A Feature Exchange and Integration-based CNN-BiLSTM Network for EEG Denoising

Electroencephalography (EEG) is a neurorecording method to record brain activity, but muscle artifacts (MAs) from actions such as talking, clenching, swallowing, and sniffing frequently contaminate EEG signals, leading to incorrect interpretations. To overcome this issue, a Feature Exchange and Integration-Based Convolutional-BiLSTM Network (FEI-Net) is proposed to suppress muscle artifacts (MAs) from single-channel EEG recordings. FEI-Net consists of an EEG Restoration Branch (ERB) for reconstructing clean EEG signals and an Artifact Restoration Branch (ARB) for artifact extraction. The two branches’ outputs are passed through an Integration Block (IB) to enhance EEG reconstruction, while a Feature Exchange Block (FEB) allows bidirectional information exchange to enhance feature learning. The effectiveness of the model is tested using the publicly available EEGDenoiseNet dataset and evaluated based on metrics such as Relative Root Mean Square Error (RRMSE), Signal-to-Noise Ratio (SNR), Percentage Reduction in Correlation Coefficient (λ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CC</inf>), Relative Error (RE), and Percentage Root Mean Square Difference (PRD). Experimental results reveal that FEI-Net surpasses existing approaches across multiple evaluation metrics, achieving superior denoising performance.