A Dynamic Evaluation-Denoising Network for Motion Artifacts Removal from Single-Channel EEG

Brain-computer interfaces (BCIs) have gained significant attention in rehabilitation research as a critical step in investigating neural remodeling techniques. However, most existing methods usually overlook the randomness and diversity of motion artifacts, thereby lacking the desired generalization ability and denoising precision, which limits their practical application. To address these limitations, we propose a Dynamic Evaluation Denoising Network (DED-Net) that incorporates an evaluation model with cross-domain feature fusion for artifact detection and classification. Then dynamically selecting Bidirectional Long Short-Term Memory (Bi-LSTM) networks with varying parameters for artifact removal, which achieves superior performance compared to state-of-the-art methods. Our experiment on a semi-simulated dataset constructed by EEGdenoiseNET demonstrates that the performance of DED-Net is advanced over the state-of-the-art method, i.e., SDNet, in terms of the signal-to-noise rate (SNR) and correlation coefficient (CC). Using our method, SNR and CC are 6.0597 dB and 95.28%, respectively increasing by 20.48% and 3.15%. Experiments on real EEG data demonstrate the superior performance of the proposed method in reconstructing EEG signals, in terms of the intent recognition tasks, achieving a remarkable accuracy of 88.89%, outperforming other methods.