EEG2GAIT: A Hierarchical Graph Convolutional Network for EEG-based Gait Decoding

TL;DR

EEG2GAIT tackles EEG-based gait decoding by introducing a hierarchical graph convolutional network to capture multi-level spatial relationships among EEG channels and a Hybrid Temporal-Spectral Reward (HTSR) loss to jointly optimize time- and frequency-domain characteristics. A new Gait-EEG Dataset (GED) with 50 participants and two lab visits, plus validation on the MoBI dataset, demonstrate that EEG2GAIT outperforms state-of-the-art baselines in joint-angle prediction and exhibits robust, lower-variance performance. Ablation studies confirm the benefits of the Hierarchical GCN Pyramid and HTSR loss, while saliency maps provide neurophysiological validation by highlighting motor Cortical areas such as Cz, FC1/FC2, and Fz as key predictors. Overall, EEG2GAIT offers a principled, data-driven approach for EEG-based gait decoding with clear implications for brain-computer interfaces in neurorehabilitation and assistive technologies.

Abstract

Decoding gait dynamics from EEG signals presents significant challenges due to the complex spatial dependencies of motor processes, the need for accurate temporal and spectral feature extraction, and the scarcity of high-quality gait EEG datasets. To address these issues, we propose EEG2GAIT, a novel hierarchical graph-based model that captures multi-level spatial embeddings of EEG channels using a Hierarchical Graph Convolutional Network (GCN) Pyramid. To further improve decoding accuracy, we introduce a Hybrid Temporal-Spectral Reward (HTSR) loss function, which combines time-domain, frequency-domain, and reward-based loss components. Moreover, we contribute a new Gait-EEG Dataset (GED), consisting of synchronized EEG and lower-limb joint angle data collected from 50 participants over two lab visits. Validation experiments on both the GED and the publicly available Mobile Brain-body imaging (MoBI) dataset demonstrate that EEG2GAIT outperforms state-of-the-art methods and achieves the best joint angle prediction. Ablation studies validate the contributions of the hierarchical GCN modules and HTSR Loss, while saliency maps reveal the significance of motor-related brain regions in decoding tasks. These findings underscore EEG2GAIT's potential for advancing brain-computer interface applications, particularly in lower-limb rehabilitation and assistive technologies.

Figures (4)

• Figure 1: Overview of EEG2GAIT Architecture and Loss Calculation. The sequential EEG signals are segmented into training samples $X \in \mathbb{R}^{C \times T}$, where $C$ is the number of channels and $T$ is the time length. The Local Temporal Learner (LTL) uses 1D convolutions along the temporal dimension to capture local temporal information. The output is then reorganised into $n \times T$ temporal graphs, where $n$ is the number of samples. These graphs are input to the Hierarchical GCN Pyramid (HGP), which transforms each graph into a token embedding. The token for all graphs and the input for graph construction will be accumulated as input to the Global Spatial Learner (GSL). GSL performs 1D convolutions across all channels. The outputs are passed into the Feature Fusion Layers and then Global Temporal Learner (GTL) for global temporal information extraction via self-attention. The final output goes through the Output Layer. Losses from both the temporal (MSE) and frequency (FFT-based L1) domains are combined via a Reward Loss block to yield the Hybrid Temporal-Spectral Reward Loss.
• Figure 2: Experiment protocol for one block.
• Figure 3: Visualization of the actual and predicted left knee joint angles for the test set in participant 2's second session. The goniometer-measured angles were segmented into gait cycles using peak detection, with each gait cycle interpolated to 400 time points. The blue solid line represents the actual mean angle, and the blue shaded area indicates the standard deviation of the measured angles across all gait cycles. The red dashed line denotes the predicted mean angle, while the red shaded area shows the standard deviation of the predictions.
• Figure 4: Saliency maps generated by EEG2GAIT: (a) individual saliency maps for each subject, and (b) the average saliency map computed across all subjects.