IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO.
3, MAY 2014
Detection of Respiratory Arousals Using
Photoplethysmography (PPG) Signal
in Sleep Apnea Patients
Chandan Karmakar, Member, IEEE, Ahsan Khandoker, Senior Member, IEEE,
Thomas Penzel, Senior Member, IEEE, Christoph Schobel, and Marimuthu Palaniswami, Fellow, IEEE
AbstractRespiratory events during sleep induce cortical
arousals and manifest changes in autonomic markers in sleep dis-
order breathing (SDB). Finger photoplethysmography (PPG) has
been shown to be a reliable method of determining sympathetic
activation. We hypothesize that changes in PPG signals are suffi-
cient to predict the occurrence of respiratory-event-related cortical
arousal. In this study, we develop a respiratory arousal detection
model in SDB subjects by using PPG features. PPG signals from
10 SDB subjects (9 male, 1 female) with age range 4375 years
were used in this study. Time domain features of PPG signals, such
as 1) PWApulse wave amplitude, 2) PPIpeak-to-peak inter-
val, and 3) Areaarea under peak, were used to detect arousal
events. In this study, PWA and Area have shown better perfor-
mance (higher accuracy and lower false rate) compared to PPI
features. After investigating possible groupings of these features,
combination of PWA and Area (PWA + Area) was shown to pro-
vide better accuracy with a lower false detection rate in arousal de-
tection. PPG-based arousal indexes agreed well across a wide range
of decision thresholds, resulting in a receiver operating characteris-
tic with an area under the curve of 0.91. For the decision threshold
(PCth re sh = 25%) chosen for the final analyses, a sensitivity of
68.1% and a specificity of 95.2% were obtained. The results showed
an accuracy of 84.68%, 85.15%, 86.93%, and 50.79% with a false
rate of 21.80%, 55.41%, 64.78%, and 50.79% at PCth re sh = 25%
or PPI, PWA, Area, and PWA + Area features, respectively. This
indicates that combining PWA and Area features reduced the false
positive rate without much affecting the sensitivity of the arousal
detection system. In conclusion, the PPG-based respiratory arousal
detection model is a simple and promising alternative to the con-
ventional electroencephalogram (EEG)-based respiratory arousal
detection system.
Index TermsHeart rate variability, photoplethysmography
(PPG), respiratory arousal, sleep apnea.
Manuscript received February 21, 2013; revised June 2, 2013 and July 24,
2013; accepted September 5, 2013. Date of publication September 23, 2013;
date of current version May 1, 2014. (Corresponding author: A. Khandoker.)
C. Karmakar and M. Palaniswami are with the Department of Electrical
and Electronic Engineering, University of Melbourne, Melbourne, Vic. 3010,
Australia (e-mail: karmakar@ unimelb.edu.au; palani@unimelb.edu.au).
A. Khandoker is with the Biomedical Engineering Department, Khalifa Uni-
versity of Science, Technology and Research, Abu Dhabi, UAE. He is also
with the Department of Electrical and Electronic Engineering, University of
Melbourne, Melbourne, Vic. 3010, Australia (e-mail: ahsank@unimelb.edu.au;
ahsan.khandoker@kustar.ac.ae).
T. Penzel and C. Schobel are with the Department for Cardiology, Interdis-
ciplinary Sleep Medicine Center, Charite Universitatsmedizin Berlin, Berlin
10117, Germany (e-mail: thomas.penzel@charite.de; Christoph.Schoebel@
charite.de).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JBHI.2013.2282338
2168-2194 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
1065
I. INTRODUCTION
LEEP-DISORDERED breathing (SDB) covers a spectrum
S from complete upper airway obstruction [obstructive sleep
apnea (OSA)] [1] to more subtle breathing abnormalities [ob-
structive hypopnea (OHA)] and even an absence of respiratory
effort [central sleep apnea (CSA)]. Termination of SDB events
requires either a change from a deeper to lighter stage of sleep
or wakefulness that restores upper airway patency [2][4]. The
arousal is important for restoring regular breathing; however,
the sleep fragmentation induced by recurring arousals has ad-
verse consequences which contribute to the pathophysiology
and clinical manifestations of SDB [5].
Arousals have been found to be associated with central au-
tonomic activation, which leads to increased sympathetic ac-
tivity and resulting peripheral vasoconstriction [6]. Electroen-
cephalography (EEG)-based arousals are generally accepted to-
day and are routinely used in clinical and research sleep labo-
ratories. However, according to the American Sleep Disorders
Association (ASDA) definition [1], only about 75% of respira-
tory events are terminated by an EEG arousal [7], while it has
been shown that stimuli that produce measurable cardiovascular
disturbances without arousal appear sufficient to produce day-
time sleepiness [8]. Therefore, there is a considerable interest
in using convenient, noninvasive autonomic markers of sleep
fragmentation, particularly in SDB. In addition, the limited-
channel polysomnography (PSG), type III portable monitors,
and recordings without electroencephalography (EEG) are in-
creasingly used for practical reasons instead of standard PSG,
which is the gold standard for SDB diagnostics [9]. This ap-
proach seems to be more cost effective, and studies have shown
sufficient sensitivity and specificity [10]. Since no EEG chan-
nels are available in the recordings, a surrogate signal is required
to score respiratory-effort-related arousals (RERA).
Autonomic activation concomitant with arousals from sleep
induces acute measurable hemodynamic changes such as ele-
vated arterial pressure and heart rate, and altered pulse transit
time (PTT) [11]. Changes in pulse wave amplitude (PWA) mea-
sured by finger photoplethysmography (PPG) have been shown
to be a reliable method of determining sympathetic activation
and are used as markers of finger vasoconstriction as well [6].
PPG noninvasively measures the relative absorption of red light
and infrared light. Arterial blood flow pulsation passing through
an artery modulates the light absorption and provokes a pulse
wave easily convertible into an electrical signal readable during
PSG. This parameter can be easily derived from conventional
1066 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 3, MAY 2014
Fig. 1. Respiratory events and arousals in PSG recording.
pulse oxymeters and does not bring any alteration of sleep, as
there is no variation of pressure at the fingertip. Catchside et al.
have shown that PWA has better correlation with acoustical-
induced arousals undetected by EEG [12]. In patients with SDB,
it has been reported that respiratory events and arousals induce
more changes in PWA than that in the heart rate (HR) [13].
The peak-to-peak time interval (PPI) measured from a PPG
signal can be used for extracting the instantaneous pulse rate
variability (PRV) signal and has been reported as a surrogate
for ECG-based heart rate variability (HRV) analysis in healthy
subjects [13]. Recently, Lazro et al. have reported that use of
PRV improves the accuracy of obstructive sleep apnea subject
classification with respect to HRV in children [14]. Therefore,
the correlation of PPI with HRV, which has a clinical relevance
in sleep studies, indicates that it could be used to detect RERAs.
In addition to PWA and PPI features, in this study we also ana-
lyzed the performance of Area (area under PWA curve) feature
in respiratory arousal detection. Although no specific physiolog-
ical relation of Area has been reported yet, we hypothesize that
it could be an important feature in detecting respiratory arousal,
since it contains information of the morphology of PPG cycle.
The purpose of this study is to investigate the ability of var-
ious PPG features to detect the respiratory arousals as a new
technique for the screening of sleep quality or fragmentation.
Examples of respiratory events and arousal in a polysomonog-
raphy (PSG) recording are shown in Fig. 1.
II. DATA AND METHODS
A. Subjects and Data
In total, ten polysomnographic recordings were used in this
study. PSGs were collected from ten patients with sleep dis-
ordered breathing at Charite Hospital, Berlin, Germany. The
research protocol was approved by Charite Hospital Ethics in
Human Research Committee. Each PSG study included elec-
troencephalogram (channel C3-A2 and C4-A1), left and right
electro-oculogram (EOG), leg movements, body positions, tho-
racic and abdominal wall expansion (by respiratory inductive
plethysmography), oronasal airflow (by nasal pressure), arterial
oxygen saturation SaO2 (by pulse oximetry) and PPG (sampling
frequency = 75 Hz with a resolution of 16 bits/sample). Both
red and infrared (IR) PPG signals were collected during PSG
recording; however, we have used only the IR channel in this
study. All subjects were free of any cardiac history. Diagnosis
was based on clinical symptoms and PSG outcomes. Arousals
were scored manually according to ASDA criteria using the
EEG signal and marked using Somnologica software. In total,
2917 arousal events were found across all ten subjects of this
study.
B. Schematic Diagram
The schematic diagram to detect arousal using features of
the PPG signal is shown in Fig. 2. At first, the features are
extracted from the PPG signal and an arousal model is developed
based on an individual feature. After detecting events, closer
events (multiple events within 10 s) are merged together as
a postprocessing step. To obtain the arousal detection using
multiple features, detected arousal events of individual features
are combined. Finally, the performance (accuracy, false positive
rate, true positive rate, and missed positive rate) of the model
is measured based on the original (EEG-based score) arousal
events.
1) PPG Features: In this study, three different time domain
features are extracted from the PPG signal. The example of the
PPG signal and the extracted features from the PPG signal are
shown in Fig. 3. A brief description of the features is given
below.
a) Peak-to-peak interval (PPI): The PPI is defined as the time
difference between two consecutive peaks of the PPG signal.
At first, the peak of each cycle of the PPG signal (peakAmp)
was detected and the time stamps of all peakAmp points were
stored in a vector. The PPI was calculated as the time differ-
ence between consecutive peakAmp points (see Fig. 3). In order
to reduce the error in calculating the PPI, the PPG signal was
KARMAKAR et al.: DETECTION OF RESPIRATORY AROUSALS USING PHOTOPLETHYSMOGRAPHY (PPG) SIGNAL IN SLEEP APNEA PATIENTS
Fig. 2. Schematic diagram of arousal detection and performance measurement.
Fig. 3. Example of PPG signal and the time domain features.
resampled using cubic spline interpolation at a sampling fre-
quency of 4000 Hz [15]. It has been reported that this resam-
pling procedure reduces the error in accuracy to less than one
sample per minute [15]. PPI < 0.5 s or PPI > 0.3 s, i.e., the pulse
rate below 46 beats/min or above 120 beats/min was considered
ectopic and removed in this study.
b) Pulse wave amplitude (PWA): The PWA is the difference
between the peak amplitude (peakAmp) and valley amplitude
(valleyAmp), which is shown in Fig. 3. The peakAmp and val-
leyAmp are the maximum and minimum amplitude points of
each PPG cycle. At first, all peakAmps and valleyAmps were
detected as the local maximum and minimum points of the
PPG signal. In the case of missing peakAmp points, the next
valleyAmp point was also discarded. Finally, PWA was calcu-
lated by subtracting valleyAmp from the immediately preceding
peakAmp. Since peakAmp and valleyAmp points were only de-
tected in pairs and otherwise discarded, there was no error in the
PWA value introduced due to one of them missing. In addition,
if there were any ectopic peakAmp points, they were discarded
by the filtering process mentioned in PPI feature extraction.
1067
c) Area: The feature Area represents the area of the triangle
that consists of one peakAmp point and two valleyAmp points
(pre- and post-valleyAmp points of the reference peakAmp
point; see Fig. 3). Similar to PWA feature extraction, all
peakAmp and valleyAmp points were detected as the local max-
imum and local minimum points in the PPG signal. However,
the time stamp (i.e., sample number of each point) was also
recorded, since for Area measurement each peakAmp or val-
leyAmp point was considered as a point in a 2-D plane that con-
sists of time and amplitude. Finally, the Area was measured for
each reference peakAmp points using pre- and post-valleyAmp
points.
2) Arousal Detection Method: As shown in the schematic
diagram (see Fig. 2), after feature extraction, each feature was
used independently to detect the arousal events. The steps of
arousal detection using single feature (PWA) are shown in Fig. 4.
Although the same steps were followed for PPI and Area fea-
tures, Fig. 4 represents the case only for PWA for simple and
clear presentation of the method. At first step, the PWA feature
was calculated from a raw IR signal and then resampled into
1-Hz signal, i.e., there is one value for each second. This was
done to accommodate the moving window used in the following
step and resolution more than 1 s was not necessary for arousal
detection. After resampling, the percentage change parameter
(PCparam ) was calculated using a moving window of 25 s with
a 24 s overlap. For calculating PCparam at the ith second, the
feature window spanned the (i 9) th to (i + 15) th second.
The steps for calculating PCparam (i) are shown in Fig. 5. For
each window PWAseg , the reference PWA value (PWAref ) was
calculated as the average of first ten PWA values of the segment.
The PC value was then calculated for each PWA value of that
segment with respect to PWAseg value (see Fig. 4).
After mapping PWA values to PC, the first average (PC1avg )
was calculated as the average feature value of the first 10 s,
1068 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 3, MAY 2014
Fig. 4. Steps for arousal detection using single feature PWA. PCp a ra m is the
PC of a parameter (PWA, PPI, or Area) and IR data is the raw PPG signal.
Fig. 5. Steps for calculating PCp a ra m (i). Top panel shows the PWAre f based
on which PC is calculated which is shown in the bottom panel. The bottom panel
also shows the average of percentage change PC1av g and PC2 av g and their
difference as PCp a ra m (i).
and the second average (PC2avg ) was calculated as the average
of feature value from the 15th to 20th seconds. Finally, the
PCparam for that window (PCparam (i)) was calculated as the
difference between PC1avg and PC2avg . Since the calculation
of PCparam at any time requires at least 9 s prior and 15 s
postfeatures, the first 9 and last 15 values of P Cparam signals
were set to zero, and therefore, it could not be used to detect
arousal for that period. After calculating the PCparam signal,
the arousal events were detected using a predefined percentage
parameter threshold value (PCthresh ). Any PCparam value larger
than PCthresh was detected as the raw arousal events. The raw
arousal events were then merged into a single event, if multiple
arousal events were within 10 s range, to generate the final
output of the arousal detection system for the individual feature.
3) Performance Analysis: In this study, two measures (ac-
curacy and false detection rate) were used to assess the perfor-
mance of the arousal detection model. Accuracy is the percent-
age of the correctly detected events with respect to total number
of actual events and it was calculated as follows:
TP
Accuracy = 100
TP + MP
where TP (true positive) is the number of events detected and
matched with the actual events and MP (missed positive) is the
number of events present in actual events but not detected. The
false rate is the percentage of false events detected by the system
with respect to actual number of events and it was calculated as
follows:
FP
False rate = 100
TP + MP
where FP (false positive) is the number of events detected but
did not match with actual events.
III. RESULTS
The Somnologica snapshot of output signals at different steps
of arousal detection using PWA feature is shown in Fig. 6. The
panel representing EEG (C4) signal represents the reference
arousal events. The PWA feature extracted from the IR signal
is shown as Raw PWA, which was resampled into a 1-Hz
signal and shown as PWA (1 Hz). The rectangular box in panel
PWA (1 Hz) represents the 25 s window used to calculate the
PCparam and the output is shown in panel %Change PWA.
The straight line in %Change PWA represents the PCthresh
value and was used to detect the arousal events, which is shown
in Raw Arousal Events PWA. Finally, Arousal Events PWA
was generated by merging multiple raw arousal events within a
10 s time period and this was the detected arousal event used to
analyze in the next step.
Briefly, the arousal detection using PPI, PWA, Area, and the
combination of PWA and Area is shown in Fig. 7. Steps in-
volving the calculation of features and resampling into a 1-Hz
signal, which is shown in Fig. 6, are not repeated in Fig. 7. Pan-
els Raw Events (PPI), Raw Events (PWA), and Raw Events
(Area) were calculated using a threshold value on %Change
PPI, %Change PWA, and %Change Area data, respec-
tively, which are shown in detail in Fig. 6 for the PWA feature.
Arousal Events PPI, Arousal Events PWA, and Arousal
Events Area were calculated by merging multiple events within
10 s of Raw Events (PPI), Raw Events (PWA), and Raw
Events (Area), respectively. Merging of Raw Events (Area)
is shown using a solid circle and an arrow in Fig. 7. Arousal
Events (PWA + Area) was calculated by combining events
detected in Arousal Events PWA and Arousal Events Area,
i.e., any event in Arousal Events (PWA + Area) was presented
only when the event was detected by both PWA and Area fea-
tures. This is shown in Fig. 7 using the dotted circle and an
arrow.
KARMAKAR et al.: DETECTION OF RESPIRATORY AROUSALS USING PHOTOPLETHYSMOGRAPHY (PPG) SIGNAL IN SLEEP APNEA PATIENTS 1069

Fig. 6. Output signals at different steps of arousal detection using the PWA feature.

Fig. 7. Arousal events detection using PPI, PWA, and Area features. Arousal events detection by filtering raw arousal events detected using PPI, PWA, and Area
features.
1070
Fig. 8. Accuracy and false rate of arousal detection for all ten subjects using (a) PPI, (b) Area, (c) PWA, and (d) PWA + Area.
The performance of PWA, Area, PPI, and PWA + Area fea-
tures (accuracy and false detection rate) with varying threshold
of PC is shown in Fig. 9. Using the PWA feature [see Fig. 9(a)],
maximum accuracy obtained was around 93% with a false rate
of around 89%. The false rate drops below 49% at 80% accuracy
level with PCthresh = 30%. Moreover, for the PWA feature, the
drop in accuracy is much slower compared to the drop in the
false rate with increasing PCthresh value. This indicates that the
use of a relatively higher threshold value will reduce the false
alarm while using PWA.
The Area feature showed a maximum accuracy value of
around 93% with a false rate of around 85%. The false rate drops
below 50% at an 80% accuracy level with PCthresh = 32%. Sim-
ilar to the PWA feature, accuracy of Area feature also decayed
slowly compared to the false rate. However, decreasing the rate
of false rate for PWA feature with increasing PCthresh value was
higher than Area feature (see Fig. 8).
Using the PPI feature, the maximum accuracy obtained was
around 88% with a false rate of around 86%. The false rate
drops below 75% at an 80% accuracy level with PCthresh = 7%.
In contrast to the PWA and Area features, the accuracy drops
faster than the false rate for PPI feature with increasing PCthresh
value. In addition, at the lowest PCthresh value, the PPI-based
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 3, MAY 2014
model showed higher accuracy than the false rate, which was
also opposite to the pattern found with PWA- and Area-based
models. As a result, with increasing PCthresh value, the false
rate becomes higher than accuracy for PPI model; however,
both values become very low, which indicates that the PPI-
based model underestimates or misses most of the events for
any PCthresh value. Using both PWA and Area features, the
maximum accuracy obtained is around 92% with a false rate of
around 85%. The false rate drops below 47% at an 80% accuracy
level with a threshold change of 28%.
PPG-based arousal indexes agreed well across a wide range
of decision thresholds (ROC = 0.91), and the decision threshold
for the final analyses (PCthresh = 25%) was chosen with a sensi-
tivity of 68.1% and a specificity of 95.2%. At PCthresh = 25%,
the results showed an accuracy of 14.68%, 85.15%, 86.93%,
and 50.79% with a false rate of 21.80%, 55.41%, 64.78%, and
50.79% for PPI, PWA, Area, and PWA + Area features, respec-
tively. The actual arousal index and number of arousals for
individual subjects are compared to that of the arousal detection
model developed in this study and summarized in Table I.
The PPI feature severely underestimated the number of
arousal events except for subject 1, which was highly over-
estimated (344 events detected against 117 actual events). This
KARMAKAR et al.: DETECTION OF RESPIRATORY AROUSALS USING PHOTOPLETHYSMOGRAPHY (PPG) SIGNAL IN SLEEP APNEA PATIENTS 1071

TABLE I
ACTUAL AND CALCULATED NUMBER OF AROUSALS AND AROUSAL INDEX OF ALL 10 SUBJECTS. COMB MEANS BOTH
PWA AND AREA FEATURES ARE USED TO DETECT THE AROUSAL

Actual Calculated
Subject Sleep duration
No of Arousal No of Arousal Arousal Index
ID (hours:mins)
Arousal Index PPI/PWA/Area/Comb PPI/PWA/Area/Comb
1 7:14 117 16.18 344/140/320/114 47.47/19.32/44.16/15.73
2 6:51 362 52.85 203/445/472/427 29.62/64.94/68.88/62.31
3 7:03 459 65.11 51/509/499/496 7.22/72.07/70.65/70.23
4 10:54 464 42.57 160/582/680/527 14.68/53.38/62.37/48.34
5 6:48 368 54.12 20/443/453/423 2.94/65.02/66.49/62.09
6 7:04 217 30.71 63/461/464/429 8.91/65.20/65.63/60.68
7 7:25 243 32.76 74/489/453/430 9.96/65.82/60.97/57.88
8 6:49 172 25.23 40/421/402/386 5.87/61.74/58.95/56.61
9 8:24 412 49.05 65/326/427/305 7.74/38.81/50.83/36.31
10 6:19 303 47.97 87/358/365/323 13.74/56.53/57.64/51.00

indicates that the PPI threshold PCthresh = 25% resulted in very
low accuracy with relatively higher false rate, which was shown
in Fig. 8(a). In contrast, the Area feature overestimated the num-
ber of arousal events for all subjects. For example, the maximum
overestimation happens for subject number 6, with 464 events
detected for 217 actual events (see Table I). This means that it
maintains a high accuracy with relatively higher false rate for
PCthresh = 25% [see Fig. 8(b)]. The PWA feature also showed
overestimation except for subject 9, where it detected only 326
events compared to 412 actual events. However, the overesti-
mation of the PWA feature is less compared to the Area feature.
This resulted in a similar accuracy value with lower false rate
than the Area feature at PCthresh = 25% for the PWA feature
[see Fig. 8(c)]. Interestingly, the combination of PWA and Area
features reduces the false rate, while keeping accuracy at the
same level [see Fig. 8(d)]. This means that the use of the com-
bination of these two features increases the reliability of the
detection of arousal events. As a result, although PWA + Area
(Comb) overestimates the number of arousal events for most of
the subjects (except subjects 1 and 9), the number of estimated
events were closest to the actual event number for all subjects
(shown in a bold format in Table I). However, the combination
of all three features decreases accuracy and increases the false
rate due to the influence of PPI feature [see Fig. 8(a)] and the
results are not incorporated in this study. The arousal index pre-
sented in Table I was calculated as the number of events/hour
and it showed the same pattern as the number of arousal events
discussed previously.
IV. DISCUSSION
The PPG-based algorithm was designed to differentiate be-
tween EEG arousal and non-EEG arousal conditions. This study
demonstrated that it is possible to detect visually scored EEG
arousals with an algorithm based on the PPG signal alone. Al-
though close association of cortical and autonomic arousals
has been repeatedly shown in the past [16], [17], some distinct
events may occur independently (i.e., cortical arousal without
autonomic arousal and vice versa). At least some of the disagree-
ment between visual EEG arousal scoring and PPG-based scor-
ing may be attributed to these independently occurring events.
The PWA feature is a well-known autonomic marker and
has been previously used for the detection of respiratory events
and arousals in patients with SAS [18]. However, in that study
PWA was used as an additional parameter along with airflow
and oxygen desaturation to score respiratory events. The PPI
feature was previously used as a surrogate of HRV analysis [13]
and recently reported to classify obstructive sleep apnoea in
children [14]. However, the PRV feature used in that study is
different from the PPI feature used in this study and no result has
been reported in detecting respiratory events or arousal in the
sleep disordered adult population. Moreover, use of another PPG
feature (Area) in arousal detection is reported for the first time
in this study to our knowledge. We are convinced that our new
analysis allows a better quantification of respiratory arousals.
Previous works have suggested that there could be other mark-
ers of arousal such as PTT, peripheral arterial tonometry, etc.,
which are much easier to measure than EEG and which could as-
sess the true extent of sleep fragmentation [19][22]. However,
in those studies PTT signal was visually analyzed for respiratory
event or arousal detection. In this study, we used PPG features
which evaluate changes in peripheral arterial volume derived
features as markers for automatic respiratory arousal detection.
On the other hand, PAT has been used with other markers for ex-
ample oximetry and actigraphy (Watch-PAT) in order to detect
sleep apnea and arousal [22].
The direct relation between the drop in PWA values and
respiratory arousals was previously reported in several stud-
ies [18], [23], [24]. In addition to the PWA feature, we have
found a similar drop in PPI and Area values. However, the main
problem in arousal event detection using PPG features lies in the
selection of a threshold value. An overestimated threshold value
reduces both the false rate and accuracy of the model. Although
reduction in false rate is desirable for any model, the accuracy
reduction is not. In contrast, the underestimated threshold value
resulted in increasing both the accuracy and false rate, which is
also undesirable. Therefore, in this study we have calculated the
accuracy and false rate over a range of threshold value for all
PPG features [see Fig. 8(a)(d)]. From the characteristic curve
of accuracy and false rate, we have found that it is not possible
to define a fixed threshold value for all features. In this study,
we have selected the threshold value for which the model shows
1072 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 3, MAY 2014
at least 80% accuracy with a minimum false rate and we have
found 30%, 32%, 7%, and 28% PCthresh values for PPG features
PWA, Area, PPI, and PWA + Area respectively.
In order to get better detection results, the following recom-
mendations in selecting thresholds are proposed. These param-
eters are as follows.
1) Percentage change parameter (PCparam ): For calculating
PCparam , the window length can be varied. In this study,
we have varied the window length from 15 to 20 s for the
PPI feature and compared their performance. From those
results, we found that window length of 20 s performed
better than other window lengths.
2) Percentage change threshold (PCthresh ): In this study,
we have used PCthresh = 25%, for all features. This may
however be changed to achieve the desired accuracy or
false rate values which may be verified with other clinical
measures of daytime sleepiness. Also, some false positives
could be autonomic arousals which need to be verified in
future.
In this study, we have used three PPG features (PPI, PWA,
and Area) for arousal detection [(see Fig. 8(a)(c)]. PWA and
Area have shown better performance (higher accuracy and lower
false rate) compared to the PPI feature at PCthresh = 25%. Af-
ter investigating the possible combinations of those features, we
have come up with a conclusion that combining PWA and Area
provides better accuracy with lower false rate in arousal detec-
tion [see Fig. 8(d)]. From the subjective analysis of the number
of arousals and arousal index, it is obvious that the developed
model overestimates the actual arousal number and index except
subjects 1 and 9 (see Table I). This is in line with the findings
reported in previous study that changes in PWA were also seen
in the absence of conventionally scored EEG arousals which
can be attributed to the subcortical arousal [24], [25].
One limitation of this study is the small number of subjects
used to validate the PPG-based respiratory arousal detection
model developed. However, most of the subjects had moderate
to high respiratory arousal index; therefore, the number of events
(2917) analyzed in this study was large. In addition, the visual
scoring of arousal depends on the scorer [26]. It was our aim to
have just one visual scorer for this validation study. However,
further validation of the developed model needs to be done with
more subjects and variation in the number of arousal per subject
(very small arousal index to very large arousal index) to get a
better understanding of the correlation between PPG features
and sleep disturbances, which would include a comparison of
different scorers on scoring the arousal [26].
In conclusion, the PWA and Area features-based model for
arousal detection showed better performance compared to the
PPI-based model. Combining the PWA and Area features in-
creased the reliability of the model, i.e., a reduced false detec-
tion rate without much effect on the accuracy. This was due to
the overestimating nature of both PWA- and Area-based mod-
els. Although PPG-based respiratory arousal detection models
mostly overestimate the number of actual events, which can
be attributed to sub-cortical arousal, future exploration of sub-
cortical and cortical arousal can provide explicit explanation
of the reason behind the overestimation found in this study.
Finally, the PPG-based respiratory arousal detection model is
a simple and could be a promising alternative to conventional
electroencephalogram (EEG)-based arousal detection systems
useful for monitoring sleep health at home and ambulatory
conditions.
REFERENCES
[1] American Academy of Sleep Medicine Taskforce, Sleep-related breath-
ing disorders in adults: Recommendations for syndrome definition and
measurement techniques in clinical research, Sleep, vol. 22, pp. 667
689, 1999.
[2] J. M. Shineerson, Sleep Medicine: A Guide to Sleep and its Disorder.
Oxford, U.K.: Blackwell, 2005, pp. 230231.
[3] E. A. Phillipson and C. E. Sullivan, Arousal: The forgotten response to
respiratory stimuli, Amer. Rev. Respir. Dis., vol. 118, pp. 807809, 1978.
[4] J. E. Remmers, W. J. deGroot, E. K. Sauerland, and A. M. Anch, Patho-
genesis of upper airway occlusion during sleep, J. Appl. Physiol., vol. 44,
pp. 931938, 1978.
[5] C. Guilleminault and M. Rosekind, The arousal threshold: Sleep depri-
vation, sleep fragmentation, and obstructive sleep apnea syndrome, Bull
Eur. Physiopathol. Respir., vol. 17, pp. 341349, 1981.
[6] L. C. Johnson and A. Lubin, The orienting reflex during waking and
sleeping, Electroencephalogr Clin. Neurophysiol., vol. 22, pp. 1121,
1967.
[7] N. J. Douglas and S. E. Martin, Arousals and the sleep apnea/hypopnea
syndrome, Sleep, vol. 19, no. 10, pp. S196S197, 1996.
[8] S. E. Martin, P. K. Wraith, I. J. Deary, and N. J. Douglas, The effect of
nonvisible sleep fragmentation on daytime function, Amer. J. Respir. Crit.
Care Med., vol. 155, pp. 15961601, 1997.
[9] N. A. Collop, Portable monitoring for the diagnosis of obstructive sleep
apnea, Curr. Opin. Pulm. Med., vol. 14, pp. 525529, 2008.
[10] D. Pitson, N. Chhina, S. Knijn, M. van Herwaaden, and J. Stradling,
Changes in pulse transit time and pulse rate as markers of arousal from
sleep in normal subjects, Clin. Sci., vol. 87, pp. 269273, 1994.
[11] P. G. Catcheside, S. C. Chiong, J. Mercer, N. A. Saunders, and
R. D. McEvoy, Noninvasive cardiovascular markers of acoustically in-
duced arousal from non-rapid-eye-movement sleep, Sleep, vol. 25, no. 7,
pp. 797804, 2002.
[12] J. Haba-Rubio, G. Darbellay, F. R. Herrmann, J. G. Frey, A. Fernandes,
J. M. Vesin, J. P. Thiran, and J. M. Tschopp, Obstructive sleep apnea
syndrome: Effect of respiratory events and arousal on pulse wave ampli-
tude measured by photoplethysmography in NREM sleep, Sleep Breath,
vol. 9, no. 2, pp. 7381, 2005.
[13] M. Nitzan, S. Turivnenko, A. Milston, A. Babchenko, and Y. Mahler,
Low-frequency variability in the blood volume and in the blood volume
pulse measured by photoplethysmography, J. Biomed. Opt., vol. 1, no. 2,
pp. 223229, 1996.
[14] J. Lazaro, E. Gil, J. M. Vergara, and P. Laguna, OSAS detection in chil-
dren by using PPG amplitude fluctuation decreases and pulse rate vari-
ability, Comput. Cardiol., vol. 39, pp. 185188, 2012.
[15] A. Zacharia, J. Haba-Rubio, R. Simon, G. John, P. Jordan, A. Fernandes,
J. M. Gaspoz, J. G. Frey, and J. M. Tschopp, Sleep apnea syndrome: Im-
proved detection of respiratory events and cortical arousals using oxymetry
pulse wave amplitude during polysomnography, Sleep Breath., vol. 12,
no. 1, pp. 3338, 2008.
[16] A. Delessert, F. Espa, A. Rossetti, G. Lavigne, M. Tafti, and R. Heinzer,
Pulse wave amplitude drops during sleep are reliable surrogate markers
of changes in cortical activity, Sleep, vol. 33, no. 12, pp. 16871692,
2010.
[17] J. Haba-Rubio, G. Darbellay, F. R. Herrmann, J. G. Frey, A. Fernandes,
J. M. Vesin, J. P. Thiran, and J. M. Tschopp, Obstructive sleep apnea
syndrome: Effect of respiratory events and arousal on pulse wave ampli-
tude measured by photoplethysmography in NREM sleep, Sleep Breath,
vol. 9, no. 2, pp. 7381, 2005.
[18] D. J. Pitson, A. Sandell, R. van den Hout, and J. R. Stradling, Use of pulse
transit time as a measure of inspiratory effort in patients with obstructive
sleep apnoea, Eur. Respir. J., vol. 8, pp. 16691674, 1995.
[19] J. Argod, J. L. Pepin, and P. Levy, Differentiating obstructive and central
sleep respiratory events through pulse transit time, Amer. J. Respir. Crit.
Care Med., vol. 158, pp. 17781783, 1998.
[20] J. Argod, J. L. Pepin, R. P. Smith, and P. Levy, Comparison of esophageal
pressure with pulse transit time as a measure of respiratory effort for
KARMAKAR et al.: DETECTION OF RESPIRATORY AROUSALS USING PHOTOPLETHYSMOGRAPHY (PPG) SIGNAL IN SLEEP APNEA PATIENTS
scoring obstructive nonapneic respiratory events, Amer. J. Respir. Crit.
Care Med., vol. 162, pp. 8793, 2000.
[21] T. Penzel, K. Kesper, I. Pinnow, H. F. Becker, and C. Vogelmeier, Periph-
eral arterial tonometry, oximetry and actigraphy for ambulatory recording
of sleep apnea, Physiol. Meas., vol. 25, pp. 10251036, 2004.
[22] S. M. Hilton, The defence-arousal system and its relevance for circulatory
and respiratory control, J. Exp. Biol., vol. 100, pp. 159174, 1982.
[23] E. Sforza, F. Chapotot, S. Lavoie, F. Roche, R. Pigeau, and A. Buguet,
Heart rate activation during spontaneous arousals from sleep: Effect of
sleep deprivation, Clin. Neurophysiol., vol. 11, pp. 244251, 2004.
[24] P. G. Catcheside, S. C. Chiong, J. Mercer, N. A. Saunders, and
R. D. McEvoy, Noninvasive cardiovascular markers of acoustically
induced arousal from non-rapid-eye-movement sleep, Sleep, vol. 7,
pp. 797804, 2002.
[25] J. Hedner, H. Ejnell, J. Sellgren, T. Hedner, and G. Wallin, Is high and
fluctuating muscle nerve sympathetic activity in the sleep apnoea syn-
drome of pathogenetic importance for the development of hypertension?
J. Hypertens Suppl., vol. 6, pp. S529S531, 1988.
[26] J. S. Loredo, J. L. Clausen, S. Ancoli-Israel, and J. E. Dimsdale, Night-
to-night arousal variability and interscorer reliability of arousal measure-
ments, Sleep, vol. 22, no. 7, pp. 916920, 1999.
Chandan Karmakar (S07M11) received the
B.Sc. Eng. degree in computer science and engineer-
ing from the Shah Jalal University of Science and
Technology, Sylhet, Bangladesh, in 1999, and the
Doctor of Engineering degree from The University
of Melbourne, Parkville, Vic., Australia, in 2012.
He is currently working as a Postdoctoral Research
Fellow in the Department of Electrical and Elec-
tronic Engineering, The University of Melbourne,
Australia. He has published one book and more than
40 peer-reviewed journal and conference articles in
the research field of physiological signal processing and modeling. His research
interests include biomedical devices and signal processing, cardiovascular and
neural system related to sleep disordered breathing, human gait dysfunctions,
cardiovascular diseases, and diabetic autonomic neuropathy.
Ahsan Khandoker (M07SM12) received the
B.Sc. degree in electrical and electronic engineering
from the Bangladesh University of Engineering and
Technology (BUET), Dhaka, Bangladesh, in 1996,
the M. Eng.Sc. from Multimedia University (MMU),
Malaysia, in 1999, and the M. Eng. and Doctor of En-
gineering degree in physiological engineering from
the Muroran Institute of Technology, Muroran, Japan,
in 2004.
He is currently an Assistant Professor in the De-
partment of Biomedical Engineering, Khalifa Uni-
versity, Abu Dhabi, UAE. He is also a Senior Research Fellow for Australian
Research Council Research Networks on Intelligent Sensors, Sensor Networks,
and Information Processing (ISSNIP), the University of Melbourne, Parkville,
Vic., Australia. He has published 35 peer-reviewed journal articles and more
than 75 conference papers the research field of physiological signal processing
and modeling in sleep disordered breathing, diabetic autonomic neuropathy,
fetal cardiac disorders and human gait dysfunction, and is passionate about re-
search helping clinicians to noninvasively diagnose diseases at early stage. He
was also with several Australian Medical device manufacturing industries, as
well as hospitals as a Research Consultant focusing on integration of technology
in clinical settings.
The author has requested enhancement of the downloaded file. All in-text references underlined in blue are linked to publications on ResearchGate.
1073
Thomas Penzel (M92SM06) received the Gradu-
ation degree from physics, human biology, and phys-
iology from the University Marburg, Marburg, Ger-
many, in 1986, 1991, and 1995, respectively.
He has been with the University of Marburg since
1984 at the Department of Physics, Physiology, and
the sleep laboratory until 2006. Since 2006 he has
been the Director of Research of the interdisciplinary
sleep medicine center, Charite University Hospital,
Berlin, Germany. He is the Editor-in-Chief of the Ger-
man journal Somnologie and on the editorial board of
other journals. His research interests include biomedical signals, biomedical
devices, the cardiovascular, and the neural system, all related to sleep.
Dr. Penzel received the Bial award for clinical medicine in Portugal, in 2001,
the Bill Gruen Award for Innovations in Sleep Research by the American Sleep
Research Society, in 2008, and the Somnus Award in Germany, in 2012. He
holds several functions in sleep medicine societies, Chair of the ESRS Sleep
Medicine Committee, Secretary of the German Sleep Society, and others.
Christoph Schobel received the Medical degree
at ChariteUniversitaetsmedizin Berlin, Berlin,
Germany, in 2008.
He is currently doing internship in internal
medicine at the Department for Cardiology + An-
giology, Interdisciplinary Center for Sleep Medicine,
Charite Universitaetsmedizin Berlin and working
in EU-funded projects (DAPHNet, HIVE).
Marimuthu Palaniswami (S84M87SM94
F12) received the M.E. degree from the Indian In-
stitute of Science, Bangalore, India, the M.Eng.Sc.
degree from The University of Melbourne, Parkville,
Vic., Australia, and the Ph.D. degree from the Uni-
versity of Newcastle, Callaghan, N.S.W., Australia.
He is currently a Professor with the Department
of Electrical and Electronic Engineering, The Univer-
sity of Melbourne. He has published more than 400
refereed research papers and leads one of the largest
funded Australian Research Council, Research Net-
work on Intelligent Sensors, Sensor Networks, and Information Processing
programme. His research interests include support vector machines sensors and
sensor networks, IoT, machine learning, neural network, pattern recognition,
signal processing, and control.
Dr. Palaniswami has been a Grants Panel Member for The National Science
Foundation, an Advisory Board Member for the European FP6 grant center, a
Steering Committee Member for National Collaborative Research Infrastructure
Strategy, Great Barrier Reef Ocean Observing System, Smart Environmental
Monitoring and Analysis Technologies, and a Board Member for Informa-
tion Technology and supervisory control and data acquisition companies. He
has been funded by several ARC and industry grants (over 40 m) to conduct
research in sensor network, Internet of things (IoT), health, environmental, ma-
chine learning, and control areas. He is representing Australia as a core partner
in European Union FP7 projects such as SENSEI, SmartSantander, Internet of
Things Initiative, and SocIoTal.