Content uploaded by Manish Sharma
Author content
All content in this area was uploaded by Manish Sharma on Mar 06, 2022
Content may be subject to copyright.
Automated detection of obstructive sleep apnea in more
than 8000 subjects using frequency optimized orthogonal
wavelet filter bank with respiratory and oximetry signals
Manish Sharmaa,, Divyash Kumbhania, Jainendra Tiwaria, T.Sudheer Kumara, U.
Rajendra Acharyab,c,d
aDepartment of Electrical and Computer Science Engineering, Institute of Infrastructure, Technology,
Research and Management (IITRAM), Ahmedabad, India
bDepartment of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore 599489,
Singapore
cDepartment of Bioinformatics and Medical Engineering, Asia University, Taiwan.
dDepartment of Biomedical Engineering, School of Science and Technology, SUSS University, Singapore
Abstract
Obstructive sleep apnea (OSA) is a common respiratory disorder marked by interrup-
tion of the respiratory tract and difficulty in breathing. The risk of serious health dam-
age can be reduced if OSA is diagnosed and treated at an early stage. OSA is primarily
diagnosed using polysomnography (PSG) monitoring performed for overnight sleep;
furthermore, capturing PSG signals during the night is expensive, time-consuming,
complex and highly inconvenient to patients. Hence, we are proposing to detect OSA
automatically using respiratory and oximetry signals.The aim of this study is to develop
a simple and computationally efficient wavelet-based automated system based on these
signals to detect OSA in elderly subjects.
In this study, we proposed an accurate, reliable, and less complex OSA automated
detection system by using pulse oximetry (SpO2) and respiratory signals including
thoracic (ThorRes) movement, abdominal (AbdoRes) movement, and airflow (AF).
These signals are collected from the Sleep Heart Health Study (SHHS) database from
the National Sleep Research Resource (NSRR), which is one of the largest repositories
∗Corresponding author
Email addresses: manishsharma.iitb@gmail.com (Manish Sharma ),
divyash.kumbhani.17e@iitram.ac.in (Divyash Kumbhani),
Jainendra.tiwari.17e@iitram.ac.in (Jainendra Tiwari), Sudheer.Kumar.20pE@iitram.ac.in
(T.Sudheer Kumar), RajendraUdyavaraACHARYA@np.edu.sg (U. Rajendra Acharya)
Preprint submitted to Journal of L
A
T
E
X Templates March 6, 2022
of publicly available sleep databases. The database comprises of two groups SHHS-1
and SHHS-2, which involves 5,793 and 2,651 subjects, respectively with an average
age of ≥60 years. The 30-s epochs of the signals are decomposed into sub-bands using
frequency optimized orthogonal wavelet filter bank. Tsallis entropies are extracted
from the sub-band coefficients of wavelet filter bank. A total 4,415,229 epochs of
respiratory and oximetry signals are used to develop the model. The proposed model
is developed using GentleBoost and Random under-sampling Boosting (RUSBoosted
Tree) algorithms with 10-fold cross-validation technique.
Our developed model has obtained the highest classification accuracy of 89.39%
and 84.64% for the imbalanced and balanced datasets, respectively using 10-fold cross-
validation technique. Using the 20% hold-out validation, the model yielded an accu-
racy of 88.26% and 84.31% for the imbalanced and balanced datasets, respectively.
Hence, the respiratory and SpO2 signals-based model can be used for automated OSA
detection. The results obtained from the proposed model are better than the state-of-
the-art models and can be used in-home for screening the OSA.
Keywords: Obstructive sleep apnea, Tsallis entropy, polysomnography (PSG),
GentleBoost, SpO2, Airflow .
1. Introduction
Sleep occupies roughly one-third of a person’s life. Good quality of sleep is very
essential for a human’s life as it aid in resetting the body, improves learning capacity,
physical development, emotional control, and overall quality of life. Apneic episodes
leads to oxygen deficit and arousals, keeping patients from sleeping and jeopardis-5
ing their health and well-being.[1]. Sleep-related disorders like narcolepsy, sleep ap-
nea, cataplexy, hypersomnia, insomnia, and others have a devastating impact on health
and quality of life [2, 3, 4]. Hence, a method for diagnosing sleep disorders must
be devised [5, 6, 7, 8, 9]. According to the American Academy of Sleep Medicine
(AASM), sleep apnea is the most frequent sleep condition [10]. Sleep apnea can be10
defined as a sleep-related condition marked by breathing problems while sleeping [11].
The Apnea-Hypopnea Index (AHI), which indicates the occurrences of apnea and hy-
2
poapnea events per hour during total sleep time (TST), and is considered as the most
important metric for diagnosing the presence and seriousness of the condition. The
hypoapnea is another type of sleep disorder in which the patient experiences shallow15
breaths due to partial (≥30%) airway blockage for 10 seconds or more duration accom-
panied by oxygen desaturation of 3% or more [12]. According to some studies [13],
around 200 million people are suffering from sleep apnea. This disease affects 4% of
adult males and 2% of adult females, making it more prevalent in men than in women
[14]. The prevalence of sleep apnea increases with age and is highest among elderly20
population [15]. Elderly patients suffering from apnea are observed to have shown a
decline in cognitive functions and are at a higher risk of stroke [16, 17] and mortality
[18].
Sleep apnea is categorised into several types based on its cause: (i) central sleep
apnea (CSA), (ii) obstructive sleep apnea (OSA) and (iii) mixed sleep apnea (MSA)25
[19]. The OSA occurs when the air passage is stopped by the throat muscles leading
to more than 90% reduction in the airflow for at least two respiratory cycles, and CSA
occurs when the impulses from our brain that regulate breathing are disrupted [20].
The MSA is characterized by the combination of OSA and CSA. It begins with CSA
and progresses towards the OSA [21].OSA is the most prevalent among all types of30
sleep apnea and is positively correlated to obesity and age [22]. Studies [22] show
that around 22% men and 17% women on average suffer from OSA if we consider an
AHI ≥5. OSA is also linked to an increased risk of stroke, arrhythmias, hyperten-
sion, preoperative problems, and myocardial involvement during sleep, in addition to
personal pain, annoyance, and tiredness during the day. Several variables can play a35
part in the growth of OSA. The presence of OSA is most commonly indicated by bone
malformation and increased soft tissue around the airway and the other factor is muscle
activity [23]. Furthermore, it plays a critical role in memory loss, obesity, impotence,
and weakness. Sleep disorders can have a significant societal cost as well as a negative
impact on one’s quality of life. Early observation of OSA may aid in the prevention of40
potentially harmful health consequences [24].
In recent years, simplifying the diagnosis of OSA has become a major research fo-
cus. The nocturnal polysomnogram (PSG) [25] is the gold standard in OSA diagnosis.
3
It is designed to diagnose the normal overnight recurrence of apneas and hypopneas,
which lead to intermittent hypoxia (insufficient gas exchange), and disrupted sleep. The45
PSG usually needs at least 11 channels which also includes electrocardiogram (ECG)
[26, 27], electroencephalogram (EEG), respiratory effort, oxygen saturation of arterial
blood through blood gas analysis (SaO2) and airflow (AF) for different bio-signal mea-
surements and twenty two wires linked to the patient who is being monitored during
the night [21, 24, 28].50
The large number of signals are needed to track patients during PSG and complex
acquisition equipment is required. PSG is a costly test due to its difficulty and the need
for qualified staff overnight. Furthermore, there is a scarcity of specialized laboratories
that can perform the PSG test. Due to the limitations of PSG, researchers are looking
for other ways to identify sleep apnea-hypopnea syndrome (SAHS). A lot of effort has55
been made on the identification of simplified method. One of the easiest methods to
simplify the diagnostic test is to assess a reduced collection of signals rather than the
whole set used in PSG. [21, 29]. Nocturnal home pulse oximetry (NHPO) is a reliable
screening method that measures a patient’s oxygenation in their normal surroundings.
Generally NHPO can be used as a preliminary step in the screening process for OSA60
[30]. Signals that measure respiratory movements are important since breathing is a
key aspect in detecting sleep apnea [23].
The SpO2signal and respiratory signals like thoracic (ThorRes), abdomen (Ab-
doRes) and AF can be considered to develop a potential low-cost method for determin-
ing apnea severity. Pulse oximetry uses photoplethysmographic pulses at two wave-65
lengths, commonly in the red and infrared, to measure SaO2 (oxygen saturation in
arterial blood) noninvasively. [31]. Respiratory inductive plethysmography (RIP) is
the most extensively used method for noninvasive quantitative and qualitative respira-
tory measurements in adults and children [32]. The ThorRes and AbdoRes signals are
measured with thoracic and abdominal belts using the RIP method. AF signals are cap-70
tured using temperature-sensitive devices called thermistors and thermocouples. The
Figure 1 represents the typical respiratory signals during OSA event.
4
Figure 1: Typical airflow, SpO2, abdominal (AbdoRes) and thoracic (ThorRes) movement signals during
OSA event.
5
1.1. Related Work
To identify OSA, several studies have been conducted in the literature. Biswal et
al. [33] published a research article on sleep stage classification and sleep apnea de-75
tection by using recurrent and convolutional neural networks (RCNN). They obtained
an accuracy of 83.2% using SHHS-1 dataset. Golroul et al. [34] used SaO2, AF, Thor-
Res and AbdoRes for the detection of sleep apnea using fuzzy logic and achieved an
accuracy of 83.7%. ´
Alvarez et al. [35] have used multivariate analysis of SaO2record-
ings to detect OSA with 148 subjects obtained from the Hospital Universitario P´
ıo del80
R´
ıo Hortega of Valladolid (Spain) and reported an accuracy of 89.7%. Barroso-Garc´
ıa
et al. [36] focused on OSA detection in children by using discrete wavelet analysis
and airflow signal. Sharma et al. [24] detected OSA using ECG signals and an opti-
mal class of anti-symmetric wavelet filter banks, however, they have used a different
dataset named Apnea-ECG with only 70 overnight recordings only. Recently, Uddin85
et al. [37] have developed an automated sleep apnea detection system using AF and
oximetry signals and SHHS dataset. They have used only 988 recordings, whereas we
have used more than 8,000 recordings. Majority of these studies have been conducted
on pediatric subjects [38]. Very few studies are performed on elderly subjects (aged
60 years or more) [39, 40, 41]. OSA is considered more severe and most prevalence90
among all types of sleep apnea. The majority of studies in the literature are on the
detection of obstructive sleep apnea [24, 37, 42, 43]. Hence, we have detected only
OSA events in this study. And, the traces of CSA events were found in only 4,191 (less
than 50% of OSA subjects) subjects, which would make the data further imbalanced.
Hence, we have considered only OSA. However, in our future work, we can take all95
four types of events.
1.2. Proposed Work
In this study, we provide a new OSA detection system based on respiratory and
oximetry signals for elderly subjects of age ≥60 years. This study uses a new class of
frequency optimized wavelet orthogonal filter bank for the subband decomposition of100
the signals considered. Feature extraction is done with Tsallis entropy, while OSA de-
tection is done using supervised machine learning classifiers. Unlike previous studies,
6
the proposed system uses a large database (SHHS) which includes SHHS-1 (5,793 Sub-
jects with average 63 years) and SHHS-2 (2,651 subjects with average age 68 years).
For the imbalanced and balanced SHHS database, the proposed work yielded the high-105
est classification accuracy of 89.39% and area under the curve (AUC) of 0.904 and an
accuracy of 84.64% and AUC of 0.932, respectively. To analyse the performance of the
proposed model, we employed RUSBoost and GentleBoost classifiers with a 10-fold
cross-validation technique. The proposed model also achieved the accuracy of 83.97%
and 84.21% for imbalanced data and balanced data with the SHHS-1 dataset. Simi-110
larly, it achieved 89.50% and 88.45% accuracy for the imbalanced data and balanced
data with SHHS-2 dataset. The proposed study, to the best of our knowledge, is the
pioneered study that used such a large number of elderly subjects for OSA detection
using wavelet subbands of respiratory and oximetry signals. This is the first study that
uses all (8,444 PSG) recording of both SHHS-1 and SHHS-2 databases. Further, a total115
4,415,229 epochs of each signals of 15 different signals are considered to develop the
model. A large number of subjects, PSG recordings, different types of signals, and
huge number epochs have been used to develop the model. The suggested model has
yielded a high classification performance, which indicates the robustness and accurate-
ness of the proposed system. The system is simple, accurate and robust. Hence, it can120
be tested for installation in home-based environment for OSA detection, specially for
elderly people.
2. Methods and Materials
2.1. Database
The proposed study used PSG data from the Sleep Heart Health Study (SHHS) [44]125
database, which is accessible through the National Sleep Research Resource (NSRR).
The SHHS database consists of 8,444 subjects and it includes two subsets namely
SHHS-1 with 5,793 subjects (avg. age 63 yrs) gathered between November 1995, and
January 1998, and SHHS-2 with 2,651 subjects (avg. age 68 yrs) collected from Jan-
uary 2001 to June 2003. This dataset is a good representation of elderly subjects. Every130
set includes one ECG channel, one EMG channel, two bipolar EEG channels (C4-A1
7
and C3-A2), two-channel EOGs, dual-channel respiratory inductance (abdominal and
thoracic), position channel, light channel, SaO2channel, and an airflow (AF) channel.
Depending on the signal nature, the sampling frequency ranges from 1 Hz to 256 Hz.
A thermistor is used to monitor nose respiration, whereas inductive plethysmography135
bands are used to assess thoracid (ThorRes) and abdominal (AbdoRes) excursions [23].
The PSG recording in the database is available in the form of a european data format
(EDF), and XML files annotated for every sleep stage as per Rechtschaffen & Kales
(R&K) rules [45]. The summary of the database is given in the Table 1 and the channel
information is given in Table 2.140
Table 1: Detailed description of sleep heart health study (SHHS) database.
Description SHHS-1 SHHS-2
No. of Subjects 5,793 2,651
Dataset Size 216 GB 137 GB
Age (year) 63 (55 - 72) 68 (60 - 76)
Male Patients 3,033 1,425
Female Patients 2,760 1,226
Body Mass Index (BMI) 27.5 (24.7 - 30.8) 27.7 (24.9 - 31.1)
Epworth Sleepiness Scale
7 (4 - 11) 7 (4 - 10)
(ESS) score
Total Sleep Time (minutes) 367.5 (321 - 406) 381.5 (336 - 417.9)
Sleep Efficiency (%) 72.8 (64.2 - 80.1) 63.6 (55.7 - 70.5)
AHI ( per hour) 24.7 (15.2 - 37.7) 14.6 (9.17 - 22.9)
** Data are median and range is IQR
2.2. Proposed method
The flow diagram for the proposed study is shown in Figure 2. It represents ex-
traction of respiratory signals from the database followed by segmentation and pre-
8
Table 2: Details of the channels and subjects used in our proposed method.
Channel
Frequency Number of Subjects
(Hz) SHHS - 1 SHHS - 2
SpO21 5,793 2,651
AF 10 3,943 2,535
AbdoRes, ThorRes 10 5,793 2,535
processing, 5-Level wavelet decomposition, feature extraction, and classification. The
respiratory signals are segmented into 30-sec epochs and then pre-processed. 5-level 1-145
D wavelet decomposition with an orthogonal wavelet filter bank is used on each epoch
to obtaind sub-bands. Tsallis entropy is extracted from each sub-band coefficients. Fi-
nally, GentleBoost and RUSBoosted tree classifiers are employed to classify OSA and
normal epochs.
2.2.1. Segmentation and Preprocessing150
In this study, we separated respiratory signals from every subject in the database.
The SHHS database consists of European Data Format (EDF) file and XML file for
every subject. The XML file contains R&K Specified annotations for each sleep stage
of 30 sec epoch and sleep disorder events. The signal data is stored as a matrix, which
is subsequently normalized for segmentation into 30 sec epoch.155
Table 3: Details of number of epochs used in imbalanced and balanced data.
SHHS - 1 SHHS - 2
Imbalanced Balanced Imbalanced Balanced
OSA 104,781 (4.7 %) 104,781 (50 %) 59,430 (2.7 %) 59,430 (50 %)
Normal 2,118,282 (95.3 %) 104,781 (50 %) 2,132,736 (97.3 %) 59,430 (50 %)
Total 2,223,063 209,562 2,192,166 118,860
9
Extraction of Respiratory
channels from SHHS dataset
Segmentation and pre-processing
Wavelet Decomposition of level 5
Feature Extraction
Classification
OSA Normal
Figure 2: Schematic diagram of the proposed method.
10
2.2.2. Design of wavelet based filter banks
In order to obtain sub-bands of the signals, we have used an optimal orthogonal
wavelet filter banks. The optimal filter bank has designed by minimizing bandwidth
of the wavelet filter of the given length for given number of zero moments of the filter
[46]. For the analysis of non-stationary signals, such as physiological signals, the160
bandwidth of localised filters is highly desirable [47, 48, 49]. Hence, in this study, we
used band-width localized filter of length 18, and with 7 number of zero-moments [46].
The frequency response of the filter with bandwidth localized filter of length 18, and
with 7 zero-moments are shown in Figure 3 and 4, respectively.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized Frequency ( rad/samples)
-50
-45
-40
-35
-30
-25
-20
-15
-10
-5
0
Magnitude in dB
Low Pass Filter
High Pass Filter
Figure 3: Magnitude response of the bandwidth localized filter of length 18, and with 7 number of zero-
moments used in the study.
2.2.3. Feature extraction165
Tsallis entropy (TE): Tsallis statistics have been used to describe a wide variety
of phenomena in physics, chemistry, biology, medicine, economics, and geophysics,
among other fields [50]. Transient feature extraction is a crucial element of signal
11
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized Frequency ( rad/samples)
0
0.2
0.4
0.6
0.8
1
1.2
Magnitude
Low Pass Filter
High Pass Filter
Figure 4: Magnitude response of the bandwidth localized filter of length 18, and with 7 number of zero-
moments used in the study.
12
analysis, and wavelet entropy has grabbed the interest of researchers around the world
to develop a new feature extraction methodology [51]. An easy approach to introduce170
Tsallis entropy (TE) is a generalization of Boltzmann-Gibbs-Shannon (BGS) entropy
and it is nonextensive [50].
T E =
n
∑
i=1
xi−xi2
where xidenotes ith sample of the N-length wavelet coefficient sequence x(n).
Because of long-range interactions, physiological signals have nonextensivity. In
nonextensive statistical mechanics, Tsallis entropy is critical. It can be used to describe175
systems with long-range interactions, multifractal space-time constraints, or long-term
memory effects, as well as non-stationary nature. [52, 53]. The Tsallis entropy has
the advantage of being able to measure the uncertainty of generalised systems. The
combination of wavelet analysis with Tsallis entropy can be considered an excellent
candidate for the analysis of transient non-stationary biomedical signals. Furthermore,180
Tsallis entropy has only a summation part in its mathematical expression. As a result,
Tsallis entropy takes significantly less time to compute than other methods, giving
it an advantage in terms of computational complexity.The physiological signals are
nonlinear, transient and non-stationary. Hence, Tsallis Entropy seems a good candidate
for their analysis.185
2.3. Classification and performance evaluation
With a 10-fold cross-validation (CV) technique, the GentleBoost and RUSBoost
classifiers are employed for the classification of normal and OSA classes. The k-
fold cross-validation approach is often used to develop a robust model with the least
amount of over-fitting and redundancy [24]. To evaluate the classification performance190
of both datasets individually and also in combination, the average classification accu-
racy (ACC), precision, sensitivity, specificity, F1-Score, and area under the ROC curve
(AUC) are computed.
In this work, the RUSBoost classifier performed the best for the detection of OSA
with imbalanced data. There are three types of resampling techniques: oversampling,195
undersampling, and hybrid. Oversampling methods generate a superset of the origi-
nal dataset by reproducing certain instances or generating new examples from existing
13
ones, whereas undersampling methods produce a subset of the original dataset by re-
moving a specific case from the majority class [54]. A hybrid method combines both
over and undersampling techniques. The numerous data sampling approaches are ex-200
plicitly developed to solve the problem of class imbalance. Boosting is an approach that
can enhance the performance of weak classifiers [55]. When training data is skewed,
RUSBoost combines data sampling and boosting to create a simple and quick for im-
proving classification performance. [56].
In the proposed study GentleBoost classifier provides optimal performance with205
balanced data. GentleBoost was introduced by Friedman et al. [57] as an alterna-
tive to AdaBoost for improving boosting performance in the face of outliers. To give
less weight to misclassified samples, GentleBoost uses a quadratic objective function
of margin, which increases at a slower rate than AdaBoost’s exponential function of
margin [58]. Furthermore, GentleBoost is a method for reducing exponential loss that210
uses decision trees as base learners. As a result, when compared to other ensemble
techniques, it has a smaller ensemble error [59].
Table 4: Tuning parameters used for the RUSBoost and GentleBoost classifier for imbalanced and balanced
data.
SHHS - 1 SHHS - 2 SHHS 1 + SHHS 2
Imbalanced Balanced Imbalanced Balanced Imbalanced Balanced
Classifier RUSBoost GentleBoost RUSBoost GentleBoost RUSBoost GentleBoost
No. of observations 2,223,063 209,562 2,192,166 118,860 4,415,229 328,422
No. of Learner 13 497 13 496 22 498
Max. no. of Splits 42,169 1,171 525,350 452 572 101,588
Learning Rate 0.168 0.399 0.391 0.051 0.893 0.013
Prediction Speed (obs/sec) 140,000 5,100 160,000 5,300 140,000 3,600
Training Time (sec) 1,333.3 7,914.7 747.8 2,883.2 539.6 25,493
14
3. Classification results
The proposed work was carried out using a computer with 8 GB of RAM, an In-
tel i5 10th generation processor running at 1.60 GHz, and MATLAB R2020a (version215
9.8.0.1323502). We conducted our assessment on both ’imbalanced’ and ’balanced’
data. In the following two subsections, the obtained results for both datasets are pre-
sented.
3.1. Classification results obtained for ’imbalanced’ data
We have collected respiratory signals from the original SHHS database to detect220
OSA. A total of 4,415,229 epochs (2,223,063 epochs from SHHS-1 and 2,192,166
epochs from SHHS-2). Table 3 shows the detailed classification of epochs employed
for the proposed method. We have used the RUSBoost classifier for imbalanced data
and the learning rate, prediction speed, and training time of the classifier are given in
the Table 4.225
Table 5 shows the classification accuracy of the proposed model using individ-
ual channels and combined channels of respiratory signals with SHHS-1 and SHHS-
2 datasets. We have obtained the maximum classification accuracy of 83.97% and
89.50% with combined channels of SHHS-1 and SHHS-2 datasets. Table 7 represents
the confusion matrix of the proposed model for both imbalanced and balanced data.230
To ensure the robustness of the proposed method, we have employed 10-fold CV tech-
nique. The performance parameters of the proposed model for both imbalanced and
balanced data are shown in Table 8.
3.2. Classification results obtained for ’balanced’ data
To overcome the bias and inaccuracy associated with classification using imbal-235
anced data, we balanced the epochs in every dataset and evaluated the performance
of our proposed method. The under-sampling approach was employed to balance the
number of respiratory epochs for healthy and OSA subjects in all datasets. For in-
stance, the number of normal and OSA epochs in the imbalanced SHHS-2 dataset are
2,132,736 and 59,430, respectively. The number of normal epochs is reduced to the240
15
number of OSA epochs by randomly selecting 59,430 normal epochs from the unbal-
anced normal epochs. As a result, the balanced data of SHHS-2 dataset includes 59,430
normal and OSA epochs. Similar approach is employed to balance normal and OSA
epochs of SHHS-1 dataset also.
For classification using balanced data, a total of 209,562 epochs (104,781 epochs245
for both normal and OSA) in the SHHS-1 dataset and 118,860 epochs (59,430 epochs
for both normal and OSA) in the SHHS-2 dataset are used. The epoch distribution for
both balanced and imbalanced datasets is depicted in Table 3.
The GentleBoost classifier is applied for balanced data, and the classifier parame-
ters are shown in Table 4. We have achieved the maximum accuracy of 84.21% and250
88.45% using SHHS-1 and SHHS-2 balanced datasets as shown in Table 5. From Ta-
ble 6, we can observe that balancing significantly improves the F1-score for OSA class.
It implies that the model is performing better to classify OSA events. The confusion
matrices for both imbalanced and balanced data using 10-fold CV are shown in Ta-
ble 8. We have obtained AUC of 0.921 and 0.953 for SHHS-1 and SHHS-2 datasets,255
respectively. The AUC values are close to one, indicating better discrimination ability.
We have also obtained maximum classification accuracy of 89.39% and AUC of
0.904 as well as the accuracy of 84.64% and AUC of 0.932 by combining both datasets
(SHHS-1 and SHHS-2) with imbalanced and balanced data, respectively. Table 7
shows the confusion matrices obtained using the RUSBoost and GentleBoost classi-260
fiers for each dataset, as well as by combining both datasets (SHHS-1 and SHHS-2)
with imbalanced and balanced data. The performance metrics of the proposed model
for each dataset and combination of datasets are shown in the Table 8.
4. Discussion
The present work focuses on automatic detection of OSA in elderly subjects using265
oximetry and respiratory signals. There are very few studies in the literature which have
considered elderly subjects (average age of more than 60 years) and which used on the
combination of SpO2and respiratory signals [60, 38]. We proposed a new orthogonal
filter bank for 5-levels using 1-D wavelet decomposition. The novelties of the present
16
Table 5: Classification accuracy (%) of the proposed method with individual and combined channels with
imbalanced and balanced data.
Channels
SHHS - 1 SHHS - 2
Imbalanced Balanced Imbalanced Balanced
SpO2+ ThorRes + AbdoRes + AF 83.97 84.21 89.50 88.45
SpO2+ ThorRes + AbdoRes + AF + ECG 79.85 82.09 83.79 86.47
SpO2+ ThorRes + AbdoRes 80.41 81.25 84.04 83.82
SpO2+ ThorRes + AbdoRes + ECG 80.36 81.63 84.21 84.51
SpO2+ ThorRes + AF 79.72 81.02 83.86 85.91
SpO2+ ThorRes + AF + ECG 79.78 81.82 83.75 86.46
SpO2+ AbdoRes + AF 79.75 80.93 83.89 85.58
SpO2+ AbdoRes + AF + ECG 79.68 81.76 83.93 86.09
SpO2+ AF 80.36 80.11 83.83 84.79
SpO2+ ThorRes 80.08 80.72 84.01 83.46
SpO2+ AbdoRes 80.02 80.64 83.93 82.89
ThorRes + AbdoRes + AF 69.27 74.02 72.01 81.89
ThorRes + AbdoRes 66.92 72.63 74.63 76.88
ThorRes + AF 59.26 69.69 71.93 80.76
AbdoRes + AF 64.25 69.67 70.95 80.11
SpO281.29 79.81 83.81 81.49
ThorRes 57.54 66.77 76.34 73.02
AbdoRes 65.14 66.76 56.48 70.83
AF 48.39 55.46 71.38 75.73
ECG 56.94 72.31 26.92 81.14
17
Table 6: F1 Score of the proposed method with individual and combined channels with imbalanced and
balanced data.
Channels
SHHS - 1 SHHS - 2
Imbalanced Balanced Imbalanced Balanced
SpO2+ ThorRes + AbdoRes + AF
N 0.909 0.843 0.943 0.884
O 0.312 0.842 0.293 0.885
SpO2+ ThorRes + AbdoRes + AF + ECG
N 0.884 0.825 0.910 0.865
O 0.246 0.817 0.206 0.864
SpO2+ ThorRes + AbdoRes
N 0.887 0.817 0.911 0.840
O 0.245 0.808 0.198 0.836
SpO2+ ThorRes + AbdoRes + ECG
N 0.887 0.821 0.912 0.847
O 0.245 0.812 0.200 0.843
SpO2+ ThorRes + AF
N 0.883 0.814 0.910 0.860
O 0.245 0.806 0.206 0.858
SpO2+ ThorRes + AF + ECG
N 0.883 0.822 0.910 0.865
O 0.246 0.814 0.206 0.864
SpO2+ AbdoRes + AF
N 0.883 0.814 0.910 0.857
O 0.244 0.805 0.206 0.855
SpO2+ AbdoRes + AF + ECG
N 0.883 0.822 0.911 0.861
O 0.243 0.813 0.206 0.861
SpO2+ AF
N 0.887 0.807 0.910 0.849
O 0.242 0.795 0.205 0.847
SpO2+ ThorRes
N 0.885 0.812 0.911 0.837
O 0.243 0.802 0.198 0.832
SpO2+ AbdoRes
N 0.885 0.812 0.911 0.833
O 0.243 0.801 0.197 0.825
ThorRes + AbdoRes + AF
N 0.812 0.741 0.833 0.818
O 0.158 0.740 0.133 0.820
ThorRes + AbdoRes
N 0.794 0.725 0.852 0.765
O 0.157 0.727 0.122 0.770
ThorRes + AF
N 0.733 0.696 0.833 0.807
O 0.141 0.698 0.133 0.808
AbdoRes + AF
N 0.774 0.699 0.826 0.800
O 0.147 0.694 0.129 0.802
SpO2
N 0.893 0.804 0.910 0.819
O 0.243 0.792 0.192 0.811
ThorRes
N 0.718 0.665 0.863 0.727
O 0.137 0.670 0.120 0.734
AbdoRes
N 0.781 0.669 0.707 0.709
O 0.146 0.666 0.088 0.711
AF
N 0.638 0.565 0.829 0.756
O 0.101 0.544 0.127 0.759
ECG
N 0.717 0.735 0.401 0.815
O 0.103 0.710 0.064 0.808
*** O: OSA; N: Normal
18
Table 7: Confusion matrices obtained for the proposed method using combination of channels with both
imbalanced and balanced data.
SHHS - 1
Imbalanced Balanced
N O N O
Actual
N 1,785,887 (84.3 %) 332,395 (15.7 %) 88,661 (84.6 %) 16,120 (15.4 %)
O 23,869 (22.8 %) 80,912 (77.2 %) 16,970 (16.2 %) 87,811 (83.8 %)
Predicted Predicted
SHHS - 2
Imbalanced Balanced
N O N O
Actual
N 1,914,329 (89.8 %) 218,407 (10.2 %) 52,262 (87.9 %) 7,168 (12.1 %)
O 11,768 (19.8 %) 47,662 (80.2 %) 6,561 (11.0 %) 52,869 (89.0 %)
Predicted Predicted
SHHS 1 + SHHS 2
Imbalanced Balanced
N O N O
Actual
N 3,829,965 (90.1 %) 421,053 (9.9 %) 139,691 (85.07 %) 24,520 (14.93 %)
O 47,620 (29.0 %) 116,591 (71.0 %) 25,930 (15.79 %) 138,281 (84.21 %)
Predicted Predicted
*** O: OSA; N: Normal
19
Table 8: Performance parameters obtained for the proposed method using imbalanced and balanced data.
Accuracy Precision Sensitivity Specificity F1-Score Kappa AUC
(%) (%) (%) (%)
SHHS - 1
Imbalanced
N
83.97
98.68 84.31 77.22 0.909
0.256 0.891
O 19.58 77.22 84.31 0.312
Balanced
N
84.21
83.93 84.62 83.80 0.843
0.684 0.921
O 84.49 83.80 84.62 0.842
SHHS - 2
Imbalanced
N
89.50
99.39 89.76 80.20 0.943
0.260 0.932
O 17.91 80.20 89.76 0.293
Balanced
N
88.45
88.85 87.94 88.96 0.884
0.769 0.953
O 88.06 88.96 87.94 0.885
SHHS 1 + SHHS 2
Imbalanced
N
89.39
98.77 90.10 71.00 0.942
0.292 0.904
O 21.69 71.00 90.10 0.332
Balanced
N
84.64
84.34 85.07 84.21 0.847
0.693 0.932
O 84.94 84.21 85.07 0.846
*** O: OSA; N: Normal
20
study are as follows: (i) automated technique for detecting OSA using respiratory and270
SpO2signals; (ii) Novel filter bank with certain desirable properties are presented; (iii)
Features are extracted using Tsallis entropy which is computationally less intensive.
Table 9 represents the comparison of the proposed method with previous study con-
ducted. Our proposed strategy has outperformed earlier studies, as seen in the table.
Biswal et al. [33] used only SHHS-1 for sleep apnea detection using convolutional275
neural networks (CNN) and recurrent neural networks (RNN) and achieved an accu-
racy of 83.2%. They have combined all respiratory events (hypopnea, central apnea,
obstructive apnea, and mixed apnea) and took them as a single class (apnea event).
They employed respiratory signals and transformed them to spectrograms to classify
presence or absence of apnea using CNN and RNN. They have computed twelve time-280
domain and eighty four frequency-domain features, resulting in high computational
complexity. However, the proposed method used the same respiratory signals with
both datasets (SHHS-1 and SHHS-2) and obtained maximum accuracy of 89.39% with
RUSBoost classifier using Tsallis entropy features.
Wavelet entropy combines wavelet decomposition and entropy, thus, it offers the285
advantages of both multi-resolution analysis and complexity evaluation for time-varying
signals. Hence, the macro and micro aspects of some signals could be captured using
wavelet entropy. Therefore, wavelet entropy has been used in analyzing physiological
signals. However, the Shannon wavelet entropy has been extensively used. Shannon
wavelet energy entropy has some limitations in processing non-stationary signals, re-290
sulting in inaccurate results. Tsallis entropy is good at expressing the uncertainty of
generalized systems; with a combination of Tsallis entropy and wavelet decomposition,
a wavelet entropy-Tsallis wavelet energy entropy is constructed. The experimental re-
sults show that Tsallis wavelet energy entropy is better than Shannon wavelet energy
entropy in analyzing transient signals[52, 53]. Thus, the combination of wavelet anal-295
ysis with Tsallis entropy can be considered an excellent candidate for the analysis of
transient non-stationary biomedical signals. The combination of wavelet filter bank
used and Tsallis entropy produced highly discriminative features. Further,we have
used frequency localized orthogonal wavelet filter bank, which possesses simultane-
ously good localization in time due to finite support and good localization in frequency300
21
due to minimum possible bandwidth. Hence, the bandwidth localized orthogonal fil-
ter banks can be regarded as an excellent tool for the analysis of transient and non-
stationary physiological signals. The minimum time-frequency product of wavelets
leads to an excellent analysis of non-stationary signals. Hence, the proposed Wavelet-
Tsallis Entropy based model perform better.305
Simpler oximetric measures, such as the mean of desaturations are used in the
OSA’s detection through oximetry. However, a systematic underestimating of the con-
dition has been found, and performance varied significantly in the study [61]. In the
present study, we proposed a simple approach using oximetry and respiratory signals.
Our experimental results indicate that the classification accuracy obtained in detecting310
OSA using a large dataset (SHHS-1 & SHHS-2) and a combination of SpO2, ThorRes,
AbdoRes, and AF signals outperformed the other study [33]. This may be because our
new filter bank and Tsallis entropy features are able to pick salient signatures from the
physiological signals accurately.
Table 9: Comparison of proposed method with state-of-the-art apnea detection systems using respiratory
signals with same SHHS database.
Study Methodology Dataset Accuracy (%)
Biswal et.al. [33] AHI detection using CNN + RNN SHHS - 1 83.20
Proposed work OSA detection using Machine learning
SHHS - 1 83.97
SHHS - 2 89.50
SHHS - 1 & SHHS - 2 89.39
Our proposed model has following advantages:315
• It is the first study to use 8,444 (5,791 from SHHS-1 and 2,535 subjects from
SHHS-2 dataset) elderly subjects for automated OSA detection.
• For the detection and diagnosis of OSA , we used respiratory (ThorRes, Ab-
doRes, AF) signals and SpO2instead of traditional ECG signal. As shown in
Table 5, we have analyzed their individual performances as well as 11 different320
22
combinations of these signals and performed more than 124 classification tasks.
It can be observed that individual SpO2signal yielded better OSA classification
performance as compared to individual respiratory signals. However, combina-
tion of these signals significantly improved the performance.
• We have used shorter epoch duration (30s) as it helps in lowering the time com-325
plexity.
• Moreover, our method is simpler than others because we just employed Tsallis
entropy to detect OSA. The method is simple and requires minimal complexity.
As a result, it can be incorporated in sleep monitoring system.
• In this study, we have also explored the possibility of using respiratory and SpO2
330
signals simultaneously rather than applying them separately. Our results show
that combination of signals yielded better results than individual signals.
• The source data is imbalanced, but a balanced dataset is constructed as well, and
both datasets are used to identify OSA. The classification performance of the
model obtained using both datasets is consistent, indicating that our model is not335
biased towards any class and 10-fold cross-validation is also employed to obtain
a robust system.We used 10-fold cross-validation to avoid possible over-fitting of
the model. Since we have used 10-fold cross-validation and the number subjects,
and number of epochs considered are massive the probability of using the same
epochs of the same subject in training and testing both is less. Similar approach340
is used in recent studies [62, 63, 64] on apnea detection using different types of
validation.
• The highest accuracy of 84.64% using balanced and 89.39% using an imbalanced
datasets is achieved by the proposed model. Using the same SHHS database,
each of these results outperformed earlier state-of-the-art systems for OSA de-345
tection.
• Obtained high classification accuracy even with huge SHHS database with ten-
fold cross-validation which confirms that, our model can be employed in real-
world situations.
23
• In order to ensure the robustness of the proposed model, we have also performed350
hold-out validation for both balanced and imbalanced data sets. In the hold-out
validation, testing data (20%) and training data (80%) are mutually exclusive,
meaning subjects involved in testing were not involved in training and vice versa.
It is to be noted that the results obtained from both 10-fold cross-validation and
20% hold-out validation are comparable. The confusion matrices correspond-355
ing to hold-out validation is given in Table 10 for the case when all signals are
combined. The results obtained using 10-fold cross-validation and 20% hold-out
validation are compared in Table 11. It can be observed from the table that there
is small difference (0.24% to 1.46%) in the average classification accuracies ob-
tained with 10-fold cross-validation and 20% hold-out validation strategies. This360
justifies that our model is robust, and accurate as it performed equally well with
different validations.
The limitations of the proposed model are explained as follows:
• Since we are using respiratory channels, the patients may experience discomfort.
Furthermore, home-based mobile sensors/recording units are relatively costly.365
• It is difficult to determine the appropriate number of decomposition stages in our
wavelet-based technique in advance.
• Rather than using a Nasal Pressure (NP) sensor, a thermistor is used to obtain the
AF signal. Thermistor’s measurements are indirectly related to only AF, which
may discard the presence of hypopneas [65].370
Deep learning algorithms have been widely used for discriminating various phys-
iological signals [66]. Moreover, while dealing with huge databases, DL-based algo-
rithms perform well. Hence, in our future study, we intend to evaluate the usage of
deep learning (DL) methodologies such as CNN, RNN, and long short-term memory
(LSTM) networks, as well as auto-encoders for detecting various sleep disorders like375
OSA, narcolepsy, restless leg syndrome, etc.
24
Table 10: Confusion matrices obtained for 20% hold-out validation method using combination of all channels
with both imbalanced and balanced data.
SHHS - 1
Imbalanced Balanced
N O N O
Actual
N 1,781,015 (84.08 %) 337,267 (15.92 %) 86,901 (82.94 %) 17,880 (17.06 %)
O 24,561 (23.44 %) 80,220 (76.56 %) 17,190 (16.41 %) 87,591 (83.59%)
Predicted Predicted
SHHS - 2
Imbalanced Balanced
N O N O
Actual
N 1,882,551 (88.27 %) 250,185 (11.73 %) 51,454 (86.58 %) 7,976 (13.42 %)
O 11,911 (20.04 %) 47,519 (79.96 %) 7,066 (11.89 %) 52,364 (88.11 %)
Predicted Predicted
SHHS 1 + SHHS 2
Imbalanced Balanced
N O N O
Actual
N 3,782,779 (88.99 %) 468,239 (11.01 %) 139,166 (84.75 %) 25,045 (15.25 %)
O 49,952 (30.42 %) 114,259 (69.58 %) 26,472 (16.12 %) 137,739 (83.88 %)
Predicted Predicted
*** O: OSA; N: Normal
25
Table 11: Comparison of classification accuracy (%) using different validation techniques for the combined
signals.
Validation type
SHHS 1 SHHS 2 SHHS 1 + SHHS2
Imbalanced Balanced Imbalanced Balanced Imbalanced Balanced
10-fold CV 83.97 84.21 89.50 88.45 89.39 84.64
20% hold-out 83.73 83.27 88.04 87.34 88.26 84.31
Difference* 0.24 0.94 1.46 1.11 1.13 0.33
*Difference between classification accuracies obtained using 10-fold CV and 20% hold-out validation techniques
5. Conclusion
In the proposed work, we have developed an automatic OSA detection system for
elderly subjects using oximetry (SpO2) and respiratory (ThorRes, AbdoRes, and AF)
signals. We employed SpO2and respiratory signals individually as well as in combi-380
nation for the detection of OSA and observed that the combination of SpO2and res-
piratory signals yields the best OSA detection performance. A new class of frequency
optimized orthogonal wavelet filter banks has been used to obtained sub-bands of the
signal. Tsallis entropy-based features are utilized to develop the model. The proposed
model has achieved a promising performance using the wavelet-based Tsallis features385
of respirator and oximetry signals. The proposed model has been developed using
a RUSBoost classifier and a GentleBoost classifiers. Using SHHS-1, the proposed
model attained maximum accuracy rate of 83.97%. Similarly, for SHHS-2 database,
the model yielded accuracy of 89.50%. When both the databases are combined the
propsed model produced the highest accuracy of 89.39%. Hence, the proposed method390
is simple and robust that can be used in the early detection of OSA. In the future, it
would be interesting to explore if our proposed model can detect other types of apnea,
such as central apnea, mixed apnea, and hypopnea. Also, the system can be tested in
the future for measuring the severity of the apnea.
26
References395
[1] G. Guti´
errez-Tobal, M. L. Alonso-Alvarez, D. Alvarez, F. del Campo, J. Ter´
an,
R. Hornero, Diagnosis of pediatric obstructive sleep apnea: Preliminary findings
using automatic analysis of airflow and oximetry recordings obtained at patients’
home, Biomedical Signal Processing and Control 18 (2015) 401–407. doi:10.
1016/j.bspc.2015.02.014.400
[2] M. Sharma, H. S. Dhiman, U. R. Acharya, Automatic identification
of insomnia using optimal antisymmetric biorthogonal wavelet filter
bank with ecg signals, Computers in Biology and Medicine (2021)
104246doi:https://doi.org/10.1016/j.compbiomed.2021.104246.
URL https://www.sciencedirect.com/science/article/pii/405
S0010482521000408
[3] M. Sharma, J. Tiwari, U. R. Acharya, Automatic sleep-stage scoring in healthy
and sleep disorder patients using optimal wavelet filter bank technique with
eeg signals, International Journal of Environmental Research and Public Health
18 (6). doi:10.3390/ijerph18063087.410
URL https://www.mdpi.com/1660-4601/18/6/3087
[4] M. Sharma, V. Patel, U. R. Acharya, Automated identification of insom-
nia using optimal bi-orthogonal wavelet transform technique with single-
channel eeg signals, Knowledge-Based Systems (2021) 107078doi:https:
//doi.org/10.1016/j.knosys.2021.107078.415
URL https://www.sciencedirect.com/science/article/pii/
S0950705121003415
[5] M. Sharma, D. Goyal, V. AchuthP., U. Acharya, An accurate sleep stages classifi-
cation system using a new class of optimally time-frequency localized three-band
wavelet filter bank, Computers in biology and medicine 98 (2018) 58–75.420
[6] M. Sharma, J. Darji, M. Thakrar, U. R. Acharya, Automated identification of
sleep disorders using wavelet-based features extracted from electrooculogram and
electromyogram signals, Computers in Biology and Medicine (2022) 105224.
27
[7] M. Sharma, V. Patel, J. Tiwari, U. R. Acharya, Automated characteriza-
tion of cyclic alternating pattern using wavelet-based features and ensemble425
learning techniques with eeg signals, Diagnostics 11 (8). doi:10.3390/
diagnostics11081380.
URL https://www.mdpi.com/2075-4418/11/8/1380
[8] H. W. Loh, C. P. Ooi, S. G. Dhok, M. Sharma, A. A. Bhurane, U. R. Acharya,
Automated detection of cyclic alternating pattern and classification of sleep stages430
using deep neural network, Applied Intelligence (2021) 1–15.
[9] M. Sharma, A. A. Bhurane, U. R. Acharya, An expert system for automated
classification of phases in cyclic alternating patterns of sleep using optimal
wavelet-based entropy features, Expert Systems n/a (n/a) e12939. arXiv:
https://onlinelibrary.wiley.com/doi/pdf/10.1111/exsy.12939,435
doi:https://doi.org/10.1111/exsy.12939.
URL https://onlinelibrary.wiley.com/doi/abs/10.1111/exsy.
12939
[10] R. B. Berry, R. Brooks, C. E. Gamaldo, S. M. Harding, C. Marcus, B. V. Vaughn,
et al., The aasm manual for the scoring of sleep and associated events, Rules,440
Terminology and Technical Specifications, Darien, Illinois, American Academy
of Sleep Medicine 176 (2012) 2012.
[11] V. Kapur, D. Auckley, S. Chowdhuri, D. C. Kuhlmann, R. Mehra, K. Ramar,
C. Harrod, Clinical practice guideline for diagnostic testing for adult obstructive
sleep apnea: An american academy of sleep medicine clinical practice guideline.,445
Journal of clinical sleep medicine : JCSM : official publication of the American
Academy of Sleep Medicine 13 3 (2017) 479–504.
[12] Q. A. Shamim-Uzzaman, S. Singh, S. Chowdhuri, Hypopnea definitions, deter-
minants and dilemmas: a focused review, Sleep Science and Practice 2 (1) (2018)
1–12.450
[13] G. C. Guti´
errez-Tobal, D. ´
Alvarez, A. Crespo, F. del Campo, R. Hornero, Eval-
uation of machine-learning approaches to estimate sleep apnea severity from at-
28
home oximetry recordings, IEEE Journal of Biomedical and Health Informatics
23 (2) (2019) 882–892. doi:10.1109/JBHI.2018.2823384.
[14] S. S. Mostafa, F. Mendonc¸a, A. G Ravelo-Garc´
ıa, F. Morgado-Dias, A systematic455
review of detecting sleep apnea using deep learning, Sensors 19 (22) (2019) 4934.
[15] G. Ernst, J. Mariani, M. Blanco, B. Finn, A. Salvado, E. Borsini, Increase in
the frequency of obstructive sleep apnea in elderly people, Sleep Science 12 (3)
(2019) 222.
[16] R. Munoz, J. Duran-Cantolla, E. Martinez-Vila, J. Gallego, R. Rubio, F. Aizpuru,460
G. De La Torre, Severe sleep apnea and risk of ischemic stroke in the elderly,
Stroke 37 (9) (2006) 2317–2321.
[17] M. Sharma, S. Patel, U. R. Acharya, Expert system for detection of congestive
heart failure using optimal wavelet and heart rate variability signals for wireless
cloud-based environment, Expert Systems (2021) e12903.465
[18] N. S. Gooneratne, K. C. Richards, M. Joffe, R. W. Lam, F. Pack, B. Staley, D. F.
Dinges, A. I. Pack, Sleep disordered breathing with excessive daytime sleepiness
is a risk factor for mortality in older adults, Sleep 34 (4) (2011) 435–442.
[19] C. A. Kushida, M. R. Littner, T. Morgenthaler, C. A. Alessi, D. Bailey, J. Cole-
man Jr, L. Friedman, M. Hirshkowitz, S. Kapen, M. Kramer, et al., Practice pa-470
rameters for the indications for polysomnography and related procedures: an up-
date for 2005, Sleep 28 (4) (2005) 499–523.
[20] T. Van Steenkiste, W. Groenendaal, D. Deschrijver, T. Dhaene, Automated sleep
apnea detection in raw respiratory signals using long short-term memory neural
networks, IEEE journal of biomedical and health informatics 23 (6) (2018) 2354–475
2364.
[21] G. Guti´
errez-Tobal, D. Alvarez, J. V. Marcos, F. del Campo, R. Hornero, Pat-
tern recognition in airflow recordings to assist in the sleep apnoea-hypopnoea
syndrome diagnosis, Medical & biological engineering & computing 51 (2013)
1367–1380. doi:10.1007/s11517-013-1109-7.480
29
[22] K. A. Franklin, E. Lindberg, Obstructive sleep apnea is a common disorder in
the population—a review on the epidemiology of sleep apnea, Journal of thoracic
disease 7 (8) (2015) 1311.
[23] M. I. R. Tokle, Comparison of missing data imputation methods for improving
detection of obstructive sleep apnea, Master’s thesis (2017).485
[24] M. Sharma, S. Agarwal, U. R. Acharya, Application of new class of antisym-
metric wavelet filter banks for obstructive sleep apnea diagnosis using ecg sig-
nals, Computers in Biology and Medicine 100. doi:10.1016/j.compbiomed.
2018.06.011.
[25] O. Yildirim, U. B. Baloglu, U. R. Acharya, A deep learning model for automated490
sleep stages classification using psg signals, International journal of environmen-
tal research and public health 16 (4) (2019) 599.
[26] O. S. Lih, V. Jahmunah, T. R. San, E. J. Ciaccio, T. Yamakawa, M. Tanabe,
M. Kobayashi, O. Faust, U. R. Acharya, Comprehensive electrocardiographic di-
agnosis based on deep learning, Artificial intelligence in medicine 103 (2020)495
101789.
[27] S. L. Oh, V. Jahmunah, C. P. Ooi, R.-S. Tan, E. J. Ciaccio, T. Yamakawa, M. Tan-
abe, M. Kobayashi, U. R. Acharya, Classification of heart sound signals using
a novel deep wavenet model, Computer Methods and Programs in Biomedicine
196 (2020) 105604.500
[28] M. Sharma, U. R. Acharya, Automated detection of schizophrenia us-
ing optimal wavelet-based l1norm features extracted from single-channel
eeg, Cognitive Neurodynamics (2021) 1–14doi:https://doi.org/10.1007/
s11571-020-09655-w.
[29] M. Sharma, D. Kumbhani, A. Yadav, U. R. Acharya, Automated sleep apnea505
detection using optimal duration-frequency concentrated wavelet-based features
of pulse oximetry signals, Applied Intelligence (2021) 1–13.
30
[30] R. Fern´
andez, G. Rubinos, C. Cabrera, R. Galindo, S. Fumero, A. Sosa,
I. Gonz´
alez, P. Casan, Nocturnal home pulse oximetry: variability and clinical
implications in home mechanical ventilation, Respiration 82 (2) (2011) 142–147.510
[31] M. Nitzan, A. Romem, R. Koppel, Pulse oximetry: Fundamentals and technology
update, Medical devices (Auckland, N.Z.) 7 (2014) 231–9. doi:10.2147/MDER.
S47319.
[32] E. Bar-Yishay, A. Putilov, S. Einav, Automated, real-time calibration of the respi-
ratory inductance plethysmograph and its application in newborn infants., Physi-515
ological measurement 24 1 (2003) 149–63.
[33] S. Biswal, H. Sun, B. Goparaju, M. B. Westover, J. Sun, M. T. Bianchi, Expert-
level sleep scoring with deep neural networks, Journal of the American Medical
Informatics Association 25 (12) (2018) 1643–1650.
[34] A. Golrou, K. Maghooli, A. M. Amiri, K. Mankodiya, K. Ghaemi, Automatic520
sleep apnea detection using fuzzy logic, 2015. doi:10.1109/SPMB.2015.
7405469.
[35] D. Alvarez, R. Hornero, J. V. Marcos, F. del Campo, Multivariate analysis of
blood oxygen saturation recordings in obstructive sleep apnea diagnosis, IEEE
Transactions on Biomedical Engineering 57 (12) (2010) 2816–2824.525
[36] V. Barroso-Garc´
ıa, G. Guti´
errez-Tobal, D. Gozal, F. Vaquerizo-Villar, D. Alvarez,
F. del Campo, R. Hornero, Wavelet analysis of overnight airflow to detect obstruc-
tive sleep apnea in children, Sensors 21 (2021) 1491. doi:10.3390/s21041491.
[37] M. Uddin, C. Chow, S. Ling, S. Su, A novel algorithm for automatic diagnosis of
sleep apnea from airflow and oximetry signals, Physiological Measurement 42 (1)530
(2021) 015001.
[38] G. C. Guti ´
errez-Tobal, D. ´
Alvarez, L. Kheirandish-Gozal, F. Del Campo,
D. Gozal, R. Hornero, Reliability of machine learning to diagnose pediatric
obstructive sleep apnea: Systematic review and meta-analysis, Pediatric Pul-
monology.535
31
[39] H. Azimi, P. Xi, M. Bouchard, R. Goubran, F. Knoefel, Machine learning-based
automatic detection of central sleep apnea events from a pressure sensitive mat,
IEEE Access 8 (2020) 173428–173439.
[40] M. Hafezi, N. Montazeri, S. Saha, K. Zhu, B. Gavrilovic, A. Yadollahi, B. Taati,
Sleep apnea severity estimation from tracheal movements using a deep learning540
model, IEEE Access 8 (2020) 22641–22649.
[41] M. L. Y. Davidovich, R. Karasik, A. Tal, Z. Shinar, Sleep apnea screening with a
contact-free under-the-mattress sensor, in: 2016 computing in cardiology confer-
ence (CinC), IEEE, 2016, pp. 849–852.
[42] L. Cen, Z. L. Yu, T. Kluge, W. Ser, Automatic system for obstructive sleep apnea545
events detection using convolutional neural network, in: 2018 40th annual in-
ternational conference of the IEEE engineering in medicine and biology society
(EMBC), IEEE, 2018, pp. 3975–3978.
[43] M. Sharma, D. Kumbhani, A. Yadav, U. R. Acharya, Automated sleep apnea
detection using optimal duration-frequency concentrated wavelet-based features550
of pulse oximetry signals, Applied Intelligence (2021) 1–13.
[44] G.-Q. Zhang, L. Cui, R. Mueller, S. Tao, M. Kim, M. Rueschman, S. Mariani,
D. Mobley, S. Redline, The national sleep research resource: towards a sleep
data commons, Journal of the American Medical Informatics Association 25 (10)
(2018) 1351–1358.555
[45] A. Rechtschaffen, A manual for standardized terminology, techniques and scoring
system for sleep stages in human subjects, Brain information service.
[46] M. Sharma, A. Dhere, R. B. Pachori, U. R. Acharya, An automatic detection of
focal eeg signals using new class of time–frequency localized orthogonal wavelet
filter banks, Knowledge-Based Systems 118 (2017) 217–227.560
[47] M. SHARMA, S. SHAH, P. V. ACHUTH, A novel approach for epilepsy de-
tection using time–frequency localized bi-orthogonal wavelet filter, Journal of
Mechanics in Medicine and Biology (2019) 1940007.
32
[48] M. Sharma, D. Bhati, S. Pillai, R. B. Pachori, V. M. Gadre, Design of time–
frequency localized filter banks: Transforming non-convex problem into convex565
via semidefinite relaxation technique, Circuits, Systems, and Signal Processing
35 (10) (2016) 3716–3733.
[49] M. Sharma, R.-S. Tan, U. R. Acharya, Automated heartbeat classifi-
cation and detection of arrhythmia using optimal orthogonal wavelet
filters, Informatics in Medicine Unlocked (2019) 100221doi:https:570
//doi.org/10.1016/j.imu.2019.100221.
URL http://www.sciencedirect.com/science/article/pii/
S2352914819301091
[50] J. M. Amig´
o, S. G. Balogh, S. Hern´
andez, A brief review of generalized entropies,
Entropy 20 (11) (2018) 813.575
[51] J. Chen, G. Li, Tsallis wavelet entropy and its application in power signal analysis,
Entropy 16 (2014) 3009–3025. doi:10.3390/e16063009.
[52] D. Zhang, X. Jia, H. Ding, D. Ye, N. V. Thakor, Application of tsallis entropy to
eeg: quantifying the presence of burst suppression after asphyxial cardiac arrest
in rats, IEEE transactions on biomedical engineering 57 (4) (2009) 867–874.580
[53] J. Chen, G. Li, Tsallis wavelet entropy and its application in power signal analysis,
Entropy 16 (6) (2014) 3009–3025.
[54] A. Sarmanova, S. Albayrak, Alleviating class imbalance problem in data mining,
in: 2013 21st Signal Processing and Communications Applications Conference
(SIU), IEEE, 2013, pp. 1–4.585
[55] C. Seiffert, T. M. Khoshgoftaar, J. Van Hulse, A. Napolitano, Rusboost: A hybrid
approach to alleviating class imbalance, IEEE Transactions on Systems, Man, and
Cybernetics-Part A: Systems and Humans 40 (1) (2009) 185–197.
[56] C. Seiffert, T. Khoshgoftaar, J. Van Hulse, A. Napolitano, Rusboost: Improving
classification performance when training data is skewed, 2008, pp. 1–4. doi:590
10.1109/ICPR.2008.4761297.
33
[57] J. Friedman, T. Hastie, R. Tibshirani, Additive logistic regression: A statistical
view of boosting, The Annals of Statistics 28 (2000) 337–407. doi:10.1214/
aos/1016218223.
[58] D.-Y. Oh, J. Gray, Ga-ensemble: A genetic algorithm for robust ensembles, Com-595
putational Statistics 28. doi:10.1007/s00180-013-0409- 6.
[59] S. Ali, A. Majid, S. Javed, M. Sattar, Can-csc-gbe: Developing cost-sensitive
classifier with gentleboost ensemble for breast cancer classification using protein
amino acids and imbalanced data, Computers in biology and medicine 73 (2016)
38–46. doi:10.1016/j.compbiomed.2016.04.002.600
[60] A. Ramachandran, A. Karuppiah, A survey on recent advances in machine learn-
ing based sleep apnea detection systems, in: Healthcare, Vol. 9, Multidisciplinary
Digital Publishing Institute, 2021, p. 914.
[61] A. Andr´
es-Blanco, D. Alvarez, A. Crespo, C. Arroyo, A. Cerezo-Hern´
andez,
G. Guti´
errez-Tobal, R. Hornero, F. del Campo, Assessment of automated anal-605
ysis of portable oximetry as a screening test for moderate-to-severe sleep apnea
in patients with chronic obstructive pulmonary disease, PLOS ONE 12 (2017)
e0188094. doi:10.1371/journal.pone.0188094.
[62] K. N. Rajesh, R. Dhuli, T. S. Kumar, Obstructive sleep apnea detection using
discrete wavelet transform-based statistical features, Computers in Biology and610
Medicine 130 (2021) 104199.
[63] M. Sharma, J. Tiwari, V. Patel, U. R. Acharya, Automated identification
of sleep disorder types using triplet half-band filter and ensemble machine
learning techniques with eeg signals, Electronics 10 (13). doi:10.3390/
electronics10131531.615
URL https://www.mdpi.com/2079-9292/10/13/1531
[64] C. M. Bishop, N. M. Nasrabadi, Pattern recognition and machine learning, Vol. 4,
Springer, 2006.
34
[65] G. Guti´
errez-Tobal, R. Hornero, D. Alvarez, J. V. Marcos, F. del Campo, Linear
and nonlinear analysis of airflow recordings to help in sleep apnoea-hypopnoea620
syndrome diagnosis, Physiological measurement 33 (2012) 1261–75. doi:10.
1088/0967-3334/33/7/1261.
[66] H. W. Loh, C. P. Ooi, J. Vicnesh, S. L. Oh, O. Faust, A. Gertych, U. R. Acharya,
Automated detection of sleep stages using deep learning techniques: A systematic
review of the last decade (2010–2020), Applied Sciences 10 (24) (2020) 8963.625
35
