Detecting motor imagery activity for EEG based brain computer interface applications

In this study, the detection of motor imagery activities related to the potential Brain-Computer Interface applications is investigated in a theoretical and experimental way. The theoretical research focuses on several approaches (Common Average Reference, Principal Components Analysis, Blind Source Separation, Generalized Eigendecomposition) of linear spatial filters to improve the classification accuracy of different motor imagery activities in Brain-Computer Interface systems. Various motor imagery activities (right hand/left hand and foot/tongue) in BCI Competition datasets are used to evaluate the spatial filter performances. It is aimed to improve the classification accuracy performance of spatial filters by regularizing and whitening operations in the data. In the experimental study, the detectability of alpha-band amplitude drop rhythms (Event-Related Desynchronization, "ERD") varying with different motor imagery behaviors (right hand/left hand) was investigated by using a two-channel EEG device. Generalized Eigendecomposition method, which is one of the spatial filtering approaches examined in the theoretical study, was used to determine the activity-dependent topographical localization and to signal patterns from the electrode measurements. Experimental studies were carried out with 6 healthy and right-handed participants, and a cue-based procedure was used. Based on the theoretical research findings, the Generalized Eigendecomposition approach yielded a significantly better outcome than other methods. The data whitening preprocesses provided a good improvement in the performance of spatial filters. Regularization for the Generalized Eigendecomposition showed a low improvement in overall classification accuracy, and it provided good classification success for some subjects. Based on the findings in the experimental studies, for 4 of 6 participants, the ERD patterns in the periods of the imagery activities and topographically associated regions of the brain were determined by the spatial filter approach, where the signal amplitudes decreased by an average of about 50% (mean: %47,3) compared to their reference periods.