BCIwiki (bciwiki.com) — A conditional adversarial denoiser called P2P Trans outperformed 5 convolutional encoder-decoder baselines at removing muscle artifacts from single-channel EEG, reaching the best mean correlation coefficient of 0.8645 and the lowest mean spectral relative root mean square error (RRMSE) of 0.3596, according to a preprint posted to arXiv on September 22, 2026. The work has not been peer reviewed. The authors are Haoyi Wang, Haowei Wang, Yihang Li, Wenjie Zhang and Yibo Wang.
Electromyographic activity can dominate scalp EEG across a broad frequency range and obscure neural structure relevant to analysis and brain-computer interfaces (BCIs), the paper notes. P2P Trans combines a U-shaped encoder-decoder, multi-head self-attention at the bottleneck, cross-attention guided skip fusion and a PatchGAN discriminator.
The experiment used EEGdenoiseNet segments represented as 1024-sample sequences, synthesized paired mixtures at 10 signal-to-noise ratios from -7 to 2 dB, and evaluated 50000 training examples and 5980 test examples.
The model's mean temporal RRMSE of 0.5797 did not lead the comparison, which the authors say indicates a tradeoff between spectral suppression and pointwise waveform fidelity. Its advantage was strongest under severe contamination. The authors identify subject-independent evaluation, direct one-dimensional modeling and downstream task validation as next steps.