CSGSA-Net: Canonical-structured graph sparse attention network for fetal ECG estimation

Abstract

Noninvasive fetal electrocardiogram (NI-FECG) plays a significant role in fetal diagnosis. However, it is challenging to estimate FECG signals from the abdominal ECG due to the following issues: (1) The FECG signals are always masked by the maternal ECG (MECG) signals; (2) The FECG waveform is often corrupted by the noise. To solve such problems, a canonical-structured graph sparse attention network is proposed for fetal ECG estimation, where the canonical spatial graph sparse attention module is designed to estimate FECG signals masked by the MECG signals by learning its waveform features, and the canonical channel graph sparse attention is devised to discriminate the characteristic waveforms from noise by capturing the FECG signal details. Experiments conducted on the two databases demonstrate the proposed CSGAS-Net outperforms the state-of-the-art deep learning methods. The project is available at https://github.com/langdecc511/CSGSA-Net.