Combination of EEG Data Time and Frequency Representations in Deep Networks for Sleep Stage Classification

Almost all of the studies in the literature of sleep stage classification are based on traditional statistical learning techniques from a set of extracted features, which need a relative amount of time and effort. Deep learning offers approaches able to automatically extract patterns and abstractions from different types of data (images, sound, biomedical signals, etc.) to perform classification. However, the application of these techniques in the automatic sleep stage scoring field is less widespread to date. This paper proposes a new approach based on a multi-state deep learning neural network architecture, which we named Asymmetrical Multi-State Neural Network. This new network is able to merge two different neural networks, based on two different architectures receiving different input data: single-channel EEG raw signal in time and the respective spectrum. The proposed Asymmetrical Multi-State Neural Network shows to enhance the separated networks' performance for the given problem on a complete wellknown sleep database.