Diff-ADN: A diffusion-guided artifact denoising network with deterministic residual refinement for EEG.

The method recovered cleaner EEG waveforms and some motor-imagery decoding accuracy in public datasets while running quickly on a CPU. This is offline engineering evidence, not proof of better clinical recordings or care.

Jamil M, Aziz MZ, Huang B, Yu X.
Journal of neural engineering · Publication date: 2026-09-22
Added to site: 2026-09-26
DOI: 10.1088/1741-2552/aeab36

Electronic publication date verified in Europe PMC; first indexed 2026-09-23. This source date is separate from the site's addition date.

Modalities

EEG, Brain-computer interface

Clinical focus

Artifact removal, Motor imagery, Signal quality

Population

Previously recorded EEG and artifact-source datasets used for offline signal-processing evaluation

Sample

Four public motor-imagery EEG datasets plus independently recorded ocular, muscular and cardiac artifact sources; participant and segment counts were not reported in the checked abstract

Intervention / method

A two-stage severity-conditioned reconstruction and diffusion-guided residual-refinement network removed ocular, muscular, cardiac and mixed artifacts from single-channel EEG

Comparator

Artifact-corrupted EEG across contamination levels and independently recorded artifact sources; the full set of competing algorithms was not available in the checked abstract

Clinician summary

Mean waveform correlations ranged from 0.839 to 0.936 by artifact type. Across four motor-imagery datasets, ECG denoising recovered 4.88 to 5.59 percentage points of decoding accuracy on average, while the authors explicitly note that waveform improvement did not produce uniform downstream gains.

How to interpret the findings

Evaluation relied on public datasets and controlled artifact contamination rather than prospective clinical acquisition. Generalization varied by artifact source, downstream decoding gains were not uniform, and the checked abstract did not provide participant denominators, all baseline comparisons, funding or conflicts.

Equipment and disclosures

Offline CPU inference was reported at 12.52 to 17.08 milliseconds for two-second segments; acquisition hardware varied across source datasets and was not detailed in the abstract.

Funding and competing-interest details were not available in the checked indexed record.

Related evidence

Compare findings from the same clinical area, including mixed results.

• Measuring motor intent for BCI control-A comparative analysis of signal quality of simultaneously recorded vECoG and scalp EEG. · Positive with limitations