Quality Control in ECG-based Atrial Fibrillation Screening

This thesis comprises an introductory chapter and four papers related to quality control in ECG-based atrial fibrillation (AF) screening. Atrial fibrillation is a cardiac arrhythmia characterized by an irregular rhythm and constitutes a major risk factor for stroke. Anticoagulation therapy significantly reduces this risk, and therefore, AF screening is motivated. Atrial fibrillation screening is often done using ECGs recorded outside the clinical environment. However, the higher susceptibility of such ECGs to noise and artifacts makes the identification of patients with AF challenging. The present thesis addresses these challenges at different levels in the data analysis chain. Paper I presents a convolutional neural network (CNN)-based approach to identify transient noise and artifacts in the detected beat sequence before AF detection. The results show that by inserting a CNN, prior to the AF detector, the number of false AF detections is reduced by 22.5% without any loss in the sensitivity, suggesting that the number of recordings requiring expert review can be significantly reduced. Paper II investigates the signal quality of a novel wet electrode technology, and how the improved signal quality translates to improved beat detection and AF detection performance. The novel electrode technology is designed for reduction of motion artifacts typically present in Holter ECG recordings. The novel electrode technology shows a better signal quality and detection performance when compared to a commercially available counterpart, especially when the subject becomes more active. Thus, it has the potential to reduce the review burden and costs associated with ambulatory monitoring.Paper III introduces a detector for short-episode supraventricular tachycardia (sSVT) in AF screening recordings, which has been shown to be associated with an increased risk for future AF. Therefore, the identification of subjects with suchepisodes may increase the usefulness of AF screening. The proposed detector is based on the assumption that the beats in an sSVT episode display similar morphology, and that episodes including detections of deviating morphology should be excluded. The results show that the number of false sSVT detections can be significantly reduced (by a factor of 6) using the proposed detector.Paper IV introduces a novel ECG simulation tool, which is capable of producing ECGs with various arrhythmia patterns and with several different types of noise and artifacts. Specifically, the ECG simulator includes models to generate noise observed in ambulatory recordings, and when recording using handheld recording devices. The usefulness of the simulator is illustrated in terms of AF detection performance when the CNN training in Paper I is performed using simulated data. The results show a very similar performance when training with simulated data compared to when training with real data. Thus, the proposed simulator is a valuable tool in the development and training of automated ECG processing algorithms. Together, the four parts, in different ways, contribute to improved algorithmic efficiency in AF screening