This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1

A Sleep Apnea Detection Method Based on

Unsupervised Feature Learning and
Single-Lead Electrocardiogram
Kaicheng Feng, Hengji Qin, Shan Wu, Weifeng Pan, Guanzheng Liu*, Member, IEEE

complications such as cardiovascular events [7], cognitive

Abstract— Sleep apnea (SA) is a harmful respiratory disorder impairment [6], stroke [8], and mortality [9]. This shows that
that has caused widespread concern around the world. SA is threatening people’s physical and mental health.
Considering that electrocardiogram (ECG)-based SA diagnostic In clinical practices, the polysomnography (PSG) based
methods were effective and human-friendly, many machine
learning or deep learning methods based on ECG have been
method is the golden standard for SA diagnosis. The PSG can
proposed by prior works. However, these methods are based on record multiple physiological signals such as
feature engineering or supervised and semi-supervised learning electrocardiogram (ECG), electroencephalogram (EEG),
techniques, and the feature sets are always incomplete, subjective oronasal air flow and pulse oximetric saturation. Experts
and highly dependent on labeled data. In addition, some related diagnose sleep apnea by analyzing those signals
studies ignored the data imbalance problem which leads to poor comprehensively. However, this diagnostic method is
performance of classifier on minority classes. In this study, a SA
detection model based on frequential stacked sparse
time-consuming, costly, and uncomfortable. Patients need to
auto-encoder (FSSAE) and time-dependent cost-sensitive (TDCS) be connected to at least 22 electrodes to measure 11 channels
classification model was proposed. The FSSAE extracts feature sleep signals in a specially equipped laboratory during several
set automatically with unsupervised learning technique, and the nights [10]. Doctors need to spend a significant amount of time
TDCS classification model is proposed by combining the hidden observing and analyzing those data to make a diagnosis
Markov model (HMM) and the Metacost algorithm to improve manually. Hence the PSG-based method is expensive,
the performance of the classifier by considering temporal
dependence and the imbalance problem. In the test set, the result
cumbersome and unfriendly. A simple, inexpensive,
of per-segment classification achieved 85.1%, 86.2% and 84.4% users-friendly alternative is needed [11].
for accuracy, sensitivity and specificity respectively, proving that In order to solve these problems, researchers have done
our method is helpful for SA detection. abundant useful exploration. Some researchers proposed
improved instruments such as pressure pad sensor and
Index Terms—Cost sensitive, Deep neural network, Hidden unobtrusive multiparameter sleep apnea monitor to replace the
Markov model, Sleep apnea, Stacked sparse autoencoder. traditional PSG [12] [13]. Taran et.al. used artificial bee
colony optimize Hermite basis function to analyze EEG
signals for feature extraction and used extreme learning
I. INTRODUCTION

s
machine and least-squares support vector machine as classifier.
LEEP apnea (SA) is a common chronic respiratory disease. They obtained encouraging performance [14]. Vimala et. al.
It features significant reduction (hypopnea) or complete used infinite impulse response Butterworth band pass filter
interruption (apnea) in respiratory airflow during sleep. and Hilbert Huang transform to analyze EEG signals and
Approximately 15% of men and 5% of women suffer from SA, produced promising results [15]. Xenia et.al. combined pulse
with ranges of influence [1]. Moreover, SA is associated with oximetry (SpO2) and airflow and proved that it was helpful to
an increasing risk of many harmful health consequence [2]. In increase the accuracy of sleep apnea detection in adults [16].
the early stage of this disease, immediate effects include Rosenwein et. al. analyzed 186 adults’ breath-by-breath audio
fatigue [3], intermittent hypoxia [4], uneven heart rate [5], and signals of respiratory cycle. They proved that sleep apnea
mood [6]. Over time, these conditions will lead to serious events can be positioned exactly and this approach is potential
to be applied in home conditions [17].
This work was supported by the natural science foundation of China Moreover, some researchers concentrated on ECG based
(61401521) (81701788), the science and technology program of Guangzhou method and achieved encouraging results. Sharma et al. used
(2017A010101035)
Kaicheng Feng, and Guanzheng Liu are affiliated with (1) School of Hermite decomposition to extract features from the ECG's Q
Biomedical Engineering, Sun Yat-sen University, Guangzhou, China; (2) the wave, R wave, S wave (QRS) complex, combined with
Key Laboratory of Sensing Technology and Biomedical Instruments of classical time-domain features, and they used the least-squares
Guangdong Province, School of Biomedical Engineering, Sun Yat-sen
University, Guangzhou 510275, China; and (3) the Guangdong Provincial
support vector machine as a classifier. The accuracy of SA
Engineering and Technology Centre of Advanced and Portable Medical segment classification is 83.8% [18]. Babaeizadeh et al.
Device, China. extracted frequency-domain features of the RR sequence and
*
Corresponding author: Guanzheng Liu (e-mail: [email protected]).

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2

adopted the quadratic classifier for SA detection. The TABLE I

classification accuracy of the SA segment reaches 84.7% [19]. DATASET INFORMATION
With the development of deep learning, some researchers Segments Released Set Withheld Set
adopted neural network method to detect SA using ECG Apnea 6464 6502
signals. Li et al. proposed a SA classification method based on Normal 10399 10611
a stacked sparse autoencoder and decision fusion. They used Total 16863 17113
stacked sparse autoencoder to perform supervised feature
TABLE II
learning on single-lead ECG signals and adopted decision TRAINING SET INFORMATION
fusion to combine multiple weak classifiers for improving the Released Set Released Set
Segments
performance of SA detection. The accuracy of per-segment (Part A) (Part B)
SA detection is 84.7% [20]. Urtnasan et al. proposed a SA Apnea 2386 4078
detection method based on convolutional neural network Normal 4313 6086
(CNN), using CNN to perform supervised feature learning on segments/subjects 6699/14 10164/21
single-lead ECG and complete the classification of SA patients
[21]. Those studies supported the assertion that single-lead
combinations. On the other hand, the TDCS classifier
ECG signal analysis and deep learning algorithm is useful for
combines the advantage of the HMM and the Metacost
SA detection.
algorithm to improve classification performance. The hidden
Some researchers also concentrated on other physiological
Markov model is beneficial to improve the performance of the
signals and adopted deep learning method to detect SA.
classifier by utilizing the temporal dependency between
Steenkiste et al. analyzed the bio-impedance of chest and
adjacent ECG signal segments [25]. Moreover, MetaCost is an
adopted a two-phase Long Short-Term Memory (LSTM)
effective algorithm-level approach to solve the data imbalance
neural network algorithm. The area under the precision-recall
problem. The cost -sensitive classifier is trained to adjust its
curve of proposed method reached the same level as automatic
preference by predetermined misclassification costs.
scoring using a polysomnography respiration channel [22].
Compared with other data-level methods such as up-sampling
Mostafa et al. used a Deep Belief Network to analyze blood
and down-sampling, MetaCost is straightforward and suitable
oxygen saturation (SpO2) for sleep apnea classification. They
for most classification algorithm [26]. Hence, we propose a
first use an unsupervised learning strategy to calculate the
novel approach to detect SA based on the FSSAE feature
initial weights. Then, a supervised fine-tuning is adopted to
extractor and the TDCS classifier.
optimize the weights [23].
Although the previous researchers have achieved great
II. METHOD
results, there still exist areas for continued development.
Firstly, traditional feature engineering methods need abundant A. Data and preprocessing
prior knowledge and consume a significant amount of time In this study, the single-channel SA-ECG signals from the
choosing the best feature sets. Deep learning algorithms PhysioNet database of Computers in Cardiology 2000
always require abundant labeled data. In the medical field, Challenge were used [27], [28]. This database contains 70
labeled data are difficult to obtain due to privacy policies or nighttime ECG recordings which were collected from 32
data quality requirements. Therefore, a deep learning network subjects. Four subjects were recorded one recording each. 22
that does not require vast amounts of labeled data and that can subjects were recorded two recordings each. Two subjects
automatically search for the best feature sets is necessary. were recorded three recordings each, and four subjects were
Second, the data imbalance problem significantly recorded four recordings [29]. This dataset was divided
compromises the performance of the learning algorithm. Data equally into two parts, the released set and the withheld set.
imbalance problem exists when one class of data (typically the The released set was given out to participants in advance for
class of interest) is smaller than the other classes in one model training and was categorized into three groups (A, B
database. This leads to a situation where the trained classifier’s and C). Group A contained at least 100 apnea events and group
performance for a certain class of data, expressed as the C had fewer than 5 apnea events. Group B was named the
specificity of the classifier is much higher than the borderline apnea group which was defined as having the
sensitivity[24]. The sensitivity metric represents the ability of number of apnea events between 5 and 100 [10] [30]. The
the classifier to recognize patients. If a classifier is only withheld set was used by organizer for scoring. After this
accurate for normal people, it is not satisfactory. competition, all data were available, and every “1 min” ECG
To solve these problems, we propose an unsupervised epoch was marked by experts as either normal or apnea. The
feature learning model based on the frequential stacked sparse sampling rate of this dataset was 100Hz, and the lengths
auto-encoder (FSSAE) and a time-dependent cost-sensitive ranged from 401 to 587 minutes.
(TDCS) classifier utilizing the hidden Markov model (HMM) First, the ECG signals were divided into 1-min segments [31]
and the MetaCost algorithm. On the one hand, FSSAE is an and we manually removed some ECG signals with severely
unsupervised feature learning neural network that requires distorted waveforms. Since the whole night’s ECG data are
unlabeled data to gain good feature representations. It aims to divided into one-minute segments, some segments less than
learn an approximation to the input to seek inner feature one minute are also removed. The number of removed

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3

Fig. 1. The overall system flow diagram.

segments is 337, and the total number of the PhysioNet sleep sequence analysis is very important and influential in ECG
apnea dataset is 34313 [29]. The Pan-Tompkins algorithm was analysis [33]. The physiological information can be easily
used to detect R-peaks for extracting the RR interval sequence extracted from the frequential ECG and the frequency-domain
[32]. This algorithm adopts a well-designed integer-coefficient sequence is more beneficial for analyzing short-term
bandpass filter which consists of cascaded low-pass and sequences such as RR interval sequences[34]. In addition, the
high-pass filters for removing baseline wander, muscle noise RR interval sequence reflects the signal components of the
and T-wave interference. An approximate derivative filter and autonomic nerves system concentrated at 0.04–0.4 Hz [35]. In
a moving window integral are then used to amplify the this range, the frequency domain signal and most noise, such
characteristics of QRS complexes. Adaptive thresholds extract as power frequency noise, are not aliased. Hence the frequency
the locations of the QRS complexes. Secondly, the median domain signals can be used as an input to better distinguish the
filter as proposed by Chen et. al. was adopted to eliminate the effective signal and noise.
physiologically uninterpretable RR intervals [11]. Finally, The stacked sparse auto-encoder (SSAE) is an unsupervised
because the deep neural network requires the same input deep neural network that consists of multiple stacked layers of
lengths, we interpolated the RR interval sequences to 100 original sparse auto-encoders (SAE) [20] [36]. SSAE is an
points by cubic spline interpolation [20]. unsupervised feature learning algorithm that only uses
As shown in Tables I and II, the total number of processed unlabeled data. It constructs loss function by minimizing the
data are 33976 (released set: 16863; withheld set: 17113). reconstruction error between the input and the reconstructed
According to the instructions of this database and to compare signal while the hidden layer is the extracted features. This
with the work of other researchers which also adopt this means that this model just uses input data without labels to
dataset, the released set was used for training, and the withheld look for the inner characteristics of input data. Supervised
set was used for evaluating the algorithm performance. In this deep learning methods always focus on associating input data
study, the released set was divided into two parts randomly: and labels to achieve better results on specific tasks, where
Part A (6699 segments/14 subjects) and Part B (10164 many input-label pairs are necessary for training. However,
segments/21 subjects). We discarded the label of data in Part A. the use of unsupervised FSSAE avoids a large amount of
Part A was used for training the unsupervised feature tedious data labeling work and can achieve satisfactory results.
extraction model and Part B was used for training the When training this entire SSAE neural network, the greedy
time-dependent cost-sensitive classifier. layer-wise training strategy [37] was adopted. In other word,
the extracted features of the last SAE network are used as the
B. Framework of the proposed model
input of the next SAE network. By repeating this step
Fig. 1 shows the overall architecture of the proposed deep iteratively, the entire neural network was trained. Finally, by
neural network for SA detection. The FSSAE module first assembling the obtained feature layers in order, a complete
receives the RR intervals as input. It converts RR intervals
from the time domain into the frequency domain and extracts a
more complete and more effective feature set automatically by
an unsupervised feature learning strategy. Following this, the
classification model is trained with this feature set. To improve
the performance of this model, The TDCS classifier module
considers the temporal dependency between adjacent
segments and the data imbalance problem of the dataset.
Finally, we validated the performance of this model with
per-segment and per-recording detection.
1) Frequential stacked sparse auto-encoder
In this study, the time-domain sequence was converted to
the frequency-domain sequence by fast Fourier transform to Fig. 2. The structure of stacked sparse autoencoder.
easily obtain abundant information. Frequency-domain

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4

SSAE model is obtained. The training process is shown in Fig.

2.
From the RR interval sequences, an unlabeled input data set
was constructed as × = { (1), (2), … , ( )}, where is
the number of 1min RR interval segments and is the length
of each data sample. The sigmoid function was adopted as the
activation function which can add nonlinear components into
model to improve the neural network’s fitting ability.
1
= (1)
1+ Fig. 3. The structure of Hidden Markov model.
The input set can then be encoded to hidden layer ℎ by the
sigmoid function as follows:
( , ) = ( , )+ (7)
ℎ = + (2)
where is the weight of the sparsity penalty. Finally, to
where is the weight matrix between the -th unit of the avoid overfitting, the regularization term was added into the
input layer and the -th unit of the encoding layer. and is the cost function [20].
bias. Correspondingly, the decoder layer can be defined as
follow: 2) Time-dependent cost-sensitive classifier
This section introduces the time-dependent cost-sensitive
= ℎ + (3) classifier that was designed to utilize the temporal dependency
between different RR intervals and to reduce the impact of the
Where is the weight matrix between the -th unit of the data imbalance problem. The temporal dependency among
encoding layer and the -th unit of the output layer, and is sleep apnea events can be observed clearly [38] and it can be
the encoding bias. The entire mapping process is summarized extracted by the hidden Markov model to improve the
as = ℎ( , , ) , where consists of and , is classifier. Metacost is a type of cost-sensitive algorithm to
consist of and , and is the unmodified original input solve data imbalance problem by adjusting the weight
data. The reconstruction error between and was defined by coefficient of different classification errors. These methods
mean square error is as follows: were combined to improve the performance of this model.
1 The hidden Markov model is a dynamic Bayesian network
( , , ) = ‖ℎ( , , ) − ‖ (4) and has a wide range of applications in time series data
2
The single-layer auto-encoder may sometimes not be able to modeling. It has been adopted to analyze physiological signals
learn available features sometimes when the number of the such as EEG, ECG in many areas [39],[40]. As shown in Fig.3,
hidden layer very large. The sparse term can be added into the the structure of the hidden Markov model can be defined as
autoencoder to find interesting structures in the data by follows: Given the observation states and the corresponding
limiting the number of active neurons. For the hidden units , unobserved Markov states , according to Markov chain
the average activation can be defined as follows: process, the depends only on the , and the is only
determined by the ,and has nothing to do with other
1 ()
= (5) unobserved states. As shown in Fig. 3, the HMM can be
defined by the following parameter set:
where ()
is the -th unit of the input layer. and ( ) = =( , , , ) (8)
( ). The initial state probability vector, = ( , , … , ) ,
To obtain a less redundant and more efficient feature set, the where = ( = ). The Markov state sequence, , is
average activation must be close to zero. Hence, the sparse defined as the different states in Markov state space. In this
penalty term is adopted to keep the average activation of study, the value of n is 2, because the segment has only two
hidden units in a small range by punishing those cases where states: apnea or normal. The parameter is the state transition
and (the spares parameter) are significantly different. The probability matrix. This parameter indicates the probability of
sparse penalty term can be defined as follows: the model transitioning between states: = [ , ] × , where
[ ,] × = ( = | = ). is the marginal probability
= ( ) (6) distribution of Markov states, = ( , , … , ), where =
( = ), and is the observation-to-state classifier, ( ) =
where is defined as Kullback-Leibler divergence. This ( = | = ).
is a common method for measuring the differences between The parameters and are determined by the nature of
probability distributions. We then add the sparse penalty term the RR intervals. = { , }, where and respectively,
in the cost functions: indicate that each RR interval is either apnea or normal. The
parameter consists of the features which are learned from
the corresponding RR intervals by FSSAE. This practice is in

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5

line with human cognition. We can observe that a segment is classifier is built. To better suit our problem, some changes
apnea or normal and the inner data characteristics are not were made to the classical Metacost algorithm. In practice,
perceptible. only the majority class’s label was changed. We used this data
In our study, the HMM was adopted to improve the set with the majority class’s label changed to train a new
performance of the classifier by considering the temporal classifier.
dependence between segments. Up to now, a time-dependent
C. Evaluation
classifier is combined through softmax and the hidden Markov
model. To evaluate the performance of our model, six indices were
Imbalanced data problems exist in multifarious practical adopted:
+
applications. The class imbalance causes a bias because = (11)
conventional classification algorithms assume all errors have + + +
an equal cost and are partial to the majority class [41]. The
MetaCost algorithm is a simple and efficient method to solve / = (12)
this problem. It takes a common unaltered classifier and +
acclimate its learning cost by a cost matrix and bagging [42].
The cost matrix indicates the cost of class samples being = (13)
misclassified into class . The cost matrix gives a larger weight +
to the few classes, making the classification model more
concerned with the high-cost samples. = (14)
+
This computational process is done through the following
steps. (1 +
)∗ ∗
1. T-timed samples with replacement are taken in the _ = (15)
∗ +
training set [ , ] = {( , ), ( , ), … }. samples
where TP, TN, FP, and FN represent “true positive”, “true
are taken out of training set each time and new subsets
negative”, “false positive”, and “false negative” respectively.
are marked as = { , , … , }.
In this study, we use the sleep apnea class as the positive class
2. Every subset is used for training a corresponding
and the normal class as the negative class.
classifier. Therefore, a classifier set =
Traditionally, the most frequently used metric is accuracy.
{ , ,…, } can be obtained.
This metric can simply evaluate the classifier’s performance.
3. Use each classifier in the classifier set to classify
However, the traditional metrics are not appropriate for
each sample in the training set [ , ]. We can obtain
evaluation with imbalanced datasets [43]. For instance, if a
the probability, ( | ; ), that the sample belongs dataset includes 10% of class A and 90% of class B, the trained
to the -th class in corresponding classifier . classifier assesses every sample to be class B. We get an
4. Setting up the cost matrix ( , ) manually which accuracy of 90%. Obviously, this accuracy is not correct.
indicates the cost of class samples being misclassified Hence, we also adopt recall, precision, and F-value to evaluate
into class . When there are only two classes, the cost our works. Recall and precision have been adopted for
matrix can be structured as follow: evaluating classification performance in the imbalanced
actual positive actual negative problem [43]. F-value combines precision and recall to
predict positive (0,0) (0,1) measure the classifier’s performance comprehensively, where
predict negative (1,0) (1,1) adjusts the relative importance between recall and precision.
In this study, is set to 1.
where positive and negative are represented by 0 and 1 Additionally, the receiver operating characteristic (ROC)
0 2 and corresponding area under the curve (AUC) were adopted
respectively. In this study, the cost matrix is .
1 0 to evaluate this method.
5. Define conditional risk as cost function：
(| )= (| ; ) (, ) (8) III. RESULTS
A. Minute-by-minute sleep apnea detection
6. Continually minimize the cost function and modify the After training the model, the performance of
label of corresponding data in the training set that meets minute-by-minute (per-segment) SA detection can judge the
the following criteria: performance of this model. As shown in Table III, the
= min ( | ) (10) proposed TDCS classifier achieved an accuracy of 85.1% for
7. The training set with modified label is used for per-segment detection. Sensitivity (recall) and specificity are
classifier training to obtain the final classification 86.2%, 84.4% respectively, and the F-value is 81.4%. These
model. metrics indicate that the proposed method used temporal
In this manuscript, all the classifier in the Meatcost dependency to improve the classification performance and
algorithm is proposed time-dependent classifier and the reduced the impact of data imbalance.
training set is Part B. So far, a time-dependent cost-sensitive

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6

TABLE III
PER-SEGMENT DETECTION PERFORMANCE
Acc Sen/Recall Spe Precision F-value
Classifiers
(%) (%) (%) (%) (%)
Softmax 77.6 68.9 83.0 71.2 70.0
Softmax-HMM 86.0 77.0 91.4 84.6 80.6
TDCS 85.1 86.2 84.4 77.2 81.4

TABLE IV
PER-RECORDING DETECTION PERFORMANCE
Classifier Acc (%) Sen (%) Spe (%) MAE
Softmax 80 100 41.7 8.13
Softmax-HMM 97.1 95.7 100 5.91
TDCS 97.1 95.7 100 5.60
MAE: the mean absolute error between estimated and true AHI. The
cut-off AHI is 5. If the estimated AHI is greater than 5, this recoding is
deemed to be ill.

TABLE V
The RESULTS of MODEL VALIDATION
Classifiers Acc Sen/Recall Spe Precision F-value
(%) (%) (%) (%) (%) Fig.4. The ROC curves of different classifiers
Softmax 76.8 67.3 82.7 70.7 68.9
Softmax-HMM 83.3 71.2 90.8 82.9 76.6
TDCS 80.7 80.1 81.0 72.4 76.0
To verify proposed model by comparing with traditional dataset practice,
the withheld set was used for model training and the released set was
used for evaluating
C. Model evaluation
To evaluate this classifier clearly, the receiver operating
characteristic curve and AUC were adopted which were shown
B. Per-recording sleep apnea detection in Fig.4. The TDCS classifier achieved the highest AUC value
According to the practices of previous researchers, it is first (0.8596), while the softmax classifier and softmax-HMM
determined whether each segment is apnea or normal, then classifier achieved lower values (0.7693 and 0.8491
each recording’s apnea-hypopnea index (AHI) is calculated to respectively).
estimate whether it is apnea or normal. AHI is the clinical In order to verify the performance of proposed method,
standard of identifying SA, it is defined as the number of apnea two-fold cross-validation is performed. The traditional
and hypopnea events per hour of sleep and can indicate the practice is to use released set as training set and withheld set as
severity of sleep apnea. test set. We exchanged both sides. The withheld set was used
60 for model training and the released set was used for evaluating.
= × (16) As shown in Table V, the accuracy, sensitivity/recall and
where represents the length of the recording (min). The F-value was 80.7%, 80.1% and 76.0% respectively. This
amount of data for each person is different. According to the proposed method has improved the classification and reduced
standard of AHI (American Academy of Sleep Medicine, the impact of the data imbalance problem.
1999), if the AHI of a patient is greater than or equal to 5, the
patient is deemed to suffer from sleep apnea. As shown in IV. DISCUSSION
Table IV, both the softmax-HMM model and the TDCS model A. Relationships between sleep apnea and ECG
achieve same accuracy, 97.1%, only one sample is identified
incorrectly. In order to evaluate which algorithm is better for Researchers have verified that sleep apnea can change the
activity of heart and we can observe corresponding
per-recording sleep apnea detection, we use the mean absolute
phenomena from ECG signals. Penzel et al. indicated that the
error (MAE) between estimated and true AHI to measure the
morphology of the ECG signal can be changed by sleep apnea
gap between estimated and true AHI. The formula is shown events [44]. Sleep apnea caused cyclic variations in cardiac
below: rhythm in the form of bradycardia during apnea followed by
1 tachycardia upon its cessation [45] [38]. This characteristic
= − (17) was used to improve the performance of proposed model by
Hidden Markov model. Additionally, ECG signals included
where = 35 is the number of test set recordings. some breathing characteristics, and respiratory event can be
We can observe that the TDCS achieves the lowest MAE in detected by ECG signal [46] [47]. Furthermore, sleep apnea
the Table IV. This shows that the difference between the induces an imbalance of autonomic nerve sympathetic (ANS).
estimated and true AHI calculated by the TDCS is the smallest. By analyzing the heart rate variability (HRV), some indices
Compared with the softmax-HMM algorithm, although the proved that sleep apnea caused the cardiac autonomic
same accuracy is achieved, the TDCS algorithm which obtains dysregulation of sleep apnea patients [48]. We can analyze
smaller MAE is better.

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 7

TABLE VI
COMPARISON of PER-SEGMENT CLASSIFICATION BETWEEN PROPOSED METHOD AND RELEVANT WORK
Feature extraction
Reference Feature type Classifier Test set results
method
Acc = 82.2%
Power spectrum
Surrel et al.[49] L-SVM Sen = 73.3%
features
Spe = 87.6%
Acc = 83.8%
Hermite basis
Sharma et al.[18] LS-SVM Sen = 79.5%
functions
Spe = 88.4%
Time-domain Acc = 87.7%
Extreme learning
Sadr et al.[50] features and power Sen = 81.3%
machine
spectrum features Spe = 91.7%
Feature engineering
Cepstrum Quadratic Acc=84.8%
Martin-Gonzalez et
coefficients, filter Discriminant Sen=81.4%
al.[51]
banks and DFA Analysis Spe=86.8%
feature pool, Acc=86.2%
Song et al.[25] including 24 SVM-HMM Sen=82.6%
features Spe=88.4%
Acc=84.7%
Babaeizadeh et IHR’s spectral
Quadratic classifier Sen=76.7%
al.[19] power
Spe=89.6%
Acc = 84.7%
Supervised learning Decision fusion
Li et al.[20] Sen = 88.9%
features classifier
Spe = 82.1%
Deep learning
Acc = 85.1%
Our proposed Unsupervised
TDCS Sen = 86.2%
method learning features
Spe = 84.4%
LS-SVM: Least square support vector machine, L-SVM: Linear support vector machine, DFA: Detrended Fluctuation Analysis,
IHR: Instantaneous heart rate, HMM: hidden Markov model, TDCS: Time-dependent cost-sensitive classifier, Acc: Accuracy, Sen:
Sensitivity, Spe: Specificity.

whether the sleep apnea occurs through the change of ECG et al. adopted supervised learning features and the decision
and this shown that ECG signals can be used for sleep apnea fusion classifier and obtained satisfactory results. However, a
detection. supervised feature learning algorithm needs large quantities
labeled data, and the decision fusion classifier needs to train
B. Comparison with existing studies
many weak classifiers.
In this part, we compare the performance of the proposed In our work, we can obtain feature combinations
method with the existing relevant studies described in Table automatically by an unsupervised feature learning algorithm
VI. Those relevant works all used the released set for training which only needs unlabeled data. Furthermore, we adopted the
and withheld set for validating. To make a fair and effective TDCS classifier to utilize temporal dependency and reduced
comparison, some relevant work that used different datasets or the impact of the data imbalance problem. The accuracy was
where there existed great disparities in the amount of their 85.1%. The sensitivity (recall) and F-value were up to 86.2%
samples and ours are not considered [52] [53]. and 81.4% respectively. The classification accuracy was
As shown in Table VI, traditional methods focused on comparable and the imbalance data problem was solved.
feature engineering and selecting a better classifier to optimize
the total classification accuracy. The accuracy ranged from C. Frequential input
82.2% to 87.7%. However, it can be observed that the It has been reported that the balance of the sympathetic
sensitivity (recall), ranging from 73.3% to 81.3% is lower than nervous system and the parasympathetic nervous system is
their specificity in most studies. In practice, the broken in SA patients [54] [48]. The time-domain RR intervals
misclassification cost between normal people and apnea of normal people fluctuate more violently, which means that
patients is not equal, and data imbalance (apnea segments: normal people have stronger self-regulation ability. In patients
6,464, normal segments:10,399) exacerbates this problem. Li with SA, more sympathetic activities cause the heart rate

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 8

Fig.5. The example of SA patient’s RR interval and normal people’s RR interval and corresponding frequency curves. This example is averaged over all
segments for a SA subject or a normal subject.

variability of patients is reduced, and its baseline changes TABLE VII

COMPARISON OF DIFFERENT INPUT
rapidly [48] [55]. Through the frequency domain sequence, it
Input Fine-tuning Acc (%) Sen (%) Spe (%)
can be found that the energy of the normal segment is evenly
no 62.01 0.00 100.00
distributed over the frequency range of 0.15–0.4Hz, which Time
indicated that the sympathetic and parasympathetic nerves are yes 77.15 61.03 87.03
both active, reflecting that the cardiopulmonary system is in a no 77.60 68.86 82.95
Frequency
healthy state. However, the energy of the patient segment is yes 75.25 62.17 83.31
concentrated at a lower frequency range (0.04–0.15Hz), ‘yes’ and ‘no’ indicate whether to use an unsupervised
indicating that the sympathetic system is more activated than fine-tuning strategy
the parasympathetic system and that the sympathetic system is
disordered [48], [54]. These changes can be observed in Fig. 5. unsupervised feature learning model based on FSSAE without
Compared with time-domain data signals, the ECGs of fine-tuning, only using unlabeled data. This model can achieve
subjects analyzed in the frequency domain show greater better performance in feature extraction. To validate that the
differences between apnea patients and normal subjects. frequential unsupervised feature learning algorithm is better,
Additionally, in the frequency domain, some stubborn noise some comparative tests are adopted. Different inputs were sent
can be removed, which is difficult to dispose of in the time into the stacked sparse auto-encoder for feature learning and,
domain. Thus, the deep neural network can easily extract more to avoid conflict with subsequent modified classifiers, the
complete feature sets. These evidences reveal that the original softmax classifier was adopted. Those results are
discriminative features can easily learn from the shown in Table VII.
frequency-domain signal. When the input is time-domain data, and there is no
supervised fine-tuning process, the accuracy is only 62.0% and
D. Unsupervised feature learning the sensitivity is zero. This means that all the segments are
Unsupervised feature learning algorithms can automatically considered to be normal. However, after fine-tuning, the
seek the internal representations of data and have been widely performance of the model increases significantly. This is
used in various aspects [43] [56]. In practice, unsupervised consistent with the results of previous studies [20]. It is
deep leaning algorithms such as the Boltzmann machine and indicated that time-domain data are not suitable as the input of
the auto-encoder often include a supervised fine-tuning stage. an unsupervised feature learning model in this task. When the
At the beginning of those algorithm, unlabeled data is used to input is frequency domain data and the unsupervised method is
adopted, the accuracy and sensitivity are improved to 77.6%
learn internal representation. This stage is called unsupervised and 68.6%, respectively; this is better than what is achieved
pre-training. In order to make the model more focuses on a with time-domain input.
specific task, labeled data is used to fine-tune the model. This To achieve satisfactory results, time-domain input must rely
stage is called supervised fine-tuning [57]. Combining on labels to adjust the feature learning network by fine-tuning.
unsupervised pre-training and supervised fine-tuning usually When the supervised fine-tuning strategy is adopted, the time
produces better results [37]. In this study, we proposed an domain input and the frequency domain input obtain similar

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 9

accuracy (77.15%, 75.25% respectively). However, when the

input is frequency-domain data and the unsupervised feature
learning algorithm is adopted, not only does the accuracy
remain at the same level as that for the time domain supervised
results, but its recognition performance of SA segments,
sensitivity, is significantly better than the other two indices in
Table VII. The frequency-domain inputs do not need to use
fine-tuning to obtain better features and reduce the dependence
on labels, indicating that frequency-domain signals are more
suitable for this unsupervised feature learning method.
E. Temporal dependency
A cyclical heart rate pattern has been confirmed to exist in
SA patients [38], [45]. Temporal dependency can be observed
frequently in real ECG recordings and can be observed to
change over time rather than appear randomly [25]. In this Fig.6. Comparison of different Metacost algorithm.
None: no Metacost algorithm used;
study, we adopted the hidden Markov model to measure the keep_pos: the modified Metacost algorithm that only changed the
temporal dependency. The performance of this mode was majority class’s label;
improved significantly. As shown in Tables III and IV, when keep_neg: the modified Metacost algorithm that only changed the
minority class’s label;
using the Softmax-HMM classifier, the accuracy of Metacost: the classic Metacost algorithm
per-segment SA detection increased from 77.6% to 86.0%, and
the accuracy of per-recording SA detection increased from 80%
TABLE VIII
to 97.1%. The sensitivity, specificity and F-value also were COST MATRIX CHOICES
improved. Hence, the use of temporal dependency was helpful Acc Sen/Recall Spe Precision F-value
(0,1)
for improving SA detection accuracy. (%) (%) (%) (%) %
1.5 80.4 78.0 82.0 72.9 75.4
F. Cost-sensitive algorithm and cost matrix choices 2 85.1 86.2 84.4 77.2 80.6
2.5 80.0 78.5 81.0 72.0 75.1
1) The modified Metacost algorithm 3 78.25 72.0 82.2 74.9 71.7
Data imbalance is widespread and commonly affect the (0,1): the cost of misclassifying sleep apnea segments as normal
segments.
quality of the classification model [58]. Collecting SA patients’
data is more difficult than obtaining normal (no-SA) data in Metacost algorithm (modified Metacost algorithm that only
medical field. Data imbalance usually causes the classifier to changed the minority class’s label), the classifier bias problem
favor the majority class [41]. This is because the minority class becomes more critical. The performance of this model is
is difficult to represent in the training model compared with greatly reduced even though the accuracy is not changed
other classes. In some previous works that used the same significantly. These two methods changed the number and
database as in our study, this problem has not been considered distribution of “normal” labels in the training set, resulting in
[18, 46, 49, 50]. Although they produced better accuracy, the the hidden Markov model’s estimating a larger “normal”
difference between sensitivity and specificity is too large. A segment transition probability when calculating the state
good SA detection model should have both high sensitivity transition probability matrix.
and high specificity.
In this study, normal ECG segments are the majority and we 2) Cost matrix choices
adopted a modified Metacost algorithm which only changed 0 2
In this study, the cost matrix is set to . Generally, the
the majority class’s label to reduce the impact of the data 1 0
severity of data imbalance in different data sets is different.
imbalance problem. The problem wherein the general
We need to select a specific cost matrix for each different
classifier is biased is caused by the classifier dividing some
unbalanced dataset. We set the cost of misclassifying normal
simple features that are difficult to correctly identify into most
segments as sleep apnea segments, (1,0), to 1, and adjust the
of the classes [41]. If only the classification cost of most
cost of misclassifying sleep apnea segments as normal
classes is adjusted, the performance of the cost-sensitive
classification can be improved [24], [26]. segments, (0,1), to analyze the effect of the cost matrix on
As shown in Fig. 6, the experimental results also proved that experimental results. In the adopted dataset, the number of
the keep-pos Metacost algorithm (the modified Metacost sleep apnea segments is less than normal. It is more difficult to
algorithm that only changed the majority class’s label) is the correctly identify SA segments. Hence, it is suitable to set
best method. Using this modified Metacost algorithm is (0,1) greater than (1,0). The performance of different cost
consistent with the classifier characteristics and data matrix is verified on the test set. The experimental
distribution. The sensitivity is close to the specificity. This results are shown in the Table VIII.
means that the cost-sensitive algorithm solves the classifier As shown in the Table VIII, when the (0,1) is set to 2, the
bias problem. Using the classic MetaCost or keep-neg highest accuracy and F-value is achieved. This shows that the

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 10

impact of data imbalance is reduced. When the (0,1) is set to [2] V. K. Somers et al., "Sleep apnea and cardiovascular
1.5, the accuracy and F-value is 80.4% and 75.4% respectively. disease: an American Heart Association/American
When the (0,1) is larger than 2, both accuracy and F-value College of Cardiology Foundation Scientific
are less than 85,1%and 80.6% respectively. Considering both Statement from the American Heart Association
accuracy and F-value, when the (0,1) is set to 2, the model Council for High Blood Pressure Research
achieves the better performance. Professional Education Committee, Council on
Clinical Cardiology, Stroke Council, and Council on
G. The limitations of the proposed method Cardiovascular Nursing," J. Am. Coll. Cardiol., vol.
Although the proposed method shows some great 52, no. 8, pp. 686-717, Aug 19 2008.
characteristics to detect sleep apnea, the limitations of the [3] W. Chotinaiwattarakul, L. M. rsquo, L. Fan, and R. D.
proposed method affect its performance. In this study, we used Chervin, "Fatigue, tiredness, and lack of energy
6699 segments/14 subjects for unsupervised feature extraction improve with treatment for OSA," J. Clin. Sleep Med.,
model training. Less feature extraction model training data vol. 5, no. 03, pp. 222-227, 2009.
result the model is not suitable for processing all types of data. [4] N. A. Dewan, F. J. Nieto, and V. K. Somers,
"Intermittent hypoxemia and OSA: implications for
The difference in the data distribution between the released set
comorbidities," Chest, vol. 147, no. 1, pp. 266-274,
and withheld set leads to lower accuracy. If more unlabeled
Jan 2015.
data are used to train the feature extraction model, the
[5] P. K. Stein and Y. Pu, "Heart rate variability, sleep
performance of proposed model can be improved. We and sleep disorders," Sleep Med. Rev., vol. 16, no. 1,
implement unsupervised feature extraction, and classifier pp. 47-66, Feb 2012.
training still requires labeled data. If complete unsupervised [6] M. Harris, N. Glozier, R. Ratnavadivel, and R. R.
methods are achieved in both feature extraction and classifier Grunstein, "Obstructive sleep apnea and depression,"
training, it is helpful to improve this model. This is one of our Sleep Med. Rev., vol. 13, no. 6, pp. 437-44, Dec 2009.
future efforts. Second, some subjects contributed multiple [7] J. Yeboah et al., "Association between sleep apnea,
nighttime recordings. These data from the same subject will snoring, incident cardiovascular events and all-cause
reduce the generalization performance. Additionally, there are mortality in an adult population: MESA,"
fewer mild patient data of which AHI is between 5 and 15. It Atherosclerosis, vol. 219, no. 2, pp. 963-8, Dec 2011.
will be very meaningful if the proposed method can accurately [8] H. K. Yaggi, J. Concato, W. N. Kernan, J. H.
identify mild patients. In the future, we will collect more data Lichtman, L. M. Brass, and V. Mohsenin,
to validate and improve this method. "Obstructive sleep apnea as a risk factor for stroke
Moreover, other illnesses may have some impact on our and death," N. Engl. J. Med., vol. 353, no. 19, pp.
algorithm. Especially, severe sleep apnea patients are likely to 2034-2041, 2005.
have some heart diseases such as arrhythmias. This may [9] T. Young, M. Palta, J. Dempsey, P. E. Peppard, F. J.
mislead our algorithm to make wrong judgments and reduce Nieto, and K. M. Hla, "Burden of sleep apnea:
the generalization performance of the model. rationale, design, and major findings of the
Wisconsin Sleep Cohort study," WMJ: official
publication of the State Medical Society of Wisconsin,
V. CONCLUSION
vol. 108, no. 5, p. 246, 2009.
In this study, we proposed an unsupervised features [10] M. Sharma, S. Agarwal, and U. R. Acharya,
extraction algorithm and a time-dependent cost-sensitive "Application of an optimal class of antisymmetric
classifier for sleep apnea detection. A more complete feature wavelet filter banks for obstructive sleep apnea
set can be learned in an unsupervised manner with the diagnosis using ECG signals," Comput. Biol. Med.,
frequential stacked sparse auto-encoder. Time dependence vol. 100, pp. 100-113, Sep 1 2018.
was added to this system to improve the performance of [11] L. Chen, X. Zhang, and C. Song, "An Automatic
classifier using the hidden Markov model. The cost-sensitive Screening Approach for Obstructive Sleep Apnea
algorithm was considered to solve the classifier bias problem Diagnosis Based on Single-Lead Electrocardiogram,"
in which some traditional SA detection methods have IEEE Trans. Autom. Sci. Eng., vol. 12, no. 1, pp.
neglected. Experimental results demonstrated that the 106-115, 2015.
proposed unsupervised features extraction algorithm and the [12] G. G. Mora, J. M. Kortelainen, E. R. P. Hernandez, M.
time-dependence cost-sensitive classifier are helpful for SA Tenhunen, A. M. Bianchi, and M. O. Mendez,
"Evaluation of Pressure Bed Sensor for Automatic
detection. However, there are some limitations that need to be
SAHS Screening," IEEE Trans. Instrum. Meas., vol.
addressed to improve the performance of this method. In the
64, no. 7, pp. 1935-1943, 2015.
future, we will improve this method for better performance.
[13] S. Ahmad, I. Batkin, O. Kelly, H. R. Dajani, M. Bolic,
and V. Groza, "Multiparameter Physiological
REFERENCES Analysis in Obstructive Sleep Apnea Simulated With
[1] P. E. Peppard, T. Young, J. H. Barnet, M. Palta, E. W. Mueller Maneuver," IEEE Trans. Instrum. Meas., vol.
Hagen, and K. M. Hla, "Increased prevalence of 62, no. 10, pp. 2751-2762, 2013.
sleep-disordered breathing in adults," Am. J. [14] S. Taran and V. Bajaj, "Sleep Apnea Detection Using
Epidemiol., vol. 177, no. 9, pp. 1006-14, May 1 2013. Artificial Bee Colony Optimize Hermite Basis

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 11

Functions for EEG Signals," IEEE Trans. Instrum. [28] T. Penzel, G. B. Moody, R. G. Mark, A. L.
Meas., vol. 69, no. 2, pp. 608-616, 2020. Goldberger, and J. H. Peter, "The apnea-ECG
[15] V. Vimala, K. Ramar, and M. Ettappan, "An database," in Computers in Cardiology 2000. Vol. 27
Intelligent Sleep Apnea Classification System Based (Cat. 00CH37163), 2000, pp. 255-258: IEEE.
on EEG Signals," J. Med. Syst., vol. 43, no. 2, p. 36, [29] P. de Chazal, C. Heneghan, E. Sheridan, R. Reilly, P.
Jan 8 2019. Nolan, and M. O'Malley, "Automated processing of
[16] X. Hoppenbrouwer, T. Fabius, M. Eijsvogel, F. De the single-lead electrocardiogram for the detection of
Jongh, and A. Garde, "Airflow from nasal pulse obstructive sleep apnoea," IEEE Trans. Biomed. Eng.,
oximetry in the screening of obstructive sleep apnea," vol. 50, no. 6, pp. 686-96, Jun 2003.
IEEE Engineering in Medicine and Biology Society. , [30] A. Nishad, R. B. Pachori, and U. R. Acharya,
vol. 2019, pp. 2572-2575, 2019. "Application of TQWT based filter-bank for sleep
[17] T. Rosenwein, E. Dafna, A. Tarasiuk, and Y. Zigel, apnea screening using ECG signals," J. Ambient
"Breath-By-Breath Detection of Apneic Events for Intell. Hum. Comput., 2018.
OSA Severity Estimation Using Non-Contact Audio [31] P. de Chazal, T. Penzel, and C. Heneghan,
Recordings," in International Conference of the IEEE "Automated detection of obstructive sleep apnoea at
Engineering in Medicine & Biology Society, 2015. different time scales using the electrocardiogram,"
[18] H. Sharma and K. K. Sharma, "An algorithm for Physiol. Meas., vol. 25, no. 4, p. 967, 2004.
sleep apnea detection from single-lead ECG using [32] J. Pan and W. J. Tompkins, "A real-time QRS
Hermite basis functions," Comput. Biol. Med., vol. 77, detection algorithm," IEEE Trans. Biomed. Eng., vol.
pp. 116-24, Oct 1 2016. 32, no. 3, pp. 230-236, 1985.
[19] S. Babaeizadeh, D. P. White, S. D. Pittman, and S. H. [33] M. Längkvist, L. Karlsson, and A. Loutfi, "A review
Zhou, "Automatic detection and quantification of of unsupervised feature learning and deep learning
sleep apnea using heart rate variability," J. for time-series modeling," Pattern Recognit. Lett.,
Electrocardiol., vol. 43, no. 6, pp. 535-41, Nov-Dec vol. 42, pp. 11-24, 2014.
2010. [34] A. J. Camm et al., "Heart rate variability: standards of
[20] K. Li, W. Pan, Y. Li, Q. Jiang, and G. Liu, "A method measurement, physiological interpretation and
to detect sleep apnea based on deep neural network clinical use. Task Force of the European Society of
and hidden Markov model using single-lead ECG Cardiology and the North American Society of
signal," Neurocomputing, vol. 294, pp. 94-101, 2018. Pacing and Electrophysiology," 1996.
[21] E. Urtnasan, J. U. Park, E. Y. Joo, and K. J. Lee, [35] J. Sztajzel, "Heart rate variability: a noninvasive
"Automated Detection of Obstructive Sleep Apnea electrocardiographic method to measure the
Events from a Single-Lead Electrocardiogram Using autonomic nervous system," Swiss Med. Wkly., vol.
a Convolutional Neural Network," J. Med. Syst., vol. 134, no. 35-36, pp. 514-522, 2004.
42, no. 6, p. 104, Apr 23 2018. [36] A. Supratak, L. Li, and Y. Guo, "Feature extraction
[22] T. Van Steenkiste et al., "Portable Detection of with stacked autoencoders for epileptic seizure
Apnea and Hypopnea Events using Bio-Impedance of detection," in 2014 36th Annual International
the Chest and Deep Learning," IEEE J. Biomed. Conference of the IEEE Engineering in Medicine and
Health. Inf., Jan 20 2020. Biology Society, 2014, pp. 4184-4187: IEEE.
[23] S. S. Mostafa, F. Mendonca, F. Morgadodias, and A. [37] Y. Bengio, "Practical recommendations for
G. Ravelogarcia, "SpO2 based sleep apnea detection gradient-based training of deep architectures," in
using deep learning," in international conference on Neural networks: Tricks of the trade: Springer, 2012,
intelligent engineering systems, 2017. pp. 437-478.
[24] C. Zhang, K. C. Tan, H. Li, and G. S. Hong, "A [38] A. H. Khandoker, M. Palaniswami, and C. K.
Cost-Sensitive Deep Belief Network for Imbalanced Karmakar, "Support vector machines for automated
Classification," IEEE Trans. Neural Networks Learn. recognition of obstructive sleep apnea syndrome
Syst., vol. 30, no. 1, pp. 109-122, Jan 2019. from ECG recordings," IEEE Trans. Inf. Technol.
[25] C. Song, K. Liu, X. Zhang, L. Chen, and X. Xian, Biomed., vol. 13, no. 1, pp. 37-48, Jan 2009.
"An Obstructive Sleep Apnea Detection Approach [39] A. D. Ricke, R. J. Povinelli, and M. T. Johnson,
Using a Discriminative Hidden Markov Model From "Automatic segmentation of heart sound signals
ECG Signals," IEEE Trans. Biomed. Eng., vol. 63, no. using hidden Markov models," in Computers in
7, pp. 1532-42, Jul 2016. Cardiology, 2005, 2005, pp. 953-956: IEEE.
[26] M. J. Lawson, L. S. Heath, H. Zhao, and L. Zhang, [40] R. Fu, H. Wang, and W. Zhao, "Dynamic driver
"Optimizing a Cost Matrix to Solve Rare-Class fatigue detection using hidden Markov model in real
Biological Problems," in Proceedings of the driving condition," Expert Syst. Appl., vol. 63, pp.
International Conference on Bioinformatics & 397-411, 2016.
Computational Biology (BIOCOMP), 2011, p. 1. [41] S. Jain, E. Kotsampasakou, and G. F. Ecker,
[27] A. L. Goldberger et al., "PhysioBank, PhysioToolkit, "Comparing the performance of meta-classifiers—a
and PhysioNet: components of a new research case study on selected imbalanced data sets relevant
resource for complex physiologic signals," for prediction of liver toxicity," J. Comput-Aided.
Circulation, vol. 101, no. 23, pp. e215-e220, 2000. Mol. Des., vol. 32, no. 5, pp. 583-590, 2018.

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIM.2020.3017246, IEEE
Transactions on Instrumentation and Measurement

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 12

[42] P. Domingos, "Metacost: A general method for severity," Medicine (Baltim.), vol. 94, no. 4, p. e516,
making classifiers cost-sensitive," in KDD, 1999, vol. Jan 2015.
99, pp. 155-164. [55] X. Gong et al., "Correlation Analysis between
[43] H. Haibo and E. A. Garcia, "Learning from Polysomnography Diagnostic Indices and Heart Rate
Imbalanced Data," IEEE Trans. Knowl. Data Eng., Variability Parameters among Patients with
vol. 21, no. 9, pp. 1263-1284, 2009. Obstructive Sleep Apnea Hypopnea Syndrome,"
[44] T. Penzel, J. McNames, P. de Chazal, B. Raymond, A. PLoS ONE, vol. 11, no. 6, p. e0156628, 2016.
Murray, and G. Moody, "Systematic comparison of [56] Z. Wang, S. Yue, K. Song, X. Liu, and H. Wang, "An
different algorithms for apnoea detection based on Unsupervised Method for Evaluating Electrical
electrocardiogram recordings," (in English), Med. Impedance Tomography Images," IEEE Trans.
Biol. Eng. Comput., vol. 40, no. 4, pp. 402-407, Jul Instrum. Meas., vol. 67, no. 12, pp. 2796-2803, 2018.
2002. [57] D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P.
[45] C. Guilleminault, R. Winkle, S. Connolly, K. Melvin, Vincent, and S. Bengio, "Why does unsupervised
and A. Tilkian, "Cyclical variation of the heart rate in pre-training help deep learning?," J. Mach. Learn.
sleep apnoea syndrome: mechanisms, and usefulness Res., vol. 11, no. Feb, pp. 625-660, 2010.
of 24 h electrocardiography as a screening [58] J. Li, H. He, and L. Li, "CGAN-MBL for Reliability
technique," The Lancet, vol. 323, no. 8369, pp. Assessment With Imbalanced Transmission Gear
126-131, 1984. Data," IEEE Trans. Instrum. Meas., vol. 68, no. 9, pp.
[46] K. Kesper, S. Canisius, T. Penzel, T. Ploch, and W. 3173-3183, 2019.
Cassel, "ECG signal analysis for the assessment of
sleep-disordered breathing and sleep pattern," Med.
Biol. Eng. Comput., vol. 50, no. 2, pp. 135-44, Feb
2012.
[47] M. Bsoul, H. Minn, and L. Tamil, "Apnea MedAssist:
real-time sleep apnea monitor using single-lead
ECG," IEEE Trans. Inf. Technol. Biomed., vol. 15, no.
3, pp. 416-27, May 2011.
[48] Y. Li, W. Pan, K. Li, Q. Jiang, and G. Liu, "Sliding
Trend Fuzzy Approximate Entropy as a Novel
Descriptor of Heart Rate Variability in Obstructive
Sleep Apnea," IEEE J. Biomed. Health. Inf., vol. 23,
no. 1, pp. 175-183, Jan 2019.
[49] G. Surrel, A. Aminifar, F. Rincon, S. Murali, and D.
Atienza, "Online Obstructive Sleep Apnea Detection
on Medical Wearable Sensors," IEEE Trans. Biomed.
Circuits Syst., vol. 12, no. 4, pp. 762-773, Aug 2018.
[50] N. Sadr and P. De Chazal, "Automated detection of
obstructive sleep apnoea by single-lead ECG through
ELM classification," in Computing in Cardiology
2014, 2014, pp. 909-912: IEEE.
[51] S. Martin-Gonzalez, J. L. Navarro-Mesa, G.
Julia-Serda, J. F. Kraemer, N. Wessel, and A. G.
Ravelo-Garcia, "Heart rate variability feature
selection in the presence of sleep apnea: An expert
system for the characterization and detection of the
disorder," Comput. Biol. Med., vol. 91, pp. 47-58,
Dec 1 2017.
[52] A. R. Hassan and M. A. Haque, "An expert system
for automated identification of obstructive sleep
apnea from single-lead ECG using random under
sampling boosting," Neurocomputing, vol. 235, pp.
122-130, 2017.
[53] H. D. Nguyen, B. A. Wilkins, Q. Cheng, and B. A.
Benjamin, "An online sleep apnea detection method
based on recurrence quantification analysis," IEEE J.
Biomed. Health. Inf., vol. 18, no. 4, pp. 1285-93, Jul
2014.
[54] E. L. da Silva et al., "Heart rate detrended fluctuation
indexes as estimate of obstructive sleep apnea

0018-9456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Carleton University. Downloaded on August 26,2020 at 03:25:00 UTC from IEEE Xplore. Restrictions apply.