Attention-Based DSC-ConvLSTM for Multiclass Motor Imagery Classification.

Area of Science:

• Neuroscience and Artificial Intelligence
• Brain-Computer Interface (BCI) Technology
• Deep Learning for Signal Processing

Background:

• Deep learning shows promise in motor imagery brain-computer interfaces (MI-BCI), outperforming traditional methods.
• Electroencephalogram (EEG) recognition rates remain a critical bottleneck for MI-BCI advancement.
• Limited availability of accurately labeled EEG data hinders the training of complex deep learning models.

Purpose of the Study:

• To enhance EEG classification accuracy for multi-class motor imagery signals within MI-BCI.
• To address the challenge of small sample sizes in labeled EEG datasets through data augmentation.
• To develop a robust model that minimizes the influence of individual specificity on classification performance.

Main Methods:

• Proposed a DSC-ConvLSTM model incorporating an attention mechanism for motor imagery EEG signal classification.
• Utilized sliding windows for data augmentation to increase training sample size and reduce global feature reliance.
• Employed depth separable convolution (DSC) for spatial feature extraction and a novel bidirectional ConvLSTM for time-domain feature extraction, integrating CNN and LSTM capabilities.

Main Results:

• Achieved an average classification accuracy of 73.7% on the BCI Competition IV Dataset 2a.
• Reached an average classification accuracy of 92.6% on the High Gamma Dataset.
• Demonstrated robust and effective performance across different subjects and datasets, mitigating individual variability.

Conclusions:

• The proposed DSC-ConvLSTM model effectively extracts significant EEG features, improving classification accuracy in MI-BCI.
• The model's performance highlights its robustness and effectiveness, showing potential for practical applications.
• This research contributes to advancing brain-computer interface technology towards marketability and widespread adoption.