A Multi-Band Self-Attention Network for Motor Imagery Classification.

Area of Science:

• Neuroscience and Biomedical Engineering
• Signal Processing and Machine Learning

Background:

• Brain-computer interfaces (BCIs) enable machine control via human thought, with motor imagery (MI) electroencephalogram (EEG) signals showing promise for applications like stroke rehabilitation and assistive device control.
• Current limitations in BCI technology stem from the decoding performance and generalization ability of MI signals, hindering widespread practical application.

Purpose of the Study:

• To introduce and evaluate a novel multi-branch self-attention network designed for enhanced classification of motor imagery (MI) electroencephalogram (EEG) signals.
• To investigate the impact of different signal decomposition techniques (EEMD, WPD, brain rhythm-based) on feature representation and classification accuracy.
• To validate the proposed network's superiority over existing benchmark models in terms of classification accuracy and generalization capability.

Main Methods:

• A multi-branch self-attention network architecture was developed, processing EEG signals decomposed into distinct frequency bands.
• Each branch utilized convolutional neural networks (CNNs) and multi-head self-attention (MHA) for spatial-temporal feature extraction, complemented by long short-term memory (LSTM) networks for temporal dependencies.
• The approach was systematically evaluated using three signal decomposition methods on the BCI Competition IV 2a dataset.

Main Results:

• The proposed multi-branch self-attention network achieved state-of-the-art performance on the BCI Competition IV 2a dataset.
• Subject-dependent accuracy reached 84.04%, and subject-independent accuracy was 71.67%.
• Comparative analysis confirmed the network's superior classification accuracy and generalization capability compared to established models like EEGNet and ShallowConvNet.

Conclusions:

• The developed multi-branch self-attention network effectively decodes motor imagery EEG signals, demonstrating significant improvements in classification accuracy and generalization.
• The study highlights the positive contribution of each network module and signal decomposition strategy to overall performance.
• Enhanced accuracy in decoding MI EEG signals holds substantial clinical relevance for advancing applications in prosthetics control, wheelchair navigation, and stroke rehabilitation.