Impact of Noise Elimination Methods on Classification Performance in Motor Imagery EEG

Classifying Motor Imagery (MI) electroencephalography (EEG) signals poses significant challenges, especially when dealing with multiple classes. EEG signals are inherently noisy, which complicates the task further. Using raw EEG signals directly in deep learning systems is possible, but without preprocessing, the system’s ability to distinguish between classes may be compromised due to the noise and variability in the data. In this study, the incorporation of electrooculography (EOG) channels alongside EEG channels was explored to facilitate the automatic extraction of cleaner data by enabling a proposed deep learning model to identify and interpret EOG as noise. The inclusion of EOG alongside with EEG method was compared with well-established preprocessing techniques such as EOG regression, wavelet-based denoising, common spatial pattern (CSP), and common average referencing (CAR). Using the BCI Competition IV dataset IIa, the integration of EOG channels demonstrated a notable performance improvement, particularly in the subject-independent model, when compared with other preprocessing methods. The subject-dependent model achieved a performance score of 0.879, while the subject-independent models reached 0.874. Additionally, testing the inclusion of EOG channels with only 3 out of 22 EEG channels showed that this setup was nearly as effective as using all 22 EEG channels, highlighting the efficiency of including EOG channels in enhancing EEG data quality and classification accuracy in BCI systems.