Complex-valued unsupervised convolutional neural networks for sleep stage classification

https://doi.org/10.1016/j.cmpb.2018.07.015Get rights and content

Highlights

  • • This study developed a novel sleep stage system based on electroencephalogram.

  • • A new unsupervised complex-valued convolutional neural network was proposed.

  • • The proposed method can automatically extract features from raw EEG signals.

Abstract

Background and objective

Despite numerous deep learning methods being developed for automatic sleep stage classification, almost all the models need labeled data. However, obtaining labeled data is a subjective process. Therefore, the labels will be different between two experts. At the same time, obtaining labeled data also is a time-consuming task. Even an experienced expert requires hours to annotate the sleep stage patterns. More important, as the development of wearable sleep devices, it is very difficult to obtain labeled sleep data. Therefore, unsupervised training algorithm is very important for sleep stage classification. Hence, a new sleep stage classification method named complex-valued unsupervised convolutional neural networks (CUCNN) is proposed in this study.

Methods

The CUCNN operates with complex-valued inputs, outputs, and weights, and its training strategy is greedy layer-wise training. It is composed of three phases: phase encoder, unsupervised training and complex-valued classification. Phase encoder is used to translate real-valued inputs into complex numbers. In the unsupervised training phase, the complex-valued K-means is used to learn filters which will be used in the convolution.

Results

The classification performances of handcrafted features are compared with those of learned features via CUCNN. The total accuracy (TAC) and kappa coefficient of the sleep stage from UCD dataset are 87% and 0.8, respectively. Moreover, the comparison experiments indicate that the TACs of the CUCNN from UCD and MIT-BIH datasets outperform these of unsupervised convolutional neural networks (UCNN) by 12.9% and 13%, respectively. Additionally, the convergence of CUCNN is much faster than that of UCNN in most cases.

Conclusions

The proposed method is fully automated and can learn features in an unsupervised fashion. Results show that unsupervised training and automatic feature extraction on sleep data are possible, which are very important for home sleep monitoring.

Introduction

Sleep is fundamental physiological activity of the human body, and its quality can seriously affect the quality of life of patients. Sleep diseases, such as insomnias, hyper-somnias, sleep-related breathing disorders, cardiovascular diseases, parasomnias, sleep movement disorders, disturb sleep quality and thus threaten human health. Effective diagnosis and treatment of sleep-related diseases mainly rely on accurate classification of sleep stages [1]. Therefore, sleep stage classification is the most important step for sleep analysis.

Sleep stage is scored by many rhythms in the electroencephalogram (EEG), electromyogram (EMG), electrooculogram (EOG) and electrocardiogram (ECG). Because different sleep stages correspond to different brain rhythms, we use EEG to score sleep stage in this paper. Table 1 shows different EEG rhythms. Sleep is scored based on the Rechtschaffen and Kales (R&K) rules [2]. First, EEG is divided into non-overlapping 30-s epochs. Second, clinicians and experts classify each of these epochs into one sleep stage. According to the R&K rules [2], sleep is categorized into six categories. They are rapid eye movement (REM), stage 1 (S1), stage 2 (S2), stage 3 (S3), stage 4 (S4) and wake (W). Usually, S3 and S4 were combined as slow wave sleep (SWS) stage [3], [4], [5]. According to our investigation, an experienced clinician should take about 40 min to score a recording. Because there are many patients every night, the manual sleep scoring will be a time-consuming task. At the same time, the quality of the scoring depends on the experience and fatigue of the experts. Therefore, many methods for automatic sleep stage classification have been developed [6]. For example, the literature provides several examples involving different techniques: artificial neural network (ANN) [7], [8], [9], [10], support vector machine (SVM) [11], [12], [13], spectral analysis [14], [15], hidden markov models [16], [17], [18], nonlinear feature analysis [19], [20], [21], and other methods [22], [23], [24]. However, all these methods need handcrafted features. The effectiveness of these features is overly dependent on the knowledge and expertise, which makes these features not robust and flexible enough to adapt different circumstances.

Convolutional neural network (CNN) [25], which is able to fuse the two major blocks (feature extraction and classification) into a single learning process, will be helpful to solve the problem. CNN allows the extraction of automatic features from recorded EEG signals within its layers [26], [27], [28]. Because complex-valued neural network (CVNN) has better classification ability and more efficient classification performance than real-valued neural network (RVNN) [29], [30], various complex-valued convolutional models have been proposed [31], [32], [33], [34], [35], [36], [37], [38]. [31], [32], [33] used complex-valued convolutional neural network (CCNN) to solve complex-valued input problems. [34] used CCNN to solve real-valued image classification. Tygert et al. [35] gave a mathematical motivation for CCNN. However, few studies have applied CCNN to classify sleep stages. Our previous work [37] has proved that the decision boundaries for the real and imaginary parts of a complex-valued convolutional neuron intersect orthogonally. The orthogonal decision boundaries help to improve the generalization ability of the complex-valued convolutional neural network (CCNN). Our another work [38] combined complex-valued backpropagation and the Fisher criterion. It can learn discriminative features and overcome the negative effect of imbalance dataset. The two works [37], [38] have shown that the CCNN can obtain state-of-the-art classification performance based on deep learning models. However, supervised training algorithm was used in the works [37], [38]. It means lots of labeled data are needed. Through literatures analysis [37], [38], [39], [40], [41], [42], [43], [44], [45], [46], apart from [40], [41], all the deep learning methods need labeled data. There are two problems for obtaining labeled data by visual inspection: quality control and throughput. Obtaining labeled data depends on specialist domain knowledge and will differ between two specialists (inter-rater variance). At the same time, even an experienced specialist requires hours to annotate the sleep stage patterns. This makes it difficult to carry out quantitative and comprehensive sleep research. According to [40], [41], the EEG, EOG and EMG were used to classify sleep stage. It needs that the polysomnography (PSG) should be used to collect these signals. PSG measurements have many disadvantages. First, patients must wear many electrodes on different parts of their body. Because of these electrodes, patients are difficult to sleep and thus produce discrepant results which are not able to reflect the actual sleep condition. Second, the operation of PSG is very complicate. Patients can't monitor sleep at home. Because of the limitation of monitoring rooms, many patients can't be diagnosed in time. In this case, an affordable, portable, and unobtrusive sleep monitoring system for at-home use would be required. Wearable sleep devices will provide a good solution for long-term monitoring at home. Lots of EEG and ECG data can be collected from these devices. Compared with EEG, ECG is easy to obtain and interpret. Although many studies have used ECG to classify sleep stages [47], the best performance is still EEG-based method. Therefore, we choose EEG to classify sleep stages. Because obtaining labeled sleep data that is derived from wearable sleep devices is impossible, unsupervised learning is important for wearable sleep monitoring system. Hence, we develop a new unsupervised method to classify sleep stages based on the single channel EEG.

The present paper proposes a novel deep learning model called complex-valued unsupervised convolutional neural network (CUCNN) to classify sleep stages. To the best of our knowledge, this study is the first to apply a CUCNN model in sleep stage classification. The paper is organized as follows. The complex-valued K-means, CCNN and CUCNN are presented in Section 2. The experiments, results, and performance comparison are presented in Section 3. In Section 4, the experimental results are discussed. Finally, Section 5 concludes the paper.

Section snippets

Sleep stage method

In the paper, the sleep stage method resembles that of [38]. It includes three parts: 1) preprocessing, 2) feature extraction, and 3) classification. The difference is that CUCNN is used to extract features.

Data preprocessing

In order to cancel out power line disturbances of the EEG signal, a notch filtering at 50/60 Hz is used. After downsampling the EEG signal to 120 Hz, an eighth-order butterworth band-pass filter of 0.5–30 Hz for EEG is used to filter the downsampled signals [4]. According to [48], [49],

Datasets

In order to show the robustness of the proposed approach, two common datasets are used in this paper. The first dataset is obtained from the UCD database [65], which is available online at http://physionet.org/physiobank/database/ucddb. The dataset consisted of the full overnight PSG recordings of 25 subjects with sleep-disordered breathing. Each subject has a unique study number. Each study number corresponded to three data files which can be downloaded from the website. The details are shown

Discussion

This paper proposes a novel deep learning model named CUCNN. To the best of our knowledge, this study is the first to apply a complex-valued unsupervised model to automated sleep stage classification. The proposed method integrated complex-valued network and unsupervised training component. The TAC of the 4-layer CUCNN applied to the single-channel EEG from UCD dataset reached 87%, and the sensitivities of each stage reached 88%, 80.2%, 94.05%, 90.07%, and 85.02%, respectively.

Tables 3 and 4

Conclusions

This paper proposes a new unsupervised sleep stage method. It can learn useful features in an unsupervised fashion, and complex-valued classifiers can demonstrate excellent classification performance. The proposed approach is robust and fully automated, and the method can be easily adapted to other physiological signal analysis and prediction problems. Our approach, which learned features from raw sleep data, is a real automatic sleep stage method. Almost all the other methods [1], [5], [7],

Conflict of interest

None

References (83)

  • T. Nitta

    Solving the XOR problem and the detection of symmetry using a single complex-valued neuron

    Neural Netw.

    (2003)
  • A. Sors et al.

    A convolutional neural network for sleep stage scoring from raw single-channel EEG

    Biomed. Signal Process. Control

    (2018)
  • Ş. Yucelbaş et al.

    Automatic Sleep Staging based on SVD, VMD, HHT and Morphological Features of Single-Lead ECG Signal

    Expert Syst. Appl.

    (2018)
  • L. Zoubek et al.

    Feature selection for sleep/wake stages classification using data driven methods

    Biomed. Signal Process. Control

    (2007)
  • Hau-tieng Wu et al.

    Assess sleep stage by modern signal processing techniques

    IEEE Trans. Biomed. Eng.

    (2015)
  • A. Rechtschaffen et al.

    A Manual of Standardized terminology, Techniques and Scoring System For Sleep Stages of Human Subjects

    (1968)
  • S.F. Liang et al.

    Automatic stage scoring of single-channel sleep EEG by using multiscale entropy and autoregressive models

    IEEE Trans. Instrum. Meas.

    (2012)
  • H. Kassiri et al.

    Electronic sleep stage classifiers: a survey and VLSI design methodology

    IEEE Trans. Biomed. Circuits Syst.

    (2017)
  • O. Aydogan et al.

    Automatic classification of sleep stages with artificial neural networks according to visual scoring rules

  • V.P. Jain et al.

    Sleep stages classification using wavelet transform & neural network

  • M.E. Tagluk et al.

    Estimation of sleep stages by an artificial neural network employing EEG, EMG and EOG

    J. Med. Syst.

    (2010)
  • V. Bajaj et al.

    Automatic classification of sleep stages based on the time-frequency image of EEG signals

    Comput. Methods Programs Biomed.

    (2015)
  • S.K. Jha et al.

    Data fusion approach for human body odor discrimination using GC–MS spectra

  • L.G. Doroshenkov et al.

    Classification of human sleep stages based on EEG processing using hidden Markov models

    Biomed. Eng.

    (2007)
  • M. Yoshimura et al.

    Predicting sleep stages based on time-dependent hidden Markov model

    Neurocomputing

    (2006)
  • T. Orestis et al.

    Automatic sleep stage scoring using time-frequency analysis and stacked sparse autoencoders

    Ann. Biomed. Eng.

    (2016)
  • J. Yang et al.

    Sleep stage recognition using respiration signal

  • S. Devuyst et al.

    Automatic sleep spindles detection-overview and development of a standard proposal assessment method

  • Y. LeCun et al.

    Gradient-based learning applied to document recognition

    Proc. IEEE

    (1998)
  • S. Stober et al.

    Using convolutional neural networks to recognize rhythm stimuli from electroencephalography recordings

  • P.W. Mirowski et al.

    Comparing SVM and convolutional networks for epileptic seizure prediction from intracranial EEG

  • H. Cecotti et al.

    Convolutional neural networks for P300 detection with application to brain-computer interfaces

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2011)
  • R. Savitha et al.

    A fast learning complex-valued classifiers for real-valued classification problems

  • R. Haensch et al.

    Complex-valued convolutional neural networks for object detection in PolSAR data

  • Z. Zhang et al.

    Complex-valued convolutional neural network and its application in polarimetric SAR image classification

    IEEE Trans. Geosci. Remote Sens.

    (2017)
  • A. Marseet et al.

    Application of complex-valued convolutional neural network for next generation wireless networks

    Image Signal Process. Workshop

    (2018)
  • C.A. Popa

    Complex-valued convolutional neural networks for real-valued image classification

  • M. Tygert et al.

    A mathematical motivation for complex-valued convolutional networks

    Neural Comput.

    (2016)
  • J. Bruna et al.

    Invariant scattering convolution networks

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2013)
  • J.M. Zhang et al.

    Automatic sleep stage classification of single-channel EEG by using complex-valued convolutional neural network

    Biomed. Eng./Biomedizinische Technik

    (2018)
  • J.M. Zhang et al.

    A new method for automatic sleep stage classification

    IEEE Trans. Biomed. Circuits Syst.

    (2017)
  • Cited by (49)

    • A review of automated sleep stage scoring

      2023, Encyclopedia of Sleep and Circadian Rhythms: Volume 1-6, Second Edition
    • Sleep classification using Consumer Sleep Technologies and AI: A review of the current landscape

      2022, Sleep Medicine
      Citation Excerpt :

      It is evident that the challenge lies in differentiating N1 and N2. The difference in accuracy between the top 2 [52], 87% and [32], 85.9% Is only 2.9% but their difference in N1 is 60.05%. Likewise the second highest N1 accuracy is [44] with 69.45% but with a significant drop in all other sleep stages.

    • A novel complex-valued convolutional neural network for medical image denoising

      2021, Biomedical Signal Processing and Control
      Citation Excerpt :

      Therefore, the credit of superior performance in the case of CVMIDNet is attributed to the complex-valued operations. The inclusion of complex operations results in additional information in terms of phase, leading to easier optimization [25] and improved generalization ability [23]. Overall, this signifies the superiority of complex-valued CNNs over real-valued CNNs.

    View all citing articles on Scopus
    View full text