A deep learning refinement strategy based on efficient channel attention for atrial fibrillation and atrial flutter signals identification

Abstract

Atrial fibrillation (AF) and atrial flutter (AFL) are the most frequent arrhythmias recently. However, given both similar physiological features, visually evaluating electrocardiogram as the most traditional diagnosis scheme is taxing and error-prone. In this work, we specifically design two network modules based on bidirectional long short term memory (BiLSTM) and gate recurrent unit (BiGRU) for automatic AF and AFL detection. Motivated from Efficient Channel Attention network, we aim to reformulate BiLSTM and BiGRU with a feature recalibration approach that enables the model to adaptively focus more on the relevant feature representations and suppress irrelevant parts while appropriately capturing cross-channel interaction for learning effective channel attention. The results lead to consistent performance gains than several published researches with an accuracy of 99.2% and 99.3% across the two publicly available data sets while demonstrating the effectiveness of both modules. In particular, various derivative gradient values of sample ECG segments are visualized to improve interpretability. To our knowledge, this work offers the first empirical investigation of existing BiLSTM and BiGRU refinements for a better performance gain, showing great potential for many computer vision tasks.