BiGRU-TFA: An Attention-Enhanced Model for EEG Signal Reconstruction Using Temporal and Frequency Features

Electroencephalogram (EEG) signals are often contaminated by artifacts from physiological and nonphysiological sources, complicating their analysis and interpretation. Traditional methods for artifact removal struggle with nonstationary noise and overlapping neural components, while deep learning approaches face challenges in preserving temporal dependencies and frequency-domain features. The article proposes BiGRU-temporal-frequency attention (TFA), an attention-enhanced bidirectional gated recurrent unit (BiGRU) model that integrates temporal and frequency features for robust EEG signal reconstruction. BiGRU-TFA employs multihead self-attention (MHSA) and a self-supervised learning model to remove artifacts without heavy reliance on labeled data. The experimental results demonstrate superior performance, with 30% improvements in signal-to-noise ratio (SNR), a structural similarity index measure (SSIM) of 0.715, a robust correlation to true signal (CC = 0.9513), a low mean squared error (MSE) of 0.0286, and better reconstruction accuracy compared to existing methods. BiGRU-TFA effectively preserves neural patterns, making it a promising tool that generalizes across EEG acquisition, enhancing brain-computer interfaces (BCIs) and neurofeedback applications.