Use of Artificial Intelligence for Prediction of Incident Atrial Fibrillation

Interview With Shaan Khurshid, MD

In this interview, Dr Khurshid discusses his recent talk at the Atrial Fibrillation (AF) Symposium, which took place February 2-4, 2023, in Boston, Massachusetts.

Can you briefly introduce yourself and the focus of your work?

My name is Dr Shaan Khurshid, and I am a senior cardiology fellow at Massachusetts General Hospital. I will be joining the faculty in electrophysiology starting this July. I have a clinical interest in the diagnosis and management of AF as well as other cardiac arrhythmias. I have a long-standing research interest in the use of statistical modeling to predict future disease risk, and more recently, incorporation of machine learning techniques for risk prediction, which has focused primarily on AF at this time. The talk I will be giving at the AF Symposium is entitled “Use of Artificial Intelligence (AI) for the Prediction of Incident AF.” I will discuss the background and rationale for AF risk prediction in general, the current landscape of AI-based methods for the prediction of future AF, and future directions about where to go next in terms of using AI.

What are the take-home messages you would like viewers to leave with?

The first message is that, despite intense interest in screening for undiagnosed AF, current mass screening approaches appear largely inefficient. It is difficult to find AF. The use of AI or other risk prediction methods has the potential to improve AF screening efficiency. The second take-home point is that AI can improve AF risk estimation above and beyond clinical factors, especially by the ability to incorporate raw data such as the 12-lead electrocardiogram (ECG), and these modalities are increasingly available at scale. The third message is that we still have a lot of work to do in terms of using AI to predict AF.

Some of the things I think we will want to focus on going forward are incorporating additional modalities into a composite model, so including other things in addition to the ECG, such as pulse wave photoplethysmography and cardiac imaging modalities. Another future direction will be validation of these models it more robust fashion. So other large and independent data sets of diverse patient populations to ensure that the models are consistently operating in the way we think they are. Finally, we can focus on integrating risk prediction with better clinical action. So, producing closed loops between risk stratification and some downstream intervention based on that risk stratification, whether that is AF screening, more intensive lifestyle modification, or other things.