analysis rather than giving some mechanistic insights that exist. In terms of the validation set for the AI, I gather you're using expert readers, but there's variability among expert readers of these electrograms. How do you determine what the truth is in order to train your AI? And I'm also interested if you're looking at ablation procedural outcomes as a factor for identifying the most predictive patterns with your AI. Indeed, if you are in a clinical configuration, there is a high level of disagreement between operators, so there is a high inter-operator variability. But what we have shown in the reader study is that if you have trained operators or if you train operators, if you put them in a room and if you give them unlimited analysis time, you will drastically reduce this variability and you will end up with agreement levels that are acceptable. So we have specific metrics that show that these agreement levels are acceptable. So EPs are able to agree on the identification of specific electrograms. So that's the first thing. So then we have shown with the reader study that we are able to have such agreement levels with operators, so between VX1 and operators. And second, we have also conducted some clinical validation in a clinical setting, notably where we have shown on a small group of patients that we were able to have maps, special temporal dispersion mapping, that was done actually in a very similar fashion when performed by the algorithm and when performed by the human operator. So basically we have been able to compare maps that were done visually and maps that were done using the software, and this was done in a blinded fashion. Now regarding outcomes and the use of outcomes, this is a topic that we are working on very intensively because we think that this is the way to go beyond human knowledge. This is the way to really improve cardiac ablation. And so we are working on it for current projects, and it's very exciting. So we try to take into account procedural events that are relevant for the outcome and for probably for the long-term outcome. Now it raises obviously a lot of questions between acute outcomes and long-term outcomes. That's fantastic. And I would imagine as you get more users, you'll be able to continually feed back and compare clinical outcomes to interpretation by your system of the electrogram pattern. So the potential for this is terrific. Exactly. Thank you very much. This has been a great presentation. We're really looking forward to future developments coming from your company. Thank you, Dr. Haynes.

artificial intelligence

cardiac arrhythmia ablation

Volta Medical

Vx1

machine learning