2026-09-22 00:00 China Papers Foundations & Methods Source language EN

Northwestern Polytechnical University Diffusion Network Cleans EEG Artifacts in 17 ms

Summary EEG is easily contaminated by blinks, muscle activity and heartbeats, which can drag down how accurately a brain-computer interface reads intent. Researchers from the School of Automation and the School of Software at Northwestern Polytechnical University in Xi'an, northwestern China, built Diff-ADN, a single-channel denoiser that first reconstructs the signal according to contamination severity, then applies one diffusion-guided residual correction. It handles ocular, muscular, cardiac and mixed artifacts, reaching mean Pearson correlations with clean EEG of 0.839 to 0.936 at contamination levels from -7 to +2 dB. Across 4 public motor-imagery datasets, removing cardiac artifacts recovered 4.88 to 5.59 percentage points of decoding accuracy, up to 21.90 points in a single case, and a 2-second segment took 12.52 to 17.08 ms on the tested CPU. Better waveform reconstruction did not always translate into equal gains in decoding accuracy.

Why it matters Diffusion-based denoisers usually need many sampling steps, which is too slow for real-time use. Diff-ADN makes a single deterministic correction at inference, finishing a 2-second segment in about 17 ms. All results, however, come from artificially contaminated public datasets, and cardiac denoising restored only about 5 percentage points of decoding accuracy.

BCIwiki (bciwiki.com) — A diffusion-guided denoising network restored 4.88 to 5.59 percentage points of motor-imagery decoding accuracy after removing cardiac artifacts from EEG across 4 public datasets, researchers at Northwestern Polytechnical University reported in Journal of Neural Engineering on September 22, 2026 (PMID 42772331; DOI 10.1088/1741-2552/aeab36).

Physiological artifacts from eye movements, muscle activity and heartbeats degrade EEG recordings and can affect downstream brain-computer interface (BCI) analysis. The Diffusion-Guided Artifact Denoising Network (Diff-ADN) is a compact single-channel framework with 2 stages: severity-conditioned reconstruction, followed by diffusion-guided residual refinement. Forward noising and velocity prediction supervise the second-stage encoder during training, but inference needs only a single deterministic residual correction, with no iterative reverse-diffusion sampling.

Tested at contamination levels from -7 to +2 dB, Diff-ADN reached mean Pearson correlations of 0.936, 0.839, 0.886 and 0.844 for ocular (EOG), muscular (EMG), cardiac (ECG) and mixed EOG+EMG artifacts. With independently recorded artifact sources, performance varied by type: EOG and EMG showed larger reductions, while PTB and INCART ECG differed from the MIT-BIH reference by 0.026 and 0.014. The largest accuracy recovery, 21.90 percentage points, came with PTB ECG artifacts on the BCI IV-2a dataset at -6 dB.

The framework processes 2-second segments in 12.52 to 17.08 ms on the tested CPU. The authors said the results show that better waveform reconstruction does not necessarily translate into uniform recovery of BCI decoding accuracy.

Compiled by BCIwiki from public sources

Sources · 1

pubmed.ncbi.nlm.nih.gov 2026-09-22

Northwestern Polytechnical University Decoding Algorithms EEG Motor Decoding China

Northwestern Polytechnical University timeline

2026-09 SEDAT: Hybrid Tokenizer Lifts EEG Foundation Model Accuracy by up to 15.3% 2026-08 Review Maps Deep Learning for Motor Imagery BCIs 2026-07 Study Involving Hangzhou Dianzi University Achieves 1024-Channel Recording in Pig Cortex 2026-06 Northwestern Polytechnical University Registers BCI Stroke Rehabilitation Trial With 4 Arms All entries →

Read original ↗

Suggest a correction Revisions · none 中 / EN