EEG Artifact Detection and Correction with Deep Autoencoders

TL;DR

The study tackles automated EEG artifact detection and correction by introducing LSTEEG, an LSTM-based autoencoder that learns a structured latent space for interpretable representations. It combines anomaly-detection-based artifact detection with supervised artifact correction, and benchmarks against CNN-based AEs (UNET and CLEEGN) on the LEMON and EOG-contaminated datasets. Key contributions include a two-task framework, unsupervised artifact screening via reconstruction error, and detailed latent-space analyses (MAD activation maps and smooth interpolations) that enable data generation and downstream EEG applications. The findings show competitive detection performance and robust correction capabilities, with a low-dimensional latent space offering interpretability and potential for augmentation in EEG research and brain-health pipelines.

Abstract

EEG signals convey important information about brain activity both in healthy and pathological conditions. However, they are inherently noisy, which poses significant challenges for accurate analysis and interpretation. Traditional EEG artifact removal methods, while effective, often require extensive expert intervention. This study presents LSTEEG, a novel LSTM-based autoencoder designed for the detection and correction of artifacts in EEG signals. Leveraging deep learning, particularly LSTM layers, LSTEEG captures non-linear dependencies in sequential EEG data. LSTEEG demonstrates superior performance in both artifact detection and correction tasks compared to other state-of-the-art convolutional autoencoders. Our methodology enhances the interpretability and utility of the autoencoder's latent space, enabling data-driven automated artefact removal in EEG its application in downstream tasks. This research advances the field of efficient and accurate multi-channel EEG preprocessing, and promotes the implementation and usage of automated EEG analysis pipelines for brain health applications.

Figures (6)

• Figure 1: LSTEEG network architecture.
• Figure 2: Artifact Detection performance of the CLEEGN, UNET and two different LSTEEG configurations ($N_{LS} = 500$ and $N_{LS}=2000$) on the two evaluation datasets, the EOG-synthetically contaminated dataset (left) and LEMON Clean/RawFiltered dataset (right). We compare the different networks' classification performance by showing their Receiver Operating Characteristic (ROC) curves. Additionally, Area Under the ROC Curve (AUC, higher is better) values are provided in the figures legends.
• Figure 3: Reconstruction Comparison for an epoch containing both muscular (high frequency burst between 0.5 and 0.75s) and ocular (high amplitude peaks at 1.50s) artifacts in the testing set. Shown networks trained with $\mathbf{X_{Br}}$. The Red line is the cleaned target EEG epoch; the Grey line is the input EEG epoch, containing the large amplitude artifact; the Blue line is the output of each network. While CLEEGN produces high amplitude outputs, unable to correct the artifact, both UNET and the two configurations of LSTEEG are able to remove the artifact from the EEG signal.
• Figure 4: Correction comparison for the same EEG epoch with muscular and ocular artifacts as in \ref{['fig:Br_muscle']}. Shown networks trained with $\mathbf{X_{Ar}}$. The Red line is the cleaned target EEG epoch; the Grey line is the input EEG epoch, containing the large amplitude artifact; the Blue line is the output of each network. While CLEEGN produces high amplitude outputs, unable to correct the artifact, both UNET and the two configurations of LSTEEG are able to remove the artifact from the EEG signal.
• Figure 5: Topographical maps indicating activation patterns of Most Activated Latent Space Dimensions (MADs) by spectral features.