Welcome to an episode of Heart Rhythm's The Lead, an abbreviated journal review podcast. I'm Deepti Vergis, coming to you live from Asia Pacific Heart Rhythm 2024 Scientific Sessions at Sydney, Australia. Joining me today is Dr. Tina Backner, Assistant Professor in the Department of Internal Medicine, Division of Cardiovascular Medicine and Electrophysiology at Stanford, and Dr. Janet Han at the VA Greater Los Angeles Health System and Associate Professor of Medicine in the Cardiac Arrhythmia Center at the University of California. Thank you both for joining us. Our study today is on Artificial Intelligence, Age Prediction Using Electrocardiogram Data, Exploring Biological Age Differences by Dr. Sean Evans and Dr. Praj Sanders from the University of Adelaide, Australia. Janet, can you get us started with a summary of the study? Yeah, sure. This was a great study to read, actually. Their study really looked at developing an AI model to predict age from ECG and basically compared baseline characteristics to identify determinants of advanced biological age. What they did was they took ECGs from their own institution-wide and collected them between the year 2000 to October 2023 with patients aged between 20 to 90 years, so it was a very large age range at that time, and then they externally validated their algorithm from ECGs that were from the UK Biobank, so again, a nice, nice study from that standpoint. From then on, what they ended up using was they did sort of the usual split for these sort of AI model training, testing, and validation, where they split the data that they obtained into a 64-16-20 ratio, and what they did was they also ensured that all the ECGs belonging to a single patient were all grouped together so that the data wasn't contaminated, so again, you don't want to use what you use in your training set for your validation set necessarily. They also made sure to oversample some of the minority age groups that they were sort of better balanced as well, and then I think we're going to talk a little bit later about the model that they developed, which was based on a convolutional neural network, so Tina's going to dive in on that a little bit later. So let's dive into the results a little bit. They included about 63,000 patients with a total of over 350,000 ECGs, and so what was nice is that almost 50% of the participants were female. We don't always see that in all studies, so I thought that was very important to see. And then their validation set consisted about 71,000 plus or minus ECGs, and what they found is that the correlation coefficient for that between the predicted age and chronological age was about 0.721, which is pretty robust. The mean absolute difference between the chronological age and the AI predicted age was about 9 years, so pretty large there. And so when they validated externally with the UK Biobank, what they used there were about 55,000 ECGs from about 51,000 patients, and of those, 52% were female, so again, pretty robust for gender, and the correlation coefficient between the predicted and chronological age was about 0.457, and the difference between age was about 7.6 years in that set. So in that internal validation set, they recalculated that as well because there was large differences in the baseline age distribution and found that the correlation coefficient there was 0.557 with the mean absolute error about 8 years. So when they looked at the true results of this study, what they found is that young people between the ages of 20 to 29 exceeded chronological age by a whopping 14 years, so that's very, very large. And when you looked at people at the oldest age set, which was 80 to 89 years, they found that these patients were actually younger by about 10 and a half years. What they also found is when they looked for gender differences, they found that men actually had older predicted age than women, so we're going to dive into that a little bit later in the study as well, and they found that the difference there was about 0.89 years plus or minus. They also found that patients who had multiple ECGs done also had an older predicted age by about 2 and a half years compared to those that only had a single ECG done. And so in the end, what they found was that older patients, women, and people with single ECGs seemed to fare a little bit better. That was an excellent summary. Thank you. Can you address some of the limitations of the study? Yeah, of course. First of all, congratulations to Dr. Sanders' group for this outstanding study. I'm always fascinated by ECG and AI studies, and this is a big one that trained a model to learn someone's biological age. The limitations are, of course, it's the original cohort of ECGs is all obtained from a single center. That's always a limitation. In our ideal world, the training set should have an international representation and ideally a representation of all ages in equal bins. I'm guessing, you know, they had an expected bell curve of they had under-representation of younger people, under-representation of older people, and you may wonder is that the reason why the prediction for those age groups were wrong, or is it truly predicting something biologically relevant for those age groups that are on the edge? In terms of a more technical limitation, maybe they use a convolutional neural network. Reading the manuscript, I didn't see too much detail on which network they chose to use, and on the, again, more technical aspect of it, there's a lot of data or there's a lot of manuscripts evaluating different kind of networks for analyzing ECG for prediction of various things. I wonder, you know, if they mentioned, authors do mention that they are presenting to us the best performing network, but I would have loved to have which networks they tried out and didn't like, which networks they didn't even think to try. I would have loved to learn that. So that's one technical limitation on my end that I can mention. Thank you for those limitations. Let's dive into some questions. How do you envision AI-ECG models evolving in the next decade? What are the key challenges and opportunities in expanding AI tools for personalized cardiovascular medicine? You want to take that? You want me to take that? Yeah. No, this is exciting. I think ECG is a perfect input for AI analysis. We've spent decades of traditional signal processing and analyzing ECG, and we already know that, gosh, 10 years worth of studies showing that machine learning networks can analyze an ECG better than human cardiologists. And that's just pure diagnosis of whether the ECG shows sinus rhythm, atrial fibrillation, SVT, et cetera. So there are so many studies to show that AI-ECG models can predict ejection fraction, age, someone's gender, someone's risk of cardiovascular mortality. These were done with very simple machine learning networks. Actually, the original study was from Stanford from 2014, predicting mortality with a very basic architecture network. So I can only see this going forward because the algorithms can see what human eye and simple signal processing can't see, and it will revolutionize our patient care. All basically from a simple ECG that we get dozens of times in one day, right? So considering the AI-predicted biological age, strong association with adverse outcomes, how might this tool be used to stratify patient risk in routine clinical practice? I think this sort of speaks to what Tina was talking about before. Is this actually looking at age, or is it actually looking at risk, right? So I could see in the future as more studies are done that if it is truly a marker of risk, we could potentially provide earlier intervention to mitigate that risk by lifestyle modifications or medications or what have you, right? What's interesting is that if you look at the study that was done several years ago, I think it's 2019, they published from Mayo Clinic, they did a similar study looking at a wide range of ECGs from their network, not just cardiovascular patients, which is what this study looked at, but in their study they looked at a very wide range of ECGs, and in their model what they found was basically that as patients had this sort of wide differential in their age, that age gap sort of got better as their cardiovascular disease got better. So I think you could be using that not just for upfront risk stratification, but sort of to follow patients to see how do they actually get better? How does their biological age improve with treatment of disease? That's a great point. Lastly, the study found gender differences in the AI-predicted biological age with women biologically younger than men. What factors do you believe contribute to this, and how could this information shape gender-specific strategies in cardiovascular electrophysiology care? This is a tough one, because you can develop algorithms that can predict someone's gender from an ECG. I'm actually trying to think if this was published or not, but I know of groups that have actually predicted biological gender from 12-lead ECGs, which is fascinating. I think this is independent of gender. I actually see this as the model is picking up signals on the ECG, or AI signals on the ECG that predict risk more than age. And we know the mortality of women differs from men, especially after a certain age. And I think the model is just predicting risk, is seeing signals of actual risk, such as it picks up signals of future AFib from a sinus ECG. I think it's picking up signals of future cardiovascular risk. And I think that's what we are seeing. I don't think, I mean, we might be seeing a gender signal, but I think we are seeing just their risk, and that might be the difference between men and women. I thought it was fascinating that they found that it was a 10-month difference, almost a year difference between men and women, right? What does that speak to? Does that speak to some hormonal changes? Does it speak to just the fact that women, in general, have better longevity than men do? Does it speak to sort of lifestyle issues, differences, you know, between men and women? Do men drink more, smoke more, you know, have maybe worse lifestyle management? And how do, actually, gender hormones, like, play into that as well? So I think that sort of still remains to be seen. Well, thank you both for joining us on this very special Live Lead podcast episode. Stay tuned for more episodes from Heart Rhythm Society. Thank you.

AI age prediction

ECG data

biological age

gender differences

cardiovascular care