Enhanced electroencephalogram signal classification: A hybrid convolutional neural network with attention-based...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery electroencephalogram (MI-EEG) signal classification is vital for brain-computer interfaces (BCI).
• Traditional machine learning methods struggle with MI-EEG's nonlinearity, low signal-to-noise ratio, and individual variability.
• Deep learning offers a promising avenue for overcoming these classification challenges.

Purpose of the Study:

• To propose an automatic deep learning-based feature extraction method for enhanced MI-EEG classification.
• To address the limitations of traditional methods in handling complex MI-EEG signal characteristics.
• To improve the accuracy and robustness of BCI systems reliant on MI-EEG signals.

Main Methods:

• Noise reduction using discrete wavelet transform and common average reference.
• Convolutional Neural Network (CNN) for extracting time-domain features and spatial information.
• Talking-heads attention mechanism to enhance critical feature sequences.
• Temporal Convolutional Network (TCN) for abstracting spatial-temporal features, followed by a fully connected layer for classification.

Main Results:

• The proposed enhanced EEG model achieved an average classification accuracy of 85.53% across subjects on the BCI Competition IV-2a dataset.
• Demonstrated significant improvements in classification accuracy compared to existing models: CNN (+11.24%), EEGNet (+6.90%), CNN-LSTM (+11.18%), and EEG-TCNet (+6.13%).

Conclusions:

• The developed deep learning model effectively extracts spatial-temporal features from MI-EEG signals.
• The proposed method offers a substantial advancement in MI-EEG classification accuracy for BCI applications.
• This approach holds significant potential for enhancing intention recognition in BCI systems.