A Filter-based GAN to Enhance Learning Frequency Information of EEG for Sleep Staging

In recent years, automatic sleep staging models based on artificial intelligence have been developed to provide physicians with more efficient diagnostic assistance support. However, the performance of automatic sleep staging models is limited by several problems such as insufficient available training data or imbalance among the classes. A common method is to manually synthesis the required data using generative adversarial networks(GANs). As an extremely informative signal in the frequency domain, existing electroencephalogram(EEG) synthesis methods pay little attention to this aspect. In this study, to enhance the learning of frequency domain information, a filter-based GAN is proposed for EEG signals synthesis. Rather than full-band learning, filtering allows the GAN to focus the learning on the frequency components containing the dominant rhythmic characteristics under different sleep stages. The results of several comparison experiments show that the filter-based GAN can synthesize higher quality EEG signals for sleep staging, with significantly improved scores on the three metrics compared to the baseline model.