Motor imagery classification performance enhancement with EEG data augmentation via wasserstein generative adversarial networks

Abstract Motor imagery electroencephalogram (MI-EEG) classification is often constrained by limited subject-specific training data. We investigated a data-augmentation framework combining continuous wavelet transform (CWT) scalograms, a Wasserstein generative adversarial network with gradient penalty (WGAN-GP), and four-class convolutional neural network (CNN) classification. For each subject, trials were divided before preprocessing into $$80\%$$ training data and a $$20\%$$ held-out evaluation partition. The WGAN-GP was trained exclusively on training-partition scalograms, and generated samples were used only to augment CNN training. The framework was compared with Gaussian noise, rotation, shifting, SMOTE, VAE, vanilla GAN, and DCGAN augmentation. WGAN-GP achieved a mean held-out evaluation accuracy of $$75.68\%$$ , compared with $$29.69\%$$ for the real-only CNN, and attained the highest performance among the evaluated methods. Qualitative inspection, feature-space Fréchet distance, and downstream classification provided complementary evidence of synthetic-data quality and utility. Alpha- and Beta-band activity showed the strongest individual band-limited discriminability, while the complete CWT representation outperformed every isolated band. Statistical analyses showed consistent gains across subjects. These findings demonstrate the potential of WGAN-GP augmentation for improving data-limited within-subject MI-EEG classification and provide a foundation for future domain-adaptive and subject-invariant decoding methods.