.png){kind=icon}
.png){kind=icon}
Return
SVD-UNet: SVD-enhanced UNet for FECG extraction in the time–frequency domain
.png){kind=icon}
PRE
AI
W
C
R
DOI：10.1016/j.bspc.2026.111186.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Fetal electrocardiogram (FECG) is crucial for the early detection of fetal health issues. Nevertheless, obtaining clear FECG morphology remains challenging due to the presence of maternal ECG interference and other background noise in the abdominal recordings. To address this, a new architecture named SVD-UNet is introduced for the extraction of clean FECG morphological signals in the time–frequency domain. The proposed approach transforms the abdominal ECG from the 1D time domain into the 2D time–frequency domain via the Short Time Fourier Transform (STFT), obtaining a joint time–frequency representation that better reveals the non-stationary characteristics of both maternal and fetal components. Within this time–frequency framework, the SVD-UNet integrates Singular Value Decomposition (SVD) into a waveform-level fidelity UNet backbone. It leverages the inherent structural patterns of FECG waveforms through SVD to enhance signal separation performance while preserving the critical morphological details. Experimental results show that the proposed SVD-UNet preserves fetal ECG morphology faithfully on the synthetic FECGSYNDB database (where ground-truth waveforms exist), and achieves superior R-peak detection on the real-world PCDB and ADFECGDB databases. Specifically, the SE, PPV and F1 of the proposed method on PCDB are 95.31%, 95.82% and 95.56%, respectively. And the SE, PPV and F1 on ADFECGDB are 97.11%, 97.53% and 97.31%, respectively. Additionally, the RMSE and SNR on the FECGSYNDB database are 0.0504 and 1.5248 dB, respectively. This approach not only significantly improves fetal ECG extraction quality, but also offers a novel architectural perspective for signal separation tasks.
Journal
IF:
4.9
Papers:
9.7K
Citations:
2.4W
Organization
No organization information available
Cited Papers