A U-Net-Based Method With Dedicated Preprocessing for Fetal ECG Extraction

The acquisition of fetal electrocardiographic (fECG) signals from electrodes placed on the maternal abdomen provides a noninvasive approach for monitoring fetal heart activity. Reliable detection of fetal QRS complexes, however, remains challenging due to the presence of dominant maternal ECG components and various disturbances in abdominal recordings. This study presents a convolutional neural network–based method for noninvasive fetal ECG extraction using a one-dimensional U-Net architecture adapted to four-channel abdominal signals. The approach combines signal preprocessing with selected architectural and training-related modifications of the baseline U-Net, including time-shifting data augmentation, squeeze-and-excitation blocks, and dropout regularization. All preprocessing and network-related parameters were determined using a cross-validation–based selection procedure. The proposed method reconstructs the fetal ECG signal with enhanced fetal QRS complexes, after which fetal QRS detection is performed using an independent detection algorithm for objective performance evaluation. The proposed method was evaluated on the publicly available ADFECGDB database, consisting of recordings from five subjects. Detection performance was assessed using Sensitivity, Positive Predictivity, and F1-score. The final configuration achieved an F1-score of 99.73% for fetal QRS detection, which places the method among the highest-performing noninvasive approaches reported for this dataset. The results confirm that combining appropriately designed preprocessing with a modified U-Net architecture can lead to highly accurate fetal QRS detection in abdominal ECG signals.