Atrial fibrillation detection using stationary wavelet transform and deep learning

Abstract

Deep learning has achieved a great success in the fields of image and audio recognition because of avoiding explicit feature extraction and attaining high classification accuracy. In this paper, we explore the application of deep convolutional neural networks (DCNNs) for automatic detection of atrial fibrillation (AF). The 2-dimension parameter input structure is essential for DCNNs and tens of thousands of samples are also needed for the proper operation. As we know, ECG is one-dimension time-varying signal, which doesn't match the requirement for the input structure of DCNNs. Furthermore the number of the marked AF samples is also limited. To address these problems, we adopt the stationary wavelet transform (SWT) for ECG preprocessing and then the processed signal is reorganized into two-dimensional parameter structure to meet the requirement of input structure of DCNNs. Besides, the original ECG signals are divided into very short data segments (namely 5-second segments) for the following reasons. On the one hand, short ECG segment is helpful for the algorithm assessment of short AF episode detection. On the other hand, it can also increase the number of AF sample for machine learning and experiment evaluation. As for DCNNs, multiple convolutional layers and fully connected layers are used for deep learning. On the MIT-BIH Atrial fibrillation data set, the proposed method can achieve sensitivity of 98.79%, specificity of 97.87% and accuracy of 98.63%, which outperforms most of other algorithms.