Dual-Attention-Based Pipelined Encoder–Decoder Network for Fetal Electrocardiogram Extraction

Monitoring fetal cardiac health during gestation is indispensable in detecting congenital heart defects. Fetal electrocardiography is considered a more reliable source for determining fetal heart rate (FHR) than many traditional approaches based on the Doppler effect. However, the isolation of fetal electrocardiogram (fECG) from the abdominal electrocardiography signal is still a challenging mix source separation (MSS) problem. The reason for this is that the fECG signal has a significantly lower amplitude in comparison to maternal electrocardiography and is also tampered with by noise from diverse sources. Therefore, we have proposed a dual-attention-based dual encoder-decoder pipelined architecture, namely, dual-attention W-Net (DAW-Net), to address these issues. The network extracts the feature maps that are related to both maternal and fetal QRS and enables the isolation of maternal electrocardiogram (mECG) and fECG signals. The masking of the relevant area of interest on the attention weights produced by the correlation attention module enhances the fetal QRS information. The self-attention module extracts the contextual features of the maternal QRS and cancels it from the abdominal electrocardiogram (aECG) features. The skip connections are used in the architecture to amplify the QRS information of the ECG signal. The efficacy of the proposed architecture was evaluated using two real databases, acquiring a mean$F1$-score of 98.13%. The proposed architecture exhibits a superior performance when compared to other cutting-edge methods. This study represents a crucial advancement in the long-duration monitoring of noninvasive fECGs.