Deep learning for automatic artefacts detection in EEG data: a comparative study between hvEEGNet and ICLabel

The identification and elimination of signal disturbances is a fundamental step in the analysis of EEG data. Manually annotating datasets through expert labelling is a laborious process, as it demands significant time and resources, along with a high level of expertise to precisely distinguish between segments affected by artefacts and those that are clean. This thesis investigates the feasibility of utilizing deep learning architectures to automate the detection of signal irregularities, thereby mitigating the challenges posed by manual labelling. To this end, two deep learning-based models, namely ICLabel and hvEEGNet, were evaluated for their performance in identifying artefactual trials within the public Dataset 2a and the in-house Cosynclab dataset. The study uncovered patterns of disruption in specific test set subjects of the Dataset 2a that were overlooked during expert review. Additionally, several classification methods were analysed, emphasizing the models’ ability to detect noisy trials. Notably, the findings introduce the concept of a tunable detection sensitivity, allowing for greater flexibility in managing artefactual segments. The results of this research highlight significant progress in EEG signal processing, addressing key challenges and paving the way for broader implementation of fully automated techniques to identify and handle noise in electrophysiological recordings.