Accurate detection of atrial fibrillation from 12-lead ECG using deep neural network. (January 2020)

• Record Type: Journal Article
• Authors: Cai, Wenjuan; Chen, Yundai; Guo, Jun; Han, Baoshi; Shi, Yajun; Ji, Lei; Wang, Jinliang; Zhang, Guanglei; Luo, Jianwen
• Is Part Of: Computers in biology and medicine. Volume 116(2020)
• Publication Date: 2020-01
• Subjects: 12-Lead ECG -- Atrial fibrillation detection -- Deep learning -- Densely connected neural network; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2019.103378
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Atrial fibrillation (AF) is the most common heart arrhythmia, and 12-lead electrocardiogram (ECG) is regarded as the gold standard for AF diagnosis. Highly accurate diagnosis of AF based on 12-lead ECG is valuable and remains challenging. In this paper, we proposed a novel method with high accuracy for AF detection based on deep learning. The proposed method constructed a novel one-dimensional deep densely connected neural network (DDNN) to detect AF in ECG waveforms with a length of 10s. A large set of 16, 557 12-lead ECG recordings collected from multiple hospitals and wearable ECG devices were used to evaluate the performance of the DDNN. In the test dataset (3312 12-lead ECG recordings), the DDNN obtained high performance with an accuracy of 99.35 ± 0.26%, a sensitivity of 99.19 ± 0.31%, and a specificity of 99.44 ± 0.17%. Its high performance and automatic nature both demonstrate that the proposed network has a great potential to be applied to clinical computer-aided diagnosis of AF or future screening of AF in wearable devices. Highlights: Two binary classifications and a three-class classification of ECG signals. Designed and trained a novel deep densely connected neural network. Accuracy of 99.35%, sensitivity of 99.19%, and specificity of 99.44%. Network can learn to use F waves-related features to make classification.