Decoding motor imagery, an imagination of body movement, is an essential component for electroencephalogram (EEG) based brain-computer interface (BCI) applications. Recent studies have shown that acquiring brain signals within immersive VR environments can enhance motor imagery performance with more discriminant neural patterns. However, applying such signals directly to train classification models for real-life BCI systems may have limitations, as exposure to different environments may produce varying neural activity. To reduce the influence from such environmental causes, we propose a channel weight re-scaling module that refines EEG features depending on the learned separability of each electrode channel. As many pre-existing EEG classification models have their own strengths in different perspectives, our module is designed to be incorporated with previously studied convolutional neural network (CNN) based models that extract spectral-spatial features, preserving their characteristics. To analyze the performance of our module, we integrated our module with three widely used CNN models and performed evaluation using two datasets, which include our dataset containing resting state, left hand, and right hand motor imagery EEG data of participants within immersive VR and non-immersive environments. For all three baseline models, our module was able to improve classification performance. Furthermore, the re-scaled channel weights incremented towards theoretically important electrode positions. These results illustrate the potential of our module to be used not only to reduce unwanted features but also to inspect which channels were more influential to the model for classification.

Discriminating motor imagery with electroencephalogram (EEG)-based brain-computer interface (BCI) poses a challenge as it involves an extensive data acquisition phase that demands a substantial amount of effort from the user. To address this issue, one approach is to use unsupervised domain adaptation, where classification models are constructed using data from multiple subjects, and only the unlabeled data from the target user is used for model calibration. However, since brain patterns from motor imagery vary between individuals, the reliability of each subject must be considered when multiple subjects are used to build the classification model. Thus in this paper, we propose Selective-MDA that performs domain adaptation on each source subject and selectively limits influences based on their domain discrepancies. To evaluate our approach, we assess our results with two public datasets, BCI Competition IV IIa and the Autocalibration and Recurrent Adaptation datasets. We further investigate the effect of source selection by comparing the discrimination performance when different numbers of source domains are selected based on discrepancy measures. Our results demonstrate that Selective-MDA not only integrates multi-source domain adaptation to cross-subject motor imagery discrimination but also highlights the impact of source domain selection when using data from multiple subjects for model training.