Discrimination of Shoulder Flexion/Extension Motor Imagery Through EEG Spatial Features to Command an Upper Limb Robotic Exoskeleton

This work presents a comparison between two methods for spatial feature extraction applied on a system to recognize shoulder flexion/extension motor imagery (SMI) tasks to convey on-line control commands towards a 4 degrees-of-freedom (DoF) upper-limb robotic exoskeleton. Riemannian geometry and Common Spatial Pattern (CSP) are applied on the filtered EEG for spatial feature extraction, which later are used by the Linear Discriminant Analysis (LDA) classifier for motor imagery (MI) recognition. Three bipolar EEG channels were used on six healthy subjects to acquire our database, composed of two classes: rest state and shoulder flexion/extension MI. Our system achieved a mean accuracy (ACC) of 75.12% applying Riemannian, with the highest performance for Subject S01 (ACC = 89.68%, Kappa = 79.37%, true positive rate (TPR) = 87.50%, and FPR < 8.13%). In contrast, for CSP, a mean ACC of 66.29% was achieved. These findings suggest that unsupervised methods for feature extraction, such as Riemannian geometry, can be suitable for shoulder flexion/extension MI to command an upper-limb robotic exoskeleton.