AI gaining ground in electrophysiology, but clinicians still run the show

Where AI is not so good is when a diagnosis depends on information outside the image or test. Patient age, medical history, timing of symptoms and other clinical circumstances can fundamentally change interpretation, Narayan said, making human judgment still quite important.

In EP, he sees significant potential for AI in remote monitoring and automated clinical workflows. Electronic health records could identify patients who are due for AFib screening or other preventive services, reducing the chance that busy clinicians overlook guideline-based recommendations.

AI also could assist physicians by documenting patient encounters. Systems that listen to conversations, summarize visits and identify medications or follow-up issues could reduce the administrative burden associated with electronic medical records and allow clinicians to spend more time interacting with patients. Narayan emphasized, however, that these summaries and recommendations still require appropriate oversight.

The limitations become particularly apparent when AI is expected to replace the clinician-patient relationship. Narayan cited research involving healthcare chatbots in areas such as stroke rehabilitation and early stroke detection, where the systems were not sufficiently accurate and patients reported concerns about a lack of empathy and standardized responses.

"I think that there's a way that we interact as humans that you can't write as an equation. And because of that, it's hard to train an AI to do it. If you were to just take all the texts on the internet, it would be very hard to thread those needles in a way that we would do in a natural human conversation. So I think there's always a big chance of missing something if we don't have the human in the loop," he explained.

EP data a perfect fit for AI applications

Narayan expects AI adoption in EP to continue growing for applications such as measuring AFib burden, identifying patients who may warrant consideration of anticoagulation and flagging individuals who could be at elevated risk of sudden cardiac death. AI is also beginning to assist with interpretation of complex electrical signals during AFib and ventricular tachycardia ablation procedures, although its current clinical adoption remains modest.

However, he said AI implementation for EP is still young, but the nature of EP clinical data makes it ripe for AI applications.

"If we continue to develop those models, I am incredibly excited about that potential," Narayan said. "It's a great time to be in computational work in cardiology, and what I love about this is so much of what we do was built for computational modeling. It was built for AI. It was built to be turned into physics."

A good example of this is the development of digital twin technology, where AI can generate a model of a patient's heart based on their testing and imaging data. This could provide another major avenue for AI to expand in cardiology. These models could eventually help clinicians understand individual patients in greater detail and potentially guide procedures and therapies. Narayan said the technology remains under development and will need to advance alongside traditional research and clinical trials.

For now, Narayan said the most appropriate path is to use AI where it can reliably automate data-driven tasks while retaining clinicians for decisions requiring context, judgment and communication. He said humans still need to be kept in the loop and make the final; decisions on care, and that is not something he said that will change anytime soon.