Wasserstein generative adversarial network with gradient penalty and convolutional neural network based motor imagery...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Acquiring high-quality motor imagery electroencephalography (MI-EEG) data is challenging, leading to insufficient training datasets.
• This data scarcity often results in overfitting and poor generalization in deep learning classification models.
• Improved decoding performance is crucial for brain-computer interfaces (BCIs) utilizing MI-EEG.

Purpose of the Study:

• To propose a novel data augmentation method and deep learning classification model to enhance MI-EEG decoding performance.
• To address the challenges of insufficient data quality and quantity in MI-EEG datasets.
• To improve the accuracy and generalization capabilities of deep learning models for MI-EEG classification.

Main Methods:

• Raw EEG signals were converted into time-frequency maps using continuous wavelet transform for model input.
• An improved Wasserstein generative adversarial network with gradient penalty was employed for data augmentation, expanding the training dataset.
• A concise and efficient deep learning model was designed to further boost decoding performance.

Main Results:

• The proposed generative network successfully produced more realistic MI-EEG data.
• Classification accuracies of 83.4%, 89.1%, and 73.3% were achieved on benchmark datasets (BCI Competition IV 2a, 2b) and a real-world dataset, respectively.
• Corresponding Kappa values were 0.779, 0.782, and 0.644, outperforming state-of-the-art methods.

Conclusions:

• The proposed method effectively enhances MI-EEG data, mitigating overfitting in classification networks.
• The approach significantly improves MI classification accuracy, demonstrating positive implications for MI-based BCIs.
• The study highlights the potential of advanced data augmentation and deep learning for robust MI-EEG signal processing.