Sleep EEG data augmentation model with Huber loss function

Currently, labeling sleep EEG data still relies on personal expertise, leading to rather insufficient labeling. Imbalance among the amount of EEG data in different sleep stages, meanwhile, could affect the accuracy of automatic assessment. We therefore propose a generative adversarial network-based data augmentation model to expand on EEG data in different sleep stages. Specifically, Huber function is introduced to assist loss function of auxiliary classifier generative adversarial network (ACGAN) in improving the quality of blurred data. The proposed model needs no specific feature extraction from EEG data. Further, generative and adversarial networks in the model are composed of 1D convolutional neural networks. One-dimensional noise and vectors representing different classes are used as input signals of the generative component. Two datasets MNIST and sleep stages are adopted to evaluate performance of the model. Other loss functions are compared with the proposed loss function to further test this model. Our experiments show that the proposed model achieves better accuracy in sleep state classification when compared to other models. It is our hope that this new model will inspire other workers to more efficiently introduce deep models into EEG-based sleep stage classification.