Motor Imagery Decoding in the Presence of Distraction Using Graph Sequence Neural Networks

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Decoding motor imagery from electroencephalograms (EEGs) is crucial for brain-computer interfaces (BCIs).
• Distractions significantly challenge the accuracy of EEG signal decoding.
• Existing models struggle to capture complex spatio-temporal dynamics in noisy EEG data.

Purpose of the Study:

• To develop a novel Graph Sequence Neural Network (GSNN) for robust motor imagery decoding from EEGs.
• To enhance the model's ability to handle distractions by exploiting brain region topology.
• To improve the accuracy and adaptability of EEG-based BCIs.

Main Methods:

• Proposed a Graph Sequence Neural Network (GSNN) model.
• Modeled EEG channel similarities using graph adjacency matrices.
• Introduced a dynamic node domain attention selection network for adaptive feature extraction.
• Utilized the Berlin-distraction dataset for extensive experimentation.

Main Results:

• GSNN achieved superior classification accuracy compared to state-of-the-art models (average Recall: 80.44%, Precision: 81.07%, F-score: 80.54%).
• The node domain attention selection network demonstrated a crucial role in enhancing model sensibility and adaptability.
• Experiments confirmed the model's effectiveness in decoding motor imagery under distracting conditions.
• Analysis revealed insights into the influence of brain regions and channels on distraction themes.

Conclusions:

• The proposed GSNN model offers a significant advancement in decoding motor imagery from EEGs, particularly in the presence of distractions.
• The dynamic attention mechanism enhances the model's robustness and adaptability to individual EEG signals.
• This research provides a promising direction for developing more reliable and accurate brain-computer interfaces.