Generalizable Beat-by-Beat Arrhythmia Detection by Using Weakly Supervised Deep Learning

• DOI: 10.3389/fphys.2022.850951
• Publication Date: 2022-03-27T23:00:02Z
• Authors: Yang Liu; Qince Li; Runnan He; Kuanquan Wang; Jun Liu; Yongfeng Yuan; Yong Xia; Henggui Zhang
• Keywords: Beat (acoustics)

Beat-by-beat arrhythmia detection in ambulatory electrocardiogram (ECG) monitoring is critical for the evaluation and prognosis of cardiac arrhythmias, however, it a highly professional demanding time-consuming task. Current methods automatic beat-by-beat suffer from poor generalization ability due to lack large-sample finely-annotated (labels are given each beat) ECG data model training. In this work, we propose weakly supervised deep learning framework (WSDL-AD), which permits training fine-grained (beat-by-beat) detector with use large amounts coarsely annotated recording) improve ability. framework, heartbeat classification recording integrated into neural network end-to-end only labels. Several techniques, including knowledge-based features, masked aggregation, pre-training, proposed accuracy stability under weak supervision. The developed WSDL-AD trained ventricular ectopic beats (VEB) supraventricular (SVEB) on five coarsely-annotated datasets performance evaluated three independent benchmarks according recommendations Association Advancement Medical Instrumentation (AAMI). experimental results show that our method improves F 1 score by 8%–290% F1 4%–11% compared state-of-the-art learning. It demonstrates can leverage abundant coarsely-labeled achieve better than previous while retaining fine granularity. Therefore, has great potential be used clinical telehealth applications. source code available at https://github.com/sdnjly/WSDL-AD .