Early Warning of Atrial Fibrillation

Abstract

ABSTRACTAtrial Fibrillation (AF) is the most common cardiac rhythm disorder. It is associated with increased hospitalization, risk of heart failure, and risk of stroke. Once on AF, it can sometimes be difficult to revert to sinus rhythm (SR), potentially requiring pharmacological or electrical cardioversion. Earlier warning of an imminent switch from SR to AF, even if by only a few minutes, could prompt patients to take actions (e.g., taking oral antiarrhythmic drugs) to avoid AF and its associated complications, thereby easing the workload of healthcare professionals and reducing costs to the health system. The question is whether there is information, even if subtle, in the minutes prior to AF to indicate an imminent switch from SR. This paper shows that, for the vast majority of patients, the answer is affirmative. On test data, our algorithm can predict the onset of AF on average 31 minutes before it appears, with an accuracy of 83% and an F1-score of 85%. Moreover, this performance was obtained from R R interval (RRI) signals, which can be obtained from common wearable devices such as smartwatches and smart bands. The predictions were performed using a deep convolutional neural network, trained and cross-validated on 24-hour RRI signals obtained from Holter electrocardiogram recordings of 280 patients, with an additional 70 patients used as test data. We further tested the model with data from two other external centers with 73 patients. Overall, the proposed method has low computational time and could be embedded in common wearable devices that capture RRI for continuous heart monitoring and early warning of AF onset.