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

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.

💡 Research Summary

NeuroDyGait introduces a two‑stage, phase‑aware framework for decoding lower‑limb gait from scalp EEG, addressing two persistent challenges in the field: (1) the loss of temporal continuity when using point‑wise prediction strategies, and (2) the inadequate handling of inter‑session and inter‑subject variability that hampers cross‑subject generalization.

In Stage I, a dual‑encoder architecture processes 2‑second EEG windows and synchronized joint‑angle trajectories. The EEG encoder (deep multi‑scale 1‑D convolutions followed by a final 1 × T convolution) yields a latent vector zₑ, while the motor encoder (stacked 1‑D convolutions plus a Transformer with positional encoding) produces zₘ. A novel relative contrastive learning scheme replaces fixed positive/negative pairs with an “all‑in‑batch” ranking objective. A cross‑attention mechanism treats zₑ as a query and zₘ as key/value, computing attention coefficients η via scaled dot‑product softmax. The attended motor feature ˆzₘ = Wₒ(η·Wᵥzₘ) is compared to the original motor embedding, defining a cross‑modal distance d(zₑ, zₘ) = ‖ˆzₘ − zₘ‖²₂. This distance drives the contrastive loss, encouraging the embedding space to reflect fine‑grained semantic similarity between EEG and gait phases.

Stage I also incorporates a reconstruction loss (L_rec) that forces the EEG embedding to reconstruct the full 2‑second motion sequence, and a prediction loss (L_pred) that focuses on the final frame of the sequence, thereby preserving information most relevant for real‑time control.

Stage II tackles domain generalization by allocating a dedicated prediction head hₛ to each training session (or subject). During fine‑tuning, a lightweight domain‑scoring network learns a weighting vector α ∈ Δⁿ (softmax over sessions) that dynamically fuses the outputs of all heads. The overall loss combines a supervised prediction term (L_sup) with a domain‑fusion term (L_df) that aligns α with the true session identity, encouraging the model to capture both session‑specific nuances and inter‑session relationships. At test time, the model produces a mixture prediction ȳ_mix = Σₛ αₛ·hₛ(zₑ), enabling robust inference on unseen subjects without any target‑domain data.

Extensive experiments on two benchmark datasets—GED (Gait Event Dataset) and FMD (Full‑body Motion Dataset)—show that NeuroDyGait outperforms the recent EEG2GAIT model by 12–18 % reduction in mean absolute error (MAE) and comparable gains in RMSE. In a leave‑one‑subject‑out scenario, performance degradation remains under 5 %, demonstrating strong cross‑subject generalization. Inference latency averages 4.7 ms per 2‑second window, comfortably satisfying real‑time BCI constraints (≤5 ms).

Interpretability analyses reveal that the cross‑attention weights highlight phase‑specific cortical regions: frontal‑central areas dominate during initial stance, sensorimotor cortex during swing, and posterior parietal regions during terminal stance. These patterns align with established neurophysiological findings on gait control, confirming that the model leverages genuine neural dynamics rather than spurious correlations.

Limitations include the exclusive use of healthy‑subject data; the framework’s efficacy on clinical populations (e.g., stroke or spinal‑cord injury) remains to be validated. Moreover, the current unimodal design could benefit from multimodal integration (EMG, inertial measurement units) to improve robustness against EEG noise. Future work will extend NeuroDyGait to rehabilitation patients, incorporate additional sensor modalities, and explore continual‑learning strategies for online adaptation in long‑term BCI deployments.

In summary, NeuroDyGait combines relative contrastive, phase‑aware representation learning with a domain‑aware head‑mixing decoder, delivering a highly accurate, generalizable, and low‑latency solution for EEG‑based gait decoding, and setting a new benchmark for real‑time neuro‑prosthetic control.