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

Physiological artifacts degrade electroencephalographic (EEG) recordings and can affect downstream brain-computer interface (BCI) analysis. This study develops a compact single-channel framework for ocular, muscular, cardiac, and mixed-artifact removal. Approach: The Diffusion-Guided Artifact Denoising Network (Diff-ADN) employs a two-stage framework comprising severity-conditioned Stage-I reconstruction followed by diffusion-guided Stage-II residual refinement. During training, forward noising and velocity prediction supervise the Stage-II encoder, whereas inference requires only a single deterministic residual correction, without iterative reverse-diffusion sampling. The framework was evaluated under contamination levels ranging from -7 to +2 dB, using independently recorded artifact sources and motor-imagery decoding across four public EEG datasets. Main results: Diff-ADN achieved mean Pearson correlations of 0.936, 0.839, 0.886, and 0.844 for EOG, EMG, ECG, and mixed EOG+EMG artifacts, respectively. When tested with independently recorded artifact sources, performance varied by artifact type: larger reductions were observed for EOG and EMG, while PTB and INCART ECG differed from the MIT-BIH reference by 0.026 and 0.014, respectively. Across the four motor-imagery datasets, ECG denoising recovered 4.88-5.59 percentage points (pp) in decoding accuracy compared with artifact-corrupted EEG. The largest recovery was 21.90 pp for PTB ECG artifacts on BCI IV-2a at -6 dB. Significance: The proposed framework combines artifact-aware EEG reconstruction with efficient deterministic inference, processing 2-s segments in 12.52-17.08 ms on the tested CPU. Results across independent artifact sources and multiple motor-imagery datasets further show that improvements in waveform reconstruction do not necessarily translate into uniform recovery of BCI decoding accuracy.