Toward calibration-free motor imagery brain-computer interfaces: a VGG-based convolutional neural network and WGAN...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) is a key Brain-Computer Interface (BCI) paradigm using electroencephalogram (EEG) signals.
• Inter-subject variability in EEG necessitates subject-dependent data for MI BCI training, hindering widespread adoption.
• Current MI BCIs face calibration challenges due to the need for personalized training data.

Purpose of the Study:

• To enhance cross-subject (CS) MI EEG classification performance.
• To develop a calibration-free MI BCI approach suitable for real-world applications.
• To improve the accuracy and robustness of MI BCIs by overcoming inter-subject variability.

Main Methods:

• Utilized EEG spectrum images for MI classification.
• Employed deep learning techniques, specifically a modified VGG-CNN classifier.
• Implemented Wasserstein Generative Adversarial Networks (WGAN) for synthetic data augmentation to expand the training dataset.
• Conducted experiments on benchmark datasets (BCI competition IV-2B, IV-2 A, IV-1) using leave-one-subject-out validation.

Main Results:

• The proposed approach demonstrated enhanced CS MI EEG classification accuracy.
• The combination of WGAN-generated data and the modified VGG-CNN classifier outperformed state-of-the-art methods.
• Achieved significant improvements in cross-subject classification without requiring target subject calibration data.

Conclusions:

• The developed framework represents a significant advancement towards calibration-free BCI systems.
• This approach has the potential to broaden the applications of MI BCIs by simplifying their implementation.
• The study highlights the efficacy of deep learning and data augmentation for robust cross-subject MI classification.