EMPT: a sparsity Transformer for EEG-based motor imagery recognition.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Transformer networks excel in signal processing but require substantial data and complex algorithms for electroencephalography (EEG) analysis.
• Self-attention mechanisms are effective for EEG feature coding but face limitations with data volume and algorithmic complexity.

Purpose of the Study:

• To develop an efficient deep learning model for decoding motor imagery (MI) EEG signals in spinal cord injury (SCI) patients.
• To address the data and complexity challenges of Transformer networks in EEG analysis.

Main Methods:

• A novel EEG MoE-Prob-Transformer (EMPT) model was developed, integrating Mixture of Experts (MoE) and ProbSparse Self-attention mechanisms.
• Time-frequency-spatial features were extracted using common spatial pattern and modified s-transform, serving as input embeddings for the EMPT model.
• The MoE layer introduced sparsity for enhanced feature reconstruction and analysis.

Main Results:

• The EMPT model achieved a high accuracy of 95.24% on a motor imagery EEG dataset from spinal cord injury patients.
• EMPT demonstrated superior performance compared to existing state-of-the-art methods in comparative experiments.

Conclusions:

• The integration of MoE layers and ProbSparse Self-attention enhances Transformer networks' applicability to EEG datasets by introducing sparsity.
• EMPT presents a novel and effective deep learning approach for decoding EEG data in motor imagery tasks, particularly for SCI patients.