Constrained Generative Model for EEG Signals Generation

Te Guo, Lufan Zhang, Rui Ding, Da Zhang, Jinhong Ding, Ming Ma, Likun Xia

Abstract: Electroencephalogram (EEG) is one of the most promising modalities in the field of Brain-Computer Interfaces (BCIs) due to its high time-domain resolutions and abundant physiological information. Quality of EEG signal analysis depends on the number of human subjects. However, due to lengthy preparation time and experiments, it is difficult to obtain sufficient human subjects for experiments. One of possible approaches is to employ generative model for EEG signal generation. Unfortunately, existing generative frameworks face issues of insufficient diversity and poor similarity, which may result in low quality of generative EEG. To address the issues above, we propose R $$^{2}$$ WaveGAN, a WaveGAN based model with constraints using two correlated regularizers. In details, inspired by WaveGAN that can process time-series signals, we adopt it to fit EEG dataset and then integrate the spectral regularizer and anti-collapse regularizer to minimize the issues of insufficient diversity and poor similarity, respectively, so as to improve generalization of R $$^{2}$$ WaveGAN. The proposed model is evaluated on one publicly available dataset - Bi2015a. An ablation study is performed to validate the effectiveness of both regularizers. Compared to the state-of-the-art models, R $$^{2}$$ WaveGAN can provide better results in terms of evaluation metrics.