RePaint High-Density Surface Electromyography Signal Using Denoising Diffusion Probabilistic Model

• Published in: IEEE transactions on biomedical engineering Vol. PP; pp. 1 - 12
• Main Authors: Zhao, Yihui, Liao, Jiawei, Fang, Xia, Wang, Hai, Jiang, Ning, He, Jiayuan
• Format: Journal Article
• Language: English
• Published: United States IEEE 02.09.2025
• Subjects: Benchmark testing; Diffusion model; Electrodes; Electromyography; Hands; High-density Electromyography; Interpolation; Muscles; Myoelectric control; Noise reduction; Recording; Signal reconstruction; Spatiotemporal phenomena
• ISSN: 0018-9294; 1558-2531; 1558-2531
• DOI: 10.1109/TBME.2025.3604527

Objective: High-density surface electromyography (HD-sEMG) has emerged as a powerful tool for myoelectric control and activation pattern analysis. However, signal loss due to poor electrode contact and channel corruption remains a significant challenge, limiting the reliability and practical applications of HD-sEMG signals. Conventional interpolation methods fail to effectively reconstruct corrupted signals, especially when multiple adjacent channels are affected. Methods: This paper proposes a novel HD-sEMG signal reconstruction approach based on the denoising diffusion probabilistic model (DDPM) with a repaint strategy. By leveraging a U-Net structure with spatiotemporal embedding modules that effectively learn the spatial and temporal characteristics of HD-sEMG signals, the proposed method achieves high-fidelity signal reconstruction without requiring prior knowledge of corruption patterns. Results: Experimental evaluations are conducted on 6 corruption patterns with varying ratios (from 12.5% to 50%) using self-collected datasets (including an amputated subject) and a benchmark dataset. Results demonstrate that the proposed approach consistently outperforms interpolation methods (linear: 0.038<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula>0.033, cubic: 0.038<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula>0.032), generative adversarial net (GAN) (0.049<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula>0.041), and variational autoencoder (VAE) (0.068<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula>0.046) in terms of <inline-formula><tex-math notation="LaTeX">nRMSE</tex-math></inline-formula> (<inline-formula><tex-math notation="LaTeX">p < 0.001</tex-math></inline-formula>), achieving the lowest error of 0.027<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula>0.027 averaged across all corruption ratios. For <inline-formula><tex-math notation="LaTeX">PSNR</tex-math></inline-formula>, the proposed approach achieves the highest mean value (35.81<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula> 17.95dB) compared to interpolation methods (linear: 33.89<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula>26.85, cubic: 33.88<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula> 26.88dB), GAN (31.08<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula> 19.14dB), and VAE (26.98<inline-formula><tex-math notation="LaTeX">\pm</tex-math></inline-formula> 18.94dB) (<inline-formula><tex-math notation="LaTeX">p < 0.001</tex-math></inline-formula>). Furthermore, the proposed method maintained robust classification accuracy, achieving statistically equivalent performance to ground truth at the lower corruption ratio. Significance: The proposed HD-sEMG signal reconstruction approach offers a new solution for enhancing the fidelity and reliability of HD-sEMG signal acquisition.