U-EEG: A Deep Learning Autoencoder for the Detection of Ocular Artifact in EEG Signal

Current methods to remove blink artifacts from electroencephalography (EEG) brain signals, such as independent component analysis (ICA), have reduced functionality when processing EEG datasets that have a relatively small number of EEG electrodes. We evaluated an approach to address this limitation by modifying a U-Net convolutional neural network (CNN) to extract blink artifacts from EEG signals contaminated with electrooculography (EOG) signal. Crucially, U-Nets have high prediction accuracy even when trained on small datasets (i.e., low EEG electrode count). Our model (U-EEG) removed ocular artifacts from contaminated signals with a higher accuracy compared to the industry standard ICA algorithm for EEG analysis. With respect to an artifact-free baseline signal, we found that U-EEG had lower root mean square error (RMSE) for artifact detection compared to ICA (U-EEG: 3.134 ± 0.893; ICA: 5.793 ± 1.769). Our findings highlight the utility of data-driven models for EEG artifact correction, and the potential these methods have for the development of a real-time braincomputer interface (BCI) for real-world solutions.