Exploring the Latent Space of Person-Specific Convolutional Autoencoders for Eye-Blink Artefact Mitigation in EEG Signals

Electroencephalography (EEG) is a non-invasive and cost-effective technique that allows the investigation of brain activity. However, EEG recordings often suffer from artefacts that complicate signal analysis. Eye-blink artefacts pose a significant challenge among these artefacts due to their frequency overlap with neural signals. Machine Learning, notably semi-supervised Autoencoders (AEs), appears promising in removing EEG artefacts. This research investigates the use of Convolutional Autoencoders (CAEs) for mitigating eye blinks in EEG signals, deviating from a previous use of Convolutional Variational AEs. This shift can offer a simpler approach with reduced computational complexity. Specifically, the latent space of CAEs, trained on spatially preserving EEG topographic maps, was explored to identify latent components selective for eye blinks. Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) were employed to evaluate each latent component's discriminative performance. The most discriminative component, determined by the highest AUC, is subsequently modified to mitigate eye blinks. Specifically, the median is chosen to mask this discriminative latent component for blink artefact removal. Visual inspections and Pearson correlation indices between the original EEG signal and the reconstructed clean version were used to evaluate the effectiveness of artefact removal. This study contributes to the knowledge for introducing an offline pipeline able to detect and remove eye blinks from EEG recordings without human intervention.