关键词: 脑电信号, 睡眠分期, 卷积神经网络, 双向门控循环单元, 合成少数类过采样技术

Abstract: Aiming at the efficiency and accuracy of single-channel EEG signal sleep automatic staging, this paper proposes to use three-scale parallel Convolutional Neural Networks to extract sleep signal features and Bidirectional Gated Recurrent Unit 3CNN-BiGRU automatic sleep staging model to learn the internal time relationship between sleep stages. First, the model performs band-pass filtering on the original single-channel EEG signal, and uses the synthetic minority oversampling technique for class balance, and then sends it to the built model for training and verification experiments. Pre-training and fine-tuning training  are used for optimizing the model, and  10-folds and 20-folds cross-validation is uses to improve training reliability. The experimental results of different models under different data sets show that the 3CNN-BiGRU model has achieved better training efficiency and better staging accuracy.

Key words: EEG signal, sleep staging, convolutional neural network, bidirectional gated recurrent unit, synthetic minority oversampling technique