Neural Noise

HD-sEMG channel reconstruction with denoising diffusion for myoelectric control

3 min read

  • High-density surface electromyography (HD-sEMG) is used for myoelectric control and activation pattern analysis; signal loss from poor electrode contact and channel corruption limits reliability and practical use.1
  • Conventional interpolation fails to reconstruct corrupted HD-sEMG effectively, especially when multiple adjacent channels are affected.1
  • A denoising diffusion probabilistic model (DDPM) is proposed to repaint or reconstruct corrupted HD-sEMG signals, improving usability for physiological time-series and neuroprosthetic control.1 1

Gardner updates

  • HD-sEMG is used for myoelectric control and activation pattern analysis; signal loss from poor electrode contact and channel corruption limits reliability. 1
  • Denoising diffusion probabilistic models can reconstruct corrupted HD-sEMG channels where conventional interpolation fails, especially when multiple adjacent channels are affected. 1

Weekly enrichment (2026-07-20)

  • The source is Zhao et al., “Repaint High-Density Surface Electromyography Signal Using Denoising Diffusion Probabilistic Model,” published in IEEE Transactions on Biomedical Engineering (2025; DOI 10.1109/tbme.2025.3604527; PMID 40892666) by a Sichuan University group including Yihui Zhao, Jiawei Liao, Xia Fang, Hai Wang, Ning Jiang, and Jiayuan He.2 3
  • The method casts corrupted-channel recovery as an inpainting (“repaint”) task using a denoising diffusion probabilistic model (DDPM), so it reconstructs signals without requiring prior knowledge of the corruption pattern.2 3
  • The network is a U-Net augmented with spatiotemporal embedding modules that jointly learn the spatial (electrode-grid) and temporal characteristics of HD-sEMG signals.2 3
  • Evaluation spanned 6 distinct corruption patterns at corruption ratios from 12.5% up to 50%, tested on self-collected datasets (including data from an amputee subject) plus a public benchmark dataset.2 3
  • On normalized RMSE the DDPM achieved the lowest error (0.027 ± 0.027, averaged across corruption ratios), significantly beating linear interpolation (0.038 ± 0.033), cubic interpolation (0.038 ± 0.032), a GAN (0.049 ± 0.041), and a VAE (0.068 ± 0.046), all at p < 0.001.2 3
  • On peak signal-to-noise ratio the method reached the highest mean (35.81 ± 17.95 dB) versus linear (33.89 dB), cubic (33.88 dB), GAN (31.08 ± 19.14 dB), and VAE (26.98 ± 18.94 dB), again p < 0.001.2 3
  • Downstream gesture-classification accuracy was preserved: reconstructed signals gave performance statistically equivalent to ground-truth signals at the lower corruption ratio, supporting practical myoelectric/neuroprosthetic control.2 3
  • Context: this fits a broader trend of conditional diffusion models for sEMG restoration—e.g., “DiffSR,” a conditional diffusion-based technique for recovering distorted sEMG signals (DOI 10.1016/j.bspc.2025.108276)—that increasingly outperform interpolation, GAN, and autoencoder baselines.4
  • Caveat/fallback: the IEEE Xplore source URL returned a JavaScript/anti-bot verification page rather than content, so these figures were drawn from the indexed IEEE TBME abstract and the matching PubMed record.2 3

Footnotes

  1. http://ieeexplore.ieee.org/document/11145804 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  2. https://doi.org/10.1109/tbme.2025.3604527 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  3. https://pubmed.ncbi.nlm.nih.gov/40892666/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  4. https://doi.org/10.1016/j.bspc.2025.108276 ↩

Graph View