Coherence Spectrum Learning for Fetal Electrocardiogram Extraction

The challenge of acquiring a reliable non-invasive fetal electrocardiogram is a complex endeavor, given the various intricate components within the data collection pipeline. These components include maternal physical parameters, hardware sensitivity, and the efficacy of signal processing algorithms in distinguishing maternal and fetal electrocardiogram sources. Our emphasis in the scope of this paper is directed towards the development of novel signal processing algorithms for fetal electrocardiogram extraction and delves deeper into understanding the effectiveness in accurately delineating fetal electrocardiogram signals. Here, we propose a data-driven method based on estimating the coherence spectrum within the frequency domain through the learning processes. Specifically, the proposed model undergoes the gradient updates in the frequency domain, differing from the model trained in a real number space where the input to the model is commonly either the amplitude derived from the power spectrum or the temporal representation of the signal. Our results demonstrate that the proposed learning strategy has superior fetal electrocardiogram extraction compared to the model solely learned for the extraction in the temporal domain. Highlighting the value of simultaneously learning the real and imaginary components, an approach preserves both amplitude and phase information throughout the learning for extraction process.