Deep Learning Approach for EEG Classification in Lower-Limb Movement Phases: Towards Enhanced Brain-Computer Interface Control

Embedded Brain-Computer Interfaces (BCI) have emerged as an alternative for rehabilitation/assistance of people with neuromotor impairments. Data-driven methods are usually implemented to design controllers to (de)activate these systems using electroencephalography (EEG), which may limit usability and degrees of freedom. Thus, the development of more adaptive classifiers remains a challenge. This study focuses on the implementation of a Deep Learning (DL) method to classify EEG signals during different phases of gait and pedaling. To address this, public and in-house databases were used, which contain information on EEG and kinematic data during walking on a treadmill and pedaling on a minibike. Kinematic labels of the phases were extracted using k-means and classified by a Convolutional Neural Network (CNN), which demonstrated accuracy above 60%. Furthermore, a negative correlation was found between velocities and classifier performance metrics, suggesting that faster movements can hinder EEG decoding. Our proposal is a starting point to identify current states of lower-limb movement, which can be used for designing state machine-based controllers of robotic BCIs with rehabilitation or assistance purposes, such as motorized mini exercise bikes or leg exoskeletons.