EEG-to-gait decoding via phase-aware representation learning.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate decoding of lower-limb motion from electroencephalography (EEG) signals is crucial for brain-computer interface (BCI) development.
• Existing methods face challenges in causal, phase-consistent prediction and handling cross-subject variability in movement intent recognition.

Purpose of the Study:

• To introduce NeuroDyGait, a two-stage, phase-aware EEG-to-gait decoding framework.
• To explicitly model temporal continuity and domain relationships for improved EEG-based motion decoding.
• To address cross-subject variability and ensure real-time inference for BCI applications.

Main Methods:

• Stage I employs relative contrastive learning with a cross-attention metric to learn semantically aligned EEG-motion embeddings.
• Stage II utilizes dynamic fusion of session-specific heads for domain relation-aware decoding.
• The framework was evaluated on two benchmark datasets (GED and FMD).

Main Results:

• NeuroDyGait demonstrated substantial performance gains over existing baseline models, including a recent 2025 model (EEG2GAIT).
• The framework exhibits generalization capabilities to unseen subjects.
• Inference latency was maintained below 5 ms per window, meeting real-time BCI requirements.

Conclusions:

• NeuroDyGait offers an effective solution for accurate EEG-based lower-limb motion decoding.
• The interpretable neural correlates of gait phases were revealed through visualization techniques.
• Future work will focus on rehabilitation applications and multimodal integration for enhanced BCI systems.