EEG-based speech imagery decoding by dynamic hypergraph learning within projected and selected feature subspaces

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) are advancing speech decoding using electroencephalogram (EEG) data.
• Current graph-based methods may not fully capture complex correlations within EEG samples.
• A more effective data structure is needed to model high-order relationships in EEG for speech imagery.

Purpose of the Study:

• To introduce hypergraphs for modeling high-order correlations in EEG data for speech imagery.
• To propose two dynamic hypergraph learning models: DHSLP and DHSLF.
• To enhance the accuracy of decoding imagined speech intentions via BCIs.

Main Methods:

• Utilized hypergraphs to represent high-order correlations between EEG samples, with feature vectors as vertices and hyperedges connecting them.
• Dynamically updated hyperedge weights, vertex weights, and hypergraph structure in projected and feature-weighted subspaces.
• Developed and applied dynamic hypergraph semi-supervised learning within projected subspace (DHSLP) and selected feature subspace (DHSLF) for speech imagery decoding.

Main Results:

• Both DHSLP and DHSLF models demonstrated statistically significant improvements in decoding imagined speech intentions compared to existing methods.
• DHSLP achieved accuracies of 78.40% and 66.64% on two independent EEG datasets.
• DHSLF achieved accuracies of 71.07% and 63.94% on the same datasets, showing competitive performance.

Conclusions:

• Learned hypergraphs effectively characterize semantic information in imagined speech content.
• The study provides interpretable insights into discriminative EEG channels for speech imagery decoding.
• This research lays the groundwork for exploring physiological mechanisms underlying speech imagery.