Biomedical Signal Processing and Control 71 (2022) 103125

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Detection of sleep apnea using deep neural networks and single-lead

ECG signals
Asghar Zarei a, Hossein Beheshti b, Babak Mohammadzadeh Asl a, *
a
Department of Biomedical Engineering, Tarbiat Modares University, Tehran, Iran
b
Department of Electrical Engineering, AmirKabir University of Technology, Tehran, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: Sleep apnea causes frequent cessation of breathing during sleep. Feature extraction approaches play a key role in
Sleep apnea the performance of apnea detection algorithms that use single-lead electrocardiogram signals. Handcrafted
ECG signal features have high computational complexity due to their large dimensions and are usually not robust. To cope
Deep learning
with the mentioned problems, in the current paper, an automatic feature extraction method is developed by
CNN
LSTM
combining the Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) recurrent network.
Also, the fully connected layers are utilized to distinguish apnea events from the normal segments. Then, the
apnea-hypopnea index (AHI) is applied to discriminate apnea subjects from healthy ones. Finally, in order to
assess the usefulness of the proposed method, some experiments are conducted on the publicly accessible Apnea-
ECG and UCDDB datasets. The results based on the sensitivity (94.41%), specificity (98.94%), and accuracy
(97.21%), indicate that our proposed method provides significant improvements compared to the other sleep
apnea detection methods. Our model also achieves an accuracy of 93.70%, sensitivity of 90.69%, and specificity
of 95.82% for UCDDB dataset. It can be inferred that using the deep-learning based algorithm for detecting apnea
patients would help physicians in making a decision more accurately.

1. Introduction Electromyogram (EMG), and Electroencephalogram (EEG) during sleep

[7]. To detect sleep apnea at least one night of the PSG recording is
Safe sleep affects the quality of our life and it is an important required. A PSG typically records at least 12 channels and is very
mechanism for maintaining mental health. When sleep apnea occurs, the expensive because it contains several costly electrophysiological tests
patient’s breathing repeatedly stops and starts due to obstruction of the [8,9]. So far, many studies have been performed on the diagnosis of
upper airways for at least 10 s during sleep [1,2]. The duration of apnea sleep apnea, using a single-lead signals, such as electrocardiogram
events is often 10 to 30 s and can occur hundred times during sleep at [10–13], electroencephalogram [14], peripheral oxygen saturation
one night. Sleep apnea can cause complete obstruction (Apnea) or par­ (SpO2) [15], or nasal airflow [16]. Some of the studies such as [17,5,9]
tial (Hypopnoea) obstruction of the upper airways [1,3]. Obstructive proved that the ECG signals can be used to detect respiratory events.
sleep apnea (OSA) is the most common type of sleep apnea [4]. In the Moreover, the results of [18,19] showed that autonomic activity during
world, roughly 4% of men and 2% of women suffer from OSA [5]. apnea events and their impact on the cardiovascular system can be
Failure to diagnose sleep apnea promptly can lead to insomnia, vehicle measured using the heart rate variability (HRV) signal. Moreover, real-
accidents, and academic under-achievement in children and time detection of OSA in patients whose physiological signals are
adolescents. constantly monitored practically is impossible. As a result, developing a
OSA can cause over-drowsiness (excessive sleepiness) throughout low cost computer-assisted OSA screening algorithm is necessary.
the day, car crashes due to sleep deprivation, mood changes, memory To tackle these problems, in literature, two categories of studies
loss, heart disease, and so on [6]. Clinically, sleep apnea is diagnosed including feature engineering-based techniques and deep learning-
using nocturnal polysomnography (PSG) recordings of the patient. The based methods have been investigated. In feature engineering-based
PSG monitors many physiological activities, such as muscle activity algorithms, the authors used handcrafted features and combined them
(EMG), Electrocardiogram (ECG), eye movements (EOG), with different classifiers to improve apnea detection accuracy. For

* Corresponding author.
E-mail address: [email protected] (B.M. Asl).

https://doi.org/10.1016/j.bspc.2021.103125
Received 6 July 2021; Received in revised form 25 August 2021; Accepted 30 August 2021
Available online 8 September 2021
1746-8094/© 2021 Elsevier Ltd. All rights reserved.
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

example, Atri and Mohebbi extracted useful information from the HRV automatically, and recognize objects in its output by the neural network.
and ECG derived respiratory (EDR) signals using the higher-order Although attempts have been made previously on the detection of the
spectrum based technique. Then, they classified the feature vectors OSA via CNN and Long Short-Term Memory (LSTM) networks [33], the
using the least-square support vector machine (LS-SVM) classifier [20]. algorithm presented here is different from the existing techniques.
Tripathy et al. utilized the intrinsic band functions algorithm to Erdenebayar et al. assessed the effectiveness of the one dimensional (1D)
decompose both the HRV and EDR signals into different subbands. Then, CNN, two dimensional (2D) CNN, and LSTM networks individually [33].
they calculated different features and fed them into the kernel extreme The authors in [33] excluded half of the existing dataset (the withheld
learning machine classifier [21]. Nishad et al. used the centered cor­ set). In the current work, both of the released and withheld sets have
rentropies and tunable-Q wavelet transform (TQWT) based filter-bank been used. This study has combined the 2D-CNN and LSTM networks to
to classify one-minute ECG segments [22]. Sharma et al. extracted extract more informative features from single-lead ECG signals at the
non-linear features from different sub-bands of ECG signal, using the same time. There are two main reasons for using this architecture:
biorthogonal antisymmetric wavelet filter bank, and classified them into
healthy and OSA groups using the LS-SVM [23]. Li et al. introduced a • Extracting deep features and improving the classification accuracy; If
novel OSA detection index using the sliding trend fuzzy approximate diverse layers are used, they can extract more deep features. Also, the
entropy (SlTr-fApEn), empirical mode decomposition (EMD) method, CNN-LSTM is a common architecture [37] for time-series
and the HRV signal [24]. Zarei and Asl presented a novel algorithm applications.
based on the discrete wavelet transform and non-linear features to • Decreasing the computational complexity; High computational
detect apnea events using ECG signals [25]. In another study, they complexity is encountered when only the LSTM layers are used.
introduced a new automatic OSA detection technique to classify one-
minute ECG segments. They extracted non-linear features such as The rest of the current study is continued as follows. All the materials
fuzzy entropy from HRV and EDR signals [2]. Gonzalez et al. calculated and methods are explained in Section 2. The results are given in Section
different features, namely Cepstrum coefficients and detrended fluctu­ 3. Finally, discussion and conclusion are presented in Sections 4 and 5,
ation analysis from the HRV signal, and fed them into different classi­ respectively.
fiers for diagnosing sleep apnea [26]. The authors in [27] implemented a
new apnea recognition framework based on the autoregressive 2. Materials and methods
modeling, spectral autocorrelation function, and ECG signals. They
employed the random forest model to classify extracted features. Faal 2.1. Proposed framework
et al. introduced a new apnea detection method using ECG signals and
statistical modeling. They applied different classifiers on the features In the present study, we developed an automatic OSA classification
calculated from the time domain [28]. Fatimah et al. extracted different approach based on deep neural networks and ECG signals. The main idea
handcrafted features in the frequency domain and investigated the of this study is developing a fully automatic (or an end-to-end) deep
capability of the Fourier decomposition approach in OSA detection [29]. learning-based OSA detection method. In the current study, all the
Mostafa et al. [30] performed two different techniques to choose the best feature extraction and classification processes have been done using
subset of features. In their study, the selected features were tested with different types of deep learning layers. Also, in order to eliminate the
different classification methods. need of additional tuning step, tuning the optimal value of different
Also, some other studies have been focused on deep learning-based parameters of classifiers, deep learning is used instead of conventional
techniques. For example, Li et al. worked on the recognition of the classifiers. In the pre-processing step, the noisy segments are detected
OSA segments using the hidden Markov model (HMM), deep neural and eliminated by the weight calculation algorithm. Then, the most
network, and ECG signals. They employed the sparse auto-encoder to prominent features are automatically extracted using an end-to-end
extract features and then classified them using the SVM and artificial deep learning structure. The schematic design of the proposed tech­
neural networks (ANN) [31]. Choi et al. developed an automatic apnea nique in this study is demonstrated in Fig. 1. As it is represented in this
detection approach using convolutional neural networks (CNN) and PSG figure, in our proposed method, the CNN and LSTM networks are used to
signals [32]. The authors in [33] investigated the effectiveness of automatically calculate the spatial and temporal features, respectively.
different deep learning structures in apnea detection application. They Also, the fully connected layers are applied to identify apnea events. At
employed the CNN and RNN based structures to automatically recognize the end of this stage, the AHI index is employed to diagnose apnea
apnea events using ECG signals. Dey et al. introduced a supervised apnea patients.
detection approach using the CNN structure and ECG signal [34]. Novak
et al. used the HRV signal and LSTM network to diagnose apnea patients 2.2. Datasets
[35]. Pathinarupothi et al. presented a new apnea detection framework
by employing the LSTM-RNN structure on instantaneous heart rates 2.2.1. Apnea-ECG dataset
[36]. It seems that two main problems remain unsolved in many studies The Apnea-ECG dataset contains divided into the released and the
related to this field of study. Firstly, the performance of previous OSA withheld sets. Each of the released and withheld sets contains 35 re­
recognition methods strongly depends on the feature calculation ap­ cordings. Expert physicians have labeled each of the one-minute ECG
proaches. Furthermore, the performance of the traditional handicraft segments as the normal or apnea ones. The length of the recordings
feature extraction methods depends on the experience of specialists and varies between 401 min to nearly 578 min. The age of the subjects (57
their prior knowledge of physiological signals. Secondly, notwith­ men, 13 women) ranged from 27 to 63 years and weights from 53 to 135
standing the researchers applied different classification models, they kg. The total number of released-set segments is 17045 one-minute
have not focused on improving the performance of the classifiers [31]. segments. Also, the withheld-set consists of 6550 apnea, and 10718
Therefore, developing algorithms with less dependency on feature en­ normal segments. In the present study, the subjects with AHI < 5 were
gineering is necessary to expand the application range of computer- considered as healthy subjects otherwise, they were defined as OSA
aided techniques. subjects. Furthermore demographic information and physiological pa­
To deal with these problems, an automatic feature extraction and rameters of subjects are summarized in Table 1. More details can be
classification technique is developed using deep neural networks. One of found in [38].
the goals of deep learning algorithms is to achieve feature learning in an
unsupervised manner. Deep learning is a fully trainable system that can
start with the raw data, extract the most informative features

2
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

Fig. 1. Flowchart of proposed sleep apnea detection algorithm based on CNN and LSTM networks.

Table 1
Demographic information and physiological parameters of poupulation. BMI: body mass index, AI: Apnea index, HI: hypopneas index, AHI: apnea-hypopnea index,
TRT: total recording time, AT: Apnea time, Non-AT: Non-Apnea time.
Parameters Released Set Withheld Set

Normal (mean ± std) Apnea (mean ± std) Normal (mean ± std) Apnea (mean ± std)

Age (years) 35.5 ± 6.32 52.17 ± 6.09 32.37 ± 6.09 49.08 ± 9.46
( )
BMI kg/m2 21.41 ± 2.20 31.41 ± 5.34 21.10 ± 1.22 31.22 ± 5.27
Height (cm) 177.5 ± 6.17 176.17 ± 3.23 172.27 ± 9.61 176.33 ± 4.18
Weight (kg) 67.58 ± 8.78 97.43 ± 16.46 62.73 ± 7.16 97.13 ± 16.63
AI (events/h) 0.68 ± 0.2 32.26 ± 20.89 0.06 ± 0.12 32.78 ± 24.85
HI (events/h) 0.03 ± 0.08 11.27 ± 12.90 0.0 ± 0.0 7.23 ± 6.50
AHI (events/h) 0.10 ± 0.21 43.53 ± 21.87 0.06 ± 0.12 40.01 ± 25.42
TRT (min) 472.5 ± 29.38 498.04 ± 30.23 461.73 ± 25.83 509.33 ± 20.68
AT (min) 3.42 ± 5.66 281.43 ± 136.22 2.27 ± 3.50 271.96 ± 144.61
Non-AT (min) 469.08 ± 30.64 216.61 ± 124.67 459.45 ± 25.71 237.38 ± 141.07

2.2.2. St. Vincent’s University Hospital/ University College Dublin dataset segments (1717 segments of Apnea-ECG dataset) were eliminated from
(UCDDB) the experimental analysis. Further details can be found in [25].
In the current work, this dataset has also been used to validate the
performance of the proposed OSA detection method. There are 25 full
overnight polysomnograms with simultaneous three-channel Holter 2.4. Convolutional neural network
ECG in this dataset and it was recorded from adult subjects with sus­
pected sleep-disordered breathing. The weights and ages of the subjects Convolutional neural networks (CNN) are one of the most common
vary between 59.8–128.6 kg and 28–68 years, respectively. Also, the and highly efficient techniques that are widely used in various signal and
AHI values are in the range of 1.7–90.9. Frequency sampling of all the image processing applications. In recent years, this method has attracted
ECG recordings is 128 Hz. For more details the readers are referred to much attention in analyzing physiological events, such as sleep stage
[39]. scoring [40], sleep apnea detection [33], or seizure detection [41]. CNN
networks consist of neurons with weights and configurable biases. Each
neuron receives a part of the ECG segment as an input and then calcu­
2.3. Pre-processing lates the multiplication of the weights in the input segment values. The
output of this process is converted to class labels using a non-linear
At first, the Chebyshev bandpass filter with the cutoff frequencies of activation function. Also, the network provides a differentiable score
0.5 Hz and 48 Hz is designed to suppress the baseline wandering and function, with inputs on one side and scores on the other side.
power line interference noises. The ECG signals are then divided into In general, CNN networks consist of three parts: convolutional,
one-minute (for Apnea-ECG dataset) and 30-s (for UCDDB dataset) pooling, and fully connected parts. The convolutional part consists of
segments. Afterward, the noisy segments are eliminated using an auto­ filters called kernels and convolve with the ECG segments. This process
matic weight calculation algorithm. This algorithm measures the simi­ produces the feature vector. Afterward, the max-pooling operation is
larity of the autocorrelation functions (ACF) of the segments with the employed to the feature vector. In this step, the values of each group of
time lags of 60 s. At the output of this process, a similarity matrix is neighbors are combined with a summarization function. The summari­
obtained, which is a symmetric matrix. At the next stage, the weights zation function performs like a low-pass operator, and common choices
vector will be obtained by summing each of the columns of the similarity for it are either maximal or average. The max-pooling layer reduces the
matrix. The threshold value of 0.8 is used to recognize and eliminate the computation burden and prevents overfitting by selecting the largest
noisy segments. This threshold value is selected by try and error after value of each feature. The fully connected part often is used as the last
conducting several experiments. In the this stage, less than 5% of the part of the CNN networks and classify the output of the max-pooling

3
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

layer (the feature vectors). series because it uses the gating mechanism [43].

2.5. Recurrent neural network 2.7. The proposed model for OSA recognition

A Recurrent neural network (RNN) commonly measures the corre­ In the literature, most of the researchers introduced handcraft
lation between the values of signals during the time. In many practical feature extraction solution. In these methods, features are extracted
situations, such as signal processing, there is a meaningful dependency manually and their performance is highly dependent on the experience
between different sequences. The RNN networks have a memory unit of specialists. To cope with the mentioned problem, an automatic end-
that can process sequences of the input signal [42]. The RNNs-based to-end deep learning structure was utilized. Moreover, the proposed
structures can extract temporal features and measure temporal automatic structure can extract the most informative features and in­
dependencies. crease the classification accuracy. As indicated in Fig. 1, our proposed
model consists of two main parts. The first part is the feature extraction
2.6. Long Short Term Memory contains four 2D CNN and three LSTM layers. These layers are utilized to
calculate the most prominent features of the ECG signals. In the second
Theoretically, RNNs show excellent performance in maintaining the part, the classification phase, the fully connected layer is applied to the
correlation of input signals whereas in practical application, they feature vector to discriminate apnea segments from normal ones. Also,
encounter a lot of restrictions to measure these time dependencies. To to keep away from the overfitting problem, the dropout layers were
cope with these challenges, the Long Short Term Memory (LSTM) applied. This method is a regularization approach for deep networks.
structure is proposed which is an extended version of RNNs [42]. The According to this solution, at each epoch, some of the input units
LSTM structures capture signal variations using long and short-term randomly were set to zero during the training time [44]. Mathemati­
memory. An example of a common LSTM cell is depicted in Fig. 2. As cally, the output of the LSTM cell can be calculated as follows:
demonstrated in this figure, each cell composed of three gates; an input ⎛ ⎞ ⎛ ⎞
sigmoid [(
gate, an output gate, and a forget gate. Each of the mentioned gates can ⎜ f ⎟ ⎜ sigmoid ⎟ xt
) ]
be defined as follows:
⎜ ⎟=⎜ ⎟
⎝ o ⎠ ⎝ sigmoid ⎠ ht− 1 .W (6)
Forget gate: tanh
( )
f t = σ g Wf xt + Uf ht− 1 + bf (1)
where W demonstrates the weight matrix which its dimension is 2k × 4k
Input gate: and k is the dimension of the input signal. Also, the dropout variation is
formulated as follows:
it = σ g (Wi xt + Ui ht− 1 + bi ) (2) ⎛ ⎞ ⎛ ⎞
sigmoid [(
Cell state update: ⎜ f ⎟ ⎜ sigmoid ⎟ xt × zx
) ]
⎜ ⎟=⎜ ⎟
⎝ o ⎠ ⎝ sigmoid ⎠ ht− 1 × zh .W (7)
Ct = f t ⊙ Ct− 1 + it ⊙ σ c (Wc xt + Uc ht− 1 + bc ) (3)
tanh
Output gate:
where zx and zh are the random filter masks which drop some weights in
Ot = σ g (Wo xt + Uo ht− 1 + bo ) (4)
each iteration. In order to tune hyperparameters, the grid search method
Output: has been utilized to achieve an optimal structure with the highest ac­
curacy and lowest computational complexity. To this end, the number of
ht = Ot ⊙ σ h (Ct ) (5)
LSTM and CNN layers was changed from 1 to 6 layers. Different values
including 32, 64, 128, and 256 were used for the number of cells and
where both of the σh and σ c are hyperbolic tangent functions. It is
filters in LSTM and CNN layers, respectively. Also, other parameters
noteworthy that weight matrices, W and U, are optimized during
such as learning rate, activation functions, and the kernel size were
training. In this expression, the output and state of the previous LSTM
determined empirically. Finally, some of the optimal parameters of our
cell are defined using the Ct− 1 , and ht− 1 , respectively. The bias vector
proposed structure are summarized in Table 2.
and input vector are denoted with b and xt , respectively. Also, σg is the
sigmoid function. Finally, the operator ⊙ denotes the Hadamard prod­
uct. In an LSTM cell, weights can be updated using the Ct− 1 and it . The 2.8. Experimental setting
LSTM cell can capture the long interval dependency of the input time
As mentioned before, in this study, an automatic OSA detection
technique was presented using CNN and LSTM networks. In the training
phase, the effect of parameters of our proposed deep structure were
investigated using the released-set (Fig. 3). In order to extract spatial
features, a 2D CNN model was employed. Therefore, one-dimensional

Table 2
The learning parameters of the proposed deep learning structure.
Task Layers Leaning Parameters

Feature CNN layers number of filters = 128, kernel size = 2 × 2,

Extraction activation = ReLU
Max-Pooling Pool size = 2 × 2
LSTM layers number of cells = 256

Classifier Optimizer Adam (learning rate = 0.0005, decay = 1e-6)

Loss function Binary cross-entropy
Fig. 2. A typical architecture of a LSTM cell. A LSTM block typically has a Activation ReLU and Sigmoid (lastest layer)
functions
memory cell, input gate (it ), output gate (Ot ), and a forget gate (f t ) in addition
Number of units 64
to the hidden state (ht ) in traditional RNNs.

4
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

Fig. 3. Dataset division for per segment and per-recording classification tasks.

ECG segments were converted to two-dimensional segments. Afterward, Tensorflow framework on GeForce GTX 1080Ti graphic card with 12 GB
temporal features were calculated using the LSTM networks. To remove memory.
any simple dependencies between the neurons, i.e., to cope with the Different criteria such as precision, specificity, F1-Score, sensitivity,
overfitting problem, the dropout function with the probability of 0.2 was and accuracy were used for performance evaluation of our proposed
applied after some layers. This method randomly sets 20 percent of input technique.
units to zero during the training time. The optimal parameters of feature
extraction and classifier layers were presented in Table 2. As detailed in 3. Results
Table 2, the Adam optimizer and binary cross-entropy loss function were
applied to optimize the values of various parameters [45]. The perfor­ In the this work, a novel end-to-end learning structure has proposed
mance of the proposed model (accuracy and loss) along 200 epochs was to diagnose OSA. The model performance was evaluated through
depicted in Fig. 4. In this figure, the number of epochs was changed from different aspects including classifier parameter tuning, per segment
1 to 200 epochs. As demonstrated in Fig. 4 (C), although the accuracy classification using different criteria, and subject-by-subject (per-
increases (or the loss value decreases in Fig. 4 (D)) with increasing the recording) OSA classification accuracy.
number of epochs in the training time, in the test phase the accuracy
(Fig. 4 (A)) and loss function (Fig. 4 (B)) converge to a constant value
after 50 epochs. Finally, the details of the CNN–LSTM based model are 3.1. Tuning classifier parameters
presented in Table 3. Our model was implemented based on the
In the present section, the effect of setting various classifier

Fig. 4. Performance of the proposed network along 200 epochs. (A) and (B) are the accuracy and loss values of the test phase, respectively. Also, (C) and (D) are the
accuracy and loss values of the training phase, respectively.

5
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

Table 3 previous methods on the Apnea-ECG dataset is listed in Table 7. The

The detailed specifications of CNN-LSTM layers. results show that (Table 7) our developed OSA detection method per­
Layer Type Number of filter/cell/ Kernel Activation forms better than existing studies in terms of accuracy, sensitivity and
unit size function specificity.
1 CNN 128 2× 2 ReLu Statistical power or the power of a hypothesis test is the probability
2 CNN 128 2× 2 ReLu that the test correctly rejects the null hypothesis. That is the probability
3 MAX- – 2× 2 – of a true positive result. It is only useful when the null hypothesis is
Polling rejected. The higher the statistical power for a given experiment, the
4 Dropout – – – lower the probability of making a Type II (false negative) error. Exper­
5 CNN 128 2× 2 ReLu imental results with too low statistical power will lead to invalid con­
6 MAX- – 2× 2 – clusions about the meaning of the results [60]. To assess the statistical
Polling
power of the presented method, the model generated by data of 35
7 Dropout – – –
8 CNN 128 2× 2 ReLu randomly selected recordings was applied to the data of remaining re­
9 MAX- – – – cordings (35 recordings) as test data. This procedure was repeated 10
Polling times and the results are presented in Table 8. Also, Fig. 5 illustrates the
10 Dropout – – – statistical power of the proposed method for a range of accuracy. As
11 Flatten – – –
indicated in Fig. 5, the statistical power of our proposed method for a
12 LSTM 256 – –
13 LSTM 256 – –
true mean (the average accuracy obtained by previous runs, i.e.,
14 LSTM 256 – – 97.21%) is one. Therefore, we can see that the proposed technique has a
15 Dropout – – – great performance in terms of statistical power analysis.
16 Dense 64 – ReLu Execution times of the proposed algorithm on the Apnea-ECG and
17 Dense 1 Sigmoid
UCDDB datasets are reported in Table 9. According to this table, it is
–

obvious the computational cost of the proposed OSA screening scheme is

parameters, the number of neurons and layers, on the accuracy of the also significantly low.
proposed model is assessed. As reported in Table 4, four different DNN
structures with different parameters were examined. The results show 3.3. Per-recording classification
that the model becomes overfit by increasing hidden layers. So, the
DNN2 which has one hidden layer with 64 neurons was employed for the In the next step, the AHI index was utilized to separate the healthy
classification task. subjects from the OSA patients. This criterion is defined as follows:
Number of apnea segments
3.2. Per-segment classification AHI Real = × 60. (8)
Total sleep time (in minutes)

The withheld-set was utilized to assess the effectiveness of the According to this criterion, the subjects with AHI more than five
introduced model. In this step, the ECG segments fed into the combi­ were classified as apneic patients. Also, the cases who have the AHI
nation of CNN and LSTM networks to automatically extract features. value less than 5 were considered as healthy subjects. The performance
Then, the extracted features were fed into fully connected layers to of the presented technique was assessed using the withheld-set which
classify one minute ECG segments. The result of the optimized CNN- contains 35 recordings. A summary of the per-recording OSA detection
LSTM structure for per segment classification were reported in performances of various researches along with the proposed method was
Table 5. In accordance with this table, our developed CNN-LSTM listed in Table 10. In accordance with this table, our proposed technique
structure provides the F1-score of 0.96, PPV of 94.42%, accuracy of outperforms most of the other studies by classifying all subjects correctly
97.21%, sensitivity of 94.41%, and specificity of 98.94%. It is obvious (an accuracy of 100%). Also, the mean absolute error (MAE) metric was
that the presented technique can improve classification performance applied to better evaluate the pre-recording classification performance.
using the combination of CNN and LSTM networks (See Table 6). This metric measures the gap between the estimated AHI and real AHI
For automatic classification of one-minute ECG segments, various and can be calculated as follows:
deep learning structures including CNN, LSTM and CNN-LSTM were
1 ∑N ⃒ ⃒
examined to find the best model. Table 11 summarizes the results of the MAE = ⃒Estimatedi − Reali ⃒
AHI AHI (9)
N i=1
designed methods.
Different deep learning structures including CNN, LSTM, and CNN-
where N is the number of recordings. Our proposed deep learning-based
LSTM were examined to find the best model for the automatic classifi­
method provides the MAE of 2.93 for per-recording OSA detection.
cation of one-minute ECG segments. The results of the designed tech­
niques are summarized in Table 11. According to Table 11, in the test
4. Discussion
set, the highest sensitivity (95.35%) is obtained using the CNN model.
The results indicated that the designed LSTM model has the worst per­
In the present paper, a novel unsupervised feature extraction
formance among the other models. The CNN-LSTM model outperforms
framework has been developed to recognize OSA patients using CNN
the other designed models by providing an accuracy of 97.21%, sensi­
and LSTM networks. In the proposed method, the spatial and temporal
tivity of 94.41%, and specificity of 98.94%. This is because the CNN and
features were automatically extracted using the 2D-CNN and LSTM,
LSTM networks extract the spatial and temporal features at the same
respectively. Moreover, a fully connected layer was adopted to classify
time. In per segment classification, the highest accuracy value of
apnea events and normal segments. As a result, a fully automatic OSA
detection algorithm was presented.
Table 4
The learning parameters of different layers of classifier.
4.1. Learning representation and temporal dependency
DNN1 DNN2 DNN3 DNN4

Hidden layer 1 32 unit 64 unit 32 unit 64 unit As shown in Table 7, most of the presented OSA detection methods
Hidden layer 2 – – 32 unit 64 unit are developed based on the handcrafted features. The results showed
Accuracy 96.41 % 97.21 % 96.16 % 96.73 %
that feature extraction techniques play a key role in the performance of

6
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

Table 5
The result of the optimized CNN-LSTM structure for per segment classification.
Approach Criteria

TP TN FP FN Sens (%) Spec (%) Acc (%) F1-score PPV (%)

CNN-LSTM 6104 10341 361 110 94.41 98.94 97.21 0.96 94.42

Table 6 Table 7
Performance evaluation of the deep learning approaches on minute-by-minute Comparison of segment-by-segment based classification results between our
classification. method and previous works on the Physionet Apnea-ECG database.
Approach Criteria Reference Method Sens Spec Acc Signal
(%) (%) (%)
Sens (%) Spec (%) Acc (%)
[13] HRV analysis in the 76.7 88.4 86.2 Lead-II
CNN 95.35 96.03 95.77
frequency domain ECG
LSTM 89.70 95.58 93.33
[1] RQA Statistics and SVM 86.37 83.47 85.26 Lead-II
CNN-LSTM 94.41 98.94 97.21
ECG
[25] Entropy-based features, 91.74 93.75 92.98 Lead-II
wavelet transform and ECG
apnea detection methods. The performance of traditional hand-crafted SVM
feature extraction methods is heavily dependent on the experts’ [17] PCA of QRS complex and 84.71 84.69 84.74 Lead-II
knowledge from different physiological signals and exclusive subjects’ HRV features ECG
characteristics. Further, some studies that used automatic feature [46] HRV, Hermite basis 79.5 89.6 84.7 Lead-II
functions, SVM and LS- ECG
extraction methods, have just considered time domain correlation in SVM
signals and they have not conducted spatial dependencies in ECG sig­ [11] QRS detection, HRV, 82.6 88.4 86.2 Lead-II
nals. In the current study, a new deep learning-based technique was and EDR-based features ECG
employed to capture the spatial and temporal dependencies using 2D- + HMM
[26] Filterbank, Cepstrum, 81.45 86.82 84.76 Lead-II
CNN and LSTM networks, respectively. The reported results in Ta­
DFA and QDA ECG
bles 7 and 10 imply that the proposed model provided outperformance [47] TQWT and RUSBoost 87.58 91.49 88.88 Lead-II
in comparison with other methods. ECG
[31] DNN + HMM 88.9 88.4 83.8 Lead-II
ECG
4.2. Comparison and summary [21] HRV, EDR, FDM, and 78.02 74.64 76.37 Lead-II
KELM ECG
In the present section, we compared the performance of our intro­ [22] TQWT, centered 90.95 93.91 92.78 Lead-II
correntropys, and ECG
duced technique with approaches that were conducted on the Physionet
random forest
Apnea-ECG dataset. The comparison of per segment classification per­ [23] BAWFB, Entropy-based 90.87 88.88 90.11 Lead-II
formance results are summarized in Table 7. As indicated in this table, features, LS-SVM ECG
our developed approach using Deep Learning can provide an accuracy of [2] HRV, EDR, Non-linear 91.52 94.36 93.26 Lead-II
97.21%, sensitivity of 94.41%, and specificity of 98.94%. The results features, GentleBoost ECG
classifier
showed that our developed OSA detection technique has a better per­
[48] NIG parameters + 81.99 90.72 87.33 Lead-II
formance compared to previous algorithms. The main drawback found TQWT + AdaBoost ECG
in most of the previous studies is that their performance strongly de­ [49] Statistical Features and 85.20 82.79 83.77 Lead-II
pends on feature extraction methods. Feature calculation and feature ELM ECG
engineering are the most important and time-consuming parts of the [50] statistical and spectral 84.14 86.83 85.97 Lead-II
features + bootstrap ECG
machine learning algorithms.Feature engineering based techniques try aggregating
to improve performance by mapping high-dimensional training data to [51] Statistical Features and – – 83.77 Lead-II
low-dimensional feature space. Whereas, in deep learning algorithms, ELM ECG
feature learning is done in an unsupervised and automatic manner. To [52] Statistical Features and – – 83.77 Lead-II
ELM ECG
accomplish these objectives, in this study, the 2D-CNN and LSTM net­
[53] Statistical and Spectral – – 85.37 Lead-II
works were applied to automatically extract the spatial and temporal Features + RUSBoost ECG
features, respectively. Then, a fully connected layer was applied to [54] 10-s ECG segment + 96.1 96.2 96.1 Lead-II
identify apnea events. CNN ECG
In Table 10, we compared per-recording classification performance [55] Fused images + CNN 92.3 92.6 92.4 Lead-II
ECG
of our developed technique with the OSA detection algorithms using the [56] RR interval + Multiscale 89.8 89.1 89.4 Lead-II
same dataset. As reported in Table 10, in per-recording classification Deep Neural Network ECG
phase, our deep learning-based algorithm detected all subjects correctly [57] ECG signal + 1D Deep 81.1 90.0 87.9 Lead-II
and provided good results (an accuracy of 100% and specificity of CNN model ECG
[58] RR interval + FSSAE + 86.2 84.4 85.1 Lead-II
100%). Based on the obtained results, it can be concluded that the
Softmax-HMM ECG
presented OSA recognition algorithm outperforms most recent studies. [59] ECG signal, EMD, CWT, 94.30 94.51 94.30 Lead-II
Moreover, in Table 11, the performance of our developed deep structure and SCNN ECG
was compared with previous deep learning-based studies. According to Proposed 2D-CNN + LSTM 94.41 98.94 97.21 Lead-II
this table, the proposed deep learning method showed better classifi­ Method networks ECG
cation performance compared to other techniques.
Also, some other studies such as [6,17,31] have obtained remarkable
outcomes. Although high classification results have been obtained in performances in the practical conditions in which physiological signals
[6], they have excluded 20 recordings with low-quality signals during have noisy nature. Some other studies, such as [17], have reached 100%
per-recording classification. Their method may provide low accuracy, but their method is not completely automatic. They have used

7
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

Table 8 Table 10
The results of 10 performances on the test data (data of 35 recordings). Comparison of subject-by-subject based classification results between our
Run Sens (%) Spec (%) Acc (%)
method and previous works on the Physionet Apnea-ECG database.
Reference Method Acc Sens Spec Signals
1 th 97.32% 97.59% 97.11%
(%) (%) (%)
2 th 96.57% 96.12% 96.91%
[56] RR interval + Multiscale 100 100 100 Lead-II
3th 95.51% 96.97% 94.42%
Deep Neural Network ECG
4th 95.95% 97.32% 94.93% [61] RR interval + LeNet-5 97.1 100 91.7 Lead-II
5th 94.11% 90.76% 96.61% ECG
[57] ECG signal + 1D Deep 97.1 95.7 100 Lead-II
6th 92.80% 83.92% 97.39%
CNN model ECG
7 th 93.08% 90.91% 95.44% [58] RR interval + FSSAE + 97.1 95.7 100 Lead-II
8th 95.70% 95.47% 95.87% Softmax-HMM ECG
[59] ECG signal, EMD, CWT, 100 100 100 Lead-II
9th 97.33% 97.80% 96.99%
and SCNN ECG
10th 96.88% 97.86% 97.90% [17] PCA of QRS complex and 100 100 100 Lead-II
HRV features ECG
Mean 95.53% 94.47% 96.36% [9] Handcrafted features + 95.7 100 87 Lead-II
Std 1.6673% 4.3376% 1.0794% machine learning ECG
[46] HRV, Hermite basis 97.14 95.8 100 Lead-II
functions, SVM and LS- ECG
SVM
Power function for sample size of 10 [11] QRS detection, HRV, and 97.1 95.8 100 Lead-II
1 EDR-based features + ECG
HMM
0.9 [31] DNN + HMM 100 100 100 Lead-II
Statistical power of the proposed method

ECG
0.8
[25] Entropy-based features, 95.71 95.83 95.66 Lead-II
0.7 wavelet transform and ECG
SVM
0.6
This ECG signal þ CNN-LSTM 100 100 100 Lead-II
work ECG
0.5

0.4

0.3 Table 11
Performance comparison with previous deep learning-based studies.
0.2
Study Acc (%) Sens (%) Spec (%) Subject Method
0.1 [36] 98 – – 35 LSTM
[62] 97.8 – – 35 LSTM
0
92.5 93 93.5 94 94.5 95 95.5 96 96.5 97 97.5 [34] 98.9 97.8 99.2 35 CNN
True mean (for accuracy metric) [32] 96.6 81.1 98.5 179 CNN
[33] 99 99 99 86 GRU
[63] 97.14 95.8 100 35 LSTM
Fig. 5. Statistical power of the proposed method based on accuracy and sample
[35] 82.1 85.5 80.1 33 LSTM
size of 10.
[61] 97.1 100 91.7 – LeNet-5
[56] 100 100 100 70 MDNN

Table 9 Our method 100 100 100 70 LSTM + CNN

Execution time of the proposed algorithm on the Apnea-ECG and UCDDB
datasets.
proposed structures. For example, a combination of 2D-CNN and LSTM
Dataset Time networks has been used in our proposed method to diagnose OSA events.
(second)
However, in the mentioned references ([56,61]), the 1-D CNN network
One-minute ECG segment with 6000 samples from the Apnea-ECG 0.0684 has been applied to detect apnea events. In summary, we can see that the
database
optimization process, feature extraction process, and classification
30-s ECG segment with 3840 samples from the UCDDB database 0.0353
process are different between the mentioned studies.
Furthermore, the performance of the proposed deep learning-based
handcrafted features for classification. However, our proposed algo­ OSA detection method was validated on UCDDB database. In our ex­
rithm obtained the same results (for per-recording classification) by periments we found that decreasing the number of the ECG segments in
employing the LSTM and 2D-CNN networks. According to the results the input leads to decreasing the performance of the deep learning-based
presented in Tables 7 and 10, the developed approach has a better techniques. Because the number of segments in the UCDDB database is
performance than the recently presented methods. less than the Apnea-ECG database, the 80–20 ratio was used to evaluate
The authors in [56] proposed a novel OSA detection technique using the performance of the proposed method on the UCDDB database.
RR intervals, multiscale dilation attention 1-D convolutional neural Consequently, to evaluate the performance of our proposed technique
network (MSDA-1DCNN) and a weighted-loss time-dependent (WLTD) 80% of recordings were randomly selected for training the model and
classification model. Also, in reference [61], the authors have developed the remaining (20% of the recordings) were used for evaluating the
a new OSA recognition system using the RR intervals, CNN, and LeNet-5 performance. This process was repeated for ten times. The average of
networks. There are a lot of differences between our study and the results is presented in Table 12. According to Table 12, the proposed
mentioned references. In our proposed method, a fully automatic OSA method obtained an average accuracy of 93.70%, an average sensitivity
detection technique has been developed using the ECG signals, 2-D CNN, of 90.69%, and an average specificity of 95.82%. In sum, it can be
and LSTM structures. Although all the mentioned studies (the proposed concluded that the proposed OSA detection technique yields better
method, [56,61]) are based on CNN, there are many differences in the performance than prior works in the literature.

8
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

Table 12 Software, Writing - original draft, Investigation. Babak Mohammad­

Performance of the proposed method on zadeh Asl: Writing - review & editing, Methodology, Supervision.
UCDDB database.
Measure Value Declaration of Competing Interest
Accuracy 93.70%
Sensitivity 90.69% The authors declare that they have no known competing financial
Specificity 95.82% interests or personal relationships that could have appeared to influence
the work reported in this paper.

In summary, based on the results, it can be concluded that the

References
combination of CNN and LSTM methods can improve the accuracy of
apnea diagnosis. In contrast to the traditional hand-crafted feature [1] H.D. Nguyen, B.A. Wilkins, Q. Cheng, B.A. Benjamin, An online sleep apnea
extraction methods, the performance of our proposed technique is not detection method based on recurrence quantification analysis, IEEE Journal of
Biomedical and Health Informatics 18 (4) (2014) 1285–1293.
dependent on the experts’ knowledge from different physiological sig­
[2] A. Zarei, B.M. Asl, Automatic classification of apnea and normal subjects using new
nals and it is robust. The proposed method can be used in the following features extracted from HRV and ECG-derived respiration signals, Biomedical
practical applications: Signal Processing and Control 59 (2020), 101927.
[3] F. Mendonca, S.S. Mostafa, A.G. Ravelo-García, F. Morgado-Dias, T. Penzel,
A review of obstructive sleep apnea detection approaches, IEEE Journal of
• Development of portable home systems for diagnosing sleep apnea Biomedical and Health Informatics 23 (2) (2018) 825–837.
• Development of mobile-based apnea diagnosis systems [4] P. De Chazal, C. Heneghan, E. Sheridan, R. Reilly, P. Nolan, M. O’Malley,
• Also, the proposed structure (with some modifications) can be used Automatic classification of sleep apnea epochs using the electrocardiogram, in:
Computers in Cardiology 2000, IEEE, 2000, pp. 745–748.
in other practical applications of physiological signal processing [5] A.H. Khandoker, M. Palaniswami, C.K. Karmakar, Support vector machines for
(such as EEG signal processing for seizure detection, ECG signal automated recognition of obstructive sleep apnea syndrome from ECG recordings,
processing for arrhythmia diagnosis). IEEE Transactions on Information Technology in Biomedicine 13 (1) (2009) 37–48.
[6] M.O. Mendez, A.M. Bianchi, M. Matteucci, S. Cerutti, T. Penzel, Sleep apnea
screening by autoregressive models from a single ECG lead, IEEE Transactions on
4.3. Limitations Biomedical Engineering 56 (12) (2009) 2838–2850.
[7] J. Han, H.-B. Shin, D.-U. Jeong, K.S. Park, Detection of apneic events from single
channel nasal airflow using 2nd derivative method, Computer Methods and
Tunning the initial weights for training different deep-learning Programs in Biomedicine 91 (3) (2008) 199–207.
models affects the performance of models, which is called transfer [8] S.Q. Ali, S. Khalid, S.B. Belhaouari, A novel technique to diagnose sleep apnea in
suspected patients using their ECG data, IEEE Access 7 (2019) 35184–35194.
learning (TL). The main advantage of TL technique is the reduction of [9] S.H. Hwang, Y.J. Lee, D.-U. Jeong, K.S. Park, et al., Apnea–hypopnea index
learning time and also, making the network with better performance, prediction using electrocardiogram acquired during the sleep-onset period, IEEE
eliminating the need for a large data, and also avoiding the overfitting Transactions on Biomedical Engineering 64 (2) (2016) 295–301.
[10] S. Wu, D. Liang, Q. Yang, G. Liu, Regularity of heart rate fluctuations analysis in
problem [64]. Therefore, in future studies, we will investigate the effect
obstructive sleep apnea patients using information-based similarity, Biomedical
of using transfer learning in OSA detection. Moreover, even though the Signal Processing and Control 65 (2021), 102370.
physiological sources of OSA and arrythmias are different, both of them [11] C. Song, K. Liu, X. Zhang, L. Chen, X. Xian, An obstructive sleep apnea detection
produce similar symptoms to some extent. Both the OSA and arrhyth­ approach using a discriminative hidden markov model from ECG signals, IEEE
Transactions on Biomedical Engineering 63 (7) (2015) 1532–1542.
mias cause some changes in the ECG signals. Therefore, development of [12] L. Chen, X. Zhang, C. Song, An automatic screening approach for obstructive sleep
an OSA detection algorithm (using the ECG signal) in the presence of apnea diagnosis based on single-lead electrocardiogram, IEEE Transactions on
arrhythmias is essential. So, the next step is to investigate the developing Automation Science and Engineering 12 (1) (2014) 106–115.
[13] S. Babaeizadeh, D.P. White, S.D. Pittman, S.H. Zhou, Automatic detection and
an OSA detection algorithm in the presence of the arrhythmias. quantification of sleep apnea using heart rate variability, Journal of
Electrocardiology 43 (6) (2010) 535–541.
5. Conclusion [14] T. Yashar, J.D. Nader, M. Keivan, S. Ali, Obstructive sleep apnea event prediction
using recurrence plots and convolutional neural networks (RP-CNNs) from
polysomnographic signals, Biomedical Signal Processing and Control 69 (2021),
In this study, a new algorithm was developed to classify apnea events 102928.
and normal segments using deep learning structures. The proposed [15] R.E. Rolón, I.E. Gareis, L.D. Larrateguy, L.E. Di Persia, R.D. Spies, H.L. Rufiner,
Automatic scoring of apnea and hypopnea events using blood oxygen saturation
LSTM-CNN model effectively learned features from training data to signals, Biomedical Signal Processing and Control 62 (2020), 102062.
discriminate apnea events from healthy ones. We have shown the reli­ [16] S. McCloskey, R. Haidar, I. Koprinska, B. Jeffries, Detecting hypopnea and
ability of the proposed method by employing the Apnea-ECG dataset. obstructive apnea events using convolutional neural networks on wavelet
spectrograms of nasal airflow, in: Pacific-Asia Conference on Knowledge Discovery
The most important advantage of the introduced technique is that it and Data Mining, Springer, 2018, pp. 361–372.
solves the need for the feature extraction/selection and classification [17] C. Varon, A. Caicedo, D. Testelmans, B. Buyse, S. Van Huffel, A novel algorithm for
algorithm. The proposed method provided the average accuracies of the automatic detection of sleep apnea from single-lead ECG, IEEE Transactions on
Biomedical Engineering 62 (9) (2015) 2269–2278.
97.21% and 100% for per segment and per-recording classification (on
[18] B. Mohammadzadeh-Asl, S.K. Setarehdan, Neural network based arrhythmia
Apnea-ECG dataset), respectively. Also, our model achieves an accuracy classification using heart rate variability signal, in: 2006 14th European Signal
of 93.70%, sensitivity of 90.69%, and specificity of 95.82% for UCDDB Processing Conference, IEEE, 2006, pp. 1–4.
[19] B.M. Asl, A.R. Sharafat, S.K. Setarehdan, An adaptive backpropagation neural
dataset. The achieved results showed that if the combination of LSTM-
network for arrhythmia classification using RR interval signal, Neural Network
CNN is used instead of the conventional feature engineering and clas­ World 22 (6) (2012) 535–548.
sification methods, classification results can be improved significantly. [20] R. Atri, M. Mohebbi, Obstructive sleep apnea detection using spectrum and
The results based on the Apnea-ECG and UCDDB datasets suggest that bispectrum analysis of single-lead ECG signal, Physiological Measurement 36 (9)
(2015) 1963.
the presented algorithm provides significant improvements over the [21] R. Tripathy, Application of intrinsic band function technique for automated
state-of-the-art techniques. Therefore, it can be concluded that the use of detection of sleep apnea using HRV and EDR signals, Biocybernetics and
deep-learning based technique to diagnose apnea patients would help Biomedical Engineering 38 (1) (2018) 136–144.
[22] A. Nishad, R.B. Pachori, U.R. Acharya, Application of TQWT based filter-bank for
physicians to make more accurate decisions. sleep apnea screening using ECG signals, Journal of Ambient Intelligence and
Humanized Computing (2018) 1–12.
CRediT authorship contribution statement [23] M. Sharma, S. Agarwal, U.R. Acharya, Application of an optimal class of
antisymmetric wavelet filter banks for obstructive sleep apnea diagnosis using ECG
signals, Computers in Biology and Medicine 100 (2018) 100–113.
Asghar Zarei: Conceptualization, Methodology, Data curation,
Writing - original draft, Validation. Hossein Beheshti: Methodology,

9
A. Zarei et al. Biomedical Signal Processing and Control 71 (2022) 103125

[24] Y. Li, W. Pan, K. Li, Q. Jiang, G. Liu, Sliding trend fuzzy approximate entropy as a [45] D.P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprint
novel descriptor of heart rate variability in obstructive sleep apnea, IEEE Journal of arXiv:1412.6980.
Biomedical and Health Informatics 23 (1) (2019) 175–183. [46] H. Sharma, K. Sharma, An algorithm for sleep apnea detection from single-lead
[25] A. Zarei, B.M. Asl, Automatic detection of obstructive sleep apnea using wavelet ECG using hermite basis functions, Computers in Biology and Medicine 77 (2016)
transform and entropy-based features from single-lead ECG signal, IEEE Journal of 116–124.
Biomedical and Health Informatics 23 (3) (2019) 1011–1021. [47] A.R. Hassan, M.A. Haque, An expert system for automated identification of
[26] S. Martín-González, J.L. Navarro-Mesa, G. Juliá-Serdá, J.F. Kraemer, N. Wessel, A. obstructive sleep apnea from single-lead ECG using random under sampling
G. Ravelo-García, Heart rate variability feature selection in the presence of sleep boosting, Neurocomputing 235 (2017) 122–130.
apnea: An expert system for the characterization and detection of the disorder, [48] A.R. Hassan, Computer-aided obstructive sleep apnea detection using normal
Computers in Biology and Medicine 91 (2017) 47–58. inverse gaussian parameters and adaptive boosting, Biomedical Signal Processing
[27] A. Zarei, B.M. Asl, Performance evaluation of the spectral autocorrelation function and Control 29 (2016) 22–30.
and autoregressive models for automated sleep apnea detection using single-lead [49] A.R. Hassan, M.A. Haque, Computer-aided obstructive sleep apnea identification
ECG signal, Computer Methods and Programs in Biomedicine 105626 (2020). using statistical features in the EMD domain and extreme learning machine,
[28] M. Faal, F. Almasganj, Obstructive sleep apnea screening from unprocessed ECG Biomedical Physics & Engineering Express 2 (3) (2016), 035003.
signals using statistical modelling, Biomedical Signal Processing and Control 68 [50] A.R. Hassan, M.A. Haque, Computer-aided obstructive sleep apnea screening from
(2021), 102685. single-lead electrocardiogram using statistical and spectral features and bootstrap
[29] B. Fatimah, P. Singh, A. Singhal, R.B. Pachori, Detection of apnea events from ECG aggregating, Biocybernetics and Biomedical Engineering 36 (1) (2016) 256–266.
segments using fourier decomposition method, Biomedical Signal Processing and [51] A.R. Hassan, Automatic screening of obstructive sleep apnea from single-lead
Control 61 (2020), 102005. electrocardiogram, in: 2015 International Conference on Electrical Engineering
[30] S.S. Mostafa, F. Morgado-Dias, A.G. Ravelo-García, Comparison of SFS and mrmr and Information Communication Technology (ICEEICT), IEEE, 2015, pp. 1–6.
for oximetry feature selection in obstructive sleep apnea detection, Neural [52] A.R. Hassan, A comparative study of various classifiers for automated sleep apnea
Computing and Applications (2018) 1–21. screening based on single-lead electrocardiogram, in: 2015 International
[31] K. Li, W. Pan, Y. Li, Q. Jiang, G. Liu, A method to detect sleep apnea based on deep Conference on Electrical & Electronic Engineering (ICEEE), IEEE, 2015, pp. 45–48.
neural network and hidden markov model using single-lead ECG signal, [53] A.R. Hassan, M.A. Haque, Identification of sleep apnea from single-lead
Neurocomputing 294 (2018) 94–101. electrocardiogram, in: 2016 IEEE Intl Conference on Computational Science and
[32] S.H. Choi, H. Yoon, H.S. Kim, H.B. Kim, H.B. Kwon, S.M. Oh, Y.J. Lee, K.S. Park, Engineering (CSE) and IEEE Intl Conference on Embedded and Ubiquitous
Real-time apnea-hypopnea event detection during sleep by convolutional neural Computing (EUC) and 15th Intl Symposium on Distributed Computing and
networks, Computers in Biology and Medicine 100 (2018) 123–131. Applications for Business Engineering (DCABES), IEEE, 2016, pp. 355–360.
[33] U. Erdenebayar, Y.J. Kim, J.-U. Park, E.Y. Joo, K.-J. Lee, Deep learning approaches [54] J. Zhang, Z. Tang, J. Gao, L. Lin, Z. Liu, H. Wu, F. Liu, R. Yao, Automatic detection
for automatic detection of sleep apnea events from an electrocardiogram, of obstructive sleep apnea events using a deep CNN-LSTM model, Computational
Computer Methods and Programs in Biomedicine 180 (2019), 105001. Intelligence and Neuroscience (2021).
[34] D. Dey, S. Chaudhuri, S. Munshi, Obstructive sleep apnoea detection using [55] S.I. Niroshana, X. Zhu, K. Nakamura, W. Chen, A fused-image-based approach to
convolutional neural network based deep learning framework, Biomedical detect obstructive sleep apnea using a single-lead ECG and a 2D convolutional
Engineering Letters 8 (1) (2018) 95–100. neural network, Plos one 16 (4) (2021), e0250618.
[35] D. Novák, K. Mucha, T. Al-Ani, Long short-term memory for apnea detection based [56] Q. Shen, H. Qin, K. Wei, G. Liu, Multiscale deep neural network for obstructive
on heart rate variability, in: 2008 30th Annual International Conference of the sleep apnea detection using RR interval from single-lead ECG signal, IEEE
IEEE Engineering in Medicine and Biology Society, IEEE, 2008, pp. 5234–5237. Transactions on Instrumentation and Measurement 70 (2021) 1–13.
[36] R.K. Pathinarupothi, R. Vinaykumar, E. Rangan, E. Gopalakrishnan, K. Soman, [57] H.-Y. Chang, C.-Y. Yeh, C.-T. Lee, C.-C. Lin, A sleep apnea detection system based
Instantaneous heart rate as a robust feature for sleep apnea severity detection using on a one-dimensional deep convolution neural network model using single-lead
deep learning, in: 2017 IEEE EMBS International Conference on Biomedical & electrocardiogram, Sensors 20 (15) (2020) 4157.
Health Informatics (BHI), IEEE, 2017, pp. 293–296. [58] K. Feng, H. Qin, S. Wu, W. Pan, G. Liu, A sleep apnea detection method based on
[37] X. Jin, X. Yu, X. Wang, Y. Bai, T. Su, J. Kong, Prediction for time series with CNN unsupervised feature learning and single-lead electrocardiogram, IEEE
and LSTM, in: Proceedings of the 11th International Conference on Modelling, Transactions on Instrumentation and Measurement 70 (2020) 1–12.
Identification and Control (ICMIC2019), Springer, 2020, pp. 631–641. [59] F.R. Mashrur, M.S. Islam, D.K. Saha, S.R. Islam, M.A. Moni, SCNN: Scalogram-
[38] T. Penzel, G.B. Moody, R.G. Mark, A.L. Goldberger, J.H. Peter, The Apnea-ECG based convolutional neural network to detect obstructive sleep apnea using single-
database, in: Computers in Cardiology 2000, IEEE, 2000, pp. 255–258. lead electrocardiogram signals, Computers in Biology and Medicine 104532
[39] C. Heneghan, St. vincent’s university hospital/university college dublin sleep (2021).
apnea databaase. ((2008). [Online]. Available: http://www.physionet.org/pn3/ [60] P.D. Ellis, The essential guide to effect sizes: Statistical power, meta-analysis, and
ucddb/). the interpretation of research results, Cambridge University Press, 2010.
[40] H. Phan, F. Andreotti, N. Cooray, O.Y. Chén, M. De Vos, Joint classification and [61] T. Wang, C. Lu, G. Shen, F. Hong, Sleep apnea detection from a single-lead ECG
prediction CNN framework for automatic sleep stage classification, IEEE signal with automatic feature-extraction through a modified lenet-5 convolutional
Transactions on Biomedical Engineering 66 (5) (2018) 1285–1296. neural network, PeerJ 7 (2019), e7731.
[41] N.D. Truong, A.D. Nguyen, L. Kuhlmann, M.R. Bonyadi, J. Yang, S. Ippolito, [62] M. Cheng, W.J. Sori, F. Jiang, A. Khan, S. Liu, Recurrent neural network based
O. Kavehei, Integer convolutional neural network for seizure detection, IEEE classification of ECG signal features for obstruction of sleep apnea detection, in:
Journal on Emerging and Selected Topics in Circuits and Systems 8 (4) (2018) 2017 IEEE International Conference on Computational Science and Engineering
849–857. (CSE) and IEEE International Conference on Embedded and Ubiquitous Computing
[42] E. Urtnasan, J.-U. Park, K.-J. Lee, Automatic detection of sleep-disordered (EUC), vol. 2, IEEE, 2017, pp. 199–202.
breathing events using recurrent neural networks from an electrocardiogram [63] O. Faust, R. Barika, A. Shenfield, E.J. Ciaccio, U.R. Acharya, Accurate detection of
signal, Neural Computing and Applications (2018) 1–10. sleep apnea with long short-term memory network based on RR interval signals,
[43] S. Rajendran, W. Meert, D. Giustiniano, V. Lenders, S. Pollin, Deep learning models Knowledge-Based Systems 212 (2021), 106591.
for wireless signal classification with distributed low-cost spectrum sensors, IEEE [64] A. Mahbod, G. Schaefer, C. Wang, G. Dorffner, R. Ecker, I. Ellinger, Transfer
Transactions on Cognitive Communications and Networking 4 (3) (2018) 433–445. learning using a multi-scale and multi-network ensemble for skin lesion
[44] Y. Gal, Z. Ghahramani, A theoretically grounded application of dropout in classification, Computer Methods and Programs in Biomedicine 193 (2020),
recurrent neural networks, Advances in Neural Information Processing Systems 105475.
(2016) 1019–1027.

10