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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Decoding gait dynamics from electroencephalography (EEG) signals is challenging due to complex motor processes and limited high-quality datasets.
• Accurate temporal and spectral feature extraction is crucial for reliable gait decoding.

Purpose of the Study:

• To introduce EEG2GAIT, a novel hierarchical graph-based model for enhanced EEG-based gait decoding.
• To improve decoding performance through a Hybrid Temporal-Spectral Reward (HTSR) loss function.
• To contribute a new Gait-EEG Dataset (GED) for advancing research in this field.

Main Methods:

• Utilized a Hierarchical Graph Convolutional Network (GCN) Pyramid to capture multi-level spatial embeddings of EEG channels.
• Developed a Hybrid Temporal-Spectral Reward (HTSR) loss function integrating time-domain, frequency-domain, and reward-based components.
• Collected and synchronized a new Gait-EEG Dataset (GED) with lower-limb joint angle data from 50 participants.

Main Results:

• EEG2GAIT with HTSR achieved high performance on the GED dataset (r=0.959, R²=0.914, MAE=0.193).
• Outperformed existing methods on the MoBI dataset (r=0.779, R²=0.597, MAE=4.384).
• Ablation and saliency studies validated the model's components and highlighted motor-related brain region involvement.

Conclusions:

• EEG2GAIT with HTSR demonstrates superior performance in decoding gait dynamics from EEG signals.
• The model shows significant potential for advancing brain-computer interface applications, especially in lower-limb rehabilitation.
• The developed dataset and model provide valuable resources for future research in gait analysis and BCI.