EXPLORING REGULARIZATION AND CHANNEL OPTIMIZATION IN MOTOR IMAGERY EEG

Abstract

Through electroencephalography (EEG) signals, motor imagery (MI) can be seen as a non-invasive channel via which people direct their psychological activity to interact with the environment in an unmanual way, so they are valuable instruments for BCIs. EEG signals, on the other hand, present difficulties in correct identification and modeling since they have a low signal-to-noise ratio and are nonstationary. Whereas earlier works center on the use of CNN as a determining factor for the emergence of multimodal selection, this study compares the efficacy of classical machine learning techniques against deep learning methods: ANNs and their hybrids on connecting motor imagery performance. Furthermore, it delves into fusion of ensemble and hybrid models that improve performance across several iterations. The outcomes reveal that although CNN-based approaches offer strong spatial and temporal birth disambiguation, better precision and usefulness result from combining conventional and deep learning models, therefore surpassing those based on CNN. The study emphasizes how critical it is for innovative techniques to address brain signal problems and in which potential ways they could be further verified in BCI uses.