Dual temporal convolutional network for single-lead fibrillation waveform extraction

The f-wave extraction (FE) is essential for analysis of atrial fibrillations. However, the state-of-the-art FE methods are model-based, and they cannot well adapt to the QRST complexes with high morphological variabilities which often appear in clinical electrocardiogram (ECG). Recently, the encoder-decoder based deep learning networks have been successfully applied to separate variable speech waveforms. However, how these networks are exploited to extract f-waves from ECG recordings is still unclear. Moreover, these networks require different sources to share the common encoder and decoder, which restricts the effectiveness of source representations. To address these issues, a dual temporal convolutional network called DT-FENet is proposed for single lead FE, which integrates the source-specific encoder-decoder mappings and the information fused attention mechanism to respectively learn the latent masks of f-waves and QRST complexes. The proposed DT-FENet can be considered as a dual-stream extension of the famous Conv-TasNet. Compared to Conv-TasNet, the source-specific encoder-decoder mappings of the DT-FENet can obtain more representative bases and sparser activations in latent feature spaces to facilitate the FE task. To the best of our knowledge, this is the first deep learning method for single-lead FE. Extensive experiments were conducted on the clinical ECG records, and the experimental results show that the proposed DT-FENet performs significantly better than the state-of-the-art FE methods with less than one tenth of unnormalized ventricular residuals and about twice spectral concentrations. The proposed DT-FENet can provide accurate f-wave information for atrial fibrillations detection using single-lead ECGs.