Enhancing Shoulder Pre-Movements Recognition Through EEG Riemannian Covariance Matrices For a BCI-based Exoskeleton

This work proposes a system to recognize motor intention during shoulder flexion/extension, in order to convey control commands towards an upper-limb robotic exoskeleton. For this purpose, two recognition systems were explored: 1) spatial features from Riemannian geometry and linear classification; and 2) common average reference pre-processing, feature extraction from time and frequency domains, and support vector machine classification. The effects of varying window sizes were also explored to anticipate shoulder flexion/extension, along with varying frequency bands, electroencephalography (EEG) channel arrangements, and classifier types. For some participants, our proposed system achieved Kappa values higher than 0.74 during shoulder movement recognition. Moreover, average accuracy (ACC) $\geq 67$%, Kappa $\geq 0.34$, and false positive rate (FPR) $\leq 33$% during shoulder motor anticipation were obtained, thus suggesting the potential usefulness of the proposed method for robotic exoskeleton control.