A MEG Study of Different Motor Imagery Modes in Untrained Subjects for BCI Applications

摘要

Motor imagery is a most commonly studied neurophysiological pattern that is used in brain-computer interfaces as a command for exoskeletons, bioprostheses, wheelchair and other robotic devices. The mechanisms of motor imagery manifestation in human brain activity include dynamics of motor-related frequency bands in various brain areas, among which the most common is sensorimotor rhythnm. In present work we consider time-frequency structure of magnitoencephalographical (MEG) motor imagery in untrained subjects. We conduct series of experiments to collect MEG motor imagery dataset in untrained subjects. We confirm the emergence of two types of motor imagery - visual (VI) and kinesthetic (KI), which cause different types of event-related potentials (ERP) dynamics and require different approaches to classification using mashine learning methods. We also reveal the impact of dataset optimization on the artificial neural network performance, which is essential topic in brain-computer interface (BCI) development. We show that developing classification stratedy based on time-frequency features of the particular MEG signal can increase classification accuracy of the VI mode to the level of the KI.