Alleviating Class Imbalance Problem in Automatic Sleep Stage Classification

For real-world automatic sleep stage classification tasks, various existing deep learning based models are biased towards the majority with high proportion. Because of unique structure, most current polysomnography datasets suffer an inherent class imbalance problem (CIP), in which number each is severely unequal. In this study, we first define factor (CIF) to describe level CIP quantitatively. Afterwards, propose two balancing methods alleviate from dataset quantity and relationship between distribution applied model respectively. The one employ data augmentation (DA) generative adversarial network (GAN) different intensities Gaussian white noise balance samples, thereinto, addition specifically tailored models, can work on raw electroencephalogram (EEG) while preserving their properties. addition, try imbalanced achieve a balanced state help neuroscience principles. We further effective convolutional neural (CNN) utilizing bidirectional Long Short-Term Memory (Bi-LSTM) single-channel EEG as Baseline. It used for evaluating efficiency approaches three (CCSHS, Sleep-EDF Sleep-EDF-V1). qualitative quantitative evaluation experimental results demonstrates that proposed could not only show superiority through confusion matrix class-wise metrics, but also get better N1 whole stages accuracies compared other state-of-the-art approaches.