A Graph-Temporal Fused Dual-Input Convolutional Neural Network for Detecting Sleep Stages from EEG Signals

Abstract

Sleep is an essential integrant in everyone's daily life. Thereby, it is an important but challenging problem to construct a reliable and stable system, that can monitor user's sleep quality automatically. In this brief, we combine complex network and deep learning to propose a novel Graph-Temporal fused dual-input Convolutional Neural Network (CNN) method to detect sleep stages by using the Sleep-EDF database. Firstly, we segment each single-channel EEG signal into non-overlapping 30s epochs to set up the network. For that, we map each epoch into a Limited Penetrable Visibility Graph (LPVG) and obtain the corresponding Degree Sequence (DS) by calculating the node degree. Finally, the DSs and the 30s EEG epochs are combined as inputs of the novel Graph-Temporal fused dual-input CNN to learn about the graph topology and about the temporal feature representations of the raw data for the purpose of classifying the sleep stages into the two-, three-, four-, five-, and six-state. Notably, the classification accuracy of six-state stage detection is 87.21% and the corresponding Kappa value is 0.80. The results demonstrate the effectiveness of our model structure in detecting sleep states, whereby they further provide a basic strategy for future sleep research.