Adaptive atrial fibrillation detection focused on atrial activity analysis

Noninvasive detection of atrial fibrillation (AF) based on electrocardiogram (ECG) is beneficial for early diagnosis, which is clinically significant in the disease management and comorbidity prevention. In this work, we present an adaptive system for rate-independent AF detection. A light-weighted convolutional neural network is first constructed to process the extracted atrial activity (AA) from single-lead ECG, and the beat-wise identification of AF and Non-AF is obtained. To enhance the performance, a set of training techniques including regularization and data augmentation are further applied in the model training. With regard to different datasets, we deploy transfer learning to deal with the heterogeneous distributions. Tested on the MIT-BIH AF database (AFDB), the designed model can directly achieve the accuracy of 86.61% in distinguishing AF beats and non-AF beats. After model adaptation, a significant boost in the performance is observed even when a small size of beats is used. Taking ventricular activity into consideration, the post-processing further improved the detection accuracy to 99.61% on AFDB. Given the short detection delay and high accuracy, the presented method could help to promote the AA-based research of AF.