Combining Density based and Linear Discriminant Approaches for Motor Imagery Classification*

For transferring brain-computer interfaces outside of the lab to clinical settings, it is necessary to have a high accuracy with models having a limited number of hyper-parameters. State of the art techniques include discriminant approaches using spatial filters, deep learning, and density based methods using Riemannian geometry. We propose a pattern recognition system for the multiclass classification of brain evoked responses corresponding to motor imagery that combines features obtained from the Riemannian geometry, with distances to the mean of each class, with a discriminant approach (Bayesian linear discriminant analysis) using 15 frequency bands from 8 to 24 Hz to cover the mu and beta bands. We investigate the impact of these different frequency bands, separated in four sets, on the accuracy and how frequency bands selection using backward elimination or forward addition can enhance the accuracy of the classification tasks. These approaches were evaluated on the publicly available dataset 2A (4 classes - 9 subjects) and (2 classes - 14 subjects). While there are differences between bands and across subjects, the best overall performance was obtained with all the bands. The kappa value for multiclass motor imagery detection is 0.60. The average binary classification across the six pairwise tasks is 80.83%.