In this paper, we present a groundbreaking method for extracting fetal electrocardiogram (fECG) from single-channel recordings using dry textile electrodes using AI techniques. We generate a novel dataset simulating abdominal recordings including real-world noise, and propose an innovative pipeline based on Complex UNet, a complex-valued noise removal network. Unlike existing methods, we process both the real and imaginary parts of the spectrogram in addition to the signal magnitude, thereby taking into account phase information and preventing inaccurate predictions. We demonstrate state-of-the-art performance in fECG extraction and R-peak detection by comparing our method with existing methods on simulated and real data. This is the first method to effectively extract fECG signals from single-channel recordings using dry textile electrodes, and represents a significant step forward toward a fully noninvasive and self-managed fECG extraction solution.
