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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Rehabilitation Technology

Background:

• Accurate decoding of lower-limb movement using electroencephalography (EEG) is critical for developing brain-computer interface (BCI) controlled exoskeletons.
• Existing methods require further refinement for real-world neurorehabilitation applications.

Purpose of the Study:

• To investigate and compare the efficacy of linear regression (LR) and a deep learning (DL) framework (CNN-LSTM) for decoding 3D lower-limb movement from EEG during overground stepping.
• To analyze brain activity patterns, including functional connectivity, associated with different stepping conditions.

Main Methods:

• Utilized EEG data from 9 healthy participants performing cued and self-paced stepping tasks.
• Applied both linear regression (LR) and a convolutional neural network-long short-term memory (CNN-LSTM) deep learning model for decoding 3D velocity at fibular markers.
• Conducted topographical and functional connectivity (FC) analyses across different frequency bands.

Main Results:

• The DL (CNN-LSTM) model significantly outperformed LR, achieving the highest decoding accuracy (DA) for forward-backward stepping (R = 0.63 ± 0.06).
• Sensorimotor cortex activity (8-40 Hz) was dominant, with additional frontal contributions in dual-cue conditions.
• Statistically significant functional connectivity (p < 0.05) was observed only in the dual-cue group (G2), involving multiple cortical regions across delta, theta, alpha/mu, and low-beta bands.

Conclusions:

• EEG-based 3D decoding of lower-limb kinematics during realistic locomotion is feasible.
• Cortical synchronization patterns differ based on movement context and cognitive load.
• The developed CNN-LSTM framework shows promise for advancing adaptive, intent-driven exoskeleton control in neurorehabilitation.