ParkJiyong1,†
An Jaewon2,†
ChoiSang Ho1,2*
-
(School of Computer and Information Engineering, Kwangwoon University / Seoul, Korea
{bjybs123, shchoi}@kw.ac.kr
)
-
(Department of Computer Engineering, Kwangwoon University / Seoul, Korea jaewonan95@gmail.com
)
Copyright © The Institute of Electronics and Information Engineers(IEIE)
Keywords
Sleep stage classification, Deep learning, Biosignals, Non-contact sensing
1. Introduction
Sleep is an essential time for physical and mental health recovery, and people spend
more than one-third of their lives sleeping. Sleep quality plays a crucial role in
learning, fatigue recovery, and daily life. Furthermore, insufficient sleep can decrease
concentration during the day, contribute to daytime sleepiness, and increase the risk
of cardiovascular disease [1,2] and obesity [3]. Furthermore, prolonged sleep deprivation causes insomnia, which interferes with
work, family, and healthy life. The paradigm of global healthcare is changing from
diagnosis and treatment to prevention, prediction, and personalization. Sleep must
be monitored accurately to prevent sleep disorders and provide customized solutions.
Polysomnography (PSG) is a systematic process for assessing sleep. PSG measures the
following biosignals during the night: electroencephalogram (EEG), electrooculogram
(EOG), electrocardiogram (ECG), electromyogram (EMG), respiration, SpO2, and other
signals. The sleep stages are classified based on these signals. They are divided
into five distinct stages based on the American Academy of Sleep Medicine (AASM) [4]: wake, non-rapid eye movement (NREM stages 1-3), and rapid eye movement (REM). The
sleep stages are scored in 30-second units based on the characteristics of the biosignals.
Sleep scoring is a time-consuming and labor-intensive process. To solve these limitations,
an automatic sleep scoring studies have been published. Automatic sleep classification
research is divided into studies that classify sleep stages into three, four, or five
stages. When the sleep stages are divided into three, they are classified as wake,
NREM, and REM. In the case of four stages, they are classified as wake, light sleep
(NREM1+NREM2), deep sleep (NREM3), and REM sleep. With five stages, the stages are
classified as wake, NREM1-3, and REM sleep. The classification of sleep stages three,
four, or five in the papers to be reviewed later means the above division.
Many studies have been conducted to classify sleep stages using conventional machine
learning algorithms. For example, Adnane et al. [5] classified the sleep-wake stages using a support vector machine (SVM) with single-lead
ECG signals; the mean classification accuracy was 79.31%. Fraiwan et al. [6] proposed a random forest-based sleep scoring method with a single EEG signal with
an accuracy of 83%.
Structural shallowness is a challenging issue in machine learning approaches. Pei
et al. [7] reported that machine-learning algorithms could not extract deeply hidden properties
from biomedical signals due to the shallow single-layer structure. Furthermore, conventional
machine learning methods require feature engineering, which is a process of converting
given data into features that can represent issues in predictive models to improve
the model performance. Feature engineering is essential in machine learning and significantly
affects model performance. For that, domain knowledge and expertise is needed. Deep
learning methods are becoming more popular because they generate features automatically
in multiple layers. Deep neural networks have recently been used to classify sleep
stages by extracting properties from biomedical signals.
As deep learning research has accelerated, deep learning approaches for sleep stage
classification studies have been published. This review paper examines the methods
of deep learning model architectures. Furthermore, sleep stage classification research
using EEG, ECG, and multi-fusion biosignals based on deep learning were reviewed.
In addition, studies using non-contact radar, which are presented as alternatives
for long-term sleep monitoring at home, were reviewed.
2. Deep Learning Architectures
Many deep learning models have been proposed for sleep research. This chapter introduces
the characteristics of representative deep learning models. Convolutional neural networks
(CNN) are deep learning architectures that learn directly from data without manual
feature extraction. Simard et al. [8] proposed a simplified CNN architecture that extracted features by converting them
to coarse resolutions through convolution with high-resolution filters of different
sizes. Furthermore, CNN is used widely in image and video classification [9]. In addition, they can be used to classify time series, audio signals, and nonimage
data. A CNN consists of dozens or hundreds of layers that extract the features of
the input. A filter was applied to the input at each layer, and the output was used
as an input at next layer. Krizhevsky et al. [9] trained high-resolution images on a large-scale CNN to classify them into 1000 different
classes from the ImageNet LSVRC-2010 contest. The authors reported a low margin error
of 15.3%. Zhao et al. [10] classified ten different time-series data groups using CNN and obtained better results
than previous methods from accuracy and noise tolerance. This model is used in sleep
stage classification research based on the characteristics of CNNs that automatically
learn features from 1D or 2D signals.
Recurrent neural networks (RNN) are deep learning architectures that take variable-length
sequential or time-series data. Furthermore, RNNs are used widely in image captioning
and natural language processing (NLP) [11,12]. The major difference between RNN and other neural networks, such as CNN, is that
RNN extracts information from previous input data that affect the current input and
output. On the other hand, RNNs have vanishing gradient problems that interfere with
the learning of long data sequences. RNN parameters are updated by gradients that
carry information, and as gradients become smaller, parameter updates become insignificant,
so there is no real learning.
A long short-term memory (LSTM) network has been developed as an RNN model to solve
the vanishing gradient issue [13]. The key to LSTM is cell state. The cell state determines which information should
be retained or neglected and which information affects the next time step. Furthermore,
LSTM can remove or add information to a cell state regulated by a structure called
a gate that moves along the entire chain through linear interaction. Biosignals are
affected by the autonomic nervous system and change over a short time or change slowly
over a certain period. Based on the AASM manual [4], there is a rule for determining the sleep stage of the current epoch (30s) by considering
the characteristics of the previous epoch signals. In this view, the LSTM is an appropriate
deep learning model to extract information from previous signals and learn meaningful
features.
3. Literature Review
3.1 Sleep Stage Classification using EEG Signal
EEG biosignals play a key role in sleep stages classification based on the AASM manual
[4]. The spectral characteristics of the EEG biosignals are used for the main features
to determine the sleep stages. The main characteristics of EEG at each stage are as
follows. In the wake stage, the alpha rhythm of the 8-13 Hz appears, and in the NREM1
stage, low-amplitude, mixed EEG activity (LAMF) of 4-7 Hz appears. The K complex and
sleep spindles, which are distinct sinusoidal waves, are seen in the NREM2. The slow
wave activity of 0.5-2 Hz and LAMF activity appears during NREM3 and REM stage, respectively.
Because of the above EEG characteristics, various sleep stage classification studies
have been performed using the EEG signals only.
Vilamala et al. [14] classified five-class sleep stages using a pretrained visual geometry group deep
convolutional network (VGGNet). The spectral image was produced from the EEG signal
and then input to the deep learning model. Classification accuracy was 86%. Simonyan
et al. [15] proposed that VGGNet improved performance in large-scale image recognition with more
layers. First, VGGNet was trained as a feature extractor (VGG-FE). Second, the author
updated all the weights in the network and obtained a fine-tuned network (VGG-FT).
The EEG signals were fed into both the VGG-FE and VGG-FT. The overall classification
accuracies of both models were 83% and 86%, respectively. Ebrahimi et al. [16] suggested an artificial neural network (ANN) model for four-class sleep stages classification.
They used the time-frequency domain of single-channel EEG signals with a relatively
high accuracy of 93.7${\pm}$3.7%. Michielli et al. [17] proposed two cascaded LSTM models. The first LSTM classified four sleep stages, in
which NREM1 and REM are merged, and the second LSTM classifies NREM1 and REM sleep
stages. An overall accuracy of 86.7% was achieved for the five-class sleep stages
classification. Yildirim et al. [18] adopted a 1-D CNN using PSG signals for the five-class sleep stage classification,
and an overall accuracy of 90.82% was obtained.
Table 1. Sleep Stage Classification Studies using EEG signals.
|
Study
|
Architecture
|
# of Sleep Stages
|
N
(PSG recordings)
|
Input Feature
|
Performance
|
|
[16]
|
ANN
|
4
|
7
|
wavelet packet transformed EEG signal
|
Acc = 93.7 ± 3.70%
|
|
[23]
|
CNN
|
4
|
18
|
raw EEG signal
|
Acc = 71.00 ± 2.71%
|
|
[14]
|
VGG-FE
|
5
|
20
|
Fourier Transformed image
|
Acc = 83%
|
|
VGG-FT
|
Acc = 86%
|
|
[17]
|
Two cascaded LSTM
|
5
|
12
|
power spectral density, energy spectral density,
amplitude spectral density
|
Acc = 86.70%
|
|
[18]
|
1-D CNN
|
5
|
8
|
Raw EEG signal
|
Acc = 90.82%
|
|
[19]
|
CNN and BiRNN
|
5
|
61
|
Raw EEG signal
|
Acc = 84.26%
κ = 0.79
|
|
[20]
|
Two CNN and bi-directional LSTM
|
5
|
62
|
Raw EEG signal
|
Acc = 86.20%
κ = 0.80
|
|
[21]
|
CNN and uni-directional LSTM
|
5
|
62
|
Raw EEG signal
|
Acc = 87.50%
κ = 0.82
|
|
[22]
|
MLP
|
5
|
154
|
Power Spectral Density
|
Acc = 89.00%
|
${\kappa}$ : cohen’s kappa coefficient
In [19,20], the authors proposed a novel model using a two-step training algorithm that extracted
features from raw EEG signals based on neural networks. Mousavi et al. [19] proposed a sequence-to-sequence model with an attention mechanism. The model used
a decoder-encoder architecture, where the decoder collects the information for the
label, and the encoder captures the dependencies between the inputs and labels. The
model classified five-class sleep stages using a CNN as a decoder and bidirectional
RNN as an encoder with an overall accuracy of 84.26%. Supratak et al. [20] proposed an architecture consisting of two CNNs with different filter sizes and bidirectional
LSTM, where the CNNs extract the time-invariant features and LSTM encodes temporal
information. The architecture classified five classes of sleep stages with an accuracy
of 86.2%. In addition, Supratak et al. [21] a deep learning model was proposed using a CNN and unidirectional LSTM to classify
five-class sleep stages. Better performance was achieved with an accuracy of 87.5%
compared to previous studies [19,20] using the same dataset.
Recently, Hussain et al. developed a multi-layer perceptron (MLP) model to classify
five-class sleep stages using three-channel EEG signals, and an overall accuracy of
89% was achieved [22]. Zhao et al. [23] trained three LSTM models using EEG, ECG, and both EEG and ECG data to classify four-class
sleep stages. The model using EEG signals classified four-class sleep stages with
an accuracy of 71.88 ${\pm}$ 2.71%.
Most studies have observed that this approach has the limitations of a single biosignal.
In particular, EEG signals cannot capture the patterns related to EOG and EMG, which
are significant for classifying the NREM and REM sleep stages. In addition, the NREM1
stage cannot be easily classified because the proportion of its stages is very low
and inter-rater variability for NREM1 scoring is higher than the other sleep stages
[24].
3.2 Sleep Stage Classification using ECG Signal
An analysis of the ECG data during sleep provides significant information on sleep
stage regulation [25]. Autonomic nervous system (ANS) is closely related to the sleep structure, and the
sympathetic activity increases during wake and REM sleep, and the parasympathetic
activity increases as the NREM stage processes [26,27]. The activity of the ANS affects the regulation of the cardiovascular system, which
can be used in the sleep stage classification models.
Urtnasan et al. [28] proposed a deep convolutional recurrent (DCR) model consisting of three CNN and two
RNN layers. A single-lead ECG signal was used with no vital intermediate features
extracted. The model classified the five- and three-class sleep stages with overall
accuracies of 74.2% and 86.4%, respectively. Radha et al. [29] trained an RNN model to classify four-class sleep stages with ECG data. The model
classified sleep stages with an accuracy of 71.88 ${\pm}$ 8.34%. The classification
was improved using photoplethysmography (PPG) signals by transferring and freezing
a subset of the ECG model weights into the PPG model with an accuracy of 76.36 ${\pm}$
7.57%. In addition, the LSTM model using ECG signals from [23] was used to classify the four-class sleep stages with an accuracy of 68.00 ${\pm}$
2.54%.
Radha et al. [30] extracted the heart rate variability (HRV) features from inter-beat intervals computed
from ECG signals. The author classified the four-class sleep stages using multilevel
bidirectional LSTM with an accuracy of 77.0 ${\pm}$ 8.90%. Li et al. [31] extracted HRV, and ECG-derived respiration (EDR) features from ECG signals. Four-
and three-class sleep stages were classified with CNN with accuracies of 75.4% and
81.6%, respectively. Sridhar et al. [32] proposed a CNN approach using instantaneous heart rate (IHR) derived from ECG to
classify four sleep stages with an overall accuracy of 77%. Wei et al. [33] classified five-class sleep stages using LSTM with single-lead ECG signals. The average
accuracy of the proposed method was 71.16%. Tang et al. [34] suggested a four-class sleep stage classification model with a CNN and gated recurrent
unit (GRU) using an ECG signal. The model exhibited the highest accuracy of 80.6%.
Table 2. Sleep Stage Classification Studies using ECG biosignal.
|
Study
|
Architecture
|
# of Sleep Stages
|
N
(PSG recordings)
|
Input Feature
|
Performance
|
|
[29]
|
RNN
|
4
|
584
|
HRV parameters
|
Acc = 71.88 ± 8.34%
κ = 0.57 ± 0.12
|
|
[23]
|
CNN
|
4
|
18
|
raw ECG signal
|
Acc = 68 ± 2.54%
|
|
[30]
|
bi-directional LSTM
|
4
|
584
|
HRV parameters
|
Acc = 77 ± 8.9%
κ = 0.61 ± 0.15
|
|
[31]
|
CNN
|
3
|
18
|
EDR and HRV parameters
|
Acc = 81.6%
κ = 0.63
|
|
4
|
Acc = 75.4%
κ = 0.54
|
|
[32]
|
CNN
|
4
|
584
|
IHR time series
|
Acc = 77%
κ = 0.66
|
|
[34]
|
CNN and GRU
|
4
|
2,056
|
raw ECG signal
|
Acc = 80.6%
κ = 0.70
|
|
[28]
|
CNN and RNN
|
3
|
112
|
raw ECG signal
|
Acc = 86.4%
|
|
5
|
Acc = 74.2%
|
|
[33]
|
LSTM
|
5
|
373
|
power spectral density
|
Acc = 71.16%
κ = 0.52
|
${\kappa}$ : cohen’s kappa coefficient
Table 3. Sleep Stage Classification Studies using multi-biosignals.
|
Study
|
Architecture
|
# of Sleep Stages
|
N
(PSG recordings)
|
Input Feature
|
Performance
|
|
[35]
|
DNN
|
2
|
18
|
Dispersion entropy
|
Acc = 73.70%
|
|
[23]
|
CNN
|
4
|
18
|
raw EEG signal and
raw ECG signal
|
Acc = 80.4 ± 1.03%
|
|
[36]
|
DBN
|
5
|
25
|
raw EEG signal, and
raw EMG signal
|
Acc = 72.2 ± 9.7%
|
Most studies using the ECG signal classify sleep stages three and four, and the performance
is lower than that of EEG studies results (Tables 1 and 2). The NREM sleep stages were scored based on transient events or frequency
band characteristics in EEG signals. It is difficult to identify these characteristics
and distinguish NREM1 and NREM2 using only the ECG signal, resulting in lower classification
performance than EEG research. In addition, the changes in the HRV parameter values
due to movement and external noise also affect the classification performance. Minimizing
such noise and extracting appropriate features related to the change characteristics
of the autonomic nervous system (ANS) could improve the classification performance.
3.3 Sleep Stage Classification using Multi-biosignals
Most studies were sleep stage classification studies using one biosignal to overcome
PSG limitations because PSG requires attaching many sensors to acquire biosignals.
Therefore, only a few studies used multi-biosignals.
The classification of sleep stages using different biosignal combinations has been
studied to improve the sleep stage classification. Tripathy and Rajendra Acharya [35] classified sleep stages between REM and NREM with DNN using the features derived
from the HRV signals extracted from the ECG and EEG signals. They reported an overall
accuracy of 73.70%. Längkvist [36] proposed a five-class sleep stage classification with feature-extracted deep belief
nets (DBN) using EEG, electrooculogram (EOG), and electromyography (EMG) signals.
The overall accuracy was 72.2 ${\pm}$ 9.7%. Furthermore, the LSTM model using the
EEG and ECG signals from [23] classified four-class sleep stages with an 80.4 ${\pm}$ 1.03% accuracy.
Table 4. Sleep Stage Classification Studies using Non-contact Signals.
|
Study
|
Radar
|
Architecture
|
# of Sleep Stages
|
N
(PSG recordings)
|
Input Feature
|
Performance
|
|
[38]
|
CW-Doppler radar
|
ANN
|
4
|
-
|
1D signal
|
Acc = 76.5%
|
|
[41]
|
IR-UWB radar
|
TCN + LSTM
|
4
|
71
|
1D signal
|
Acc = 76 ± 0.07%
|
|
[43]
|
Wi-Fi
|
CNN +
Bi-LSTM
|
4
|
19
|
Channel State Information signal
|
Acc = 81.8%
|
|
[40]
|
FMCW
|
GBM + LDA
|
4
|
32
|
1D signal
|
Acc = 63.5 ± 0.08% κ = 0.49 ± 0.12
|
|
[44]
|
RF
|
CNN + LSTM
|
4
|
100
|
RF Spectrogram
|
Acc = 79.8%
κ = 0.70
|
|
[39]
|
CW-Doppler radar
|
k-NN
|
3
|
13
|
1D signal
|
Acc = 85.8%
|
|
4
|
Acc = 81%
|
|
[42]
|
IR-UWB radar
|
ATT-LSTM
|
4
|
51
|
1D signal
|
Acc = 82.6 ± 6.7% κ = 0.73 ± 0.11
|
${\kappa}$ : cohen’s kappa coefficient; RF : radio frequency;
3.4 Sleep Stage Classification using Non-contact Signal
Unconstrained biosignal measurements are required to monitor sleep stages in a home
environment for a long time. This chapter reviews non-contact methods for sleep stage
classification. A representative method for non-contact biosignal monitoring is radar,
which transmits a signal from a transmitter to the human body and receives a reflected
signal from a receiver. Radar is appropriate for non-contact biosignal measurements
and does not require direct contact with the human body to measure breathing and heart
rates [37]. Radar-based sleep stage classification studies have been conducted using these characteristics.
There are a few sleep stage classification studies based on deep learning. Therefore,
conventional machine learning methods and deep learning-based studies will be reviewed.
The different kinds of radars include continuous wave (CW), frequency-modulated continuous
wave (FMCW), and impulse-radio ultra-wideband (IR-UWB). CW radar has advantages, such
as its simplicity and potential minimal spread in the transmitted spectrum to detect
heart rate, respiration rate, and body movement. The FMCW radar is characterized by
the linear change in the frequency of the output signal in the system over time. In
impulse-based radar, pulses are transmitted to a person using a radar transmitter.
A receiver captures the reflected echo and processes it in the time domain to obtain
human vital signs. A typical impulse-based radar system is an IR-UWB radar system.
Zhang et al. [38] classified four sleep stages using bagged trees, a machine learning algorithm utilizing
CW Doppler radar. CW Doppler radar was used to extract six features, which were respiratory
and heartbeat signals. After selecting m samples randomly as training sets, the sleep
stages were classified with a bagged tree algorithm that was repeated n times. Compared
with the PSG results, an accuracy of 78.6% was achieved when tested with a 10-fold
cross-validation. Hong et al. [39] extracted data from 13 healthy people using a CW Doppler radar. It was characterized
by KNN, a non-parametric method that uses classification and regression by extracting
BPM and RPM features from radar signals. At this time, the K-nearest neighbor algorithm
was used, and the results were 81% when the sleep stage was classified into four stages
and 85.8% when classified into three. Tataraidze et al. [40] extracted four stages of sleep using frequency-modulated CW radar-based GBM+LDA.
For sleep stage classification, the gradient-boosted machine (GBM) predicted the epoch
class probabilities in the first step and linear discriminant analysis (LDA), a probability
between the target epoch and six neighborhood epochs in the second step. This result
was verified using the data from 32 people without sleep disturbance breathing, and
the accuracy was 63.5${\pm}$ 0.08% and Kappa = 0.49 ${\pm}$ 0.12.
Toften et al. [41] used IR-UWB radar, a model combining temporal convolutional networks (TCN) and LSTM,
to derive four sleep stages using Somnofy, a sleep algorithm. The model showed high
accuracy in healthy individuals with an epoch-specific accuracy of 0.76 ${\pm}$ 0.07,
a sensitivity of 0.97 ${\pm}$ 0.03, and a specificity of 0.72 ${\pm}$ 0.19. Kwon et
al. [42] proposed an attention Bi-LSTM model that could remotely monitor vital signs using
an IR-UWB radar. In this study, the LSTM and Attention layers were combined to allow
the model to focus on specific portions of the time-varying vital sign information
and assign heavier weights. They classified four sleep stages to derive the results
of Acc = 82.6 ${\pm}$ 6.7% and kappa = 0.73 ${\pm}$ 0.11, which performed much better
than conventional LSTMs.
Sleep stage studies have been conducted using other non-contact and radar-based methods.
Zhao et al. [40] analyzed radiofrequency (RF) signals to extract sleep-specific features and applied
the CNN+LSTM model to capture the temporary progression of sleep. This model converted
the RF signal into an RF spectrogram and featured an input value for 30 seconds. Sleep
was classified into four stages with an accuracy of 79.8% and kappa = 0.70. Bohan
et al. [43] used Wi-Fi to propose WiFi-sleep, which combined CNN+BiLSTM with duration and transition
constraints. The proposed CNN + BiLSTM model used a multiscale CNN. The model can
simultaneously capture features with long and short durations, thereby obtaining extended-duration
features and minute features around the present moment. Using Wi-Fi sleep, 81.8% accuracy
was achieved compared to PSG data with 19 sleep data points from 12 subjects.
Recently, with the rapid development of IoT technology, studies on non-contact methods
for monitoring sleep in a home environment have been reviewed. The main issue when
applying non-contact radar signals is noise removal. Compared to the biosignals measured
based on a contact sensor, the noise of the external environment has a larger effect
on the non-contact signal. Non-contact sleep monitoring techniques could be improved
if a deep learning method is applied to an algorithm that removes noise and extracts
the user's location and biosignals.
4. Conclusion
This study reviewed automatic sleep stage classification studies using biosignals
and non-contact methods. Deep learning methods were examined in sleep monitoring research
to overcome the limitations of existing feature engineering-based machine learning
methods. In addition, to replace the time-consuming and restrictive PSG method, studies
based on single bio- and fusion multi-signals and studies on non-contact methods have
been reviewed.
A large number of biosignals accumulated in the sleep research field overnight, and
deep learning research using these data will be accelerated. Deep learning models
are known as black box models, and it is difficult to interpret the results. Recently,
however, the development of explainable AI is increasing the understanding of the
model results. Applying this method to sleep research will provide additional sleep
information. In addition, research on lightweight AI models that can be used on-devices
is needed for the long-term monitoring of sleep in the home environment. Deep learning
approaches to sleep research is expected to expand to the dimensions of managing health
and preventing disease.
ACKNOWLEDGMENTS
This work was supported by the National Research Foundation of Korea(NRF) grant
funded by the Korea government(MSIT) (No. NRF-2021R1F1A1052460). This research was
supported by the MSIT(Ministry of Science and ICT), Korea, under the ITRC(Information
Technology Research Center) support program(IITP-2022-RS-2022-00156225) supervised
by the IITP(Institute for Information & Communications Technology Planning & Evaluation).
This research was also supported by the MIST(Ministry of Science and ICT), under the
National Program for Excellence in SW (2017-0-00096), supervised by the IITP(Institute
of Information & Communications Technology Planning & Evaluation). The authors would
like to thank Dr. Seok-heon Cho, a member of the Qualcomm Institute, University of
California, San Diego, for the kind guidance.
REFERENCES
W. M. A. van Leeuwen et al., “Sleep restriction increases the risk of developing cardiovascular
diseases by augmenting proinflammatory responses through IL-17 and CRP,” PLoS One,
Feb. 2009.
B. Faraut, K. Z. Boudjeltia, L. Vanhamme, and M. Kerkhofs, “Immune, inflammatory and
cardiovascular consequences of sleep restriction and recovery,” Sleep Med Rev, Apr.
2011.
J. E. Gangwisch, D. Malaspina, B. Boden-Albala, and S. B. Heymsfield, “Inadequate
sleep as a risk factor for obesity: analyses of the NHANES I,” Sleep, Oct. 1992.
R. B. Berry et al., “The AASM manual for the scoring of sleep and associated events:
rules, terminology and technical specifications, Version 2.0,” American Academy of
Sleep Medicine, 2012.
M. Adnane, Z. Jiang, and Z. Yan, “Sleep-wake stages classification and sleep efficiency
estimation using single-lead electrocardiogram,” Expert Systems with Applications,
Aug. 2011.
L. Fraiwan, K. Lweesy, N. Khasawneh, H. Wenz, and H. Dickhaus, “Automated sleep stage
identification system based on time-frequency analysis of a single EEG channel and
random forest classifier,” Computer methods and programs in biomedicine, Sep. 2009.
W. Pei, Y. Li, S. Siuly, and P. Wen, “A hybrid deep learning scheme for multi-channel
sleep stage classification,” Computers, Materials and Continua, Feb. 2022.
P. Y. Simard, D. Steinkraus, and J. C. Platt, “Best practices for convolutional neural
networks applied to visual document analysis,” Icdar, Aug. 2003.
A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep
convolutional neural networks,” Commun ACM, Jun. 2017.
B. Zhao, H. Lu, S. Chen, J. Liu, and D. Wu, “Convolutional neural networks for time
series classification,” Journal of Systems Engineering and Electronics, Feb. 2017.
W. Yin, K. Kann, M. Yu, and H. Schütze, “Comparative Study of CNN and RNN for Natural
Language Processing,” 2017.
O. Vinyals, A. Toshev, S. Bengio, and D. Erhan, “Show and tell: A neural image caption
generator,” Proceedings of the IEEE Computer Society Conference on Computer Vision
and Pattern Recognition, June. 2015.
S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural computation, Nov.
1997.
A. Vilamala, K. H. Madsen, and L. K. Hansen, “Deep convolutional neural networks for
interpretable analysis of EEG sleep stage scoring,” Computers, Materials and Continua,
Sep. 2021.
K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image
Recognition,” arXiv preprint arXiv:1409.1556, Sep. 2014.
F. Ebrahimi, M. Mikaeili, E. Estrada, and H. Nazeran, “Automatic Sleep Stage Classification
Based on EEG Signals by Using Neural Networks and Wavelet Packet Coefficients,” IEEE
Engineering in Medicine and Biology Society, Aug. 2008.
N. Michielli, U. R. Acharya, and F. Molinari, “Cascaded LSTM recurrent neural network
for automated sleep stage classification using single-channel EEG signals,” ScienceDirect
Computers in Biology and Medicine, Mar. 2019.
O. Yildirim, U. B. Baloglu, and U. R. Acharya, “A Deep Learning Model for Automated
Sleep Stages Classification Using PSG Signals,”, International journal of environmental
research and public health, Feb. 2019.
S. Mousavi, F. Afghah, and U. R. Acharya, “SleepEEGNet: Automated sleep stage scoring
with sequence to sequence deep learning approach,” PLoS ONE, May. 2019.
A. Supratak, H. Dong, C. Wu, and Y. Guo, “DeepSleepNet: A Model for Automatic Sleep
Stage Scoring Based on Raw Single-Channel EEG,” IEEE Transactions on Neural Systems
and Rehabilitation Engineering, Jun. 2017.
A. Supratak and Y. Guo, “TinySleepNet: An efficient deep learning model for sleep
stage scoring based on raw single-channel EEG,” 2020 42nd Annual International Conference
of the IEEE Engineering in Medicine & Biology Society (EMBC), Jul. 2020.
I. Hussain et al., “Quantitative Evaluation of EEG-Biomarkers for Prediction of Sleep
Stages,” Sensors, Apr. 2022.
R. Zhao, Y. Xia, and Q. Wang, “Dual-modal and multi-scale deep neural networks for
sleep staging using EEG and ECG signals,” Biomedical Signal Processing and Control,
Apr. 2021.
C. W. Whitney et al., “Reliability of scoring respiratory disturbance indices and
sleep staging,” Sleep, 1998.
T. Penzel et al., “Modulations of heart rate, ECG, and cardio-respiratory coupling
observed in polysomnography,” Frontiers in Physiology, Oct. 2016.
R. P. Bartsch et al., “Phase transitions in physiologic coupling,” Proc Natl Acad
Sci USA, Jun. 2012.
P. A. Lanfranchi and V. K.Somers, “Cardiovascular Physiology: Autonomic Control in
Health and in Sleep Disorders,” Principles and Practice of Sleep Medicine: Fifth Edition,
2010.
E. Urtnasan, J. U. Park, E. Y. Joo, and K. J. Lee, “Deep Convolutional Recurrent Model
for Automatic Scoring Sleep Stages Based on Single-Lead ECG Signal,” Diagnostics,
May. 2022.
M. Radha et al., “A deep transfer learning approach for wearable sleep stage classification
with photoplethysmography,” NPJ digital medicine, Sep. 2021.
M. Radha et al., “Sleep stage classification from heart-rate variability using long
short-term memory neural networks,” Scientific Reports, Oct. 2019.
Q. Li et al., “Deep learning in the cross-time frequency domain for sleep staging
from a single-lead electrocardiogram,” Physiological measurement, Dec. 2018.
N. Sridhar et al., “Deep learning for automated sleep staging using instantaneous
heart rate,” NPJ digital medicine, Oct. 2020.
Y. Wei al., “A multi-class automatic sleep staging method based on long short-term
memory network using single-lead electrocardiogram signals,” IEEE Access 7, Jun. 2019.
M. Tang et al., “Deep adaptation network for subject-specific sleep stage classification
based on a single-lead ECG,” Biomedical Signal Processing and Control 75, Feb. 2022.
R. K. Tripathy and U. R. Acharya, “Use of features from RR-time series and EEG signals
for automated classification of sleep stages in deep neural network framework,” Advances
in Artificial Neural Systems, May. 2012.
M. Längkvist, L. Karlsson, and A. Loutfi, “Sleep Stage Classification Using Unsupervised
Feature Learning,” Advances in Artificial Neural Systems, Jul. 2012.
M. Kebe et al., “Human Vital Signs Detection Methods and Potential Using Radars: A
Review,” Sensors, Mar. 2020.
L. Zhang et al., “Sleep stages classification by CW Doppler radar using bagged trees
algorithm,”2017 IEEE Radar Conference, May. 2017.
H. Hong et al., “Noncontact Sleep Stage Estimation Using a CW Doppler Radar,” IEEE
Journal on Emerging and Selected Topics in Circuits and Systems, June. 2018.
A. Tataraidze et al., “Bioradiolocation-based sleep stage classification,” 2016 38th
Annual International Conference of the IEEE Engineering in Medicine and Biology Society(EMBC),
Aug. 2016.
S. Toften et al., “Validation of sleep stage classification using non-contact radar
technology and machine learning(Somnofy),” Sleep Medicine, Feb. 2020.
H. B. Kwon et al., “Attention-Based LSTM for Non-Contact Sleep Stage Classification
Using IR-UWB Radar,” IEEE Journal of Biomedical and Health Informatics, Oct. 2021.
B. Yu et al., “WiFi-Sleep: Sleep Stage Monitoring Using Commodity Wi-Fi Devices,”
IEEE Internet of Things Journal, Mar. 2021.
M. Zhao et al., “Learning sleep stages from radio signals: A conditional adversarial
architecture,” arXiv:1706.03762, 2017.
Author
Jiyong Park is pursuing a B.S. in the School of Computer and Information Engineering
from Kwangwoon University, Seoul, South Korea. His research interests include deep
learning and computer architecture.
Jaewon An received his B.S. degree in the School of Computer and Information Engineering
from Kwangwoon University, Seoul, South Korea. He is currently pursuing an M.S degree
in the Department of Computer Engineering from Kwangwoon University, Seoul, South
Korea. His research interests include biomedical signaling and machine learning algorithms.
Sang Ho Choi received his B.S. degree in biomechatronics and electronic&electrical
engineering from Sungkyunkwan University, Suwon, South Korea, and his Ph.D. degree
in bioengineering from Seoul National University, Seoul, South Korea. He worked as
a senior researcher in the Smart Device Team, Samsung Research, Seoul, South Korea.
He is currently an assistant professor in the School of Computer and Information Engineering,
at Kwangwoon University, Seoul, South Korea. His research interest includes biomedical
signal processing and artificial intelligent algorithm for biomedical applications.