Computer Methods and Programs in Biomedicine 180 (2019) 105001

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

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

Deep learning approaches for automatic detection of sleep apnea

events from an electrocardiogram
Urtnasan Erdenebayar a, Yoon Ji Kim a, Jong-Uk Park a, Eun Yeon Joo b, Kyoung-Joung Lee a,∗
a
Department of Biomedical Engineering, College of Health Science, Yonsei University, Wonju 26493, Korea
b
Department of Neurology, Samsung Medical Center, School of Medicine, Sungkyunkwan University, Korea

a r t i c l e i n f o a b s t r a c t

Article history: Background and Objective: This study demonstrates deep learning approaches with an aim to find the
Received 5 March 2019 optimal method to automatically detect sleep apnea (SA) events from an electrocardiogram (ECG) signal.
Revised 20 July 2019
Methods: Six deep learning approaches were designed and implemented for automatic detection of
Accepted 29 July 2019
SA events including deep neural network (DNN), one-dimensional (1D) convolutional neural networks
(CNN), two-dimensional (2D) CNN, recurrent neural networks (RNN), long short-term memory, and gated-
Keywords: recurrent unit (GRU). Designed deep learning models were analyzed and compared in the performances.
Sleep apnea The ECG signal was pre-processed, normalized, and segmented into 10 s intervals. Subsequently, the sig-
Deep learning nal was converted into a 2D form for analysis in the 2D CNN model. A dataset collected from 86 patients
Convolutional neural network
with SA was used. The training set comprised data from 69 of the patients, while the test set contained
Recurrent neural network
data from the remaining 17 patients.
Long short-term memory
Gated-recurrent unit Results: The accuracy of the best-performing model was 99.0%, and the 1D CNN and GRU models had
99.0% recall rates.
Conclusions: The designed deep learning approaches performed better than those developed and tested in
previous studies in terms of detecting SA events, and they could distinguish between apnea and hypopnea
events using an ECG signal. The deep learning approaches such as 1D CNN and GRU can be helpful tools
to automatically detect SA in sleep apnea screening and related studies.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction snoring signals [22-24], and respiratory signals [25-27]. The respi-
ratory and snoring are directly effected by SA and upper airway
Sleep apnea (SA) is an abnormal respiratory pattern occurring obstruction. The ECG and SpO2 are indirectly effected by SA events
during sleep. It includes apnea and hypopnea and is caused by re- because the responses to SA include both an increase in the sym-
current episodes of reduced or absent respiratory airflow caused phathetic and the parasymphathetic tone to the heart rate and sys-
by upper airway collapse or other airway obstruction [1]. The like- tolic blood pressure [28]. These physiological changes of the ECG
lihood of developing SA is positively correlated with age [2]. SA it- signal during SA can be noticed by the time, frequency, and non-
self may increase the risk of heart disease [3], diabetes [4], chronic linear domain analysis. Therefore, most alternative studies were fo-
kidney disease [5], stroke [6], depression [7], and cognitive impair- cused on extracting the temporal, spectral and nonlinear features
ment [8]. Generally, SA screening and diagnostic methods require from the physiological signals, and various methods of feature se-
that various physiological signals be recorded by polysomnography lection including principal component analysis, statistical evalua-
(PSG) during overnight sleep in sleep centers [9]. These methods tion, and wrapper methods to reduce the dimension of the feature
are labor-intensive, time-consuming, come at high cost, and are in- space. Many types of supervised learning methods have been em-
convenient for patients. Moreover, it is also difficult to manually ployed in those extracted features to improve the performance of
annotate PSG recordings. SA detection. In particular, the support vector machine [13,23] and
To replace PSG, many simple and minimized methods have neural networks [15,29,30] were used widely for SA detection, as
been proposed to detect SA. These methods are based on physi- well as the k-nearest neighbor [11,12], linear/quadratic discriminant
ological recordings such as a single-lead ECG [10-16], SpO2 [17-21], analysis [18], AdaBoost [17], and fuzzy logic [31]. All these studies
can be described within canonical supervised learning that com-
posed data processing, feature extraction, feature selection, and
∗
Corresponding author. classification. In supervised learning, the discrimination power of
E-mail address: [email protected] (K.-J. Lee).

https://doi.org/10.1016/j.cmpb.2019.105001
0169-2607/© 2019 Elsevier B.V. All rights reserved.
2 U. Erdenebayar, Y.J. Kim and J.-U. Park et al. / Computer Methods and Programs in Biomedicine 180 (2019) 105001

feature is significant, but it can require domain knowledge, labor- Table 2

SA dataset specifications.
intensive, and can be limited for complex data [32].
Recently, convolutional neural network (CNN) and recurrent Events Training set Test set Total
neural network (RNN) systems have become increasingly popu- Normal 21,405 4561 25,966
lar. Although these represent supervised learning systems, they Apnea 15,933 1623 17,556
perform an excellent performance across a wide range of appli- Hypopnea 26,103 5486 31,589
cations. CNN has become an invaluable technology in the field Total 63,441 11,670 75,111
of image signal processing and computer vision [33], while RNN
performs well in speech signal processing and speech recognition
fields [34,35]. Few sleep studies have applied deep neural network from each subject group. The training set included 63,441 events
(DNN), CNNs and RNNs to automatically detect SA [36-38]. They from 69 subjects of the training group, while the test set contained
have achieved higher performances for the automatic detection 11,670 events from 17 subjects of the test group (Table 2).
of sleep-breathing disorder events, including apnea and hypopnea,
than conventional machine learning methods. However, they did
2.2. Data processing
not eliminated processes in the complex signal analysis and the
hand-crafted feature extraction. In addition, they did not deliver
The designed deep learning approaches were trained and vali-
the appropriate function of the deep learning approach, which is
dated on the SA dataset obtained from 86 patients diagnosed with
automatic detection of SA events by using ECG.
SA. The single-lead ECG signals were recorded at 200 Hz over ap-
In this study, we demonstrated the comprehensive analysis of
proximately 6 h during the PSG recording. ECG signals were fil-
the representative deep learning approaches that were optimally
tered by an FIR bandpass (0.5–30 Hz) to remove the noise and
designed for automatic detection of SA using an ECG signal. Deep
baseline drift. Subsequently, the ECG signals were segmented to
learning approach consists of the six different models including
match event-based classification. Non-overlapping segmentations
DNN, CNN, and RNN-based methods. The CNN-based model con-
were applied to entire ECG signals, which were divided into 10-s
sists of the 1D CNN and 2D CNN which were designed and opti-
intervals. ECG signals were converted into 2D spectrogram images
mized by employing 1D and 2D convolutions, respectively. 1D CNN
to generate 2D input signals using the following short-time Fourier
was designed for the time domain characteristics of the ECG signal
transformation:
whereas 2D CNN model was intended for the spectral components
L−1

of the ECG signal during the SA events. Because of the SA events
x[n, k] = w[m] · x[n + m] · e− jm(2π k/N ) (1)
are occurring in the sequentially and repeatedly during sleep, long
m=0
short-term memory (LSTM) and gated-recurrent unit (GRU) mod-
els were used to construct the RNN-based approach. Finally, we where n and k denote the time the signal and the signal fre-
designed two basic models such as DNN and vanilla RNN model quency were received, respectively, and w[m] is a window function
for comparison purposes. The clinical dataset of the SA patients where the window length was 128 points with a 127-point over-
was used to the training and testing each model, and to com- lap. Each segment of the 1D and 2D approaches was formatted
pare strengths and weaknesses for each constructed deep learning into (1 × 20 0 0) and (129 × 1873) file format, respectively (Fig. 1).
models. All abovementioned signal processing was performed by the signal
processing toolbox of MATLAB software (Mathworks, U.S.A.).
2. Material and method
2.3. CNN-based approaches
2.1. Subjects and datasets
A CNN consists of three main parts: convolution layer, pooling
For this study, nocturnal PSG recordings from 86 patients with layer, and classification layer. In the convolution layer, the feature
SA were analyzed (Table 1). PSG recordings were measured by an map was extracted by applying a filter kernel to produce the con-
Embla N70 0 0 amplifier device (Embla System Inc., U.S.A.) at the volution integral of the input data activation function, thereby en-
Samsung Medical Center (Seoul, Korea). Accordance with the AASM hancing discrimination. In the pooling layer, the feature map has
guidelines [39] the PSG recording was labeled by an experienced reduced and restricted the dimensions of input data. Finally, the
sleep technician. The institutional review board of the Samsung classification layer is performed the final discrimination of the in-
Medical Center (IRB:2012-01063) authorized this study protocol. put data by using the fully-connected network. At this stage, and
All patients that enrolled in this study provide written consent. the learning process is performed through feed-forward and back-
The SA datasets comprised normal breathing and SA events, in- propagation algorithms.
cluding hypopnea (H) and apnea (A). Moreover, SA datasets com- Two CNN models were designed and optimized for automatic
posed of a balanced number of events were randomly selected detection of SA events, namely the 1D CNN and 2D CNN (Fig. 2).
Primarily, the 1D CNN model can be used in biomedical engineer-
ing and speech recognition applications that use time-series signal
Table 1
The demographics of the training group and the test group.
as input. We compared the 1D and 2D CNN model to evaluate the
differences and determine which is more appropriate in the appli-
Measures Training group Test group p-value Total
cation of physiological signal such as ECG. The 1D CNN model used
Subjects (M: F) 69 (53: 16) 17 (12: 5) = 86 (65: 21) a pre-processed ECG signal as input, which was batch-normalized
Age (years) 58.48 ± 10.74 56.88 ± 12.45 NS 58.18 ± 11.02 before input to the CNN-based approach. Next, the deep learning
BMI (kg/m2 ) 25.63 ± 3.05 25.33 ± 2.20 NS 25.57 ± 2.89
model proceeds with the 1D convolution operation, which is ex-
AHI (per h) 28.93 ± 19.00 19.28 ± 11.12 NS 27.02 ± 18.08
TRT (h) 7.46 ± 0.61 7.15 ± 1.04 NS 7.40 ± 0.72 plained in detail below, and 1D pooling at the convolution and
TST (h) 5.86 ± 1.03 5.72 ± 1.07 NS 5.83 ± 1.04 pooling layers followed by the activation function and dropout. Fi-
SE (%) 78.40 ± 13.51 80.19 ± 11.64 NS 78.77 ± 13.10 nally, the 1D CNN model uses a fully-connected layer to discrimi-
Note: data are presented as mean ± SD; BMI: body mass index; AHI: apnea- nate SA events (Fig. 2A).
hypopnea index; TRT: total recording time; TST: total sleep time; SE: sleep effi- The 2D CNN model is a general method of deep learning ap-
ciency; NS: no significant difference between training and test set (p-value > 0.01). plied to computer vision and the image recognition fields, where
 U. Erdenebayar, Y.J. Kim and J.-U. Park et al. / Computer Methods and Programs in Biomedicine 180 (2019) 105001 3

Fig. 1. Example of input signal for the designed deep learning approaches. Pre-processed ECG segment as 1D input (top), spectrogram of ECG segment as 2D input (bottom).

Fig. 2. CNN−based approaches for automatic detection of SA events. (A) Flowchart of 1D CNN model, six−layer convolution, three kernels with sizes of 50 × 1, 30 × 1, and
10 × 1, pooling size 1 × 2, dropout p = 0.25. (B) Flowchart of 2D CNN model, seven−layer convolution, three kernels with sizes of 50 × 2, 30 × 2, and 10 × 2, pooling size 2 × 2,
dropout p = 0.25.

the image is used as input. In this study, the 2D CNN model re- be expressed as:
ceived a whole spectrogram of the pre-processed ECG segment as Mi −1 Ni −1
 
the input image. The 2D convolution and 2D pooling were per- C ( j, k ) = I (m, n ) ∗ f ( j − m, k − n ) (3)
formed with ReLu activation and dropout, to avoid overfitting and m=0 n=0
divergence. This routine was repeated seven times, upon which
the CNN model generated the feature maps that were used at the where 0 ≤ j ≤ Mi + Mf − 1 and 0 ≤ k ≤ Ni + Nf − 1.
fully-connected layer for final classification (Fig. 2B).
2.3.2. CNN architecture
The CNN-based approaches used almost same architecture for
automatic detection of SA events. In Table 3, we presented the ar-
2.3.1. 1D and 2D convolution
chitecture of the 1D CNN and 2D CNN models. The 1D CNN model
CNN-based approaches were performed to a 1D and 2D convo-
used three kernels with sizes of 50 × 1, 30 × 1, and 10 × 1. Each
lution operation, according to the inputs. Each convolutional oper-
pooling layer in the 1D CNN model computed 1 × 2 max-pooling
ation was calculated at the convolutional layer of CNN model. First,
regions. The 2D CNN model employed kernels with sizes of 50 × 2,
the 1D convolution operation of convolutional layer was computed
30 × 2 and 10 × 2 and 2 × 2 max-pooling regions. It used one more
according to the following expression:
convolutional layer than the 1D-CNN model.
 

N
x = f b+ ∗(w, y ) (2) 2.4. RNN-based approaches
i=1
A recurrent neural network (RNN) is regarded as most common
where x is the total feature map, b is the bias, y is a feature map, type of the conventional neural network that can accept variable
w is the convolutional kernel, N is the total number of features, (∗) and sequence input [40]. RNN is ideally suited to sequential data,
and f(.) are vector convolution and the activation function, respec- which proves robust for time series, as they have a memory as
tively. well. Consequently, the current input data and previous state af-
The spectrogram, or 2D conversion, of the ECG signals for the fect the output of the next state.
2D CNN approach had a straightforward application to the 2D con- RNN-based approaches consist of three basic sections: input
volutional kernels. Input data I had the dimensions (Mi , Ni ) and blocks, the deep learning model, and output blocks (Fig. 3). LSTM
the filter function f(.) had the dimensions (Mf , Nf ), such that the and GRU models were used for RNN-based approaches, and the
2D convolution kernels which calculated the full output size could same architectures were applied to the corresponding datasets
4 U. Erdenebayar, Y.J. Kim and J.-U. Park et al. / Computer Methods and Programs in Biomedicine 180 (2019) 105001

Table 3
Architectures of the 1D CNN and 2D CNN models.

Layers 1D CNN 2D CNN

Filter size Output shape Parameters Filter size Output shape Parameters

batchnorm_1 = 2000 × 1 4 = 129 × 1873 4

conv_1 16@50 × 1 1985 × 16 272 32@50 × 2 128 × 1858 × 32 1056
maxpool_1 2×1 992 × 16 2×2 64 × 929 × 32
conv_2 16@50 × 1 977 × 64 16,448 64@50 × 2 63 × 914 × 64 65,600
conv_3 64@30 × 1 962 × 64 65,600 64@30 × 2 62 × 899 × 64 131,136
maxpool_2 2×1 481 × 64 2×2 31 × 449 × 64
batchnorm_2 = 481 × 64 256 = 31 × 449 × 64 256
dense_1 32@30 × 1 481 × 32 2080 32@30 × 2 31 × 449 × 32 2080
maxpool_3 2×1 240 × 32 2×2 15 × 224 × 32
conv_4 24@10 × 1 233 × 32 8224 24@10 × 2 14 × 217 × 32 16,416
maxpool_4 2×1 116 × 32 2×2 7 × 108 × 32
conv_5 16@10 × 1 109 × 16 4112 12@10 × 2 6 × 101 × 16 8208
maxpool_5 2×1 54 × 16 2×2 3 × 50 × 16
flatten_1 2 864 × 2 1730 2 2400 × 2 4802
dense_1
Total 168 filters 98,726 229,558

Note: ReLu activation was used for all deep learning models.

Fig. 3. RNN−based approaches for automatic detection of SA events. (A) LSTM model, (B) GRU model. Each model has the same architecture of three−layer RNN with 60,
80, and 120 memory cells.

to compare their performance. Detailed explanations of LSTM ct = ft · ct−1 + it · gt (8)

(Fig. 3A) and GRU (Fig. 3B) models follow.
ht = ot · ϕ (ct ) (9)
2.4.1. Long short-term memory where c is the cell activation vector. Terms σ and τ are the non-
Long short-term memory (LSTM) is an updated version of a ba- linear hyperbolic and tangent functions. xt is the input to the
sic RNN with memory cells that facilitates the learning of temporal memory cell layer at time t. W is weight matrices, bi , bf , bc , and
correlations of data over time. The concept of LSTM is based on a bo are bias vectors.
memory cell that handles the read, write, and reset functions of
its internal state through an input gate (it ), an output gate (ot ), 2.4.2. Gated-recurrent memory
and a forget gate (ft ), respectively. Each gate works to remember A gated-recurrent unit (GRU) is a regarded as a simplified ver-
when and to what extent the weights in the memory should be sion of the LSTM that contains two gates, namely the update (zt )
updated. The input and output gates controls the flow of input and and reset gate (rt ), and controls the flow of information simi-
output of the memory cell activations. The forget gate deals the in- larly to LSTM without the memory unit. Despite these simplifi-
ternal state of the cell, therefore adaptively forgetting or resetting cations, GRUs have shown similar performance to those of LSTMs
the cell’s memory (Fig. 4A). The following expressions depict the [41]. In addition, GRU can reduce the number of calculations per
functions of these gates. training phase (Fig. 4B). These calculations are can be represented
  as:
it = σ W xi xt + W hi ht−1 + bi (4)  
zt = σ W xz xt + W hz ht−1 + bz (10)
 
ft = σ W x f xt + W h f ht−1 + b f (5)  
rt = σ W xr xt + W hr ht−1 + br (11)
 
ot = σ W xt + W ht−1 + bo
xo ho
(6) ˜ t = tanh (W xt + W ht−1 · rt )
h (12)
 
gt = σ W xc xt + W hc ht−1 + bc (7) ˜t
ht = (1 − zt ) · ht−1 + zt · h (13)
 U. Erdenebayar, Y.J. Kim and J.-U. Park et al. / Computer Methods and Programs in Biomedicine 180 (2019) 105001 5

Fig. 4. Unit structure of (A) LSTM and (B) GRU. (A) ft , it , and ot are the forget, input, and output gates, respectively. ct- 1 and ct denote the memory cell and the new memory
cell content. (B) rt and zt are the reset and update gates, and ht and ht depict the activation and the candidate activation.

Table 4
Architectures of the LSTM and GRU models.

Layers LSTM GRU

Memory cells Activation Parameters Memory cells Activation Parameters

batchnorm_0 = = 4 = = 4
rnn_1 120 sigmoid 58,560 120 sigmoid 43,920
batchnorm_1 480 480
rnn_2 100 sigmoid 88,400 100 sigmoid 66,300
batchnorm_2 400 400
rnn_3 80 57,920 80 sigmoid 43,440
batchnorm_3 sigmoid 320 320
dense_1 6480 6480
rnn_4 60 sigmoid 33,840 60 sigmoid 25,380
batchnorm_4 240 240
rnn_5 40 sigmoid 16,160 40 sigmoid 12,120
batchnorm_5 160 160
rnn_6 20 4880 20 3660
batchnorm_6 sigmoid 80 sigmoid 80
dense_2 420 420
dense_3 2 softmax 42 2 softmax 42
Total 268,386 229,558

2.4.3. RNN architecture was computed using this configuration for the parameters after be-
RNN-based approaches were designed for this study as shown ing trained on every mini-batch.
in Table 4. The architecture consisted of three layers of RNNs, each
of which had either 60, 80, or 120 memory cells. After the RNN 2.6. Performance evaluation
layers, output feature maps conducted batch normalization and
dropout to avoid overfitting and divergence. Then the optimal ar- The accuracy, sensitivity, and specificity were calculated as the
chitecture for automatic detection of SA events was determined evaluation measure to the performance of the designed deep learn-
empirically and used in LSTM and GRU models to compare their ing approaches. In addition, Cohen’s kappa coefficient (k) was cal-
performance. culated for comparison purposes. The evaluation measures are ex-
pressed as below:

2.5. Implementation accuracy = (TP + TN )/(TP + TN + FP + FN ) (14)

The deep learning approaches were implemented using Python sensit ivit y = TP/(TP + FP ) (15)
2.7, the Keras library, and a TensorFlow background [42]. Keras is a
speci f icity = TP/(TP + FN ) (16)
deep learning library that used to build and evaluate the designed
deep learning approaches. Each approach training and testing was
k = (accuracy − Pe )/(1 − Pe ) (17)
conducted on a hardware specification with a GTX1080 Ti (3584
CUDA cores) in the Win10 environment. Training of the deep learn- where true positive (TP), the number of normal events is classi-
ing approaches was entirely supervised by back-propagating algo- fied as normal, true negative (TN), the number of events abnor-
rithm. The model parameters were optimized by minimizing cross- mal counted as abnormal, false positive (FP), the number of events
entropy loss functions based on the Adam update rule [43]. Data abnormal detected as normal, and false negative (FN), the num-
were segmented into mini-batches of 256 data segments to opti- ber of events normal presented as abnormal. Pe is the hypothetical
mize the training and testing processes. An accumulated gradient probability of agreement by chance.
6 U. Erdenebayar, Y.J. Kim and J.-U. Park et al. / Computer Methods and Programs in Biomedicine 180 (2019) 105001

3. Results

The sensitivity, specificity, and accuracy of the designed deep

learning approaches was evaluated in the DNN, 1D CNN, 2D CNN,
RNN, LSTM, and GRU models for the automatic detection of SA
events. In the test set, the 1D CNN model exhibited an accuracy of
98.5%, 96.4%, and 96.3% for apnea, hypopnea, and A + H events, re-
spectively. The 2D CNN model performance on the test set showed
an accuracy of 95.9%, a sensitivity of 96.0%, and a specificity of
96.0% for apnea, 95.8%, 96.0%, and 96.0%, for hypopnea, and 91.2%,
92.0%, and 91.0% for combined A + H events, respectively. DNN
model showed a lower accuracy of 93.1%, 82.3%, and 85.3% for ap-
nea, hypopnea, and A + H events, even though it has the same ar-
chitecture of the 1D CNN model (Table 5).
Furthermore, the LSTM model obtained an accuracy of 98.0%,
97.0%, and 96.0% for apnea, hypopnea, and A + H events, respec-
tively. The GRU model had the accuracy, sensitivity, and specificity
with 99.0%, 99.0%, and 99.0% for apnea events, 97.0%, 97.0%, and
97.0%, for hypopnea events, and 95.0%, 95.0%, and 96.0% for com-
bined A + H events, respectively (Table 6). In contrast, simple RNN
model achieved the accuracy, sensitivity, and specificity with 85.4%,
97.0%, and 87.0% for apnea events, 80.7%, 95.0%, and 79.0% for hy-
popnea events, and 83.2%, 96.0%, and 82.0% for A + H events, re-
spectively.
The accuracy of the designed deep learning approaches (1D
CNN, 2D CNN, LSTM, and GRU) is compared in Fig. 5. A more sig-
nificant difference was found in the performance between the 1D
and 2D CNN models than between the RNN-based approaches in
all experiments (Fig. 5). The 1D CNN model performed best in de-
tecting combined A + H events. LSTM and GRU models were not
shown a significant difference in the terms of accuracy. However,
they did require at least 20 iterations to achieve optimal perfor-
mance and showed some spikes and fluctuations thereafter.

4. Discussion

The deep learning approaches were designed and found the op-
timal method for automatically detecting SA events based on an

Table 5
Performance evaluation of the CNN-based approaches and DNN.

Event Approach Accuracy (%) Sensitivity (%) Specificity (%) Kappa

Apnea DNN 93.1 93.0 94.0 0.85

1D CNN 98.5 99.0 99.0 0.98
2D CNN 95.9 96.0 96.0 0.92
Hypopnea DNN 82.3 85.0 83.0 0.67
1D CNN 96.4 96.0 96.0 0.92
Fig. 5. Comparison of the accuracy of the deep learning models. (A) Graph of accu-
2D CNN 95.8 96.0 96.0 0.92
racy for apnea events. (B) Graph of accuracy for hypopnea events. (C) Accuracy for
A+H DNN 85.3 88.0 85.0 0.74 combined A + H events.
1D CNN 96.3 96.0 96.0 0.92
2D CNN 91.2 92.0 91.0 0.83

ECG signal. Six deep learning models, based on DNN, CNN and
RNN, were designed and evaluated for their effectiveness. Finally,
Table 6 we obtained very high performances, with accuracies of 98.5%,
Performance evaluation of the RNN-based approaches.
95.9%, 98.0%, and 99.0% for 1D CNN, 2D CNN, LSTM, and GRU, re-
Event Approach Accuracy (%) Sensitivity (%) Specificity (%) Kappa spectively.
Apnea RNN 85.4 97.0 87.0 0.81 CNNs are popular models used in image recognition, bioinfor-
LSTM 98.0 98.0 98.0 0.96 matics, and medical imaging applications. They are capable of rep-
GRU 99.0 99.0 99.0 0.98 resenting the low to high-level features of input data through lin-
Hypopnea RNN 80.7 95.0 79.0 0.75 ear and nonlinear data abstraction. CNNs perform a morphology-
LSTM 97.0 97.0 97.0 0.94 based recognition, such that they can accept one-, two-, or
GRU 97.0 97.0 97.0 0.94
three-dimensional input data. The 1D CNN and 2D CNN models
A+H RNN 83.2 96.0 82.0 0.78 were compared in the context of automatic detection of SA events.
LSTM 96.0 96.0 96.0 0.92
GRU 95.0 95.0 96.0 0.91
The designed 1D CNN model exhibited higher performances, and
it is the lightest and simplest among the designed deep learning
 U. Erdenebayar, Y.J. Kim and J.-U. Park et al. / Computer Methods and Programs in Biomedicine 180 (2019) 105001 7

Table 7
Performance comparison with previous studies.

Study Subject Method Accuracy (%) Sensitivity (%) Specificity (%)

Jafari [14] 35 SVM 94.8 95.4 94.1

Chen [15] 90 SVM 97.4 98.9 92.9
Hassan [44] 35 RUSBoost 88.8 87.5 91.4
Sharma [45] 35 LS-SVM 90.1 90.8 88.8
Nishad [46] 35 RF 92.7 93.9 90.9
Viswabhargav [47] 35 SVM = 85.4 92.6
Pathinarupothi [36] 35 LSTM 98.0 = =
Cheng [48] 35 LSTM 97.8 = =
Dey [49] 35 CNN 98.9 97.8 99.2
Choi [38] 179 CNN 96.6 81.1 98.5
Our method 86 DNN 93.1 93.0 94.0
1D CNN 98.5 99.0 99.0
2D CNN 95.9 96.0 96.0
RNN 85.4 97.0 87.0
LSTM 98.0 98.0 98.0
GRU 99.0 99.0 99.0

approaches. In addition, it is straightforward to apply and enables detect specific events like hypopnea that cannot be discriminated
facile analysis of physiological information, such as the ECG signal. using conventional machine learning methods. Hypopnea events
Meanwhile, the 2D CNN model showed the loweer performance have consistently been the most difficult to detect from an ECG
in comparison to the 1D CNN model. The 2D CNN model has the signal. They are associated with reduced airflow and respiratory
bigger number of parameters and requires a high computational effort, such that the ECG signals they produce may appear similar
cost. Moreover, the conversion or domain transformation process to those produced in normal sleep. For this reason, AASM guide-
is needed to apply the time series into the 2D CNN model. Hence, lines recommend the use of additional physiological signals such
the 1D CNN outperformed the 2D CNN model in every respect. In as airflow, SpO2 , and CO2 measurements in order to reliably deter-
the case where the input signal is the physiological signal as ECG mine whether hypopnea events have occurred [39]. However, the
or time series, we recommend the 1D CNN model for any applica- deep learning approaches designed and implemented in this study
tion in deep learning and artificial intelligence algorithm. proved able to effectively detect hypopnea events using ECG signal.
RNNs are familiar models used with sequence data such as The designed deep learning approaches were compared with
speech signal, physiological signal, and gene sequences. They have previous studies, which used RNN models listed in the lower rows
memory gates that store the previous sequences and use those of Table 7. Pathinarupothi et al. [36] developed an LSTM algorithm
to predict the next sequence. RNNs conduct gate-based classifi- that was 98.0% accurate; however, the pulse-to-pulse interval fea-
cation using the input, update, and forgot gates. With respect to tures had to be extracted manually before it could be trained.
the designed LSTM model, not only was the computational cost Cheng et al. [48] likewise applied an LSTM algorithm but used
the highest, but also the amount of training/learning time, since an ECG and an SpO2 signal from which and instantaneous heart
it contains the biggest number of parameters. No significant differ- rate data was extracted to train the algorithm. Despite involving
ences were found between the performances of the LSTM and GRU a larger amount of data from the two-channel signal, this model
models, though the LSTM model performed slightly worse at ap- performed worse than the designed deep learning approaches in
nea and hypopnea event detections than the GRU model. Further- this study. Hence, the designed deep learning models using an ECG
more, there is a significant difference between the computational signal without any feature extraction demonstrated better perfor-
cost and training time of the LSTM and GRU models in terms of the mance in comparison to other methods and were able to distin-
automatic detection of SA events. This is because the LSTM model guish between apnea and hypopnea events.
has one additional memory cell in comparison to the GRU model. In this study, our designed deep learning models have some
Meanwhile, the GRU model exhibited robust performance and rel- advantages in automatically detecting SA events. First, we did not
atively lower computational cost for the training and test phases. use any hand-crafted or hand-extracted features for the automatic
However, GRU model demonstrated comparable and higher perfor- detection of SA events. Therefore, there is no need of the prepro-
mances in comparison to the 1D CNN model, for all SA events. For cessing such as feature extraction and selection. Second, our study
this reason, the GRU model is deemed to be appropriate for auto- population consists of the diverse and equally sized SA patients
matic detection of SA events using an ECG signal. including mild, moderate and severe groups. Third, we have de-
For automatic detection of SA events, conventional studies per- signed and evaluated six different deep learning models by the
formed well and obtained high scores. In those studies, they fol- same datasets to demonstrate the differences between each model.
low the canonical procedure of supervised learning, which con- Fourth, the designed deep learning models have greater ability to
sists of data processing, feature extraction, feature selection, and classify or detect events in comparison with conventional classi-
classification. They extract a number of features to classify SA by fication methods listed in Table 7. Finally, designed deep learn-
using several signal processing techniques that analyze ECG signal ing models can more accurately detect SA and better distinguish
at the time, frequency, and non-linear domain [10-13,17-19]. Then events including apnea, hypopnea, and normal breath using an ECG
they obtain many well-known and powerful features to automati- signal. Nevertheless, this study has some limitations, which are de-
cally detect SA events from an ECG signal. However, all processes scribed as follows. Primarily, this study does not cover some types
are hand-crafted, require heavy computation, and require domain of SA, including central and mixed apnea events. In addition, the
knowledge. Deep learning approaches, such as CNN and RNN mod- ECG signal was contaminated by position changing, loudy snor-
els, do not require a separate hand-crafted feature extraction pro- ing, and coughs during sleep. Finally, a relatively small number of
cess. They are capable of learning which features are significant patients and SA dataset was used for training and evaluation of
directly from raw data, using the convolution processes and mem- this study. In further research, we should be conducted the study
ory cells. Another important point of deep learning is the ability to that overcomes these limitations of the automatic detection of SA
8 U. Erdenebayar, Y.J. Kim and J.-U. Park et al. / Computer Methods and Programs in Biomedicine 180 (2019) 105001

events by using multi-modal and multi-class classification based [12] A. Yildiz, M. Akın, M. Poyraz, An expert system for automated recognition of
on deep reinforcement learning. patients with obstructive sleep apnea using electrocardiogram recordings, Ex-
pert Syst. Appl. 38 (2011) 12880–12890 https://doi.org/10.1016/j.eswa.2011.04.
080.
5. Conclusions [13] H. Al-Angari, A. Sahakian, Automated recognition of obstructive sleep ap-
nea syndromeusing support vector machine classifier, IEEE Trans. Inf. Technol.
In this study, the comprehensive analysis of the representative Biomed. 16 (2012) 463–468 https://doi.org/10.1109/TITB.2012.2185809.
[14] A. Jafari, Sleep apnoea detection from ECG using features extracted from re-
deep learning models was performed for automatic detection of SA constructed phase space and frequency domain, Biomed. Signal Process. Con-
events using an ECG signal. We designed six different deep learn- trol 8 (2013) 551–558 https://doi.org/10.1016/j.bspc.2013.05.007.
ing models (DNN, 1D CNN, 2D CNN, RNN, LSTM, and GRU) for de- [15] L. Chen, X. Zhang, C. Song, An automatic screening approach for obstructive
sleep apnea diagnosis based on single-lead electrocardiogram, IEEE T. Autom.
tecting apnea, hypopnea and normal event from ECG signal. De- Sci. Eng. 12 (2015) 106–115 https://doi.org/10.1109/TASE.2014.2345667.
signed deep learning models did not use any hand-extracted fea- [16] R.K. Tripathy, Application of intrinsic band function technique for automated
tures, and they were trained and tested using a single-lead ECG detection of sleep apnea using HRV and EDR signals, Biocybern. Biomed. Eng.
38 (2018) 136–144 https://doi.org/10.1016/j.bbe.2017.11.003.
signal of the clinical PSG study. We obtained robust performance [17] J.V. Marcos, R. Hornero, D. Álvarez, F. Del Campo, C. Zamarrón, M. López, Util-
for all SA events than conventional studies that used the ECG sig- ity of multilayer perceptron neural network classifiers in the diagnosis of the
nal. Also, 1D CNN and GRU model were more appropriate to use obstructive sleep apnoea syndrome from nocturnal oximetry, Comput. Meth-
ods Programs Biomed. 92 (2008) 79–89 https://doi.org/10.1016/j.cmpb.2008.05.
for automatic detection of SA events using a single-lead ECG sig-
006.
nal than other models that we have designed. We can recommend [18] D. Álvarez, R. Hornero, J. Marcos, F. del Campo, Feature selection from noc-
1D CNN and GRU for automatic detection based on the time series turnal oximetry using genetic algorithms to assist in obstructive sleep ap-
signals such as ECG and other physiological signals. noea diagnosis, Med. Eng. Phys. 34 (2012) 1049–1057 https://doi.org/10.1016/j.
medengphy.2011.11.009.
[19] B. Xie, Hlaing Minn, Real-Time sleep apnea detection by classifier combina-
Funding tion, IEEE Trans. Inf. Technol. Biomed. 16 (2012) 469–477 https://doi.org/10.
1109/TITB.2012.2188299.
[20] J. Marcos, R. Hornero, D. Álvarez, F. Del Campo, M. Aboy, Automated detection
This work was supported by the Ministry of Trade, Industry and
of obstructive sleep apnoea syndrome from oxygen saturation recordings using
Energy (MOTIE) and Korea Institute for Advancement of Technol- linear discriminant analysis, Med. Biol. Eng. Comput. 48 (2010) 895–902 https:
ogy (KIAT) through the National Innovation Cluster R&D program //doi.org/10.1007/s11517-010-0646-6.
[P0 0 06697]; and supported by the Yonsei University Research Fund [21] J. Marcos, R. Hornero, D. Álvarez, M. Aboy, F. Del Campo, Automated prediction
of the apnea-hypopnea index from nocturnal oximetry recordings, IEEE Trans.
of 2019. Biomed. Eng. 59 (2012) 141–149 https://doi.org/10.1109/TBME.2011.2167971.
[22] J. Solà-Soler, J. Fiz, J. Morera, R. Jané, Multiclass classification of subjects with
Declaration of Competing Interest sleep apnoea–hypopnoea syndrome through snoring analysis, Med. Eng. Phys.
34 (2012) 1213–1220 https://doi.org/10.1016/j.medengphy.2011.12.008.
[23] A.M. Benavides, R.F. Pozo, D.T. Toledano, J.L.B. Murillo, E.L. Gonzalo,
None. L.H. Gómez, Analysis of voice features related to obstructive sleep apnoea and
their application in diagnosis support, Comput. Speech Lang. 28 (2014) 434–
Supplementary materials 452 https://doi.org/10.1016/j.csl.2013.08.002.
[24] J. Solé-Casals, C. Munteanu, O. Martín, F. Barbé, C. Queipo, J. Amilibia, J. Durán-
Cantolla, Detection of severe obstructive sleep apnea through voice analysis,
Supplementary material associated with this article can be Appl. Soft Comput. 23 (2014) 346–354 https://doi.org/10.1016/j.asoc.2014.06.
found, in the online version, at doi:10.1016/j.cmpb.2019.105001. 017.
[25] B. Koley, D. Dey, Automatic detection of sleep apnea and hypopnea events
References from single channel measurement of respiration signal employing ensemble
binary SVM classifiers, Measurement 46 (2013) 2082–2092 https://doi.org/10.
[1] E.J. Olson, W.R. Moore, T.I. Morgenthaler, P.C. Gay, B.A. Staats, Obstructive sleep 1016/j.measurement.2013.03.016.
apnea-hypopnea syndrome, Mayo Clin. Proc. 78 (2003) 1545–1552 https://doi. [26] H. Lee, J. Park, H. Kim, K. Lee, New rule-based algorithm for real-time detecting
org/10.4065/78.12.1545. sleep apnea and hypopnea events using a nasal pressure signal, J. Med. Syst.
[2] N.M. Punjabi, The epidemiology of adult obstructive sleep apnea, Proc. Am. 40 (2016) 282 https://doi.org/10.1007/s10916- 016- 0637- 8.
Thorac. Soc. 5 (2008) 136–143 https://doi.org/10.1513/pats.200709-155MG. [27] S. Grover, S. Pittman, Automated detection of sleep disordered breathing using
[3] D.J. Gottlieb, G. Yenokyan, A.B. Newman, G.T. O’Connor, N.M. Punjabi, S.F. Quan, a nasal pressure monitoring device, Sleep Breath 12 (2008) 339–345 https:
S. Redline, H.E. Resnick, E.K. Tong, M. Diener-West, E. Shahar, Prospective //doi.org/10.10 07/s11325-0 08-0181-y.
study of obstructive sleep apnea and incident coronary heart disease and [28] V.A. Rossi, J.R. Stradling, M. Kohler, Effects of obstructive sleep apnoea on heart
heart failure: the sleep heart health study, Circulation 122 (2010) 352–360, rhythm, Eur. Resp. J. 41 (2013) 1439–1451 https://doi.org/10.1183/09031936.
doi:10.1161/CIRCULATIONAHA.109.901801. 00128412.
[4] N. Botros, J. Concato, V. Mohsenin, B. Selim, K. Doctor, H.K. Yaggi, Obstructive [29] M.E. Tagluk, M. Akin, N. Sezgin, Classıfıcation of sleep apnea by using wavelet
sleep apnea as a risk factor for type 2 diabetes, Am. J. Med. 122 (2009) 1122– transform and artificial neural networks, Expert Syst. Appl. 37 (2010) 1600–
1127 https://doi.org/10.1016/j.amjmed.2009.04.026. 1607 https://doi.org/10.1016/j.eswa.2009.06.049.
[5] Y.T. Chou, P.H. Lee, C.T. Yang, C.L. Lin, S. Veasey, L.P. Chuang, S.W. Lin, Y.S. Lin, [30] L. Almazaydeh, M. Faezipour, K. Elleithy, A neural network system for detection
N.H. Chen, Obstructive sleep apnea: a stand-alone risk factor for chronic kid- of obstructive sleep apnea through SpO2 signal features, Int. J. Adv. Comput.
ney disease, Nephrol. Dial. Transplant. 26 (2011) 2244–2250 https://doi.org/10. Sci. Appl 3 (2012) https://doi.org/10.14569/ijacsa.2012.030502.
1093/ndt/gfq821. [31] D. Álvarez-Estévez, V. Moret-Bonillo, Fuzzy reasoning used to detect apneic
[6] H. Yaggi, V. Mohsenin, Sleep-disordered breathing and stroke, Clin. Chest Med. events in the sleep apnea-hypopnea syndrome, Expert Syst. Appl. 36 (2009)
24 (2003) 223–237 https://doi.org/10.1016/s0272-5231(03)00027-3. 7778–7785 https://doi.org/10.1016/j.eswa.2008.11.043.
[7] P.E. Peppard, M. Szklo-Coxe, K.M. Hla, T. Young, Longitudinal association of [32] C. Angermueller, H.J. Lee, W. Reik, O. Stegle, DeepCpG: accurate prediction of
sleep-related breathing disorder and depression, Arch. Intern. Med. 166 (2006) single-cell DNA methylation states using deep learning, Genome Biol 18 (2017)
1709–1715 https://doi.org/10.1001/archinte.166.16.1709. 67 https://doi.org/10.1186/s13059- 017- 1189- z.
[8] S.J. Kim, J.H. Lee, D.Y. Lee, J.H. Jhoo, J.I. Woo, Neurocognitive dysfunction asso- [33] D. Ravı, C. Wong, F. Deligianni, M. Berthelot, J. Andreu-Perez, B. Lo, G.Z. Yang,
ciated with sleep quality and sleep apnea in patients with mild cognitive im- Deep learning for health informatics, IEEE J. Biomed. Health Inform. 21 (2017)
pairment, Am. J. Geriatr. Psychiatry 19 (2011) 374–381 https://doi.org/10.1097/ 4–21 https://doi.org/10.1109/JBHI.2016.2636665.
JGP.0b013e3181e9b976. [34] A. Graves, A.R. Mohamed, G. Hinton, Speech recognition with deep recurrent
[9] V. Kapur, K. Strohl, S. Redline, C. Iber, G. O’Connor, J. Nieto, Underdiagnosis neural networks, in: Acoustics, Speech and Signal Processing (ICASSP), 2013,
of sleep apnea syndrome in U.S. communities, Sleep Breath 6 (2002) 49–54 pp. 6645–6649. https://doi.org/10.1109/ICASSP.2013.6638947.
https://doi.org/10.10 07/s11325-0 02-0 049-5. [35] T. Mikolov, M. Karafiát, L. Burget, J. Černocký, S. Khudanpur, Recurrent neural
[10] A. Khandoker, C. Karmakar, M. Palaniswami, Automated recognition of patients network based language model, In INTERSPEECH-2010 1045–1048.
with obstructive sleep apnoea using wavelet-based features of electrocardio- [36] R.K. Pathinarupothi, E.S. Rangan, E.A. Gopalakrishnan, R. Vinaykumar, K.P. So-
gram recordings, Comput. Biol. Med. 39 (2009) 88–96 https://doi.org/10.1109/ man, Single sensor techniques for sleep apnea diagnosis using deep learning,
TITB.2012.2185809. IEEE J. Biomed. Health Inform (2017 August) 524–529 https://doi.org/10.1109/
[11] M. Mendez, J. Corthout, S. Van Huffel, M. Matteucci, T. Penzel, S. Cerutti, ICHI.2017.37.
A. Bianchi, Automatic screening of obstructive sleep apnea from the ECG based [37] R.K. Tripathy, U.R. Acharya, Use of features from RR-time series and EEG signals
on empirical mode decomposition and wavelet analysis, Physiol. Meas. 31 for automated classification of sleep stages in deep neural network framework,
(2010) 273–289 https://doi.org/10.1088/0967-3334/31/3/001. Biocybern. Biomed. Eng. 38 (2018) 890–902 https://doi.org/10.1016/j.bbe.2018.
05.005.
 U. Erdenebayar, Y.J. Kim and J.-U. Park et al. / Computer Methods and Programs in Biomedicine 180 (2019) 105001 9

[38] S.H. Choi, H. Yoon, H.S. Kim, H.B. Kim, H.B. Kwon, S.M. Oh, K.S. Park, Real- [45] M. Sharma, S. Agarwal, U.R. Acharya, Application of an optimal class of an-
time apnea-hypopnea event detection during sleep by convolutional neural tisymmetric wavelet filter banks for obstructive sleep apnea diagnosis using
networks, Comput. Biol. Med. 100 (2018) 123–131 https://doi.org/10.1016/j. ECG signals, Comput. Biol. Med. 100 (2018) 100–113 https://doi.org/10.1016/j.
compbiomed.2018.06.028. compbiomed.2018.06.011.
[39] R.B. Berry, R. Brooks, C.E. Gamaldo, S.M. Harding, R.M. Lloyd, C.L. Marcus, [46] A. Nishad, R.B. Pachori, U.R. Acharya, Application of TQWT based filter-bank for
B.V. Vaughn, The AASM Manual for the Scoring of Sleep and Associated sleep apnea screening using ECG signals, J. Ambient Intell. Humaniz. Comput.
Events: Rules, Terminology and Technical Specifications: Version 2.3, American (2018) 1–12 https://doi.org/10.1007/s12652- 018- 0867- 3.
Academy of Sleep Medicine, 2016. [47] C.S. Viswabhargav, R.K. Tripathy, U.R. Acharya, Automated detection of sleep
[40] T. Mikolov, M. Karafiát, L. Burget, J. Černocký, S. Khudanpur, Recurrent neural apnea using sparse residual entropy features with various dictionaries ex-
network based language model, INTERSPEECH, 2010. tracted from heart rate and EDR signals, Comput. Biol. Med. 108 (2019) 20–30
[41] K. Cho, B. Van Merriënboer, D. Bahdanau, Y. Bengio, On the properties of neu- https://doi.org/10.1016/j.compbiomed.2019.03.016.
ral machine translation: encoder-decoder approaches, arXiv preprint 2014 (Ac- [48] M. Cheng, W.J. Sori, F. Jiang, A. Khan, S. Liu, Recurrent neural network based
cessed 12 Jan 2015). classification of ECG signal features for obstruction of sleep apnea detection,
[42] F. Chollet, Keras (2015) http://keras.io/. in: Computational Science and Engineering (CSE) and Embedded and Ubiqui-
[43] D.P. Kingma, J. Ba, Adam: a method for stochastic optimization. arXiv preprint tous Computing (EUC), 2, 2017, pp. 199–202. https://doi.org/10.1109/CSE-EUC.
(2014). 2017.220.
[44] A.R. Hassan, M.A. Haque, An expert system for automated identification of [49] D. Dey, S. Chaudhuri, S. Munshi, Obstructive sleep apnoea detection using con-
obstructive sleep apnea from single-lead ECG using random under sam- volutional neural network based deep learning framework, Biomed. Eng. Lett.
pling boosting, Neurocomputing 235 (2017) 122–130 https://doi.org/10.1016/j. 8 (2018) 95–100 https://doi.org/10.1007/s13534- 017- 0055- y.
neucom.2016.12.062.