Multi-scale self-attention approach for analysing motor imagery signals in brain-computer interfaces

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Advancements in electroencephalogram (EEG) brain-computer interface (BCI) technology are significant.
• Deep learning models show promise but face challenges in extracting features from EEG data for motor imagery (MI).
• Accurate EEG data classification remains a fundamental challenge in BCI research.

Purpose of the Study:

• To develop a novel deep learning model for accurate four-class motor imagery EEG signal classification.
• To leverage attention mechanisms and multi-scale spatiotemporal networks for enhanced feature extraction.
• To improve the performance of BCIs in recognizing distinct motor intentions.

Main Methods:

• A model utilizing multi-scale spatiotemporal self-attention networks for EEG signal processing.
• Spatial self-attention to prioritize relevant brain channels and reduce noise.
• Parallel multi-scale Temporal Convolutional Network (TCN) layers for temporal feature extraction.

Main Results:

• Achieved 85.09% accuracy on the BCI Competition IV-2b dataset.
• Reached 96.26% accuracy on the HGD datasets (IV-2a and IV-2b).
• Demonstrated superior accuracy in single-subject classification compared to existing methods.

Conclusions:

• The proposed model exhibits high performance and resilience in motor imagery classification.
• The approach shows significant potential for transfer learning in BCI applications.
• This method offers a promising advancement for next-generation neurotechnologies.