W-Net Feature-based Attention Diffused Convolutional Network Architecture for the Detection of Fetal Arrhythmia

This paper proposes a Fetal Arrhythmia detection architecture that also separates the fetal and maternal electrocardiography (ECG) components present in the mother's abdominal ECG. The ECG that was collected from the abdomen contains the ECG of the mother and fetus, along with other artifacts. Detection of Fetal Arrhythmia is a challenging task, since the strength of the fetal ECG (FECG) is much less than the maternal ECG signal. To address this challenge, we propose a Fetal Arrhythmia detection architecture with the use of W-Net and a convolutional neural network (CNN) architecture termed W-Net attention diffused CNN (W-AD-CNN). The W-Net model is initially trained to separate the FECG and the mother's ECG (MECG). The CNN structure was integrated with the FECG branch U-Net using an attention diffusion mechanism, which enhances the complex Fetal Arrhythmia patterns while eliminating the mother's ECG components. The evaluation of the proposed Fetal Arrhythmia detection approach was conducted using the measures, namely accuracy, Specificity, precision, F1-score, and recall, with the use of the NIFEADB dataset. The proposed approach results in an F1-score, accuracy, recall, and Specificity of 97.93%, 98.24%, 98.01%, and 98.13%, respectively. From the extracted FECG and MECG signals, the fetal and mother's heart rates can be further computed with the use of the Pan-Tompkins algorithm. The proposed Fetal Arrhythmia detection approach improves maternal and fetal healthcare through early diagnosis using non-invasive ECG analysis.