False Alarm Reduction in Atrial Fibrillation Screening

Abstract

Early detection of AF is essential and emphasizes the significance of AF screening. However, AF detection in screening ECGs, usually recorded by handheld and portable devices, is limited because of their high susceptibility to noise. In this study, the feasibility of applying a machine learning-based quality control stage, inserted between the QRS detector and AF detector blocks, is investigated with the aim to improve AF detection. A convolutional neural network was trained to classify the detections into either true or false. False detections were excluded and an updated series of QRS complexes was fed to the AF detector. The results show that the convolutional neural network-based quality control reduces the number of false alarms by 24.8% at the cost of 1.9% decrease in sensitivity compared to AF detection without any quality control.