Channel- and Frequency-Independent EEG Artifact Removal Transformer

Electroencephalography (EEG) is a widely used method for recording brain activity, but it is highly susceptible to artifacts. This makes artifact removal essential for reliable downstream tasks such as epileptic seizure detection or Brain-Computer Interface (BCI). However, existing artifact removal methods are often inflexible with respect to input channel configurations and sampling frequencies, limiting their adaptability and range of application. Therefore, this thesis introduces CLEAN - a ChanneL- and frequency-independent Eeg Artifact removal traNsformer. The proposed approach reconstructs EEG data for any channel configuration and at any sampling frequency. This adaptability is achieved by sampling data points at random channels and time steps during training, combined with positional and temporal encoding. Moreover, the encoder of the proposed model learns a field representation of EEG data in the latent space, enabling a perceiver block in the decoder to query denoised signals at arbitrary positions and frequencies. This approach ensures high flexibility across diverse datasets and downstream tasks. Experimental results demonstrate that the model outperforms existing deep learning-based artifact removal methods while offering the added benefits of channel- and frequency-independence