Optimizing ML Algorithms Under CSP and Riemannian Covariance in MI-BCIs

Motor imagery brain-computer interface (MI-BCI) systems face a multitude of challenges, one of which is optimizing multiclass classification of electroencephalography (EEG) signals. Hersche et al. (2018) extracted features from the 4-class BCI competition IV-2a data using Common Spatial Patterns (CSP) and Riemannian Covariance methods which resulted in improved performance speed and accuracy when fed to Support Vector Machines (SVM). We propose testing a variety of classifiers for both feature extraction methods to see their relative performance compared to SVM and to observe the impact of the two different feature extraction methods aforementioned on the different classifiers. SVM performed best, and ensemble algorithms had poor performance- especially AdaBoost. CSP feature extraction resulted in improved accuracy for most algorithms, but consumed more time, whereas the Riemannian feature extraction was twice-faster runtime for all algorithms, as expected. These results provide better understanding of feature extraction using CSP or Riemannian Covariance for MI-BCI data.