Deep Identity Confusion for Automatic Sleep Staging Based on Single-Channel EEG

Sleep Staging (SS) is a vital step in sleep neurobiology. Though many previous approaches have been proposed to solve it, most of them suffer from poor generalization for unknown identity. In this paper, we proposed a deep identity confusion method to extract powerful task-specific and identity-invariant feature and then score sleep stages with non-linear machine learning model. With an unified CNN-LSTM structure employed for feature extraction, we implement identity confusion with an extra identity prediction branch and apply inversed gradients to frontal layers during back-propagation. Then the deep feature is used to train a XGBoost classifier. Experiments on Sleep-EDF benchmarks achieve classification accuracy and macro F1 score of 84.1% and 78.9%, and it suggests proposed method boost performance of origin deep learning base model and show competitive result comparing to state-of-the-art methods.