Bridging Artificial Intelligence Research and Real-World Clinical Practice for Atrial Fibrillation Management

Discussion With Bradley Knight, MD, and Tina Baykaner, MD

© 2025 HMP Global. All Rights Reserved.

Any views and opinions expressed are those of the author(s) and/or participants and do not necessarily reflect the views, policy, or position of EP Lab Digest or HMP Global, their employees, and affiliates. 

In this discussion at the 2025 Western Atrial Fibrillation Symposium, Bradley Knight, MD, talks with Tina Baykaner, MD, about the potential for artificial intelligence (AI) in EP, how AI can help in the management of patients with heart rhythm disorders in the near future, and take-home messages from the Western AFib roundtable on bridging AI research and real-world clinical practice for AFib management. 

Bradley Knight, MD: Hi, I am Dr Brad Knight, editor-in-chief of EP Lab Digest, published by HMP Global, which produces the Western AFib Symposium, where I am currently joined by my friend and colleague, Dr Tina Baykaner from Stanford. Tina, you were just part of a roundtable discussing artificial intelligence (AI) and bridging AI research with clinical needs and atrial fibrillation (AFib). Could you provide some highlights of that roundtable?

Tina Baykaner, MD: Yes, it was moderated by Nic Peters and Hamid Ghanbari. The highlight was there are a lot of AI models out there—which ones are produced because we can, and which ones are actually impacting our day-to-day clinical life as physicians or improving outcomes for patients? That was the highlight of the discussion.

Bradley Knight, MD: There has been a lot of hype and excitement about AI for probably 10 years, and I think that discussion brought up some issues that there is not that many applications that are being used by electrophysiologists. It has changed the world in many ways, but could you give us some examples of how it is being used in AFib management?

Tina Baykaner, MD: Yes, a few things we highlighted was it is used a lot in screening. Mayo Clinic has done some preliminary work on that and developed a model that is quite successful in screening the electrocardiograms (ECGs) of the general population of patients who are in sinus rhythm and who do not have a history of AFib and identifying those who will develop AFib in the near future. 

Bradley Knight, MD: Yes, all our ECGs now go through that screening process at Northwestern, I believe Sanjiv Shah put this into place, and we get a report on the likelihood of a patient developing AFib. Many of our patients already have AFib, so it does not really apply to those. Have these models been applied to patients who have had an ablation procedure? 

Tina Baykaner, MD: That is a great question. I was involved in developing one model specifically looking at that question. These are patients undergoing AF ablation by different providers with different modalities, and we tried to see if we could come up with an AI model that could predict which patient was more likely to have success so we would not have to worry about them, including which patients were likely to have recurrence so we could adjust both our arrhythmic management and monitoring frequency for those patients. There are existing clinical models that you might be aware of, such as the APPLE score and CHA₂DS₂-VASc score, to see if we can identify those patients. Their accuracy has been in the 0.6-0.7 range in terms of area under the ROC curve, so clinically they have not been that high. In my model, I took all the predictors in those existing models that we know somehow help and then added the preprocedural ECGs and intracardiac signals to see which ones had the best incremental role. Honestly, the preprocedural ECG had the most contribution to the overall value of the model. 

Bradley Knight, MD: One of the audience members brought up the point that just P wave duration itself can be effective. So, how much does AI add to the usual measures on an ECG? 

Tina Baykaner, MD: It adds a lot, because there has been pure ECG signal processing developed models in terms of these predictions, and they are about the same range as clinical factors, since there is only so much you can do. You can measure the traditional tools, like you can measure the P wave duration amplitude area under the P wave curve, PR-related metrics in terms of duration, but there is so much more in that signal that is hard to measure that the AI models can pick up and learn things from just the shape of the signals, which is hard to quantify. 

Bradley Knight, MD: Do these apply to patients who are in AFib preablation too? Can you analyze AFib ECGs or just sinus ECGs? 

Tina Baykaner, MD: As far as I know, there have not been any models to work on AFib signals. 

Bradley Knight, MD: My colleagues and I have been talking about other ways to use this concept, maybe just post cardioversion. Is there any data on an ECG post cardioversion in predicting recurrence just to give a patient prognostic information? 

Tina Baykaner, MD: There have not been any models on them. 

Bradley Knight, MD: There are other applications for AI that are being utilized in EP. Can you give us some other examples, such as from your device clinic? 

Tina Baykaner, MD: Exactly. I think we are already using it every day in our day-to-day life and it is already improving clinicians' quality of life, if I may say. For implantable loop recorders, a big problem has been false-positive alerts for pauses and false-positive alerts for AFib. A lot of them have been false because of frequent premature atrial contractions (PACs) or pauses have been from amplitude-related issues. But every company now has their AI algorithm to filter these alerts, which eliminates 80%-90% of these false alerts. So, we are now looking at much less events in our device clinic reports. 

Bradley Knight, MD: That is a good example of how AI has been incorporated into some of these things. We may not appreciate it. Maybe the industry brought that onto the table, but it has improved our practice. I guess other examples of applications of AI are ways to improve documentation in clinic, for example. Can you talk about that? 

Tina Baykaner, MD: Yes, we discussed it in the session too that a lot of different models are hearing you in the background and transcribing your clinic visit as if you are writing it yourself. You list present illness, the things discussed, and your assessment and plan. These are the pilot versions, and there is room for improvement. Some people shared their real-life frustrations with it costing a little bit more time, but I know that with time it will hopefully save us. 

Bradley Knight, MD: Yes, I think in the future it will make these things much better. I am aware of a company that records video and audio, and using that and AI, can generate an operative report. I favor just dictating, but if it helps you be more efficient, then I think it is good. 

The moderator of your panel, Nic Peters, can be quite provocative. He ended with an anecdote of going to the bank with a roll of coins and being told they no longer took coins and that he could now do everything on his phone. The irony was that he was on his way to clinic to see a patient in person, and that EP should be much further along as well. Any comments on how we as physicians can push this further? 

Tina Baykaner, MD: Yes, that was a great anecdote highlighting that we are a little bit behind everyone else in the world in embracing this. One point they highlighted was that we should not leave AI and model development to just the AI engineers—we should be involved as clinicians to highlight what is important for us and our patients, such that we can create more relevant models. 

Bradley Knight, MD: Well, that was a great session. Anything else you want to add? 

Tina Baykaner, MD: No, I think we covered everything! 

Bradley Knight, MD: Thank you very much. I appreciate your time, and great discussion!

The transcripts were edited for clarity and length.