OCRNN: An orthogonal constrained recurrent neural network for sleep analysis based on EEG data

摘要

This paper introduced an end-to-end mixed deep learning model for automatic sleep analysis based on the EEG signal. Unlike some existing machine learning models for EEG analysis, we did not rely on any hand-crafted feature engineering and elaborate pipeline design. Furthermore, apart from some existing deep learning frameworks based on some off-the-self existing modules, we introduced the orthogonal constrained recurrent neural network (OCRNN) as the downstream module after the spatial-temporal expansion and representation provided by the one dimensional convolutional neural networks. We evaluated our model using the EEG-based sleep datasets for sleep stage scoring. We compared the performances of four types of RNN frameworks, where three of them are OCRNNs. The results show that OCRNN can achieve competitive better F1 score, accuracy and AUC score compared to the previous baseline results. Moreover, our model (orRNN and pdRNN) can achieve the above results with less number of parameters and less number of training epochs, which demonstrate its potential usage to launch approximate real-time medical diagnosis. (C) 2020 Elsevier B.V. All rights reserved.