LPGGNet: Learning from Local-Partition-Global Graph Representations for Motor Imagery EEG Recognition

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Existing motor imagery electroencephalography (MI-EEG) decoding methods struggle with limited graph representations, underutilization of multi-scale information, and poor adaptability.
• Current approaches often rely on single connectivity graph representations, hindering comprehensive analysis of brain dynamics.

Purpose of the Study:

• To develop a novel Local-Partition-Global Graph learning Network (LPGGNet) to overcome the limitations of existing MI-EEG decoding techniques.
• To enhance MI-EEG decoding by integrating multi-view brain connectivities and dynamically constructed graph structures.

Main Methods:

• Proposed LPGGNet utilizes Partial Directed Coherence (PDC) for local functional adjacency matrices and temporal convolutions for feature extraction.
• Employed a partition learning module with Gaussian median distance and graph filtering for intra-partition feature consistency.
• Integrated a global learning module with dynamically computed adjacency matrices and residual graph convolutions for task-relevant representations.

Main Results:

• LPGGNet achieved high classification accuracies of 82.9% on the BCI Competition IV-2a dataset and 87.5% on a laboratory dataset.
• The proposed model outperformed several state-of-the-art MI-EEG decoding methods.
• Ablation studies confirmed the significant contribution of each module within the LPGGNet architecture.

Conclusions:

• Integrating multi-view brain connectivities with dynamically constructed graph structures significantly improves MI-EEG decoding performance.
• The LPGGNet model presents a novel, efficient, and superior solution for decoding electroencephalography (EEG) signals.
• This research advances the field of brain-computer interfaces through enhanced EEG signal analysis.