Multi-Scale convolutional neural networks integrated with self-attention for motor imagery EEG decoding.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) are advancing rapidly, with motor imagery (MI) electroencephalography (EEG) classification being crucial for applications like assistive technology.
• Traditional convolutional neural networks (CNNs) struggle with long-term dependencies in EEG data, potentially limiting decoding performance.

Purpose of the Study:

• To propose a novel deep learning network that enhances MI-EEG classification by effectively capturing temporal information and global dependencies.
• To overcome the limitations of CNNs in extracting long-range temporal features for improved EEG decoding.

Main Methods:

• A multi-scale convolutional neural network combined with an attention mechanism was developed.
• The network incorporates a multi-scale structure for spatial-temporal and multimodal feature extraction (mean and variance).
• A squeeze-excite-compress module and a multi-head attention encoder were utilized to refine feature extraction and highlight critical information.

Main Results:

• The proposed method achieved high classification accuracies of 85.26% on the BCI Competition IV-2a dataset and 95.86% on the High Gamma Dataset (HGD).
• The network demonstrated state-of-the-art performance in MI-EEG decoding tasks.
• Data augmentation using signal reorganization improved the network's generalization ability.

Conclusions:

• The novel deep learning network effectively captures temporal information and global dependencies for superior MI-EEG classification.
• The proposed method shows significant potential as a new baseline for general EEG decoding, offering improved accuracy and generalization.