Multi-view graph fusion of self-weighted EEG feature representations for speech imagery decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Electroencephalogram (EEG)-based speech imagery is an emerging brain-computer interface (BCI) for communication.
• Current decoding performance in speech imagery research is limited due to a lack of consensus on discriminative features.
• Identifying optimal EEG features is crucial for advancing BCI technology.

Purpose of the Study:

• To propose a novel model for adaptively capturing complementary information from different domain features in EEG.
• To enhance the decoding performance of speech imagery by integrating multi-view EEG features.
• To identify critical EEG channels and frequency bands for improved speech imagery decoding.

Main Methods:

• A multi-view graph fusion of self-weighted EEG feature representations (MVGSF) model was developed.
• The model learns a consensus graph from multi-view EEG features for effective intention decoding.
• A view-dependent feature importance exploration strategy was incorporated to address varying discriminative abilities of features.

Main Results:

• MVGSF demonstrated outstanding performance on two public speech imagery datasets.
• The learned consensus graph effectively characterized relationships among EEG samples.
• Feature importance analysis identified critical EEG channels and frequency bands for speech imagery decoding.

Conclusions:

• MVGSF effectively integrates multi-domain features to enhance EEG-based speech imagery decoding capabilities.
• The model contributes to identifying spatial-frequency patterns in EEG for speech imagery.
• This approach advances the potential for natural communication using brain-computer interfaces.