Multi-head noise regression for single-channel EEG: estimating ocular and muscle contamination to guide artifact removal.

Abstract

<i>Objective.</i>Electroencephalography (EEG) is often contaminated by ocular electrooculographic (EOG) and muscle electromyographic (EMG) artifacts, yet many pipelines apply uniform denoising, risking distortion of clean neural activity. We propose a two-head, single-channel regressor that estimates EOG and EMG noise-to-signal ratio (NSR, dB) from short segments and test whether it can guide selective artifact reduction, including downstream brain-computer interface decoding.<i>Approach.</i>Using EEGdenoiseNet clean EEG and artifact exemplars, we synthesized 2 s single-channel mixtures with known EOG/EMG NSR spanning -10 to +10 dB and trained several model families to jointly regress both NSRs. Generalization was evaluated on an independent eyeblink dataset via agreement with regression-based ocular-reference topographies, and in two applications: (i) gating stationary wavelet blink removal on a P3 event-related potential (ERP) dataset and (ii) gating the same denoiser on a 55-subject RSVP P300 speller dataset (FP1/FP2).<i>Main results.</i>A dilated temporal convolutional network (TCN) performed best (EOG: MAE ≈ 1.8 dB,<i>R</i>²≈ 0.82; EMG: MAE ≈ 1.0 dB,<i>R</i>²≈ 0.94) with low bias across NSR. The EOG head recovered blink topographies (median spatial correlation ≈ 0.91). On the P3 dataset, indiscriminate wavelet denoising reduced significant ERP channels, whereas TCN-guided gating preserved 22-23 of 24 while processing ∼9%-20% of segments. On the speller dataset, denoising all epochs reduced decoding, while selective denoising improved AUC (<i>θ</i>= 9 dB: ΔAUC = 0.327,<i>p</i>= 0.0040) while denoising 12.45 ± 9.29% of test segments.<i>Significance.</i>Multi-head noise regression provides interpretable, continuous ocular and muscle contamination estimates that can act as control signals for conservative, noise-aware artifact handling under constrained-lead conditions.