A Dual-Discriminator GAN for Sleep EEG Signal Synthesis

Abstract: The interpretation of one’s overnight sleep process based on EEG signal is of importance for 
the inspection and treatment of various sleep-related disorders. The automatic sleep staging models have 
benefits as the assistant computerized tools to release the clinicians from the laborious task of manual 
sleep stage scoring. However, the issues of data insufficient and class imbalance are common for clinical 
data. The data problem is crucial to be solved when applying the automatic sleep staging models for real 
clinics. In this research, the network architecture of GAN (Generative Adversarial Network) is 
investigated by using one generator and two discriminators for the synthesis task of sleep EEG signals. 
The data augmentation performance by the dual-discriminator GAN and the sleep stage classification 
performance by combining with a typical machine learning classifier are evaluated on the sleep recording 
of subjects. The obtained results showed that the constructed dual-discriminator GAN is effective to 
generate samples which are closer to the time and frequency characteristics of real sleep EEG signal. It 
would be a useful method to solve the data problems for the training and optimization of automatic sleep 
stage classifiers in the field of sleep staging.
Keywords: Generative adversarial network; Electroencephalograph; Data augmentation; Sleep staging; 
Dual-discriminator