EEG-to-Gait Decoding via Phase-Aware Representation Learning

TL;DR

NeuroDyGait addresses the challenge of decoding lower-limb kinematics from EEG across diverse subjects and sessions by introducing a two-stage, phase-aware framework. Stage I learns semantically structured EEG–gait embeddings through biomechanical supervision and relative contrastive learning with a cross-attention distance, while Stage II performs domain-relations-aware decoding by dynamically fusing session-specific heads. The approach achieves robust cross-subject and cross-dataset generalization, supports real-time inference with sub-5 ms latency per 2-second window, and offers interpretable insights via phase-aware embeddings and cortical saliency maps. The work demonstrates strong performance gains over baselines, effective cross-dataset transfer, and practical potential for clinical rehabilitation, with future directions including clinical populations and multimodal integration.

Abstract

Accurate decoding of lower-limb motion from EEG signals is essential for advancing brain-computer interface (BCI) applications in movement intent recognition and control. This study presents NeuroDyGait, a two-stage, phase-aware EEG-to-gait decoding framework that explicitly models temporal continuity and domain relationships. To address challenges of causal, phase-consistent prediction and cross-subject variability, Stage I learns semantically aligned EEG-motion embeddings via relative contrastive learning with a cross-attention-based metric, while Stage II performs domain relation-aware decoding through dynamic fusion of session-specific heads. Comprehensive experiments on two benchmark datasets (GED and FMD) show substantial gains over baselines, including a recent 2025 model EEG2GAIT. The framework generalizes to unseen subjects and maintains inference latency below 5 ms per window, satisfying real-time BCI requirements. Visualization of learned attention and phase-specific cortical saliency maps further reveals interpretable neural correlates of gait phases. Future extensions will target rehabilitation populations and multimodal integration.

Figures (6)

• Figure 1: Overview of NeuroDyGait architecture. Stage I: The dual-encoder model consists of an EEG encoder, a motor encoder, and a decoder. It is trained with reconstruction, contrastive, and prediction losses to extract biomechanically meaningful EEG representations. Stage II: During training, session-specific heads and a domain weighting layer are optimized to predict final motor state. Test: The model computes a normalized mixture of all source-domain heads for unseen-domain generalization.
• Figure 2: t-SNE visualization of EEG embeddings from before (upper panel) and after (lower panel) Stage I training. Colors represent different gait phases as defined in Section\ref{['sec:kinematic_segmentation']}.
• Figure 3: Scatter plot showing the relationship between domain attention entropy and L1 prediction error across test sessions. All results are aggregated over all cross-validation folds.
• Figure 4: Saliency maps generated by NeuroDyGait computed across all folds.
• Figure 5: Aggregated attention map shows that each session learns a distinct weighting pattern, indicating that the model captures session-specific domain relationships rather than converging to a uniform structure.