Due to the non-stationary nature of electroencephalography (EEG) signals, a Brain-computer Interfacing (BCI) system requires frequent calibration. This leads to intersession inconsistency which is one of the main reason that impedes the widespread adoption of non-invasive BCI for realworld applications, especially in rehabilitation and medicine. Domain adaptation and deep learning-based techniques have gained relevance in designing calibration-free BCIs to solve this issue. EEGNet is one such deep net architecture that has been successful in performing inter-subject classification, albeit on data from healthy participants. This is the first paper, which tests the performance of EEGNet on data obtained from 10 hemiparetic stroke patients while performing left and right motor imagery tasks. Results obtained on implementing EEGNet have been promising and it has comparably good performance as from expensive feature engineering-based approaches for both withinsubject and cross-subject classification. The less dependency on feature engineering techniques and the ability to extract generalized features for inter-subject classification makes EEGNet a promising deep-learning architecture for developing practically feasible solutions for BCI based neuro-rehabilitation applications.
|Title of host publication||IEEE WORLD CONGRESS ON COMPUTATIONAL INTELLIGENCE (WCCI) 2020|
|Publication status||Accepted/In press - 20 Mar 2020|