Neighbourhood-clipped latent space of VAEs with spatially preserved EEG topographic maps for ocular artefact reduction

Electroencephalographic (EEG) recordings are often affected by artefacts such as eye blinks, which complicate their analysis. Although multiple techniques exist to detect and remove artefacts, many require manual intervention. This study presents a novel, self-supervised, and fully automated approach to identify and reduce artefacts in EEG signals using a Variational Autoencoder (VAE) framework. In this approach, subject-specific VAEs with convolutional layers are trained on spatially preserved EEG topographic maps. An anomaly detection strategy based on the negative log-likelihood of activated latent vectors from the training data is employed to identify abnormal topomaps, assigning each input an anomaly score. Input topomaps exceeding a defined threshold, together with their neighbouring topomaps, are clipped using a standard IQR-based method, combining clipping and mitigation to influence surrounding regions affected by eye blinks. The reconstructed EEG signals are then compared against a baseline created using an offline Independent Component Analysis (ICA) method with automated detection of artefact components inspired by the FASTER methodology. Results indicate improved signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR) in channels such as FP1 and FP2, while other channels show comparable performance to ICA Fast. Additionally, mean absolute error (MAE), normalised root mean square error (NRMSE), and correlation coefficients demonstrate that the reconstructed signals maintain comparable quality to the baseline. The findings further show that the method preserves non-artifactual segments, maintaining their neural dynamics. Overall, this study introduces a fully automated, subject-specific approach for EEG artefact identification and denoising using latent space neighbourhood clipping of the latent space representation of anomalous topomaps.