Multi-feature Fusion of Deep Neural Network for Screening Atrial Fibrillation Using ECG Signals

Abstract

Atrial fibrillation (AF) is the most common cardiac arrhythmia, and it can cause a variety of cardiovascular diseases. This brings great hidden danger to people’s health and life safety all over the world. Electrocardiogram (ECG) is one of the most important noninvasive diagnostic tools for heart disease. Accurate interpretation of ECG is particularly important for the detection and treatment of AF. It is valuable to develop an efficient, accurate, and stable automatic AF detection algorithm in clinical settings. Therefore, this article proposes a novel integrated module, which combines densely connected convolutional network (DenseNet) module and bidirectional long short-term memory (BLSTM) module, based on the excellent ability of BLSTM on extracting the time series features, while DenseNet on capturing local features. Furthermore, we also propose a novel network architecture (MF-DenseNet–BLSTM) based on the integrated module mentioned above and multi-feature fusion for automatic AF detection using the ECG signals. The proposed model employs the architecture of dual-stream deep neural network to fusing multiple features. Specifically, the network of each stream structure consists of two parts with DenseNet module and BLSTM module. The data sets used to validate and test the proposed model are from the MIT-BIH Atrial Fibrillation Database. The experimental results show that the proposed model achieved 98.81% accuracy in training set, and achieved 98.04% accuracy in the testing set which is unseen data set. The proposed MF-DenseNet–BLSTM has shown excellent robustness and accuracy in automatic AF detection.