Optimizing Time-Dependent Discriminative Spatial Filter for EEG-Based Multi-class Motor Imagery Classification

Neuronal power attenuation or enhancement in the specific frequency bands over the sensorimotor cortex, called Event-Related Desynchronization (ERD) or Event-Related Synchronization (ERS), is one of the major phenomena in brain signals evoked by imagination of body parts movement. So many research groups have devoted their efforts to extract discriminative features by utilizing these phenomena and to classify different motor imageries. It is, however, known that the nature of the motor imagery related EEG signals is highly dependent on the subjects and variable over trial to trial. To address this issue, in this paper, we propose a novel method of optimizing discriminative spatial filters on a time domain in each frequency band. It effectively reflects changes of subject-specific discriminative spatial distributions of the ERD/ERS patterns on a time domain in different frequency bands. We assess the proposed method with experiments of 4-class motor imagery (left-hand, right-hand, foot, and tongue) classification using the BCI Competition IV dataset 2-a. Experimental results show that the proposed method outperformed previous methods in the literature.