Temporal and frequency feature extraction with canonical variates analysis for multi-class imagery task

The objective of this study is to improve the accuracy of classification for a multi-imagery task by using canonical variates analysis in a brain-computer interface (BCI). Electroencephalogram (EEG) is recorded from subjects performing a four-class motor imaginary task, left hand, right hand, foot and tongue. Temporal features are extracted as squared band pass filtered EEG, and frequency features are extracted as energy in specific rhythms. Features in both domains are projected into a canonical discriminant spatial feature space provided by canonical variates analysis (CVA), and classified by support vector machines (SVM) with different kernel functions and parameters. The classification accuracy is assessed using 10-fold cross-validation. The maximum estimated accuracy is 82.8% at temporal domain using C-SVM with radial basis kernel. The results show that this approach achieves a good performance in multi-class motor imagery task, and has the potential in the application of complicated control device, such as brain-control based meal assistance system.