Welcome to an episode of Heart Rhythms The Lead, an abbreviated journal review podcast. We're at HRX 2024 in Atlanta, Georgia. I have with me Dr. Sana Al-Khatim, Professor of Medicine, Division of Cardiology at Duke University Medical Center in North Carolina. Welcome. Thank you very much. Thank you for having me. Of course. Today we'll be discussing two of your papers. The Potential of AI to Revolutionize Healthcare Delivery Research and Education in Cardiac Electrophysiology and Understanding AI Bias in Clinical Practice. Can you give us a summary of the paper that you led? Yeah, of course. First of all, thank you for having me here. I do want to acknowledge the contributions of all the authors to this paper. This is the one on how AI can actually revolutionize EP clinical care research and education. We felt it was important to touch on all three. We start the paper by basically defining artificial intelligence, and then we give an idea regarding the different types of artificial intelligence, how machine learning, deep learning, as well as natural language processing all fit under artificial intelligence. We talk about generative AI. As you know, there's a lot of interest in that now in terms of the artificial intelligence being able to generate text or music or something in response to a query. Then we really delve into all the challenges that we as electrophysiologists and other clinicians, of course, face in clinical practice as well as in research and education. Starting with clinical research, or let me talk about clinical care first. Starting with that, we talk about how the patients that we see in clinical practice are changing. People are aging. They are getting sicker. They have multiple comorbidities. We all use EMR, so there is a deluge of data that we receive, be it test results, be it messages from patients, from colleagues. We are getting inundated with data. The other thing is, as you very well know, there's so much inefficiency and redundancy in clinical care. We see a lot of potential for AI to help us in all those areas in terms of streamlining these processes, improving communication. Could we perhaps have AI address some questions? Of course, you can look at those responses before they go back to the patient or before they go back to your colleagues, so there needs to be some supervision, of course. I can see a lot of potential for AI to help us communicate more, be more efficient in our processes, maybe document more. We all love seeing patients in clinic. I think the hurdle that we face is when you see so many patients, do you have enough time to complete notes and have everything that's pertinent in the notes? That's where I think AI can really help. If you want to go beyond improving efficiencies and workflows, I do think that there is great potential in AI to help us select patients for best therapies for them. How do you customize each therapy to fit each person's needs? We all talk about how do you deliver personalized medicine because it's not one-size-fits-all. We really need to delve more into models that so far, before AI, we haven't been able to get to that in terms of personalizing care. Now with a lot of data, applying AI appropriately and correctly to those data, I think we will be better positioned to identify best patients for therapies and then be able to predict actually how people are going to respond to therapies. That's really the biggest thing for me. As I shared with you earlier, I'm very passionate about sudden cardiac death prevention. As you know, we've been working on this for decades, trying to predict who is at an increased risk of sudden cardiac death. I see great potential in AI to really help us use all the data that we have now, from clinical data to demographics, to social determinants of health, to omics, be it genomics, proteomics, to imaging. You have so much, EKG, whatever, pulling all those data together and helping us really predict who is at risk for sudden cardiac death so we can really better tailor therapies to fit those patients' needs. Of course, there are many examples there. In terms of clinical research, of course, the biggest concerns in clinical research are number one, lack of recruitment, slow recruitment. I think AI can help us there in terms of applying it to EMR, to other databases, to really identify those patients easily and then be able to reach those patients and tell them about the research. Something we're going to talk about in the second paper is bias, like trying to extrapolate the results of studies that are developed in largely white patients, for example, to other patients, underrepresented minorities, women who are not well-represented in those clinical trials. AI can help us enhance representation of different groups who need to be well-represented and then conducting research more efficiently and reporting research and disseminating it more efficiently so you can appreciate there are so many applications there. Finally, with regard to education, we talk about how, yes, I think the system that we have now is reasonable. People learn, but can we improve? How can we get people more engaged? We talk about virtual reality. We talk about augmented reality, especially for a procedural specialty like EP. You can see the potential there. One thing we do highlight is certainly all the concerns around not implementing AI appropriately. We do talk about regulatory concerns about all those things and things that people have to keep in mind when you try to implement AI in clinical practice research and education. All of those are listed in the paper. I certainly encourage people to look at the paper if you're interested in how AI can be used in clinical care research or education. I think this is such a great start for people. What specific AI applications do you see as most immediately impactful in enhancing clinical workflows and patient care in EP? I think the easiest one, honestly, is streamlining workflows and creating efficiencies so that we're not redundant. As I said, trying to embed AI in EMR where you can just, at the push of a button, you can get the data that you want. You're not trying to sift through so many notes and things like that to try to find information about patients. As I said, AI in terms of responses to queries, to questions, things like that. Again, as I said, with supervision. I think that's where we're going to see the most progress, in my view, but I sincerely hope that we can do more with prediction and selecting patients for therapies, things like that. How can we ensure that AI models used in clinical care can be developed with diverse data set to avoid bias? That's an excellent question. As you very well know, and as I stated earlier, unfortunately, a lot of the data that we have come from clinical trials and registries that don't have enough. Women don't have enough underrepresented minorities, and it's not just about sex and race and ethnicity, but even people of different literacy levels, socioeconomic status, geographic locations. We really need to take all of these things into account when we build those registries and clinical trials to enhance recruitment of those patients in those data sources because at the end of the day, we really want to be able to generate data that represent all those populations. That's the biggest concern when you talk about avoiding bias in clinical care when you're using AI. When you use those models, the output very much depends on the input. If you're not inputting data that are valid in terms of the patient populations that you're looking at, you worry that the output may not apply to those people, may not generalize to them. That's what we really need to make sure is we need to have great representation, but also in the second paper that you alluded to in terms of how to avoid that AI bias in clinical care, we talk about labeling bias as well. There's the inherent bias, which is lack of underrepresented minorities, but also there's the labeling bias where you just infer something about certain groups of people based on incorrect data. For example, you may say, well, if you look at healthcare utilization, you look at cost, and you say, okay, those are low. This means that these people are not as sick. Well, that may not be true because if you think of these underrepresented minorities, the costs may be low, but it's because of they can't afford it. They don't have good access. You really need to avoid those things, and the paper talks about how we can really avoid and prevent algorithmic bias. I know you mentioned that you are passionate about sudden cardiac death. Can you conclude with a little more expansion on that? Yeah, of course. Biggest thing for me having been in this area for a couple of decades now, devoting most of my career to sudden cardiac death prevention, I couldn't be more disappointed with the lack of progress that we've made in terms of identifying patients who are at risk for sudden cardiac death. As many people know, if you look at the incidence of sudden cardiac death, it's highest among high-risk groups, so patients with heart failure, patients with a low EF, post-MI. As we all know, the absolute number of sudden cardiac death is highest in the general population, meaning among people who don't appear to be at an increased risk of sudden cardiac death. Where I see great potential in AI is to help us identify low-risk patients among these high-risk groups, the heart failure patients, post-MI, blah, blah, blah, as well as high-risk patients in these low-risk groups, the general population, maybe people with some comorbidities like diabetes, like other metabolic syndromes, things like that. That's where I see great potential, honestly, for AI. Of course, as we all know, our therapies such as the ICD are starting to implement AI in terms of distinguishing sinus tachycardia from SVT, from VT, ensuring that ATP is programmed only when it's effective and things like that. I truly expect that we're going to see more of that. Well, thank you, Dr. Al-Khatib, for joining us on this very special episode of The LEAD. Stay tuned for more episodes from the Heart Rhythm Society. Thank you very much. Thank you.

AI in cardiac electrophysiology

clinical care enhancement

diverse datasets

sudden cardiac death

algorithmic biases