SPAR-EEG: Selective Pass-Wise Artifact Reduction for Wearable Single-Channel EEG Denoising
Abstract
Single-channel electroencephalography (EEG) is attractive for wearable neurotechnology and brain–computer interface (BCI) applications, including assistive interfaces and clinical monitoring, but artifact suppression is difficult when auxiliary channels, artifact labels, or user-selected clean baseline segments are unavailable. We introduce SPAR-EEG, a self-contained framework for Selective Pass-wise Artifact Reduction in single-channel EEG. The framework applies three artifact-specific attenuation passes to each EEG epoch: a variational mode decomposition (VMD)-based pass for high-frequency electromyographic (EMG) bursts, a singular spectrum analysis (SSA)-based pass for blink-like electrooculographic (EOG) transients, and an SSA-based pass for slow motion-related drift. Rather than rejecting components globally, each pass estimates artifact-dominant regions and attenuation strength directly from the input channel. SPAR-EEG was evaluated using controlled EEGdenoiseNet and PhysioBank benchmarks, pass ablations, task-locked event-related potential (ERP) preservation, runtime diagnostics, dry-electrode exercise EEG, and a downstream rapid serial visual presentation (RSVP)/P300 speller task using only FP1 and FP2. Across 26 EEGdenoiseNet input signal-to-noise ratio (SNR) levels, it obtained the largest average artifact-region SNR improvement among the tested wavelet, empirical mode decomposition (EMD), and artifact-label-guided wavelet quantile normalization (WQN) baselines for EMG, EOG, and combined EOG+EMG contamination (9.05, 8.28, and 8.00 dB, respectively). In exercise EEG, denoising reduced high-amplitude artifact burden and increased alpha and steady-state visual evoked potential (SSVEP) spectral-prominence metrics. In the P300 validation, the full SPAR-EEG sequence increased repetition-curve area under the curve (AUC) by 0.048 (Holm-adjusted ${p}={0}.{0069}$ ) and improved final Letter@15 accuracy by 7.8 percentage points. These results suggest that artifact-specific selective attenuation can provide a practical self-contained alternative for single-channel EEG denoising in low-burden and movement-prone settings.