Publication

Abstract : EEG-based brain computer interfaces (BCI) and neurophysiological research take advantage of electroencephalography (EEG) signals although they are vulnerable to various physiological and environmental artefacts. Classic preprocessing artifacts rejection methods (e.g., ICA, wavelet and so on) require manual attention and are computationally expensive, so they can hardly be applied in real-time systems. In this paper a lightweight fully convolutional 1D autoencoder network structure is presented to perform automatic artifact removal directly in the time domain. The model trained from end-to-end with segment window preprocessing on the EEG Motor Movement/Imagery Dataset (EEGMMIDB). Experiments showed significant reduction of reconstruction error, with mean squared error reduced from 0.9473 to 0.1824 in 100 epochs, and a final test loss of 0.1765, demonstrating effective artifact suppression and good generalization. The resulting approach does not require any handcrafted feature nor reference channel, thus allowing to scale up EEG denoising in embedded systems in real time.