Unsupervised multi-subepoch feature learning and hierarchical classification for EEG-based sleep staging

Highlights

• An unsupervised feature learning method is proposed to extract EEG features.

• A hierarchical classification model is established for EEG-based sleep staging.

• A novel feature evaluation criterion is presented for feature subset selecting.

• Extensive experiments are conducted to evaluate the proposed method.

Abstract

As the medium of developing brain–computer interface system, the recognition of EEG signals is complicated and difficult due to the complex nonstationary characteristics and the individual difference between subjects. In this paper, we investigate the EEG signal classification problem and propose a novel unsupervised multi-subepoch feature learning and hierarchical classification method for automatic sleep staging. First, we divide the EEG epoch into multiple signal subepochs, and each subepoch is mapped to amplitude axis and time axis respectively to obtain two kinds of feature information with amplitude–time dynamic characteristics. Then, the statistical classification features are extracted from the mapped feature information. Furthermore, we conduct unsupervised feature learning for consistent and specific classification features from the perspective of time series. Finally, according to the differences and similarities of EEG signals in different sleep stages, a hierarchical weighted support vector machine-based classification model (H-WSVM) is established, which can use different feature subsets at each classification level and different weighting parameters for unbalanced data samples. To select the optimal feature subset for detecting each sleep stage, we propose a novel evaluation criterion for feature classification ability based on rough set theory. Experimental results on the most commonly used dataset show that the proposed method has better sleep staging performance and can effectively promote the development and application of EEG sleep staging system.

Introduction

EEG is an electrical signal generated by synchronous activity of a large number of neurons and has become an important monitoring method to record electrical activities of the brain, which can also reflect the human mental and health states (Casson, 2019, Gaxiola-Tirado et al., 2017, Lim et al., 2018). Since the EEG signals of various brain states can provide abundant data for research on human computer interaction, the use of EEG in brain–computer interface system has attracted widespread attentions, but it is still difficult to develop a generalized pattern recognition framework. Sleep, a natural behavior of human life, is an indispensable part of the body’s work and rest. It not only has an impact on the physical and physiological recovery, but also plays an important role in people’s cognition, emotion and health (Adnane et al., 2012, Chriskos et al., 2019, Motamedi-Fakhr et al., 2014). Moreover, Sleep quality directly affects people’s learning, thinking and other behaviors. Despite the growing awareness of the importance of sleep, there are still many people experiencing inadequate sleep and related diseases (Kang et al., 2017, Liang et al., 2012). Sleep staging is the classification of various sleep stages, including six stages: W, S1, S2, S3, S4 and REM, and each sleep stage is closely related to different brain behaviors. Through monitoring and analyzing sleep EEG signals, we can study the brain function of subjects with different cognitive ability, estimate fatigue state in continuous driving (Bose et al., 2019, Li et al., 2016), and conduct the diagnosis, treatment and prevention of sleep diseases (Boostani et al., 2017, Liu and Sun, 2017, Peker, 2016, Zhu et al., 2014). In this work, we are motivated to study the EEG-based sleep staging, by developing a novel classification method to recognize the EEG signals of different sleep stages.

With the rapid development of brain–computer interface (BCI) and artificial intelligence, the use of EEG signals for sleep staging has become a feasible and effective sleep monitoring technology (Ghimatgar, Kazemi, Helfroush, & Aarabi, 2019). Although EEG, ECG, EMG, EOG and other physiological signals can be used for sleep staging (Yan et al., 2019), the sleep EEG is the most reliable and effective way, it contains much information related to age, gender and mental states. In particular, the use of single-channel EEG is more suitable for wearable system development and application (Maddirala and Shaik, 2017, da Silveira et al., 2017). Considering that traditional EEG sleep staging requires the sleep specialist to manually classify the signal stages, it is not conducive for real-time wearable sleep monitoring (Ronzhina et al., 2012). Through research and development of automatic sleep staging system, real-time monitoring and analysis of sleep stages can effectively reduce the workload of medical staff and promote the development of the smart health industry (Hassan & Bhuiyan, 2017). However, due to the complex nonstationary characteristics of EEG signals and the signal similarity between special sleep stages, the key to EEG-based sleep staging is that how to obtain the optimal signal features. Moreover, the design of classification model should consider the existing class imbalance problem between different sleep stages.

To obtain effective classification features from sleep EEG signals, many feature extraction methods have been studied , and the commonly used features can be generally categorized into four kinds: (1) time domain features, (2) frequency domain features, (3) time–frequency domain features and (4) complex nonlinear features. In the study of Liu, Sun, Zhang, and Rolfe (2016), a feature extraction method named multi-domain analysis was proposed, including multifractal detrended fluctuation analysis, natural visibility graph, frequency analysis, etc. Şen, Peker, Çavuşoğlu, and Çelebi (2014) extracted more complex features of sleep EEG signals, such as time domain features, frequency domain features, time–frequency domain features, nonlinear features, and entropy features. Generally, the time-domain statistical features of EEG signals have the advantages of simplicity, intuition and easy calculation, so they are suitable for the development of wearable sleep staging system. However, it is difficult to obtain the frequency-domain and complex nonlinear variation information of EEG only using the time-domain features. Thus, it is necessary to decompose EEG into some signal sub-bands with different frequency ranges, and extract the time-domain statistical features from each signal sub-band. Memar and Faradji (2018) extracted 13 signal features from each of 8 signal sub-bands with different frequency range, and a total of 104 classification feature were obtained for sleep staging. Moreover, authors in the recent study of Liu et al. (2021) used ensemble empirical mode decomposition (EEMD) to decompose the sleep EEG epochs, and then extracted various features using statistical, time-domain and nonlinear dynamics characteristics from the original EEG signal and the decomposed intrinsic mode functions. When the same feature is extracted from each signal sub-band, a large number of redundant features will be obtained, so it needs to select the optimal classification features. Tabar, Mikkelsen, Rank, Hemmsen, and Kidmose (2021) investigated the low dimensional feature spaces for automatic sleep staging based on a bootstrapping approach guided by Gini ranking and mutual information between the features, and attempted to represent sleep EEG patterns using a minimum number of features, without significant performance loss. Furthermore, to overcome the dependence of feature extraction on professional knowledge, many studies conduct automatic feature learning for sleep EEG signals by using deep learning methods (Loh et al., 2020, Zhang and Wu, 2017), and the methods based on deep learning use the 1-D CNN to extract signals features for sleep staging (Fernandez-Blanco et al., 2019, Sokolovsky et al., 2019), which can discover more new features of different sleep stages. In Mousavi, Rezaii, Sheykhivand, Farzamnia, and Razavi (2019), a network architecture includes 9 convolutional layers followed by 2 fully connected layers was proposed to extract features from raw EEG signals, this automatic recognition method used single-channel EEG signals to classify 2–6 class sleep stages. Authors in Supratak, Dong, Wu, and Guo (2017) proposed a deep learning model for automatic sleep staging from raw single-channel EEG, the time-invariant features were extracted by convolutional neural networks, and bidirectional-long short-term memory was used to learn transition rules among sleep stages automatically. Zhu, Luo, and Yu (2020) proposed a multi-branch convolutional neural network model with embedded stage refinement and a residual attention channel fusion method for multi-channel sleep staging. In the study of Khalili and Asl (2021), a neural network architecture is used to extract features, followed by Temporal Convolutional Neural Network to extract the temporal features from the extracted feature vector of CNN. Although the features learned by CNN network obtain a good result for EEG-based sleep staging in many studies, the direct use of CNN that designed for obtaining the spatial characteristics of images may not be able to detect the complex variation characteristics of EEG signals. In addition, deep learning methods require much more computation resources compared with traditional methods, and it is very hard to explain that what the automatically learned features represent and why they are effective in differentiating sleep stages from the view of physiology (Jiang, Lu, Yu, & Yuanyuan, 2019).

According to the used feature extraction methods for EEG-based sleep staging, most existing studies extract classification features from the whole epoch of sleep EEG signals, and the features obtained from various signal sub-epochs may have different sleep staging performance. In the study of Diykh and Li (2016), authors used segmentation technique to divide the EEG signal epoch of 30S into 75 sub-epochs, and then extracted 12 statistical features from each sub-epoch. This study obtained better classification results of 6-class sleep staging, while taking all the features as a whole for model learning is lack of the ability to learn the temporal information. In the study of Seo et al. (2020), authors used CNN network to extract features from signal sub-bands, and BiLSTM was used to learn the temporal context of the representative features. The learned features from consecutive signal sub-bands in this study can represent the temporal characteristic of EEG signal, while the important sleep-related events may only appear in some special sub-bands, it still needs to consider the feature learning of different brain activities from the EEG signal sub-bands. In view of the existing problem of feature extraction for EEG-based sleep staging, we propose a novel feature extraction approach to learn the time series variation features from a new perspective of automatic unsupervised multi-subepoch feature learning, which can conduct the consistent and specific feature learning of sleep-related brain activities by fusing the conventional handcrafted feature extraction and unsupervised learning for multi-subepoch EEG signals. For instance, K-complex is a major component of EEG signal in REM sleep stage, which consists of two waves: a large negative wave and a small positive wave, lasting more than 0.5s, and with an average of about 0.63 s, and thus, we can obtain the effective classification features from the specific signal subepoch. As discussed above, the key to EEG-based sleep staging lies in effectively utilizing classification features that can reflect the variation characteristics of sleep EEG signals. Compared with the existing studies, our work is to fulfill the purpose of learning the consistent and specific time series features from multi-subepoch EEG signals.

In this paper, we classify the sleep stages by using a hierarchical weighted support vector machine-based classification model (H-WSVM) for the learned consistent and specific time series features. Through the multi-subepoch partition of an sleep EEG epoch, the time series features can be obtained from the consecutive signal subepochs. Furthermore, according to the ability of the change of frequency and amplitude for characterizing the power and oscillations of EEG signals in various sleep stages, we exploit an amplitude–time mapping method to extract the time-domain statistical features from each signal subepoch. Then, we conduct unsupervised feature learning for the obtained classification features from all the consecutive signal subepochs while properly handling the redundancy, and the learned consistent and specific features are selected for the classification model based on H-WSVM. Moreover, the design of H-WSVM model considers the similarity and distribution of sleep stages, in that S1 and REM sleep stages have the relatively low detection rate because of the small number of samples and the indistinguishability between EEG signals. Compared to WSVM, the proposed H-WSVM model can select different feature subsets for various classification tasks, and each sleep stage can be gradually detected according to the signal similarity of EEG signals in different sleep stages. Thus, the H-WSVM model can overcome the influence of detecting different sleep stage on each other and can improve the overall sleep staging performance by optimizing the class imbalance problem. To better illustrate the work of this paper, we present an EEG-based sleep staging system for wearable device based on our study, the systems structure is shown in Fig. 1.

The contributions of our work are summarized as follows:

(1) To the best of our knowledge, we are the first to exploit unsupervised multi-subepoch feature learning method to extract the classification features for EEG-based sleep staging, and the consistent and specific variation features of sleep EEG signals are obtained for classifying various sleep stages.

(2) We propose a statistical feature extraction method for each signal subepoch, named amplitude–time mapping, which can obtain better classification features related to EEG variation characteristic without any signal filtering, transformation technology.

(3) We propose a novel feature evaluation criterion based on rough set theory to select the optimal feature subsets for different sleep stage classification tasks, which helps to design real-time application system.

(4) A hierarchical classification model of H-WSVM is presented to classify various sleep stages gradually, which can use different feature subsets at each classification level, and optimize the classification results according to class imbalance.

The rest of this paper is organized as follows: Section 2 presents the experiment data used in this study. The following is Section 3 which describes the unsupervised multi-subepoch feature learning method. In Section 4, we present the H-WSVM classification model in detail. The experiments and results are discussed in Section 5. Finally, we conclude this paper in Section 6.