Testing and optimizing hvEEGNet: a deep learning-based solution for EEG automatic artefacts detection

Electroencephalography (EEG) is a technique used to record electrical brain activity (e.g. electrical field generated by neurons' activities). However, the signals produced by this method are susceptible to interference from external sources, as other electrical phenomena can superimpose with the electrical fields generated by neurons, or other electrical fields generated within the body. When cleaning EEG signals, there is no universally accepted procedure for achieving optimal results. The state-of-the-art semi-automatic pipelines for EEG preprocessing and artifact removal vary significantly in the sequence of individual steps and the specific settings of parameters. This lack of standardization creates challenges in achieving consistent and reproducible results across different studies and applications. To address these issues, there is a growing interest in leveraging deep learning and AI techniques to automate and standardize EEG preprocessing. A promising approach in this regard is the employment of generative AI, especially variational autoencoders. These are a class of generative models that learn a probabilistic representation of the data by capturing its statistical distribution. They are designed to reconstruct input data from a lower-dimensional latent space, where the data is transformed into a more compact and informative representation. In the context of EEG signal processing, variational autoencoders can be employed to automatically identify artifacts from genuine brain activity signals. Their ability to model the statistical distribution of clean EEG data allows them to distinguish between signal components that represent true brain activity and artifactual ones. By learning from the distribution of clean EEG data, variational autoencoders can generate a model that can identify artifacts, leading to more consistent and objective preprocessing of EEG signals. In particular, hvEEGNet, a hierarchical variational autoencoder that was originally tested on the BCI competition dataset 2a dataset, has demonstrated the ability to identify artifacts. In this study, hvEEGNet is tested on the Temple University EEG corpus dataset to further evaluate its effectiveness in artifact detection. The reconstruction performance of hvEEGNet was observed to vary. The underlying reasons for this behaviour are examined in the current work, and comprehensive testing with diverse parameter configurations is conducted to enhance comprehension of the network's response to this particular dataset. A hypothesis is that the presence of artifacts is so pervasive that the model struggles to differentiate between artifactual and clean components. As a result, some artifactual components are mistakenly reconstructed, since they were interpreted as clean during the reconstruction phase of hvEEGNet.\\ The objective of this study is to optimise the training phase of hvEEGNet to enhance its flexibility for datasets characterised by high variability and the absence of preprocessing. Key improvements include the substitution of the initial Soft Dynamic Time Warping with Soft Dynamic Time Warping Divergence, as the network's loss function, and the incorporation of gradient clipping to effectively address the issue of exploding gradients. Moreover, the study suggested an investigation into the precise percentage split of cleaned and artifactual signal required to ensure the model's efficiency for automatic cleaning.\\ Overall, the integration of deep learning techniques, like variational autoencoders, into EEG signal cleaning processes represents a significant step forward in addressing the challenges of reproducibility and subjectivity in EEG data analysis. The automation of artifact detection and removal has the potential to enhance the accuracy and reliability of EEG research.