Attention-based convolutional neural network with multi-modal temporal information fusion for motor imagery EEG...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) are common for decoding electroencephalography (EEG) signals in motor imagery (MI)-based brain-computer interfaces (BCIs).
• However, CNNs have limited perceptual fields, hindering their ability to capture long-term dependencies and multi-modal temporal information crucial for accurate EEG decoding.

Purpose of the Study:

• To propose a novel deep learning network that integrates CNNs with a self-attention mechanism.
• To effectively capture multi-modal temporal information and global dependencies in EEG signals for improved MI-BCI performance.

Main Methods:

• A novel deep learning network combining CNNs with a self-attention mechanism was developed.
• The network extracts multi-modal temporal information (average and variance) and uses a shared self-attention module to capture global dependencies.
• A convolutional encoder fuses these features, and a signal segmentation and recombination method enhances generalization.

Main Results:

• The proposed method achieved a 4-class average accuracy of 85.03% on the BCI Competition IV-2a dataset.
• Experimental results on BCI Competition IV-2a and IV-2b datasets demonstrated superior performance compared to state-of-the-art methods.
• The approach highlights the effectiveness of multi-modal temporal information fusion in attention-based deep learning networks.

Conclusions:

• The novel network effectively decodes MI-EEG signals by integrating multi-modal temporal information and global dependencies.
• This work offers a new perspective for MI-EEG decoding, emphasizing the benefits of attention-based deep learning architectures.
• The proposed method shows significant potential for advancing brain-computer interface technology.