deep learning-based signal source separation for enhanced fetal cardiac assessment: a non-invasive approach

1. introduction accurately extracting fetal electrocardiogram (ecg) signals from noninvasive abdominal recordings remains a significant clinical challenge in prenatal care [1]. the primary difficulties arise from overlapping maternal ecg components and various noise sources that contaminate the desired fetal signals [4]. traditional filtering methods have shown limited success, often degrading essential waveform details and compromising their diagnostic value [5]. this study introduces a novel w-net architecture, comprising dual interconnected u-nets, specifically designed to address the critical need for precise fetal ecg extraction in non-invasive monitoring scenarios [1]. 2. methods and materials our methodology centers on a w-net architecture, where two u-nets are cascaded to form a w-shaped network structure [1]. the first u-net focuses on initial signal decomposition and maternal ecg extraction, while the figure 1: maternal-fetal ecg signal separation using the proposed architecture. top: combined maternal-fetal ecg recording at 12db snr. bottom row (left to right): target fetal ecg signal compared with the model predicted fetal ecg output; target maternal ecg signal compared with the model predicted maternal ecg output. second u-net refines the fetal ecg component through specialized feature maps. we developed a custom loss function that combines scale-invariant signal-to-distortion ratio (si-sdr) with a novel morphological preservation term [2], ensuring both signal separation quality and physiological waveform integrity. the model was evaluated using the fecgsyn dataset [3], providing comprehensive simulated fetal and maternal ecg data across multiple signal-to-noise ratios (snr) and measurement scenarios. 3. results the w-net architecture, coupled with our custom loss function, demonstrated exceptional performance in separating fetal ecg signals. as illustrated in figure 1, our method successfully separates the combined maternalfetal ecg recording (snr = 12db) into its constituent components, achieving high fidelity in both maternal and fetal signal extraction. key achievements include: • superior signal separation quality with an average improvement of 2.3% compared to baseline methods [5] • enhanced qrs complex detection accuracy by1.8%, directly attributable to the dual u-net structure [1] • morphological preservation score of 93.1%, validated through evaluation assessment. • consistent performance across various snr levels (5db to 20db range) 4. conclusion and discussion our w-net based approach with custom loss function establishes a new benchmark in fetal ecg extraction [1, 5]. the dual u-net architecture effectively leverages hierarchical feature learning, while the custom loss function ensures clinical fidelity of the extracted signals. these advancements demonstrate significant potential for integration into practical clinical applications, particularly in non-invasive fetal monitoring and early detection of cardiac abnormalities.