Using Machine and Deep Transfer Learning for Classification of EEG Signals from Embodied and Non-embodied Priming in a Motor Imagery Training in Virtual Reality

Motor imagery (MI) frameworks have a long history of being used in different motor training applications. Since the emergence of brain-computer interfaces (BCI), MI techniques have been integrated into BCI frameworks to enable controlling external devices through interpreting neural signals into executable commands. Virtual reality (VR) has the capacity to introduce novel ways of improving the performance of MI-BCI trainings through enabling embodiment prior to and during the tasks. While the performance improvements achieved by using VR-based embodiment during the MI training has been investigated previously, the effects of VR-based motor priming prior to the training needs to be further addressed. This study uses machine learning (ML) to find out whether or not VR-induced avatar embodiment before the actual MI-BCI training is capable to make significant differences in electroencephalography (EEG) signals recorded from the users. Detecting the most relevant features which best represent such differences enables the introduction of biomarkers of VR-based motor priming embodiment in MI-BCI applications. Relating these biomarkers to the specific neurophysiological functions which facilitate MI can help in design and development of MI-BCI applications with improved accuracy and performance.