ReHA-Net: a ReVIN–hybrid attention network with multiscale convolution for robust EEG artifact removal in brain–computer interfaces

Abstract

Electroencephalography (EEG) is a non-invasive technique for monitoring brain activity, but its signal quality is frequently degraded by artifacts from ocular movements, muscle activity, and environmental noise. ReHA-Net is a deep learning framework for robust EEG denoising, combining a U-Net-based encoder–decoder with three core modules. (1) Hybrid Attention integrates temporal, spatial, and frequency attention to emphasize neural patterns while suppressing structured noise. (2) The Multiscale Separable Convolution (MSC) block employs dilated and parallel depth-wise separable convolutions with varying kernel sizes to capture both short-term and long-term temporal dependencies. (3) Reversible Instance Normalization (ReVIN) enhances cross-subject generalization while retaining subject-specific features. The model trains on an enhanced EEGdenoiseNet dataset with a wider signal-to-noise ratio range, combined artifact conditions, and tailored normalization strategies. ReHA-Net achieved strong denoising performance, with a PSNR of 27.10 dB, an SNR of about 17.06 dB, and a correlation coefficient of 0.976 with clean signals and a relative root mean square error (RRMSE) of 0.165. These outcomes demonstrate effective artifact reduction while maintaining neural activity, highlighting its suitability as a preprocessing step for tasks such as seizure detection, imagined speech decoding, and cognitive state monitoring.