Decoding the Variable Velocity of Lower-Limb Stepping Movements From EEG

Accurate decoding of lower-limb movement from electroencephalography (EEG) is essential for developing brain-computer interface (BCI) controlled exoskeletons in neurorehabilitation. This study investigates 3D velocity decoding at three fibular anatomical markers during overground stepping in healthy participants ( ${N}={9}$ ), using two approaches: (1) linear regression (LR) and (2) a deep learning (DL) framework combining convolutional neural networks (CNNs) and long short-term memory (LSTM) units. Participants were divided into two groups: G1 ( ${n}={5}$ ) performed cued forward and self-paced backward steps; G2 ( ${n}={4}$ ) performed cued forward and backward steps. The DL model significantly outperformed LR, achieving highest decoding accuracy (DA) in the forward-backward direction at the fibular head (R $= 0.63\pm 0.06$ , M±SD). Topographical analysis identified dominant contributions from the sensorimotor cortex (coupled with frontal regions in G2) within the 8-40 Hz band. Functional connectivity (FC) analysis revealed significant differences: only G2 showed statistically significant FC ( ${p}\lt {0.05}$ ), likely reflecting increased cognitive and sensorimotor demands under dual-cue conditions. In G2, FC occurred across delta (0-4 Hz), theta (4-8 Hz), alpha/mu (8-12 Hz), and low-beta (12-18 Hz) bands, linking motor areas associated with lower- and upper-limb control to other cortical regions, including the middle temporal gyrus (MTG), superior frontal gyrus (SFG), posterior cingulate cortex (PCC), superior parietal lobule (SPL), and supramarginal gyrus (SMG). These findings demonstrate that EEG-based 3D decoding of lower-limb kinematics is feasible during realistic locomotor tasks and suggest that cortical synchronization patterns vary with movement context. Our CNN-LSTM framework may support adaptive, intent-driven exoskeleton development for personalized neurorehabilitation.