Automatic Detection of Sleep Apnea from Single-lead ECG Signal Using Machine Learning

Abstract

Sleep apnea is a common disorder that reduces sleep quality which can lead to serious health consequences. It is identified by a pause in breathing during sleep. At present, polysomnography (PSG) is used to diagnose sleep disorders at sleep centers. PSG is non-invasive but expensive and inconvenient as a clinician needs to observe the subjects overnight. Therefore, development of low-cost and most comfortable techniques is being developed by the researchers. This research presents an implementation of a classification algorithm for identifying sleep apnea episodes using an electrocardiogram (ECG) signal. It also provides a comparison of the accuracies achieved by different classifiers and a comparison between the performance of individual features. A Sgolay filter was utilized for baseline correction and denoising the ECG signal and used to calculate the Heart Rate Variability (HRV) sequence using RR intervals. Ten features were extracted from the ECG and HRV signals, and the performance of the proposed algorithm has been estimated for five classifiers, support vector machine (SVM), linear discriminant analysis (LDA), K-nearest neighbor (KNN), random forest (RF) and decision tree (DT). For the ECG signal, the average accuracies depicted by SVM, LDA, KNN, RF, and DT are 86.01±1.16,68.16±0.76,84.99±1.24,87.00±1.40, and S2.56±1.05, respectively. Further, for the HRV signal, the average accuracies depicted by SVM, LDA, KNN, RF, and DT are 83.53±1.36, 72.69±2.06, 82.60±1.07, 85.10±1.44, and 78.60±1.15, respectively. This research highlights the significance of selecting the appropriate classifier for a given issue and establishing and using the most relevant characteristics to achieve a higher level of accuracy. These encouraging early-stage findings might pave the way for additional research to refine the classifiers in preparation for potential use in the real world. The effectiveness of the suggested model was evaluated alongside the performance of other methods already in use for the detection of sleep apnea.

Full text

Automatic Detection of Sleep Apnea from Single-
lead ECG Signal Using Machine Learning
Amit Bhongade*
Department of Electrical Engineering,
Indian Institute of Technology Delhi,
Delhi, India
amitbhongade46@gmail.com*
Rohit Gupta
Department of Electrical Engineering,
Indian Institute of Technology Delhi,
Delhi, India
rohit.udai@yahoo.co.in
Tapan Kumar Gandhi
Department of Electrical Engineering,
Indian Institute of Technology Delhi,
Delhi, India
tgandhi@iitd.ac.in
Abstract— Sleep apnea is a common disorder that reduces
sleep quality which can lead to serious health consequences. It
is identified by a pause in breathing during sleep. At present,
polysomnography (PSG) is used to diagnose sleep disorders at
sleep centers. PSG is non-invasive but expensive and
inconvenient as a clinician needs to observe the subjects
overnight. Therefore, development of low-cost and most
comfortable techniques is being developed by the researchers.
This research presents an implementation of a classification
algorithm for identifying sleep apnea episodes using an
electrocardiogram (ECG) signal. It also provides a comparison
of the accuracies achieved by different classifiers and a
comparison between the performance of individual features. A
Sgolay filter was utilized for baseline correction and denoising
the ECG signal and used to calculate the Heart Rate
Variability (HRV) sequence using RR intervals. Ten features
were extracted from the ECG and HRV signals, and the
performance of the proposed algorithm has been estimated for
five classifiers, support vector machine (SVM), linear
discriminant analysis (LDA), K-nearest neighbor (KNN),
random forest (RF) and decision tree (DT). For the ECG
signal, the average accuracies depicted by SVM, LDA, KNN,
RF, and DT are 86.01±1.16, 68.16±0.76, 84.99±1.24, 87.00±1.40,
and 82.56±1.05, respectively. Further, for the HRV signal, the
average accuracies depicted by SVM, LDA, KNN, RF, and DT
are 83.53±1.36, 72.69±2.06, 82.60±1.07, 85.10±1.44, and
78.60±1.15, respectively. This research highlights the
significance of selecting the appropriate classifier for a given
issue and establishing and using the most relevant
characteristics to achieve a higher level of accuracy. These
encouraging early-stage findings might pave the way for
additional research to refine the classifiers in preparation for
potential use in the real world. The effectiveness of the
suggested model was evaluated alongside the performance of
other methods already in use for the detection of sleep apnea.
Keywords—ECG signals, RR/HRV signals, sleep disorders,
classification.
I. INTRODUCTION
Approximately one-third of a person's life is spent
sleeping, based on average longevity. A sleeping problem is
characterized by an inability to sleep, which results in
impaired bodily functions. Sleep is essential for every person
since it enables the body to recover and keep itself healthy
[1]. In the same way as getting enough rest may have a
variety of positive impacts on a person's body, including
mental, emotional, and physical consequences, not getting
enough sleep can have a negative effect on the body
physically, emotionally, and physiologically. At present,
restless leg syndrome, narcolepsy, sleep apnea, and insomnia
are some of the most common sleep disorders identified out
of 84 existing sleep disorders [2].
Sleep apnea is a clinically significant and possibly life-
threatening sleep disorder defined by periods during sleep in
which a person stops breathing for seconds or even minutes.
Apnea refers to pauses in breathing and might vary in terms
of the frequency with which it occurs and the amount of time
it lasts. On the other hand, hypopnea refers to an absence of
breathing during sleep [3]. Sleep apnea is a primary cause of
morbidity and death because it directly affects the
cardiovascular system by increased sympathetic activity and
systemic hypertension. This causes the heart and arteries to
work harder than they should [1]. Obstructive sleep apnea
(OSA) and central sleep apnea are the two categories of sleep
apnea. Because of a narrowing of the upper respiratory
airway, OSA is well known, affecting 1-3% of preschoolers
and 2-4% of middle-aged people [4]. Moreover, CSA is
mixed, produced by a lack of or suppressed respiratory drive,
and commonly coupled with OSA cases, and the instances of
CSA by itself emerge exceptionally rarely. OSA cases
generally occur in conjunction with CSA cases [5].
The traditional approach is PSG, which diagnoses sleep
disorders at sleep centers. PSG records breathing, heart rate,
the oxygen level in the blood, brain waves (electromyogram
(EMG), electrooculogram (EOG), and electroencephalogram
(EEG)), and eye and leg movements during sleep [6].
However, these physiological data recordings are time-
consuming, cumbersome, and complex because it is
necessary to conduct an extensive test in a well-calibrated
laboratory environment, such as a hospital setting. Hence,
this diagnosis is not only unrealistic for a broad
population but is also exceedingly expensive. Therefore, the
development of surrogate procedures that are capable of
being administered to the patient in a manner that is
comfortable for them and provides a solution that facilitates a
more straightforward diagnosis and more suitable treatment
for sleep apnea. In addition, it seems that automatic, portable,
at-home gadgets used while sleeping is becoming more
desirable, promising, and highly on demand.
In literature, various techniques have been proposed to
diagnose sleep apnea disease over the last few years. The
study by N. Pombo et al. [7] performed a systematic review
on computerized systems and the classification of ECG
signal-based apnea detection. However, the accuracy of the
classifiers is highly dependent on feature selection from
multiple data sources. In addition, the existing method of
PSG requires extensive data collection combined with
various modalities of data such as oxygen saturation (SpO2),
electromyogram (EMG), electrooculogram (EOG), and
2022 International Conference on Futuristic Technologies (INCOFT)
K
arnataka, India. Nov 25-27, 2022
978-1-6654-5046-1/22/$31.00 ©2022 IEEE 1
2022 International Conference on Futuristic Technologies (INCOFT) | 978-1-6654-5046-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/INCOFT55651.2022.10094459
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on April 17,2023 at 05:43:08 UTC from IEEE Xplore. Restrictions apply.

electroencephalogram (EEG). Therefore, as an alternate
approach for diagnosing sleep apnea, the use of fewer
physiological signs is now favored. [8]. The sleep apnea
detection algorithm using RR and HR signals retrieved from
single-lead ECG data was introduced by R. K. Tripathy et al.
[9]. Distinguishing sleep apnea and normal episodes were
accomplished with the help of the SVM and RF classifiers.
The inspection findings showed that each classifier's
accuracy was improved when they used features derived
from the HR signal rather than the RR signal. Furthermore,
the results obtained from 10-fold cross-validation
demonstrated that the SVM had performed better than the
RF. Similarly, the study conducted by the A. Pinho et al.
implemented the SVM and Artificial Neural Network (ANN)
to determine the system performance. In this, SVM
performed better with higher accuracy than ANN. The
outcomes of the experiments revealed that various aspects of
the system are responsible for multiple aspects of the
system’s performance. The study by A. Zarei et al. [10]
presented a study based on ECG signals to detect sleep
apnea. The classifiers used for classification between normal
and sleep apnea events were RF, Naive Bayes (NB), Logistic
Regression (LR), kNN, RF, SVM, and LDA. The author
concluded that RF performed better with high accuracy than
other classifiers. Similarly, a study by H. Sharma et al. [11]
utilized single-lead ECG electrodes and used SVM, LS-
SVM, Multilayer Perceptron Neural Network (MLPNN),
KNN, etc. Based on the findings of the experiments, it was
determined that the RBF kernel-based LS-SVM performed
better than other classifiers.
In this research, we exhibited a complete analysis by
employing two distinct sets of features derived from the
filtered ECG and the HRV signal. Both sets of features were
used to examine the data. In addition to that, the
effectiveness of the proposed method is determined by five
different classifiers (SVM, KNN, LDA, RF, DT). The main
contribution of this research is to provide a cost-efficient and
simple diagnostic tool to detect sleep apnea.
II. MATERIALS AND METHODS
Based on the literature above, we have formulated the
study that contained; sleep apnea detection using ECG signal
alone; appropriate feature selection is a prominent factor in
determining the accuracy of the classifier and presented the
comparison between the performance of different classifiers
for the same ECG signals. The proposed data processing and
system model is depicted in Figure 1, and a detailed
explanation is provided in the below sections.
A. Dataset
The dataset used to conduct this study is available in
PhysioNet [12] database. It is the most popular dataset, and
many studies have already been conducted on this dataset.
The considered dataset consists of ECG records of 35
subjects (Age:27 to 63, body mass index: 19.2 to 45.33 kg/m,
and apnea-hypopnea index (AHI): 0 to 83).
Out of these recordings, only ten subjects’ datasets were
selected based on sleep apnea and non-apnea events in the
recordings to perform this research. A total of 10 single-lead
ECG recordings (a01, a02, a03, a04, a06, a08, a09, a10, a14,
b01) were analyzed from this data set. These ten recordings,
each of which has been evaluated by a specialist in the field
of sleep apnea, include a total of 3032 minutes of apnea and
2044 minutes of non-apnea events. Each ECG record was
sampled at 100 Hz (12-bit resolution) and is of a different
length. However, details regarding waking and sleeping
times, as well as awakenings caused by trembling and
jerking limbs during sleep, are lacking. Furthermore, the
dataset does not include information on other sleep-related
diseases or cardiac difficulties, nor does it include specifics
about obstructive apnea, hypopnea, and central apnea.
Fig. 1. Proposed data processing and system model
B. Pre-processing of ECG signal
The dataset used contains data files in. apn format. The
.apn files were imported and processed in the MATLAB
2020b software. The database utilized for this study has
artifacts and noise incorporated with it, which occurred in a
clinical setting. It contains a variety of P-, T-wave
morphologies, and QRS complexes. The Sgolay filter [13]
was utilized for the baseline corrections and to denoise the
ECG signal. The filtered signal was subtracted from the
original ECG signal to obtain the noise-free and baseline
corrected ECG signal, as shown in Figure 2.
In addition, the R-peaks and QRS complexes can be
detected using the filtered ECG signal without misclassifying
or missing a pulse. The cyclic fluctuation in pulse duration is
also known as the RR interval. According to the research, the
RR interval comprises bradycardia during apnea and
tachycardia after it ends because of sleep apnea [6].
Fig. 2. (a) Raw ECG waveform, (b) Filtered ECG waveform with baseline
correction, (c) RR peak marked waveform.
2
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on April 17,2023 at 05:43:08 UTC from IEEE Xplore. Restrictions apply.

Several studies were conducted to identify sleep apnea by
utilizing the characteristics calculated using HRV signal and
RR interval. These characteristics include the standard
deviation, inter-quartile range (IQR), mean, and median of
the HRV signal [6] [14][15]. In this research, thresholding of
30% of maximum values is utilized to find the RR peaks in
filtered ECG signals because it was able to improve the
performance of the system. Then, the RR time intervals using
filtered ECG signals can be calculated using the formula
shown in equation 1 [14]. In addition, the sequence formed
using RR intervals represents the HRV signal.
ܴܴሺ݆ሻൌܴܴ
ሺ݆൅ͳ
ሻെܴܴ
ሺ݆ሻǡሺ݆ ൌ ͳǡ ʹǡ ǥ Ǥ Ǥ ǡ ݊ െ ͳሻ ǥ ǥ ሺͳሻ
Where n: Number of RR peaks in the ECG signal.
C. Feature extraction and classification
The feature extraction was performed using a moving
window of a one-minute on ECG and the HRV signal. A
total of ten-time domain features were extracted to perform
the comprehensive study, as shown in Table 1. To perform
the classification, all recordings from the one-minute section
were classified as either non-apnea or apnea episodes by
assigning them the values 0 or 1, respectively. Further, to
tackle the class imbalance issue, the under-sampling
technique has opted. As mentioned before, only ten subject
databases have been used for this study. So, the database of
these subjects contained 5076 records, of which 59.73%
related to sleep apnea events, whereas 40.27% related to non-
apnea events. The database was divided into two groups,
testing and training sets. In addition, the k-fold cross-
validation technique, with k=10, was utilized to enhance the
classifiers' training. Five classifiers (SVM, LDA, KNN, RF,
and DT) were applied during the classification phase, and
their comparative performances were assessed. Finally, the
six parameters, accuracy, precision, sensitivity, specificity,
recall, and F1 score, were evaluated for all these classifiers.
TABLE I. LIST OF EXTRACTED FEATURES FROM ECG AND HRV
SIGNALS
Signal
Features
ECG Signal
RMS, variance, kurtosis, skewness, mean, std,
power, entropy, minimum, maximum.
HRV Signal
HR, mean, standard deviation, mean of peak
values
in HRV,
standard deviation of HRV,
HRV
>50 ms, %HRV>50 ms,
RMS of HRV,
entropy
of HRV, interquartile range
of HRV
[16][17].
III. RESULTS AND DISCUSSION
The present research explored the two different feature
extraction methodologies applied to ECG signals to detect
sleep apnea. The first approach utilized the time- domain and
statistical features of the ECG signal. Whereas the second
approach utilized the features derived from the R-peaks and
HRV signal. To detect the RR peaks, the filtered ECG signal
has been passed through the peak detector module. The
performance of the peak detector module was highly
dependent on the threshold value. Hence, to select the
appropriate threshold value, the threshold was varied from 20
to 50 with an increment of 5, and subsequently, the features
were estimated for each threshold value. Further, the
performance of all the classifiers was estimated.
Figure 3 depicts the average accuracies illustrated by the
classifiers for different RR threshold values. Apart from the
DT and LDA classifier, the performance of all the classifiers
was found similar (p-value>0.05) for all the threshold values
less than 40. For the LDA classifier, the best performance
was depicted with a threshold value of 25. Whereas, for the
DT classifier, the best performance was achieved with a
threshold value of 20. So, for further analysis, the threshold
value of 25 was considered.
Fig. 3. Classifiers performance for different RR threshold values
Figure 4 (a) and (b) show the box plot of individual
features estimated from the ECG signal and RR/HRV signal,
respectively. For ECG features, none of the features depicted
very significant differences between the groups, normal(N)
and apnea (A). Further, for features estimated from RR and
HRV signals, only the HR depicted a substantial difference
between the groups, whereas the rest of the features were
found comparatively similar for both groups. Table 2
presents the performance of all the classifiers for the
individual feature. For the first type of feature set, the lowest
performance was depicted by all the classifiers for the mean
of the ECG signal. Whereas, for the second type of feature
set, StdRR feature emerged as the least performing feature for
all the classifiers. Among the classifiers, LDA performs the
least for most of the features. However, it depicted the least
variation in performance for 10-fold cross-validation.
Further, the performance of all the classifiers was
estimated for the complete feature vector formulated through
the feature extraction approach individually. Figure 5 depicts
the % classification accuracy of the sleep apnea detection
model for all compared classifiers using ECG and HRV
signals, respectively. The quantitative analysis used the
performance measures —namely, accuracy, precision,
sensitivity, specificity, recall, and F1 score of each class. The
average classification accuracies using ECG signals for
SVM, LDA, KNN, RF, and DT are 86.01±1.16, 68.16±0.76,
84.99±1.24, 87.00±1.40, and 82.56±1.05, respectively.
Furthermore, the average classification accuracies of sleep
apnea detection using HRV signal for SVM, LDA, KNN,
RF, and DT are 83.53±1.36, 72.69±2.06, 82.60±1.07,
85.10±1.44, and 78.60±1.15, respectively. The RF classifier
was found to be the best performer, whereas the LDA is the
least performed.
3
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on April 17,2023 at 05:43:08 UTC from IEEE Xplore. Restrictions apply.

Fig. 4. Performance of individual features. (a) For ECG signal, (B) For HRV signal
TABLE II. PERFORMANCE OF DIFFERENT CLASSIFIERS FOR INDIVIDUAL FEATURES (ACCURACIES)
Performance of classifiers for the individual feature of the ECG signal
Features
SVM
LDA
KNN
RF
DT
RMS
64.22±2.04
62.08±0.86
67.61±1.93
64.54±2.32
65.56±2.13
Variance
64.03±3.18
60.82±0.74
67.75±1.94
64.01±2.73
65.13±2.47
Kurtosis
61.41±2.33
59.73±0.08
64.76±1.43
60.74±2.03
62.53±2.06
Skewness
64.28±1.45
63.87±1.34
68.19±2.46
64.82±2.11
65.86±1.35
Mean
53.58±1.34
60.07±0.55
58.63±1.24
55.46±2.17
57.52±1.93
Standard Deviation
63.42±2.78
62.12±0.87
67.77±1.94
63.99±2.76
65.13±2.47
Power
63.14±1.90
60.76±0.73
67.59±1.90
64.48±2.30
65.56±2.13
Entropy
63.24±2.55
60.80±0.99
65.27±1.90
61.98±1.72
63.93±1.61
Minimum
63.24±1.99
60.70±0.81
65.82±1.21
63.28±2.42
64.36±2.13
Maximum
66.31±1.05
60.70±0.56
69.64±0.73
65.44±2.34
66.73±1.82
Performance of classifiers for the individual feature of RR and HRV signal
Features
SVM
LDA
KNN
RF
DT
HR
70.84±1.66
70.76±2.05
66.07±1.42
71.21±1.75
71.07±1.70
MeanHRV
63.27±1.11
60.06±0.20
65.36±1.70
63.21±1.33
63.78±2.21
StdHRV
61.06±1.19
59.74±0.07
62.76±2.27
57.67±2.26
60.12±2.94
MaxHRV
62.56±1.85
59.74±0.07
65.34±2.05
62.31±2.28
63.82±1.42
StdRR
58.76±1.29
59.74±0.07
58.60±1.02
55.90±1.26
57.30±2.15
HRV>50ms
62.78±1.18
59.74±0.07
55.84±1.67
62.72±1.30
62.86±1.22
%HRV>50ms
63.66±1.51
62.40±0.74
60.55±1.63
64.41±1.67
64.27±2.00
RMSHRV
62.09±1.38
59.74±0.07
66.36±1.66
62.32±2.52
64.08±2.36
EntropyHRV
61.83±1.14
58.96±1.04
65.46±1.96
60.87±1.48
63.59±2.15
IQRHRV
67.25±1.50
60.10±0.26
61.18±1.83
66.86±1.32
67.25±1.70
Fig. 5. Performance of classifiers for ECG and RR/HRV signal
4
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on April 17,2023 at 05:43:08 UTC from IEEE Xplore. Restrictions apply.

IV. CONCLUSION
In the present research work, an ECG-based model for
analyzing sleep apnea on a minute-by-minute basis is
presented. The primary objective is to develop an accurate
and effective alternative technique to the traditional PSG. In
addition, a standard that has considered all five types of
classifiers, namely SVM, LDA, KNN, RF, and DT, has been
constructed. As a consequence of the findings that have been
provided, it is clear that various classifiers use different
strategies while attempting to solve the same problem. In
addition, the model developed in this research is reliable,
feasible, and appropriate to detect sleep apnea using ECG
signals. The performance of the proposed algorithm has been
estimated for five classifiers, SVM, LDA, KNN, RF, and
DT. The average classification accuracies using ECG signals
for SVM, LDA, KNN, RF, and DT are 86.01±1.16,
68.16±0.76, 84.99±1.24, 87.00±1.40, and 82.56±1.05,
respectively. Furthermore, the average classification
accuracies of sleep apnea detection using HRV signal for
SVM, LDA, KNN, RF, and DT are 83.53±1.36, 72.69±2.06,
82.60±1.07, 85.10±1.44, and 78.60±1.15, respectively.
Possible directions for future research include incorporating
feature selection to identify an optimal set of characteristics
for sleep apnea detection, increasing sensitivity to capture all
apnea moments, simulating the study in real patients to
assess its feasibility, and calculating and comparing the
performance of the various techniques used in the research.
REFERENCES
[1] G. Atkinson and D. Davenne, ‘‘Relationships between sleep, physical
activity and human health,’’ Physiol. Behav., vol. 90, nos. 2–3, pp.
229–235, Feb. 2007. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0031938406003957
[2] Sleep Disorder Overview. www.neurologychannel.com
[3] S. Isa, M. Fanany, W. Jatmiko and A. Arymurthy, “Sleep Apnea
Detection from ECG Signal, Analysis on Optimal Features, Principal
Components, and Nonlinearity,” in Proceedings of the 5 th IEEE
International Conference on Bioinformatics and Biomedical
Engineering (iCBBE), pp. 1-4, May 2011.
[4] Sleep Apnea: What Is Sleep Apnea? www.nhlbi.nih.gov
[5] Apnea guide. www.apneaguide.com
[6] P. Chazal, T. Penzel, and C. Heneghan, “Automated Detection of
Obstructive Sleep Apnea at Different Time Scales Using the
Electrocardiogram,” Institute of Physics Publishing, vol. 25, no. 4, pp.
967-983, Aug. 2004.
[7] N. Pombo, N. Garcia, and K. Bousson, ‘‘Classification techniques on
computerized systems to predict and/or to detect apnea: A systematic
review,’’ Comput. Methods Programs Biomed., vol. 140, pp. 265–
274, Mar. 2017. [Online]. Available:
www.sciencedirect.com/science/article/pii/S016926071630611.
[8] A. M. da Silva Pinho, N. Pombo, and N. M. Garcia, ‘‘Sleep apnea
detection using a feed-forward neural network on ECG signal,’’ in
Proc. IEEE 18th Int. Conf. e-Health Netw., Appl. Services
(Healthcom), Sep. 2016, pp. 1–6.
[9] R. K. Tripathy, P. Gajbhiye, and U. R. Acharya, ‘‘Automated sleep
apnea detection from cardio-pulmonary signal using bivariate fast and
adaptive EMD coupled with cross time–frequency analysis,’’
Comput. Biol. Med., vol. 120, May 2020, Art. no. 103769. [Online].
Available: http://www.
sciencedirect.com/science/article/pii/S0010482520301414.
[10] A. Zarei and B. M. Asl, ‘‘Performance evaluation of the spectral
autocorrelation function and autoregressive models for automated
sleep apnea detection using single-lead ECG signal,’’ Comput.
Methods Programs Biomed., vol. 195, Oct. 2020, Art. no. 105626.
[Online].Available:http://www.sciencedirect.com/science/article/pii/S
0169260720314590.
[11] H. Sharma and K. K. Sharma, ‘‘An algorithm for sleep apnea
detection from single-lead ECG using Hermite basis functions,’’
Comput. Biol. Med., vol. 77, pp. 116–124, Oct. 2016. [Online].
Available: http://www.sciencedirect.com/science/article/pii/
S0010482516302086.
[12] T. Penzel, G. B. Moody, R. G. Mark, A. L. Goldberger, and J. H.
Peter, ‘‘The apnea-ECG database, ’’ in Proc. Comput. Cardiol., vol.
27, Sep. 2000, pp. 255–258. PhysioNet, www.physionet.org.
[13] A. Savitzky and M. J. E. Golay, ‘‘Smoothing and differentiation of
data by simplified least squares Procedures.,’’ Anal. Chem., vol. 36,
no. 8, pp. 1627–1639, Jul. 1964, doi: 10.1021/ac60214a047.
[14] S. Isa, M. Fanany, W. Jatmiko and A. Murini, “Feature and Model
Selection on Automatic Sleep Apnea Detection Using ECG,” in
International Conference on Computer Science and Information
Systems, ICACSIS 2010, pp. 357-362, 2010.
[15] P. Langley, E. Bowers and A. Murray, “Principal Component
Analysis as Tool for Analyzing Beat-to-Beat Changes in ECG
Features: Application To ECG-Derived Respiration,” in IEEE
Transactions on Biomedical Engineering, vol. 57, no. 4, pp. 821- 829,
Apr. 2010.
[16] P. Schmidt, A. Reiss, R. Duerichen, C. Marberger, and K. V.
Laerhoven. 2018. Introducing WESAD, a Multimodal Dataset for
Wearable Stress and Affect Detection. In Proceedings of the 20th
ACM International Conference on Multimodal Interaction (ICMI '18).
Association for Computing Machinery, New York, NY, USA, 400–
408. https://doi.org/10.1145/3242969.3242985.
[17] Bertini M, Ferrari R, Guardigli G, Malagù M, Vitali F, Zucchetti O,
D'Aniello E, Volta CA, Cimaglia P, Piovaccari G, Corzani A, Galvani
M, Ortolani P, Rubboli A, Tortorici G, Casella G, Sassone B, Navazio
A, Rossi L, Aschieri D, Rapezzi C. Electrocardiographic features of
431 consecutive, critically ill COVID-19 patients: an insight into the
mechanisms of cardiac involvement. Europace. 2020 Dec
23;22(12):1848-1854. doi: 10.1093/europace/euaa258. PMID:
32944767; PMCID: PMC7543398.
5
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on April 17,2023 at 05:43:08 UTC from IEEE Xplore. Restrictions apply.