Artificial Intelligence and Inherited Arrhythmias

Podcast Discussion With Lior Jankelson, MD, PhD, and Zachi Attia, PhD

In this episode, we are highlighting a discussion on artificial intelligence (AI) and inherited arrhythmias. Lior Jankelson, MD, PhD, is the director of the inherited arrhythmia program, PI of the computational cardiology research lab, and associate professor of medicine at NYU Langone Health in New York. Zachi Attia, PhD, is the codirector of AI in cardiology and an assistant professor of medicine at the Mayo Clinic in Rochester, Minnesota.

This episode of The EP Edit is also available on Spotify and Apple Podcasts!

Lior Yankelson: Hi, my name is Lior Yankelson. I am a cardiologist and cardiac electrophysiologist with NYU Langone Health in New York City. I am also the PI of the Computational Cardiology Research Lab and director of the Inherited Arrhythmia Program in the Heart Rhythm Center. Today with me here is my colleague and friend, Dr Zachi Attia.

Zachi Attia: Hi, my name is Zachi Attia. I am the codirector of AI in cardiology at the Mayo Clinic in Rochester. I am an AI scientist by training and lead a team of 8 PhD-level engineers doing all kinds of interesting AI work with different signals relevant to cardiology, such as echocardiograms and electrocardiograms (ECGs). We hope to find and diagnosis diseases earlier and faster.

Yankelson: Zachi, I thought we would start by having you tell us briefly about your general, more preliminary work that you have done with Mayo on AI and ECGs.

Attia: Sure. So, when we started our work on ECGs, the most common way of doing things was to look at an ECG as a human. We would take the ECG, break it apart, and describe it. We said the P wave was this tall and the PR interval was this wide. We took all of the features or descriptors of the ECG and trained a model to take those features and try to predict the disease. The first disease we focused on was hyperkalemia or high levels of potassium in the blood that are dangerous and can lead to devastating arrhythmias.

At some point, we hypothesized if we could get a better model by reading the whole ECG. At the time, the whole area of AI and deep learning that allows one to feed in raw signals or images without describing them at all was just starting. We fed in ECGs in the same way, and when you do that, you basically replace the human features by patterns that the machine learning algorithms find by itself. So, all you show it are groups of “good” and “bad” ECGs: one with the disease and the other without the disease. The AI itself doing the training finds the patterns or features that are relevant to it. Once you do that, you do not understand which features become a black box, but you have to spend much less time on engineering those features, and most of the time, you get more accurate and faster results.

So, we did that, and then we wondered if we could extend it to other diseases. The first disease that we focused on was low ejection fraction (EF) or systolic dysfunction, which is basically a weak heart pump. We took 100,000 patients with and without the disease, and we showed that even though a human cannot look at an ECG and detect low EF, a machine could, which we found exciting and interesting. My colleague here also did a lot of interesting things about the concept of black boxes and what are we missing. So, I think we really see the benefit of using an unbiased approach using AI to detect diseases.

Yankelson: Before we dive into that corner, tell me what you think about this statement. I think a good way of thinking about AI is like an intuition, such that when cardiologists or clinicians are taught to interpret an ECG, we generally have a lot of rules and spend a lot of time in learning these rules. Then we look at the ECGs, measure them, look at the duration of each interval, and make a determination. But ultimately, after seeing a bunch of ECGs, and with time and experience, we kind of stop doing that and just look at the ECG and say, “We think the diagnosis is X or Y.” I think this intuition is what is embedded in our brain, and that is, if we have to identify one’s function of the machine, of the AI, that kind of maps to what we are doing as humans. I am curious to hear what you think about this.

Attia: I completely agree. As I said earlier, humans cannot detect low EF. One of the other things that we detected was the patient’s age and sex, which is, again, not something that most people would know, but there is always a cardiologist who has seen so many of them that they can definitively look at the ECG and guess the patient’s age and sex. It is not a magic box. There are patterns, though. We just have to see hundreds of thousands of those, probably, to develop that intuition. By using AI, we can speed up the process of cardiologists developing this strong intuition from 40 years to a few hours.

Yankelson: So, if we agree that we are basically teaching the computer an intuition, how is it that we can solve with AI the problems that are not solvable to the human? Is it because we are essentially combining the intuitions of maybe millions of human beings? Or, is it that the computer does something completely different than replicating the human intuition?

Attia: I think it is divided into a few areas. Some things we can just do faster and in a more scalable and consistent way. Think of reading a rhythm from an ECG. It is something that cardiologists can do, but we can get AI to do it in the same way all the time. This is one level of AI that would allow us to provide better health care to patients in rural areas, for example, where there is often one cardiologist for thousands of people.

The second level is the things that humans cannot do. Again, I agree with you, it is just probably combining knowledge from thousands of clinicians or combining multimodal. So, we take information that only exists in the echocardiogram and match it with the inputs of the ECG. I think that the patterns are always there. It is a black box, it is not a magic box, but you really have to look at many of them to understand those.

Now, these features might be nonlinear, and humans tend to think in a linear fashion. So, maybe these are things that humans can spot, but the reason that most humans cannot take those findings of the ECG and convert them to the diagnosis, is that linear thinking. We do have these savants or mavericks who can look at the ECG and guess it because in their head, they are doing nonlinear thinking. They see how these features combine and maybe that is the reason they sometimes cannot explain why they see it. Sometimes you look at an ECG and do not know what bothers you, but something bothers you.

Yankelson: Right. I think what you are touching on are the features or components that we have actually learned by the machine. The way I think about it is that any kind of input—whether it is an image, a signal of an ECG, or an echo—can be ultimately reduced to crucial components that make the difference between, let’s say, 2 examples. So, if we look at 2 faces, one of the reasons we are so good at recognizing the difference between human beings is because our brains are wired and have developed to extract some important features, and by using these features that are critical, assign an identity and say, “This is Zachi. This is Lior. This is not the same person.”

But I think the difference between machines and humans, in this regard, is that we have a lot of other layers that work in parallel to that pattern recognition or feature extraction that protects us from making mistakes. For example, one of the studies that we have done is we essentially trained an efficient model for identifying the rhythm on an ECG. Then, once we understood the pattern that the model was actually looking at, we were able to inject tiny artifacts into the signal. The artifacts were so tiny to the human reading that the ECG was not altered at all, it was exactly the same ECG. However, because it included some tiny artifact, but an artifact that was important as a feature, it completely disrupted the AI model for the ECG. So, I think that a lot of the forefront of AI research these days is centered on the question of how to encode information in a way that is robust and efficient. Can you tell us a little about encoding and decoding?

Attia: Yes, that is a nice way to describe what AI is doing. Because in a way, it converts a signal to a mathematical space. You can imagine a 2-dimensional space—X and Y and dots in different colors—but in real life, it is multidimensional and maybe thousands of dimensions. We encode ECG to a single dot in that multidimensional space, and then we slice that space and say that everything around this area is a disease and everything outside of that area is not a disease.

There are many ways to do it. One of them is to do it directly. You compress the ECG, take the input and try to generate the same output, but put a bottleneck in the middle that makes it only focus on the most important things. Because if you can describe the whole ECG by 10 variables, and you can take those 10 variables and get back the ECG, that means that those 10 variables contain all of the information you can extract from it. Maybe using those, I am throwing out the number 10, it can be 32 or 64, it does not matter, maybe those 10 parameters encode all of our knowledge.

Now, most of these networks are trying to do one specific thing. Most of them detect low EF or atrial fibrillation (AF). So, these parameters are adjusted for that space. But a more global presentation of the ECG is super exciting, because if you can encode the whole ECG, not a specific disease, not with that specific area, you can think about things like your digital twin. Can we take someone with a very similar ECG to your ECG—even though visually they look different, the heart rate is different, and maybe the morphology is different, but the essence of the ECG is similar—could we can use that to learn what will happen in the future? Maybe we know that the same patient is going to have a heart attack in 5 years, because we are looking at that retrospectively and we can adjust your treatment. Maybe they have a genetic disease that we have not found the gene for, but there is the same fingerprint in the ECG.

Yankelson: Yes, exactly. I think that is a nice segue to newer frontiers of AI. But another example to what you are alluding to is that if we do a deeper dive into the electrophysiology (EP) angle of the ECG, many of the problems that we have in EP are that we are limited in our ability to move forward because of a lack of understanding of a mechanism. For example, when you are trying to train a model or teach a machine to identify low EF, then you have a lot of examples, we have a lot of ECGs and we marry them with the echos. We then have a lot of examples of ECGs married to normal echos, abnormal echos, and echos of people with heart failure. By that process, we are able to teach the computer to determine between the two. Many of the problems in EP are different in nature. For example, if you think about AF, or even more specifically about persistent AF or AF that does not necessarily derive from pulmonary vein triggers, the problem is we do not know how to explain the reason for that phenomena or characteristic that most importantly describes the nature of the physiological phenomena of the ECG, or of the echo, or anything objective that we can measure.

If we are trying to predict an outcome, that is difficult, because we do not really have a label. So if you want to do something like look at an intracardiac map and try to predict who is going to benefit and not benefit from an ablation, it is difficult on multiple levels. One problem is that we do not have a lot of examples. Even if each of our large centers, NYU and Mayo has maybe thousands and thousands of examples, it is still not at the caliber of ECGs and echos, when you can reach many hundreds or even millions of examples. The other problem is that we do not have that label. We do not know what to say about the AF map—how is it different from another map? It looks the same. It is a small, distorted, high-frequency signal.

I think that is one of the places where the more general encoding frontier can help us very much. Because if we can teach the computer something more substantial or profound about the signal, then we can potentially ask questions that we do not have good labels for. The way we differentiate these approaches is between what we call supervised and unsupervised, or self-supervised learning. Zachi, what do you think? Where can we get with self-supervised or unsupervised learning for the ECG?

Attia: I believe that self-supervised is probably the future of AI in general, because we are running out of labels. Even with the amount of data we have, there is just not enough. When you think about rare diseases, when .01% of the population has a disease, you will never have enough labels to find it. When we think about humans, how do you teach your daughter or son to think? We do not go and label everything in life. They just see things. They try it, see the response, and learn from it. We show them few labels to learn language. But most of our learning is by trying or trying understand the basic structure of things.

So, I think it is important to find a more global approach in the encoding and decoding of ECGs in that regard. We see the same in language models.

I think we are close to doing the same in cardiology. By having the same understanding of the ECG as we have of language models, we can then, with a small label data set, take this global understanding and focus it on a single disease. Eventually, we will have to have a small label data set for the end of the model. But I think it is a great opportunity.

Yankelson: Yes, and particularly, if I want to take it to the inherited arrhythmia space, that is a great example, where trying to use a supervised learning approach could be frustrating, because we do not have enough examples. For our most common inherited arrhythmia, which is arguably long QT syndrome (LQTS), how many patients with LQTS are there? For how many do we have access to their ECGs? If you want to go even deeper, you can start to self-divide into their genetic background and into smaller segments and populations. So, it is difficult.

Another interesting publication that I saw recently was related to reproducing the polygenic risk score based on the ECG, meaning that we can essentially use an ECG to, for example, identify a certain condition by labeling the condition and training a neural network to identify that. But in genetics, there has recently been a significant interest in the polygenic risk score, in which we are composing these risk scores made of many genetic variances in many different locations along the DNA. Essentially, if you think about it, it might be possible to reproduce these polygenic risk scores from the ECG, just like we detect EF. So, in this example, instead of diagnosing a low EF from the ECG, we would diagnose a high polygenic risk score from the ECG.

I think if you combine that approach with a nonsupervised learning approach, that is where we are going to get a lot of power for AI, because on one hand, we are going to use methods that not necessarily require a significant amount of examples. On the other hand, on the genetic plane, we are going to go a little bit away from the clear monogenic labels to polygenic labels or the polygenic spectrum. Then, essentially, we are broadening the example dates from the diagnosis end of the rope. So, the future is promising.

Attia: Do you think the information in the polygenic scores is all included in the ECG, or do you think there are still things that we will only be able to detect by having both the ECG and the actual sequencing?

Yankelson: That is a great question. On a broader level, the question is, what is the resolution of the ECG as a reflection of the genetic substrate? If that is what you are alluding to, I think that it is not as precise as that genetic data encoded in the actual DNA, but it is much deeper than what we think we see. So, between what we think we see in the ECG or know how to explain or identify as humans, and the breadth of information that is embedded in the DNA, there is a large space where we can dive into and extract value from.

At the end of the day, it is going to be a moving scale where we are going to use different combinations of genetics and ECG, depending on the condition, but I think that just by using the ECG, we can already extract a lot of value in terms of, for example, identifying risk that is basically encoded in the DNA. LQTS is a great example, because 80% of the LQTS population is what we call gene positive. They have a monogenic mutation that is the pathogenic substrate and the pathophysiological path to clinical expression and clinical outcome. But 20% are not explained by a monogenic finding, and that population has been demonstrated to have what we call a high polygenic risk score, meaning they probably have a lot of small mutations in many different spots that combine together to increase the risk of having a long QT.

Now, this is difficult to identify for the human eye, and definitely in such a diverse population. But I think at least a significant portion of the high polygenic risk score is reproducible from the ECG. And just like we do that for LQTS, we can probably do that for many other conditions. There is also the space of time, the temporal space, where I think you have nicely demonstrated in the EF project, where at least part of the false-positives were not really false-positive, they just predicted something in the future.

Attia: Yes, this is an amazing concept. We can think about all of these patients who have some clinical disease that AI sees but that has not yet manifested. So, we count them as false-positive, but then we detect something on AI that we do not know yet. I think the concept of concealed LQTS that you mentioned, in patients with the genetic mutation and risk of disease that is invisible to the human eye—if you and other experts can look at an ECG and detect the same pattern, but make it scalable to AI in ways that you can record ECGs with your watch or phone, it is a big deal for community health. We might think of cascade testing where we say, send us a singular ECG—we will run a quick analysis using AI and decide if genetic testing is needed.

Yankelson: Yes, that is an interesting opportunity for AI and ECG, because we can both use it as a cheap and readily available screening tool. The ECG is so abundant now—almost everyone has an ECG in their hand. Another great opportunity is for preclinical diagnosis, where we do not have any manifestation of symptoms or disease, but we can maybe flag the high-risk population that would require different types of follow-up, such as more frequent or deeper follow-up. So, these are 2 interesting frontiers.

Zachi, it was a huge pleasure as always to speak with you! I hope it was interesting for our audience, and I hope we can continue collaborating. Thanks again for doing this with me.

Attia: Thank you for inviting me. It is really a pleasure to talk to you, exchange opinions, and learn so much about the cardiology side of it. n

Editor’s Note: The transcripts have been edited for clarity and length.