EEGPARnet: codificador de transformador de atención tiempo-frecuencia y decodificador GRU para la eliminación de...

Effective Electroencephalogram (EEG) signal processing necessitates the mitigation of physiological artifacts. While deep learning frameworks have demonstrated superior performance over traditional methods for this task, their high complexity and computational demands hinder deployment on resource-constrained platforms. In this work, denoising network called EEGPARnet is proposed to address this limitation. The proposed architecture integrates transformer encoders equipped with temporal and spectral attention modules and a Gated Recurrent Unit (GRU)-based decoder. This fusion enables the model to learn time-frequency long-range similarities, facilitating efficient feature extraction and a reduced number of trainable parameters. Experimental validation of the proposed model on the EEGDenoiseNet dataset revealed an average temporal relative root mean square error ([Formula: see text]) of 0.289, spectral relative root mean square error ([Formula: see text]) of 0.312, and a correlation coefficient (CC) of 0.942 for ocular artifact removal. For muscular artifact removal, the proposed method achieved competitive results against state-of-the-art techniques, with mean [Formula: see text], [Formula: see text], and CC values of 0.458, 0.428, and 0.855, respectively. Compared to state-of-the-art model, the proposed EEGPARnet demonstrated a significant reductions in computational complexity with [Formula: see text] fewer trainable parameters, [Formula: see text] less FLOPS, and [Formula: see text] smaller storage, making it a step closer towards deployment on resource-constrained devices for real-time EEG denoising without compromising performance.