Automatic Classification of Sleep Stages Based on Raw Single-Channel EEG

Abstract

Electroencephalogram (EEG) is a common signal for monitoring people’s sleep quality. Manual sleep stage classification on EEG is a time-consuming task. In this paper, we design a model for automatic sleep stage classification based on raw single-channel EEG. This model can preserve the information, broaden the network and enlarge the receptive field as much as possible to extract appropriate time invariant features and classify sleep stage well. For the class-imbalanced problem in sleep stage classification, most of the exsisting methods rely on cross entropy loss and adjust model hyperparameters by experience, leading to poor performance. We implement a two-step training algorithm. The first is pre-training the model with the hyperparameters obtained by Bayesian Optimization after rebalancing datasets by over-sampling. The second is using feedback loss in model fine-tuning to reduce the impact of class-imbalanced problem. The loss weights dynamically change with the per-class F1-score which is used as feedback information. We evaluate our method on Fpz-Cz channel from the Sleep-EDF dataset. The overall accuracy, macro F1-score, Cohen’s Kappa coefficient are 85.53%, 81.18%, 0.80 respectively, showing our method has better classification performance than the state-of-the-art methods and is an efficient tool for automatic sleep stage classification.