Finding Discriminant Lower-Limb Motor Imagery Features Highly Linked to Real Movements for a BCI Based on Riemannian Geometry and CSP

Stroke is a neurological syndrome that may affect severely lower-limb movements and the normal gait. The complete or partial restoration may be achieved through alternative rehabilitation therapies, such as Motor Imagery (MI)-based Brain Computer Interfaces (BCIs). Although these systems have shown promising results on post-stroke patients with severe disability, their performance recognizing MI may be reduced for people executing MI tasks with high difficult or producing weak brain activation. This study presents a proposal to improve the calibration stage of a low-cost electroencephalographic (EEG) based MI BCI with pedal end-effector, which integrally aims to activate continuously the central and peripheral mechanisms related to lower-limbs, and obtain the best feature vectors for MI recognition. This setup enables users to perform pedaling MI and receive passive pedaling into a Calibration phase. Consequently users can produce related EEG signals useful to obtain those more discriminant MI feature vectors through a probability analysis combining patterns from pedaling MI and passive pedaling. Here, Riemannian geometry and Common Spatials Patterns (CSP) for feature extraction were used independently or combined in our approach. Preliminary results show that the proposed method may improve the BCI performance. For healthy subjects, the approach using CSP achieved accuracy (ACC) up to 98.43%, whereas for PS1 and PS2 obtained ACC of 71.07% and 79.24%, respectively. However, Riemannian geometry plus CSP using LDA reached better results for healthy subjects and patients (mean ACC of 73.84%).