Atrial fibrillation screening through combined timing features of short single-lead electrocardiograms

Abstract

Atrial fibrillation (AF) is the most common cardiac arrhythmia, as well as a growing healthcare burden worldwide. It is often asymptomatic and usually starts with very brief episodes, thus making its early detection an interesting challenge. For that purpose, the present work introduces a novel method exploiting the variability presented both by ventricular and atrial activities reflected on the surface electrocardiogram (ECG). Thus, time series from the RR intervals and the fibrillatory waves morphology contained by the TQ intervals are first generated and, then, their regularity is estimated making use of the Coefficient of Sample Entropy (COSEn). The collected information is finally combined through a multi-class support vector machine (SVM) approach to discern among short episodes of AF, normal sinus rhythm (NSR) and other rhythms (OR). The algorithm has been validated in the context of the PhysioNet Computing in Cardiology Challenge 2017, thus reporting a global F1 measure of 0.73 for the training set and 0.71 for the testing group. Nonetheless, to evaluate the method in a common scenario for previous works, the widely used MIT-BIH AF database has also been considered. A notably higher F1 score of 0.87 has been provided in this case. The significantly different balance between the number of AF, NSR and OR recordings in both databases could justify the obtained outcomes.