Uncertainty-aware circular residual U-Net for electrocardiogram denoising under realistic nonstationary noise
Abstract
Electrocardiogram (ECG) signals in ambulatory monitoring are frequently degraded by nonstationary baseline wander (BW), muscle artifact (MA), and electrode-motion (EM) artifact. This study proposed an uncertainty-aware circular residual U-Net integrating circular convolution, residual-noise prediction, a morphology-aware boundary loss, and Monte Carlo (MC) dropout. Clean QT database segments were contaminated with Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) noise stress test database noise at input signal-to-noise ratios (SNR) from -6 to 24 dB. Evaluation on 2,560 test segments compared the proposed model with db4 wavelet denoising, convolutional neural network-denoising autoencoder (CNN-DAE), a zero-padded U-Net, and a circular U-Net. The proposed model achieved an output SNR of 11.6789 dB and waveform correlation of 0.9309, with the lowest root mean square error (RMSE), mean absolute error (MAE), percentage root-mean-square difference (PRD), derivative MAE, and boundary MAE of 0.1820, 0.1382, 32.5572, 0.0236, and 0.1678, respectively. Its peak-amplitude error of 0.1998 was comparable to the circular U-Net. Output SNR exceeded CNN-DAE and wavelet denoising by 3.1449 and 2.8925 dB. Predictive uncertainty strongly correlated with reconstruction error (r=0.8239, p<0.001), supporting its use as an indicator of reconstruction difficulty.
Keywords
Bayesian deep learning; Boundary artifact; Electrocardiogram denoising; Monte Carlo dropout; Morphology preservation; Predictive uncertainty
Full Text:
PDFDOI: https://doi.org/10.11591/eei.v15i6.13968
Refbacks
- There are currently no refbacks.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Bulletin of Electrical Engineering and Informatics (BEEI)
ISSN: 2089-3191
,
e-ISSN: 2302-9285
This journal is published by the
Institute of Advanced Engineering and Science (IAES)
.