A dual-stream deep neural network integrated with adaptive boosting for sleep staging

Abstract

Humans spend about one-third of their time in sleep. Sleep is closely related to human physical and mental health and is a very important life activity. Automatic sleep stage classification is an important tool for analyzing sleep quality. However, due to different difficulties such as poor signal quality, the minor difference between different stages, and class imbalance problem, developing sleep staging systems with high performance is still a hot topic in current sleep-related research. In this study, we propose a novel adaptive-boosting-based dual-stream neural network to classify sleep stages using different modalities of single-channel EEG signals. The proposed network is designed based on two different forms of the single-channel EEG signal, that is, the original raw one-dimensional EEG signal and the two-dimensional time–frequency signal obtained by Continuous Wavelet Transform (CWT). The network for one-dimensional signal consists of a double-branch convolution neural network in order to extract low-frequency and high-frequency features separately. The backbone network for the two-dimensional signal is the modified ResNet50 network structure, which is enhanced by a multi-scale attention mechanism for learning more discriminate features. Furthermore, adaptive boosting algorithm is utilized to alleviate the class-imbalance problem that existed in many sleep-related datasets. Specifically, multiple weak classifiers are trained, and then a boosting-based ensemble strategy is used to obtain a strong classifier for sleep stage classification. In our experiments, two public datasets are utilized to evaluate the performance of the proposed network. Compared with other state-of-the-art methods, the method proposed in this study obtained competitive performance in terms of metrics such as Accuracy (Acc) and Cohen Kappa (κ), which are 85.8%, and 0.80 respectively on the Sleep-EDF-20 dataset and 81.0%, 0.73 respectively on Sleep-EDF-78 dataset.