Automatic Sleep Staging for EEG Signals Using GoogLeNet Deep Architecture and Continuous Wavelet Transform

Majority of automatic sleep stage scoring studies rely on the use of feature engineering and classical machine learning algorithms. Recently, several automatic sleep stage classification methods have been proposed based on convolutional neural networks (CNN). However, the state of the art CNN methods are still complex which usually require significant time and considerable computational resources in order to set-up and sufficiently train a deep CNN from scratch. More importantly, many of the previous deep learning studies were trained in limited scenarios which might limit their ability to generalize in different settings. This study eliminates the need to establish and train a deep CNN from scratch by leveraging a pre-trained deep architecture that has been previously trained from sufficient labeled data in a different context. The proposed approach uses a GoogLeNet deep architecture to set-up and train an end-to-end CNN model for automatic feature extraction and sleep scoring using data from a single EEG source. In order to obtain imaging data inputs needed for the CNN, the continuous wavelet transform (CWT) is used to obtain time-frequency images from EEG sleep epochs. The proposed model was evaluated over the three primary EEG signals recommended by the American Academy of Sleep Medicine (AASM) from data sets that were collected by different devices that are used by two major sleep centers in USA and Germany, respectively. Evaluation results show that this method can achieve a promising performance and can be further deployed in real-time for online sleep monitoring and wearable systems.