DocTalk Podcast: Using AI to Read ECGs with Dr. Kapa


Suraj Kapa, MD, takes part in an episode of DocTalk where he discusses the results of a recent study he led and how to implement technology into cardiology in a way that benefits patients and physicians.

doctalk podcast

As technology in medicine continues to grow by leaps and bounds, the onus on incorporating it in a manner that benefits both patients and physicians has never been greater.

A recent study into use of artificial intelligence (AI) to predict overall health using 12-lead ECG readings is one of many examples in cardiology where physicians are looking to do just that.

Suraj Kapa, MD, cardiac electrophysiologist at the Mayo Clinic in Minnesota, was the lead investigator of that study and he is the subject of this episode of the DocTalk Podcast.

MD Mag: Hello, everybody, and welcome to the DocTalk podcast. I'm Patrick Campbell, associate editor with MD Magazine, and I will be your host for this edition of DocTalk as we discuss using AI to determine overall physical health from 12-lead ECG readings with Dr. Kapa. Welcome to DocTalk, Dr. Kapa, if you wouldn't mind introducing yourself to our audience and telling us a little bit about your background, and listing any relevant disclosures you have before we begin, then we can dive into our chat.

Kapa: Yes, absolutely. So, my name is Suraj Kapa. I'm a cardiac electrophysiologist at the Mayo Clinic in Rochester, Minnesota and in addition to being a practicing electrophysiologist, performing cardiac ablation, I also do research related to both artificial intelligence, while also directing innovation related to an augmented and virtual reality. In terms of disclosures, related we're talking about today and all relevant disclosures.

I do serve on advisory boards for Boston Scientific and Abbot, but otherwise no other significant disclosures to speak of.

MD Mag: Just to start, I think the obvious question for our listeners who don't know, Dr. Kapa recently played a major role in a study that examined use of an artificial intelligence program that utilizes EKG or ECG data to measure a patient's overall health status. Now, Dr. Kapa if you could sort of take us through the key findings of this study and why you feel they're clinically impactful.

Kapa: Absolutely. So, as many people who are starting to get interested in this area have probably seen in the last, I would say few years, with improving computational abilities, we've actually seen the opportunity to interpret large scale amounts of data and actually not just recapitulate what we're able to do as physicians, such as reading a chest X-ray, or figuring out what a bundle branch block looks like on an electrocardiogram, but we've actually been able to obtain new insights.

So, those have actually been demonstrated in a couple of papers we've published in the last year, the first one being in Nature Medicine, showing that we can use the simple 10-second, 12-lead electrocardiogram to determine if somebody has low ejection fraction or not without any other data. And subsequently, in The Lancet that we can actually determine if somebody has atrial fibrillation just from a normal sinus electrocardiogram. But these are things we expect to be manifested within electrocardiogram because they relate to cardiac health and cardiac activity. But one of the key things that we were able to find when we were doing these analyses previously, was that despite the fact that in cardiology standards, age and sex, really predict a lot of what we have in atrial fibrillation, for example, or ejection fraction, prediction models. Age and sex always play a major role. But for whatever reason, for us, when we're using the simple form factor of the ECG, it didn't, which led us to the question of is the electrocardiogram actually seeing age and sex? And if so, how is it seeing it? And what role might that play?

So, that's actually what led us to evaluating on a large scale number of patients on several hundred thousand people, whether or not the ECG was actually seeing the age of the individual and the sex of the individual. And what we were able to see from that is, with almost a 97%, accuracy could determine if somebody was male versus female. And on average, it could determine somebody's age, within about 7 years of their actual I would call it chronologic age.

Now, where this becomes interesting, and I mean, a lot of people can say, "Well, can you just ask a patient?". Are you male or female? And how old are you? But the thing we're finding interesting, and we're excited about is in the outliers in the people where the ECG thought they were something that, in fact, they weren't supposed to be — that the ECG predicted age, for example, is significantly older than their actual chronological age. And the deeper, interesting aspect of that was the fact that when they're easy to predict age was much higher, they had a lot more comorbidities. They had a lot of other diseases, and they weren't necessarily the standard array of cardiovascular diseases. There were things along the lines of diabetes without cardiovascular complications, or obesity — and that's where we find a lot of this interest in what we've been able to see on this initial foray.

MD Mag: Looking at the data from this recent study, what are the effective ways— Or what are the next steps a physician has to take to properly implement AI like this into their own practice?

Kapa: So, I would look at this in a three prong way. Now I'll start with the last question for us, such as how do we apply AI to our practice? Now, one thing we need to realize is whenever we identify these models and develop them, and just the high levels of accuracy, we always need to make sure when we're deploying them, that they're similarly accurate when deployed amongst different populations.

Whether that be because of differing racial or ethnic, racial or ethnic proportions within the population, their consideration, or even socio economic factors that might lead to differences and environmental factors that may lead to differences in the raw signal. So, both things always need to be considered and the question of implementation science, how are we going to take these developed algorithms and deploy them amongst our patients or amongst people in our community is still an active area of consideration, I think, at the FDA level, as well as the researcher level.

Now, when you're talking about people interested in doing research in the space, what's interesting is the engineers and the people who work with an AI and especially with lower cost computing, are really democratizing the ability to develop algorithms and to use large amounts of data to obtain new insights into data. And the thing I would actually most point out is that, again, a lot of physicians and a lot of papers, and these articles about AI and medicine, always raise the fear factor of "Oh, is AI going to replace our jobs," and I would actually say the exact opposite.

While, AI can augment our insights into individuals, the fact is that we still have to be training the systems. And in many cases, the systems will actually allow us to understand patients better and maybe treat them more efficiently or effectively. For example, in this case, if we have a patient we're seeing as a primary care doctor, and there ECG as a simple form factor identifies it, you know what, there might be something more going on, maybe we do need to check the fasting glucose, if it hasn't been checked in a while or check the panel or drive them to reconsider certain lifestyle factors that are driving that age higher, or the predicted age higher? Those are things that might come out of that.

Now, another factor that you actually brought up was the question about the ECG. And, you know, we live in a very interesting time now, because while in this study, all of the ECG is were clinically indicated. They were, you know, people done retrospectively what ECG is, for some reason, certainly many of them are healthy, and they just got an ECG as a "screening tool," but that being said, the ECG as a current and evolving form factor is really going to be available beyond what a clinician prescribes. And this gets to the point that the majority of people in the world are not patients. In fact, the vast majority people don't manifest or show themselves to an institution or hospital or a clinic, or critical institution, until they actually have something wrong with them, or until they're driven to do so for some reason.

So well, smart watches and smartphones and even wearable underwear and T shirts are increasingly embedding the ability to get an ECG in real time, without the need for a clinician prescribing it, we might be able to use these insights to allow people to be better dedicated to present to their doctor at the time they need it, or when a disease is first starting maybe before it's become so critically apparent that our opportunities to intervene and prevent progression might not be quite as good.

MD Mag: All right. And now sort of a follow up to that last question. We're seeing a lot of progress being made and wearable tech technology specifically in the cardiology field. Apple watches the big name out there and live course recent approval of their device. Is this AI system that you guys have developed, another way of technology carving out its own special niche in cardiology?

Kapa: I believe it is and that does bring up two points. The first one is, you know, all ECG is aren't necessarily created equal in terms of our ability to deploy these algorithms. This goes back to that implementation science question we brought up earlier, you know, two different institutions, for example, use two different ECG acquisition systems, does the tool look the same on both? We are doing research into that in terms of validating our algorithms on different ECGs obtained through different form factors, including AliveCor, and others such devices.

So that's one element of it. And the second element of it in terms of be evolving form factors, you know, what I would say is, this has been a natural evolution of, I would say, the last 20 to 30 years in medicine. You know, if we go back 30, 40 years ago, we were all very much built out of, some research, some data, some statistics on largely retrospective populations, a few clinical trials and a lot of anecdotal expert-based opinion. Then came the era of evidence-based medicine, where we started working more and more with increasingly large data sets to obtain insights into how best to manage patients. When we saw the explosion of guidelines, scientific statements, and whatnot to help synthesize all these data together. And really, we have a data explosion, as well as an information explosion derived from that data.

Now, with that increasing those increasing big data sets, we have data such as Optum, Medicare, and whatnot, that are allowing us to get even bigger, larger skill insights. But the limitation of all of them has traditionally been the fact that they were based on claims data. They weren't based on the nuance things that we as experts look at when we're just staring at an ECG ourselves. They're based on whether or not somebody orders an ECG, but didn't give us the raw clinical signals of pain from just getting that piece of data — and that's, I think, where we're coming into this current era. It's a natural extension of the progression that we've seen. Where improved computational analysis, increased computer speed is now allowing us to really adjudicate very complex data, images, ECG is and whatnot and gain insights. That is, again, the next step in basically making what we're working on a little bit more systematic, and the way we interpret data.

MD Mag: Okay, that was really about it from questions on my end, was there anything else you want add about this particular AI device or system, or anything related to artificial intelligence and cardiology moving forward that I may not have touched on?

Kapa: So, you know, the only thing I would mention beyond all the standard approaches, it's important to realize that, while we call this artificial intelligence, we're not creating necessarily an intelligent machine. Really, if you want to, you know, think about this, it's really a much more complex version of mathematics. And there's nothing magical about it.

The reality is that when we're doing these neural networks as were described in the paper, it's basically allowing a system to learn by repeated exposure, and you're just exposing you to the same thing over and over and over and over again.

It's the same way we treat we treat our students when we expose them to many, many, many, many examples of an admirable electrocardiogram versus the normal ones. They eventually understand what happens normal versus normal is and that's what the machine is doing here. It's essentially building this network based on this repetitive exposure.

So, the most important thing to realize is that it really is nothing specifically magical or totally out there, but something that is, now, things we've been doing for centuries, and teaching and training. So, I think that's probably the most important thing to realize out of all of this.

MD Mag: Okay, now before we go, I just like to take the opportunity to say thank you, Dr. Kapa for joining us on the DocTalk podcast.

Kapa: Absolutely, no, no problem. Thank you so much for the questions.

MD Mag: That's it for this edition of DocTalk for the latest cardiology news, be sure to head to Thanks for listening.

Related Videos
Video 6 - "Evaluating Safety of Novel LDL Management Mechanism"
Video 5 - "Optimizing PCSK9 Inhibitors and Analyzing Plaque Reduction Data"
Video 4 - "Innovations in Small Interfering RNA (siRNA) Therapy"
Video 3 - "Ongoing Lp(a) Trials and Clinical Approaches to Treatment"
Roger S. McIntyre, MD: GLP-1 Agonists for Psychiatry?
Payal Kohli, MD | Credit: Cherry Creek Heart
Matthew Nudy, MD | Credit: Penn State Health
Kelley Branch, MD, MSc | Credit: University of Washington Medicine
Kelley Branch, MD, MS | Credit: University of Washington Medicine
David Berg, MD, MPH | Credit: Brigham and Women's
© 2024 MJH Life Sciences

All rights reserved.