Measurement 150 (2020) 107048

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Computerized approach for cardiovascular risk level detection using

photoplethysmography signals
Divya Ramachandran a,⇑, Vanathi Ponnusamy Thangapandian b, Harikumar Rajaguru c
a
Department of Electrical and Electronics Engineering, PSG College of Technology, Tamil Nadu, India
b
Department of Electronics and Communication Engineering, PSG College of Technology, Tamil Nadu, India
c
Department of Electronics and Communication Engineering, Bannari Amman Institute of Technology, Tamil Nadu, India

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

Article history: Cardiovascular disease (CVD) is the leading cause of death globally. In order to decrease the medical cost
Received 31 December 2018 for treating the heart related pathologies, this paper proposes a computer-aided diagnostic system to
Received in revised form 13 July 2019 classify various risk level of cardiovascular disease utilizing inexpensive and non-obtrusive diagnostic
Accepted 24 July 2019
tool called photoplethysmography (PPG). In this study, features such as singular value decomposition
Available online 13 September 2019
(SVD), statistical features and wavelets (Haar, Daubechies, and Symlet) are extracted from the photo-
plethysmography signals. These feature vectors are then applied to the softmax discriminant classifier
Keywords:
(SDC) and Gaussian mixture model classifier (GMM) for classification of various risk phases of CVDs.
Cardiovascular disease
Gaussian mixture model
The classification performance of the proposed model incorporating SDC with SVD and statistical feature
Singular value decomposition vectors increases with sensitivity of 97.24%, specificity of 99.09% and an accuracy of 97.88%.The method
Softmax discriminant classifier presented in this paper assist cardiologists to validate their diagnosis.
Statistical features Ó 2019 Elsevier Ltd. All rights reserved.
Wavelets

1. Introduction Allen examined photoplethysmography signals and has vali-

Cardiovascular disease (CVD) is becoming one of the life threat-
ening diseases in most of the countries. Likely 29.2% or 16.7 million
of total world-wide deaths resulted from the different forms of
CVD [1]. Many forms of CVD can be preventable by taking prime
actions on risk factors like smoking, unhealthy food and lack of
physical activity. The problems due to non-communicable diseases
likely to be increased over the next decades significantly in low and
middle income countries [2,3].To increase the life expectancy of
the people and to cut down the medical cost, PPG has been emerg-
ing as a promising technique for the early screening of heart
related problems [4,5]. The PPG signals are pulsatile in nature
and from the PPG waveforms relevant features of the blood flow
activities can be identified and can be used to measure the cardiac
output [6,7]. PPG has been widely acknowledged by both the Inter-
national Standard Organization (ISO) and European committee for
measurement of oxygen saturation level. A high quality PPG signal
can be easily recorded when the fingers are placed on the PPG
device [8–11]. PPG signal require only less hardware and becomes
more accessible when compared to current electrocardiogram
(ECG) monitoring systems [12].
⇑ Corresponding author.
E-mail address: divyame86@gmail.com (D. Ramachandran).
dated its latent ability for use in clinical measurements in a wide
range especially for the valuation of the cardiovascular system
[13]. Gil et al. measured the pulse rate variability (PRV) from the
PPG signals as a substitute measurement of the non-stationary
heart rate variability (HRV). Indices of time-varying analysis
derived from the PRV showed no static differences with the indices
derived from HRV. Thus for assessing the changes in the autonomic
modulation of heart rate, PRV can be measured [14]. Kenji Taka-
nawa et al. measured second derivative of the PPG waveform for
the assessment of vascular aging [15]. Gonzalez et al. computed
fourth derivative from the acquired PPG signal and computed the
photoplethysmographic augmentation index for cardiovascular
assessments. Techniques like Artificial Neural Networks (ANN),
Support Vector Machine (SVM), Genetic Algorithm (GA), Extreme
Learning Machine (ELM) and K-Nearest Neighbor (KNN) have been
used in the PPG signal classification [16].
Soltane et al. adopted multilayer perceptron neural network
employing Back Propagation (BP) algorithm to classify PPG signals
as normal or pathological and its performance is compared with
that of GMM classifier. Recordings of PPG signal are taken from 2
groups of volunteers consisting of 48 patients (11 pathologies
and 37 normal having age between 21 and 64) [17]. Rohan Baid
et al. proposed a model for extracting features known as an
Auto-regressive exogenous input (ARX) linear parametric model
which characterizes the circulatory system and adopted SVM for

https://doi.org/10.1016/j.measurement.2019.107048
0263-2241/Ó 2019 Elsevier Ltd. All rights reserved.
2 D. Ramachandran et al. / Measurement 150 (2020) 107048
assessing the CVD risk level. SVM classifies the signal depending on
the four selection policies [18]. Polania et al. carried out a tech-
nique for detecting and classifying cardiac arrhythmias through
morphological analysis of PPG signals. Methods starts with pre-
processing stage followed by extracting discriminative features
based on time-domain signal processing, Fourier analysis, and
non-linear dynamics of the heart rate variability. Arrhythmia clas-
sification is performed using SVM [19].
Shobitha et al. classified the PPG signals as healthy or at risk of
CVD using ELM and its performance is compared with other super-
vised learning algorithms such as SVM and back propagation algo-
rithm. These algorithms are tested for 30 pathological and 30
healthy signals which are obtained from Biomedical Research
Lab. Also, ELM uses only less feature vectors as inputs to detect
the risk of CVD and thus reducing the computation time [20]. Hos-
seini et al. proposed finger PPG obtained before and after hypermia
to differentiate between subjects with normal-to-mild and serious
coronary artery disease. PPG and ECG signals were recorded for 37
patients. Many time-domain features were examined using K-
nearest neighbor classifier and the better results have been
obtained for combination of features like pulse transit time, rising
time, and crest time [21].
In this paper, an investigation on various risk groups of cardio-
vascular disease from PPG signal has been carried out. Statistical
features, singular value decomposition and wavelets are extracted
from the PPG signals and given to a simple but efficient classifier
called softmax discriminant classifier (SDC). SDC takes nonlinear
transformation of the distance between the training and testing
samples and assigns the label information to the new testing sam-
ple. Then the performance of the SDC is compared with the Gaus-
sian mixture model classifier (GMM). The performance of SDC and
GMM are assessed using measures like sensitivity, specificity,
accuracy, error rate, false alarm, precision and F-measure.
The paper is structured as follows: Section 2 explains in detail
the methods for identifying the various risk stages of cardiac using
PPG which includes feature extraction process, Softmax discrimi-
nant classifier and Gaussian mixture model classifier. Section 3
explains the results and discussions. The Section 4 of this paper
gives conclusion followed by references.
2. Methods
In this study, classification of various risk levels of cardiovascu-
lar disease has been carried out. The input given to the proposed
system is the photoplethysmography (PPG) signal. The PPG signal
Statistical Features
PPG SVD
signal Wavelets
Daubechies (db4), Haar
Symlet (sym8)
Feature
extraction
Cardiovascular disease
risk level analysis
Fig. 1. Flow diagram of the proposed CVD risk levels detection system.
dataset has been sampled at a sampling rate of 300 Hz and the data
sample length obtained from the PPG signal is 1,44,000 which has
720 segments in total. These segments are of equal intervals com-
prising of 200 samples per segment.
In the first stage, features extraction process is carried out using
statistical analysis, singular value decomposition (SVD) and differ-
ent types of wavelets from the PPG signal dataset. These extracted
features are given as input to the SDC and GMM classifier. Finally
classification of the PPG signal into normal and three different
levels of abnormality are labelled. The illustration of the proposed
CVD risk level detection system is shown in Fig. 1. Each block of
Fig. 1 is explained briefly in the following sub-sections.
2.1. SVD and statistical features
In the context of machine learning or signal processing or pat-
tern recognition, feature extraction aims to build features from
the original data set which are instructive and non-redundant.
Appropriate information will be obtained from the selected fea-
tures and the expected task can be performed from this reduced
dataset instead of the comprehensive original data. SVD and nine
statistical features are extracted from the PPG signals are defined
in Table 1.
The SVD provides a systematic way to determine the dominant
patterns underlying in a high-dimensional system and also pro-
vides a low-rank approximation to high-dimensional data. The
SVD results in numerically stable computation. By extracting above
features, data has been compressed to 10 samples out of 200 sam-
ples per segment. Since there are 720 segments in the recorded
PPG signal, data size has been reduced to 7200 samples from
1,44,000. Table 2 depicts the sample SVD and statistical features
extracted from abnormal and normal PPG signal of Capnobase
dataset.
2.2. Discrete wavelet transform
Discrete Wavelet Transform (DWT) uses a discrete set of wave-
let scales and transformations obeying some defined rules [22].
Wavelet transform can be applied to PPG signal and convert that
to wavelet parameters or coefficients. The acquired coefficients
depict the behavior of the PPG signal. There are different types of
wavelet transforms such as Daubechies, Haar, Symlets, Coiflets,
Biorthogonal etc [23]. In this paper, db4, haar and sym8 wavelets
are chosen based on their properties like symmetry, smoothness,
orthogonality and symmetry for feature extraction.
Softmax discriminant Performance metrics
classifier
Sensitivity, Specificity,
Gaussian mixture Accuracy, Error rate, FA,
model classifier Precision and F-measure
Classification
 D. Ramachandran et al. / Measurement 150 (2020) 107048 3

Table 1 nonlinear classifiers like SDC and GMM are selected for further
SVD and Statistical features extracted from PPG signals. classification.
Features Definition
Arithmetic mean l ¼ Rnx Where, x – observed data values, n – 2.3. Target selection based on SVD parameters
Complete number of values in that particular set of
observations, m - Arithmetic mean. The target TN is set for normal case based on the condition
Variance r2 ¼ Rðxn lÞ Where, r2 – Variance, x – observed
2

1 XM
data values, n- Complete number of values in that l  TN ð1Þ
particular set of observations. M k¼1 k
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Standard deviation r ¼ Variance
Skewness
Pn  3
ðl lÞ =n where lk represents mean for normalised features of normal signals
g 1 ¼ i¼1 ri3 Where, n – No. of data points, l –
for M number of normal signals.
mean and r – standard deviation
Kurtosis
Pn  4 The target TC is fixed for abnormal case having cardiac of risk
ðl lÞ =n
K ¼ i¼1 ri4 Where, n – No. of data points, l - level 1 depends on the constrain
mean, and r – standard deviation
P
Energy E ¼ 1- 1 jxðnÞj2 1 XQ
l  TC ð2Þ
Approximate entropy It is a regularity statistic that quantifies the Q i¼1 i
unpredictability of variations in a time-series of
data. The time-series is deterministic for a low
where li is the mean for normalised features of abnormal signals
value of the entropy
Peak maximum It refers to the location of the maximum having risk level 1for Q number of abnormal signals having risk
concentration of a peak in a data sample level 1.
Maximum slope y ¼ mx þ c The target TB is set for abnormal case having cardiac of risk level
The slope is positive when m > 0 and it is negative 2 based on the condition
when m < 0
SVD Real or complex matrix is a factorization of the 1X S
form U RV  l  TB ð3Þ
where, U - c  c real or complex unitary matrix, S j¼1 j
P
- c  b rectangular diagonal matrix and V - b  b
real or complex unitary matrix where lj is the mean for normalised features of abnormal signals
having risk level 2 for S number of abnormal signals having risk
level 2.
In signal analysis using wavelet, appropriate selection of wave- The target TA is fixed for abnormal case having respiratory dis-
let and the number of decomposition level is vital. Approximation order based on the condition
coefficients at level 5 decomposition have been extracted for these 1 XU

wavelets and they are used as feature vector for classification. For l  TA ð4Þ
U l¼1 l
200 samples per segment, data size has been compressed to 13
using db4, 7 using Haar and 20 using sym8. In total, space required where ll is the mean for normalised features of abnormal signals
for storing data using db4, haar and sym8 is only 9360, 5040 and with respiratory disorder for U number of abnormal signals with
14,400 samples respectively. The easiest of all discrete wavelet respiratory disorder.
transform is Discrete Haar Wavelet Transform (HWT). Haar wave- The target TA should be greater than the average of mk and ll .
let is compact, symmetrical and differentiable but not continuous After the selection of targets TN and TA, we must ensure that
[24]. For the analysis of signals with sudden transitions, this prop- k TN  TA k  0:5. TC and TB target values are intermediate between
erty will be an advantage. Daubechies wavelet is more appropriate TN and TA.
for its orthogonality nature and applicable for smoothly changing The target TN, TC, TB and TA for normal and abnormal signals
signals [25]. Symlet is near symmetrical, compactly supported having various risk levels of cardiac are fixed at 0.11, 0.45, 0.65
and suits to both orthogonal and biorthogonal families [26]. Tables and 0.85 respectively. In order to map the TN, TC, TB and TA to
3–5 shows the sample features extracted using db4, haar and SVD value, following condition should be satisfied.
sym8. Condition (1): ll > lj > li > lk
The non-linearity behavior of SVD and statistical features are    
Condition (2): 12 ðlk  li Þ < 12 li  lj < 12 lj  ll
depicted in the Fig. 2 as a histogram. From the Fig. 2 it is observed
that there exists non-linearity in the PPG signals. Condition (3): Considering 16 neighbourhood points
Fig. 3 depicts the histogram of haar wavelet features in Chi 1 1 2 1  2
square PDF for 10 degree freedom of normal PPG signals and PPG ðlk  li Þ2 < li  lj < lj  ll
16 2 2
signals of CVD Patient. Fig. 3 is also non-linear in nature. Therefore

Table 2
Depiction of sample SVD and statistical features vectors extracted from PPG data set.

Signal Mean Variance Stand. Dev. Skewness Kurtosis Energy App. entropy Peak Max. Max.Slope SVD
abnorm1 0.34 0.47 0.69 0.62 0.75 0.3 0.78 0.87 1 0.55
abnorm2 0.36 1 1 0.87 0.91 1 0.83 1 0.97 1
abnorm3 0.1 0.1 0.32 0 0.01 0.1 1 0.63 0.33 0.31
abnorm4 0.07 0.8 0.9 0.29 0.91 0.79 0.95 0.96 0.82 0.89
abnorm5 0.37 1 1 0.03 1 1 0.74 0.96 0.39 1
abnorm6 1 0.75 0.87 0.08 0.54 1 0.42 1 0.67 1
norm1 0.36 0.39 0.63 0.86 0.15 0.27 0.39 0.82 0.31 0.52
norm2 0.92 0.36 0.6 0.14 0.67 0.48 0.92 0.89 0.92 0.69
norm3 0.59 0.88 0.94 0.42 0.15 0.91 0.94 0.92 0.83 0.95
norm4 0.54 0.82 0.91 0.41 0.96 0.84 0.98 0.93 0.86 0.92
4 D. Ramachandran et al. / Measurement 150 (2020) 107048

Table 3
Depiction of sample features extracted using wavelets-db4 from PPG data set.

db4, wavelet Level 5

Signal 1 2 3 4 5 6 7 8 9 10 11 12 13
abnorm1 0.34 0.34 0.34 0.34 0.36 0.29 0.09 0.40 0.48 0.48 0.47 0.46 0.45
abnorm2 0.42 0.42 0.42 0.42 0.39 0.56 0.35 0.79 0.56 0.13 0.75 0.53 0.40
abnorm3 0.58 0.58 0.58 0.58 0.57 0.57 0.88 0.20 0.63 0.45 0.08 0.03 0.04
abnorm4 0.69 0.69 0.69 0.69 0.67 0.75 0.67 0.23 0.20 0.35 0.59 0.65 0.59
abnorm5 0.69 0.69 0.69 0.68 0.70 0.66 0.12 0.73 0.97 0.08 0.50 0.26 0.48
abnorm6 0.70 0.71 0.70 0.70 0.73 0.64 0.01 0.52 0.56 0.53 0.52 0.47 0.45
norm1 0.36 0.37 0.36 0.36 0.41 0.20 0.78 0.70 0.08 0.20 0.32 0.44 0.47
norm2 0.07 0.06 0.07 0.07 0.03 0.20 0.38 0.06 0.13 0.10 0.07 0.08 0.08
norm3 0.28 0.28 0.28 0.28 0.26 0.33 0.03 0.32 0.09 0.29 0.20 0.15 0.14
norm4 0.13 0.13 0.13 0.13 0.12 0.15 0.59 0.82 0.48 0.03 0.15 0.22 0.23

Table 4
Depiction of features extracted using wavelets-sym8 from PPG data set.

Sym8, wavelet Level 5 abnorm1 abnorm2 abnorm3 abnorm4 abnorm5 norm1 norm2 norm3
1 0.34 0.68 0.68 0.87 0.60 0.09 0.07 0.27
2 0.34 0.68 0.69 0.88 0.59 0.09 0.08 0.29
3 0.35 0.69 0.68 0.88 0.59 0.09 0.07 0.29
4 0.34 0.68 0.69 0.88 0.59 0.09 0.08 0.29
5 0.34 0.68 0.69 0.88 0.59 0.09 0.08 0.29
6 0.35 0.69 0.67 0.89 0.59 0.08 0.05 0.27
7 0.33 0.71 0.72 0.86 0.57 0.11 0.15 0.34
8 0.05 0.00 0.69 0.12 0.86 0.25 0.42 0.11
9 0.38 0.69 0.28 0.60 0.39 0.44 0.10 0.35
10 0.49 0.96 0.13 0.71 0.64 0.12 0.13 0.04
11 0.50 0.12 0.43 0.67 0.48 0.17 0.11 0.29
12 0.48 0.57 0.53 0.65 0.13 0.10 0.08 0.22
13 0.47 0.19 0.62 0.61 0.03 0.27 0.08 0.16
14 0.47 0.26 0.64 0.60 0.04 0.29 0.08 0.15
15 0.47 0.28 0.63 0.60 0.04 0.30 0.08 0.15
16 0.47 0.28 0.63 0.60 0.04 0.30 0.08 0.15
17 0.47 0.31 0.62 0.60 0.04 0.30 0.08 0.15
18 0.47 0.27 0.63 0.60 0.04 0.30 0.08 0.15
19 0.47 0.26 0.63 0.60 0.04 0.29 0.08 0.15
20 0.47 0.36 0.60 0.60 0.04 0.31 0.08 0.14

The network will be described by mapping from inputs to the out-

Table 5
Depiction of features extracted using wavelets-Haar for PPG data set.
puts. SDC uses ‘h’ test class samples to minimize the reconstruction
error. Reconstruction of the input is the mapping from the outputs
Haar, wavelet Level 5 to the inputs in a network. To measure the correctness of the out-
Patient 1 2 3 4 5 6 7 put, reconstruction error is calculated. Reconstruction error is
abnorm1 0.62 0.03 0.96 0.41 0.41 0.90 0.69 determined by taking the difference of actual input and recon-
abnorm2 0.57 0.21 0.67 0.65 0.62 0.61 0.57 struction input. The non-linear transformation value of the dis-
abnorm3 0.22 0.02 0.32 0.61 0.18 0.63 0.58 tance between h class samples and the testing samples is
abnorm4 0.52 0.47 0.42 0.56 0.69 0.19 0.75
maximized to achieve the aim of the SDC.
norm1 0.28 0.32 0.02 0.14 0.09 0.08 0.08
norm2 0.28 0.26 0.01 0.09 0.13 0.15 0.16 Now, SDC can be defined as
norm3 0.26 0.10 0.22 0.17 0.26 0.17 0.13 i
norm4 0.30 0.36 0.04 0.81 0.38 0.17 0.25 zðyÞ ¼ arg max ky ð5Þ
i

!
X
ni

and subsequently lk  40, li ¼ 41  50;lj ¼ 51  60 and ¼ arg max log expðkk y  yij k2 Þ ð6Þ
i
j¼1
ll > 60 are mapped to target values 0.11, 0.45, 0.65 and 0.85
i
respectively. where ky , z (y) shows the distance between testing sample and ith
class. When the parameter k > 0, a relative penalty cost is provided.
If y belongs to ith class, then y and yj has same characteristics,
2.4. Softmax discriminant classifier i
thus k y  yij k2 is chosen close to zero and ky can asymptotically
The main objective of SDC is to correctly identify the class to attains the maximum value.
which a new testing sample belongs to [27]. SDC simply weights
the distance between training and testing samples from that class. 2.5. Gaussian mixture model
Let the training set Y ¼ ½Y1 ; Y2 ; :::::Yh  2 Rmxn ; taken from the ‘h’
Gaussian mixture model (GMM) is a combination of numerous
different classes. Yh ¼ ½Yh1 ; Yh2 ; ::::::Yhnh  2 Rmxnh denotes nh samples
Ph Gaussian distributions and can therefore denote different sub-
from the hth class, where i¼1 ni ¼ n. Let the testing samples be classes inside one class. Given a data point , Gaussian mixture
m
z 2 R and the outputs of the network be Q ¼ ½Q 1 ; Q 2 ; :::::Q l . model is defined as [28]
 D. Ramachandran et al. / Measurement 150 (2020) 107048 5

100 450

90 400

80
350

70
300
60

Occurence
Occurence

250
50
200
40
150
30

100
20

10 50

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 1 2 3 4 5 6
-4
SVD and statistical features of normal PPG Signal Haar wavelet features of normal PPG Signal x 10
(a) (a)
45 160

40
140

35
120
30
100
Occurence

Occurence

25
80
20
60
15

40
10

5 20

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 1 2 3 4 5 6 7 8
-4
SVD and statistical features of abnormal PPG Signal Haar wavelet features of abnormal PPG Signal x 10
(b) (b)
Fig. 2. Histogram of SVD and statistical features of normal and abnormal PPG Fig. 3. Histogram of Haar wavelet features in Chi square PDF for 10 degree freedom
signal. of normal and abnormal PPG Signal.

X
Q
pðxjhÞ ¼ ac Nk ðx j lc ; Rc Þ ð7Þ
c¼1

where lc ; Rc and ac are the mean, covariance and mixture compo-

Pk
nents respectively. In addition, c ¼ 1 ac ¼ 1, ac  0 and
 
h ¼ a1 ; l1 ; R1 ; :::::::ac ; lc ; Rc .
Nk represents the D-dimensional Gaussian distribution:
 
1 1
Nk ðXjl; RÞ ¼ exp  ðx  lÞT R1 ðx  lÞ ð8Þ
ð2pÞ jRj
D 1
2 2 2

GMM parameters are determined using Expectation Maximiza-

tion (EM) algorithm by the following steps.
E step: estimate the posterior probability ptic at t iterations as:

at p xi jltc ; Rtc
ptic ¼ PQ c
ð9Þ
c ¼ 1 ac p xi jlc ; Rc
t t t

M step: based on the probabilities from E-step, the parameters

P
ac ; lc ; c are updated at t + 1iteration.

1 XM
atþ1
c ¼ pt ð10Þ
M i¼1 ic Fig. 4. Normal signal.
6 D. Ramachandran et al. / Measurement 150 (2020) 107048

PM
ptic xi 3. Results and discussions
lctþ1 ¼ Pi¼1
M t
ð11Þ
i¼1 pic
To show that the proposed system are effective in medical
applications, the method has been validated using several PPG
PQ
 ltc Þðxi  ltc Þ
T
C¼1 pic ðxi
t recordings with a wide variety of waveform morphologies from
Rctþ1 ¼ PM t ð12Þ IEEE TMBE pulse oximeter benchmark data set, which is obtained
C¼1 pic
from capnobase database. The data set contains recordings of
Repeat these two steps until the parameters stabilize at a speci- raw PPG signal of 8 min duration. The database contains annotated
fic value. Since EM is assured to converge, locally optimal solution respiratory signals such as inhaled and exhaled carbon-dioxide
is always guaranteed. (CO2) also known as capnogram, respiratory flow, and pressure.
Capnobase contains consistent recordings of 13 adults and 29 pedi-
atric cases. The cases comprise observable events like hyper venti-
lation, hypo ventilation, cardiac oscillations, rebreathing,
inadequate ventilation, apnea and changes in respiratory rate
[29,30]. To ensure consistency of the results, total number of sam-
ples of the entire IEEE bench mark dataset from Capnobase data-
base (42 records) has been considered for the experiments. Out
of 42 subjects, 14 subjects are normal and 28 subjects are in risk
of CVD. Figs. 4 and 5 shows normal and abnormal PPG signals for
the record number 36 and 5 respectively.
The wavelets, SVD and statistical feature vectors extracted from
the PPG signals are normalized and the given as input to SDC and
GMM. Classification performance of the SDC and GMM using quan-
tifiers are discussed below.

True Positive (TP): Normal instances correctly identified as

normal
False Negative (FN): Normal instances that are identified as
abnormal
True Negative (TN): Abnormal instances that are identified as
abnormal
False Positive (FP): Abnormal instances that are identified as
Fig. 5. Abnormal signal. normal

Table 6
Performance analysis of SDC for PPG signals.

Learning rate k = 0.5

Feature Extraction Sensitivity (%) Specificity (%) Accuracy (%) Err (%) FA (%) Precision (%) F-measure (%)
Statistical Features + SVD 97.237 99.094 97.880 2.120 0.906 96.568 97.795
db4 91.495 96.854 94.911 5.089 3.146 93.350 93.528
Haar 94.547 98.098 96.465 3.535 1.902 95.191 95.668
Sym8 88.519 95.697 90.081 9.919 4.303 81.599 88.881
k = 0.6
Statistical Features + SVD 97.089 98.965 97.650 2.350 1.035 96.483 97.657
db4 92.471 96.795 94.970 5.030 3.205 94.318 94.338
Haar 94.437 98.067 96.236 3.764 1.933 94.988 95.505
Sym8 88.151 96.130 90.373 9.627 3.870 80.589 89.115
k = 0.7
Statistical Features + SVD 98.043 98.840 97.387 2.613 1.112 97.353 97.300
db4 90.140 96.463 94.415 5.585 3.045 92.199 93.600
Haar 94.014 97.816 95.972 4.029 2.256 94.994 95.198
Sym8 89.100 96.109 90.677 9.323 3.968 80.305 89.500
k = 0.8
Statistical Features + SVD 98.102 98.888 97.585 2.415 1.112 97.572 97.521
db4 91.106 96.955 94.517 5.483 3.045 92.256 93.478
Haar 95.036 97.744 95.971 4.029 2.256 95.994 94.912
Sym8 89.154 96.032 90.399 9.601 3.968 79.394 88.116
k = 0.9
Statistical Features + SVD 97.396 98.504 97.452 2.548 1.497 97.698 97.144
db4 91.040 96.917 94.254 5.746 3.083 92.701 92.947
Haar 93.747 97.656 96.037 3.963 2.344 95.029 94.838
Sym8 90.501 96.620 91.951 8.049 3.380 80.450 89.574
k=1
Statistical Features + SVD 97.491 98.537 97.551 2.449 1.463 97.778 97.250
db4 90.897 96.853 94.220 5.780 3.147 92.592 92.829
Haar 93.842 97.664 96.037 3.963 2.337 94.927 94.819
Sym8 91.063 96.427 92.183 7.817 3.573 81.061 89.509
 D. Ramachandran et al. / Measurement 150 (2020) 107048 7

Based on values of these parameters, Sensitivity, Specificity, 120

Accuracy, Error rate (Err), False Alarm (FA), Precision and F- Statistical
measure are calculated [22,31]. 100 features + SVD
The mathematical formula used to define metrics is defined

Sensitivity (%)
80 db4
below.
TP 60
Sensitivity ðRecallÞ ¼  100 ð13Þ Haar
TP þ FN
40
TN
Specificity ¼  100 ð14Þ 20 Sym8
TN þ FP
0
TN þ TP
Accuracy ¼ TP þ TN þ FP þ FN
 100 ð15Þ N C B A
Risk Levels
FN þ FP (a)
Err ¼  100 ð16Þ
TP þ TN þ FP þ FN
105

FP Statistical
FA ¼  100 ð17Þ 100
TN þ FP features + SVD

Specificity (%)
TP 95
Precision ¼  100 ð18Þ db4
TP þ FP
90
ðRecall  PrecisionÞ
F  measure ¼ 2   100 ð19Þ Haar
ðRecall þ PrecisionÞ 85

Performance of the SDC classifier is tabulated by varying its

80
learning rate ‘k’ from 0.5 to 1 for different feature vectors in Sym8
Table 6.
It is worth noting that when compared to wavelets as feature 75
N C B A
vectors, SVD and statistical features resulted in the high values of
Risk Levels
accuracy for SDC. The values of the parameter ‘k’ of SDC have been
(b)
changed to analyze the effect of it on the performance metric ‘‘ac-
curacy”. Learning rate ‘k’ values are varied from 0.5 to 1 in steps of 120
0.1, the accuracy for statistical features + SVD achieves the maxi-
Statistical
mum value of 97.88% for k = 0.5. The average sensitivity (recall) 100 features + SVD
of 97.24%, specificity by 99.09%, and false alarm by 0.91% is
F-measure (%)

obtained at k = 0.5. For wavelet features also there is only small
change in accuracy values. Therefore, when ‘k’ values are changed
from 0.5 to 1, relatively small change in the accuracy is observed.
Precision is measured to determine the correctly predicted CVD
patients to the total predicted CVDs. Precision is a good measure
to determine, how the non-CVD patients are classified as CVD
patients. Similarly sensitivity (recall) is measured to determine
the correctly predicted CVDs to all observations in the actual
class-CVDs.
Statistical features combined with SVD feature when fed to soft-
max classifier shows high precision of 96.57%. Precision and recall
can be combined to have an overall score that depicts how well the
model is performing. F-measure is the weighted average of preci-
sion and recall. F-measure for SDC (97.8%) is found to have low
false positives and low false negatives, therefore this model is able
to correctly identify the real CVD patients and it is not disturbed by
false alarms. SDC with db4, haar and sym8 feature vectors have a
better accuracy for the learning rate k = 0.6, 0.5 and 1 respectively.
Out of 3 types of wavelets, haar wavelet features achieves better
accuracy and low error rate when applied to SDC. The performance
of the SDC in the classification of cardiovascular risk levels from
PPG signals is shown in Fig. 6. SDC classifies the PPG signals into
normal (N) and abnormal signals. The abnormal signals are further
classified as cardiac of risk level 1 (C), cardiac of risk level 2 (B) and
respiratory disorder (A). Abnormality is measured as an average of
C, B and A.
From the Fig. 6, it is inferred that SDC with statistical features
and SVD classify the PPG signal into N, C, B and A with the high
sensitivity, Specificity and F-measure compared to wavelets. If
there is any variation in the amplitude of the PPG signal, singular
80 db4
60
Haar
40
20 Sym8
0
N C B A
Risk Levels
(c)
Fig. 6. Performance comparison of feature vectors using SDC for various risk levels
of CVD.
values are highly sensitive and they can be used for the prognosis
of CVD risk groups. If the SVD  40, it is identified as normal signal
and if 40 > SVD  50 denotes an abnormality (C). If 50 > SVD  60,
then it is identified as abnormality (B). If the SVD > 60, it indicates
abnormality (A). Performance of SDC is compared with the Gaus-
sian mixture model classifier. Table 7 depicts the GMM based clas-
sification using different feature vectors from the PPG data set for
different variance r values ranging from 0.1 to 1.
Average classification accuracy of 96.64% and F-measure of
96.8% for the SVD combined statistical features is obtained for
r = 0.3 using Gaussian mixture model classifier. Haar wavelet is
also found to have better accuracy of 95%. Fig. 7 shows the perfor-
mance analysis of GMM classifier to identify the various risk
groups of CVD. Also comparative analysis has been made on the
8 D. Ramachandran et al. / Measurement 150 (2020) 107048

Table 7
Performance analysis of GMM classifier for PPG signals.

Variance r = 0.1
Feature Extraction Sensitivity (%) Specificity (%) Accuracy (%) Err (%) FA (%) Precision (%) F-measure (%)
Statistical Features + SVD 70.790 88.221 68.674 31.327 12.095 69.444 71.085
db4 53.140 63.900 43.199 56.801 36.100 46.302 61.144
Haar 53.064 62.165 42.962 57.038 37.836 47.283 60.573
Sym8 52.503 63.869 43.066 56.934 36.132 44.656 61.356
r = 0.2
Statistical Features + SVD 93.673 99.16 96.313 3.687 0.840 94.509 95.130
db4 54.333 69.044 45.630 54.37 30.956 51.262 65.717
Haar 57.852 68.699 45.925 54.075 31.301 53.287 62.799
Sym8 54.333 68.997 45.498 54.502 31.003 55.224 62.452
r = 0.3
Statistical Features + SVD 93.801 99.654 96.642 3.358 0.344 96.379 96.799
db4 58.121 72.699 49.632 50.368 27.301 58.082 69.199
Haar 61.971 75.961 54.806 45.194 24.039 63.650 67.816
Sym8 59.328 77.586 52.820 47.180 22.414 62.979 63.986
r = 0.4
Statistical Features + SVD 91.678 99.300 95.475 4.525 0.721 93.633 95.676
db4 68.597 80.882 65.746 34.254 19.118 68.987 75.472
Haar 71.476 84.459 71.792 28.208 15.541 73.089 77.525
Sym8 71.618 83.965 71.877 28.123 16.035 69.166 77.376
r = 0.5
Statistical Features + SVD 88.310 98.846 92.873 7.123 1.153 89.421 94.646
db4 75.301 88.337 76.102 23.898 11.663 80.275 82.823
Haar 79.192 92.400 83.655 16.345 7.600 83.345 87.521
Sym8 76.167 88.907 78.459 21.541 11.093 79.707 82.033
r = 0.6
Statistical Features + SVD 72.234 98.652 70.822 29.178 15.695 73.312 78.561
db4 84.154 94.906 90.762 9.238 5.094 87.635 91.799
Haar 85.299 96.784 95.073 4.927 3.216 84.655 95.223
Sym8 83.563 95.600 91.970 8.030 4.400 82.178 91.138
r = 0.7
Statistical Features + SVD 81.244 96.445 92.024 7.976 3.599 79.701 90.615
db4 84.794 95.670 93.795 6.205 4.330 91.008 93.724
Haar 82.124 95.931 93.158 6.842 4.069 83.514 93.558
Sym8 65.011 84.580 58.855 41.145 15.420 73.548 68.420
r = 0.8
Statistical Features + SVD 76.437 94.617 86.602 13.398 5.672 74.927 86.873
db4 79.325 93.286 89.353 10.647 6.714 84.692 90.133
Haar 82.319 93.848 85.516 14.484 6.153 86.297 90.285
Sym8 78.389 94.794 90.128 9.872 5.206 82.487 88.895
r = 0.9
Statistical Features + SVD 86.785 95.640 85.829 14.171 6.942 82.294 87.890
db4 74.380 90.704 82.035 17.965 9.296 78.341 84.854
Haar 69.529 86.307 79.398 20.602 13.693 70.073 86.616
Sym8 73.027 91.036 82.914 17.086 8.964 72.692 83.986
r=1
Statistical Features + SVD 87.062 94.766 85.515 14.485 6.904 80.460 87.661
db4 64.941 83.528 72.005 27.995 16.472 68.147 83.793
Haar 60.924 81.752 70.279 29.721 18.248 61.541 80.153
Sym8 66.024 86.008 74.676 25.324 13.992 65.594 79.387

parameters (F-measure and Accuracy) of SDC and GMM using sta-
tistical and SVD features and it is depicted in Fig. 8.
Since better performance is obtained for statistical features
2
+ SVD feature vectors when compared to wavelet features, Fig. 7
depicts the comparative analysis of accuracy and F-measure of
Softmax discriminant classifier and Gaussian mixture model using
statistical and SVD features. From the results it can be inferred that
both the classifiers shows improvement in classification using SVD
and statistical feature vectors in determining the risk levels of car-
diac. Also combination of SDC with SVD and statistical features
vectors may be better pairs in diagnosing risk stages of CVD. Finally
from the research work carried, it can be concluded that these
results are pre non-invasive diagnosis to identify whether to go
in for a detailed diagnosis.
SDC and GMM models are trained using 34 subjects and tested
with 8 subjects. Each model has been trained using wavelets (db4,
haar and sym8), SVD and statistical features. The number of train-
ing set for one subject using different feature vectors are tabulated
in Table 8.
The testing process is monitored by the Mean Square Error
(MSE) which is defined as
MSE ¼ Ti  Oj ð20Þ
where T i and Oj are the target and observed values at time i and j
respectively. Since the results found to be better for SDC with
SVD and statistical features, Table 8 depicts the average training
and testing MSE value of SDC with SVD and statistical feature vec-
tors for different learning rate. As from Table 9, training and testing
MSE is found to be very less for k = 0.5.
The proposed work aims to identify a better SDC GMM system
based on the parameters like learning rate ‘k’ and variance ‘r’. Also
wavelet features with GMM and SDC produces an accuracy of>90%.
By means of SVD based labelling, various risk phases of the CVD
can be identified. One of the limitations of the study is that the
accuracy of SDC depends on the nature of the testing samples. This
is one of the limitations of SDC. Inherently wavelet is unable to
capture the flatness in the statistical features which leads to low
 D. Ramachandran et al. / Measurement 150 (2020) 107048 9

120 100
Statistical
90

Performance Measures (%)

100 features + SVD
80
Sensitivity (%)

80 db4 70
60
60
50 F-Measure
40 Haar 40 Accuracy
30
20
20
0 Sym8
10
N C B A 0
Risk Levels SDC GMM
(a) Fig. 8. Comparative analysis of F-measure and Accuracy of SDC and GMM for
statistical and SVD features.
120
Statistical
100 features + SVD Table 8
Different training set for modeling SDC and GMM classifier.
Specificity (%)

80 db4
Type of features Training dataset per subject
60 SVD and statistical features 7200
Haar db4 9360
40 Haar 5040
Sym8 14,400
20 Sym8
0
N C B A Table 9
Risk Levels Training and testing MSE for SDC.

(b) SDC-learning rate ‘k’ Training MSE Testing MSE

0.5 2.45E-07 2.52E-07
120 0.6 7.73E-06 7.82E-06
0.7 2.05E-05 2.13E-05
100 Statistical features 0.8 8.88E-06 8.10E-06
+ SVD 0.9 4.01E-05 4.24E-05
1.0 9.56E-06 9.72E-06
F-measure (%)

80
60
40
20
0 MLP was very volatile for small size data. Even though pre-
N C B A
Risk Levels
(c)
Fig. 7. Performance comparison of feature vectors using GMM for various risk
levels of CVD.
accuracy using these features. Characteristics of the mixture model
produce low results invariable to underlying features. This is due to
outlier problems in a singular point or cluster centre. Therefore an
intelligent clustering algorithm such as firefly, dragon fly and har-
monic search can be utilized to remove these limitations.
Table 10 represents the summary of existing works reported for
detecting cardiovascular diseases with different PPG database
using various feature extraction techniques and classifiers.
Soltane et al. (2004) collected high dimensional PPG data from
healthy and pathological subjects and the recording time of the
subject is considered to be 60 s. Authors reduce the dimensionality
by performing successive smoothing of the original signal and 14
feature vectors are extracted from the UKM/Malaysian hospitals
database. Here 170 data samples were used for training and testing
and the data samples were testing using MLP and GMM classifica-
db4
Haar tion methods. In their study, MLP neural network has one input
layer, two hidden layers and one output layer. Back-propagation
training algorithm has been used to train the MLP network. The
Sym8 cumulative classification rate achieved is 94.7% for MLP and
91.17% for GMM. Examination of the MLP results showed that
processing algorithms employed to reduce dimension and
extracted features are effective, feature recognition was limited
to small size data.
Hosseini et al. proposed a classification method using PPG to
discriminate between different coronary artery stenosis conditions
using a set of time-domain features. A total number of 48 patients
(16 females and 32 males) for coronary angiography were consid-
ered. Authors used standard occlusion and investigated the flow-
mediated dilation (FMD) test via PPG, using their proposed PPG-
FMD technique. The ECG was also recorded using a 3-lead system.
PPG signals were pre-processed using downsampling, detrending
and left–right AC normalization to minimize the errors and the
time domain features namely amplitude, crest time, stroke volume,
rising velocity, heart rate, pulse transit time are extracted from the
single PPG pulse.
The K-Nearest Neighbor classifier (K-NN) was used to classify
the 48 subjects as high risk and low risk (CLA1 approach). Then
48 subjects were regrouped slightly differently by moving the
single-vessel disease SVD subjects from the low-risk to the first
high-risk class (CLA2 approach). The proposed technique is able
to non-invasively determine patients at a greater risk of CAD (sen-
sitivity 82.0% and negative predictive value 90.0%), and as such can
10 D. Ramachandran et al. / Measurement 150 (2020) 107048
Table 10
Qualitative performance comparison for automated prediction algorithms for CVD.
S.No Authors Classifier used
1. Soltane et al. (2004) Multilayer perceptron Neural network (14 feature vectors)
2. Baid et al. (2008) Support Vector Machine (Auto-Regressive eXogenous input (ARX)
linear parametric model for feature extraction)
3. Hosseini et al. (2015) K-nearest neighbor (Systolic amplitude, PPI, pulse width,
Augumentation Index)
4. Shobitha et al. (2016) Extremely Learning Machine (Crest time, rising velocity and pulse
transit time)
5. Proposed Method Softmax Discriminant Classifier (SVD and statistical feature vectors)
Gaussian Mixture Model Classifier (SVD and statistical feature vectors)
assist the medical caretakers in scheduling them for earlier coro-
nary angiography. But the limitation of the study is the lesser num-
ber of subjects forcing to group subjects with different conditions
into the same classes. Also, risk factors for different CAD groups
subjects in high-risk and low-risk groups for both CLA 1 and CLA
2 classifiers have similar health status.
Shobitha et al. proposed a technique to identify the CVDs risk in
subjects using PPG signals. Authors use ELM for classifying PPG sig-
nals as healthy or at a risk of CVDs. The database contains 30
healthy and 30 pathological signals from both the arms and each
signal is recorded for 90 s. In the pre-processing stage, detrend
and low pass filter are applied to remove baseline drift and high
frequency noise. The features such as systolic amplitude, peak to
peak interval, pulse width, augmentation index are extracted from
the PPG signal and, a-a interval, ratio b/a, ratio c/a, ratio d/a are
derived based on heights of four points in the second derivative
of the PPG signal, called acceleration plethysmogram (APG). The
various combinations of the features are given as input to each
classifier. Kernel-based extreme learning machine (KELM) was
designed to improve the stability of the ELM for non-linear input
data by introducing a positive regularization coefficient. Among
supervised learning algorithms used, ELM gives the best results
with sensitivity of 89.33% and specificity of 90.33%, with only five
input features and minimum computation time. ELM can train
really fast with very slow evaluation. For most applications, evalu-
ation speed is more important than training speed.
The performance of the classification system depends on the
suitable choice of feature vectors and the classifiers. In the pro-
posed method, SVD was found to be the best feature in the risk
level classification of PPG signals. Thus, improved accuracy in
CVD classification is obtained by using SVD with statistical features
in SDC compared to conventional methods.
4. Conclusion
PPG signal is widely utilized in various applications like mea-
suring blood oxygen saturation level, vascular assessment, blood
pressure and heart rate variability etc. The biomedical signals
recorded from the PPG enable the specialists to gain substantial
information about the patient’s medical condition. However, it is
very tedious for doctors to examine all the generated signals man-
ually. Henceforth, it is essential for an automated classifier system
which saves the medical costs and speeds up the doctors’ pre-
surgical assessments as well. The proposed system identifies the
various risk levels of cardiovascular ailment namely normal, car-
diac risk level 1, cardiac risk level 2 and respiratory disorder from
the PPG signal. Compared to multilayer perceptron neural network
as reported in [17], the proposed method accuracy is improved by
3.18%. By choosing the appropriate features from the PPG signals,
the performance of the classifier is enhanced. After analysing the
performance of the classifier it is concluded that SDC outperforms
Accuracy (%) Sensitivity (%) Specificity (%)
94.70% – –
– 84.62% 92.31%
81.50% 82% 80.90%
– 89.33% 90.33%
97.88% 97.24% 99.09%
96.64% 93.8% 99.65%
the GMM classifier in the CVD classification from the PPG signals.
In this research, only time-domain features are considered, how-
ever this work can be scaled up by taking into account the fre-
quency domain features of the PPG for better diagnosis of CVDs.
Also the current work focus only on the general risk level classifi-
cation of CVDs, future focus is to investigate the various risk levels
of coronary artery disease which is one of the most significant
CVDs. PPG signals can be analysed in depth to detect the patients
with coronary artery disease using deep learning based architec-
tures like AlexNet convolution neural network.
Declaration of Competing Interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared
to influence the work reported in this paper.
References
[1] D.I. Uchenna, T.E. Ambakederemo, D.E. Jesuorobo, Awareness of heart disease
prevention among patients attending clinic in Southern Nigeria, Int. J. Preven.
Treat. 1 (2012) 40–43, https://doi.org/10.5923/j.ijpt.20120103.01.
[2] Wan-Hua Lin, Heye Zhang, Yuan-Ting Zhang, Investigation on cardiovascular
risk prediction using physiological parameters, Computat. Mathemat. Method.
Med. (2013) 1–21, https://doi.org/10.1155/2013/272691, Article ID 272691.
[3] J. Wtorek, A. Bujnowski, J. Ruminski, A. Polinski, M. Kaczmarek, A.
Nowakowski, Assessment of cardiovascular risk in assisted living, Metro.
Meas. Syst. 19 (2) (2012) 231–244, https://doi.org/10.2478/v10178-012-0020-0.
[4] E. Miranda, E. Irwansyah, A.Y. Amelga, M.M. Maribondang, M. Salim, Detection
of cardiovascular disease risk’s level for adults using Naive Bayes classifier,
Health Care Inform Res. 22 (3) (2016) 196–205, https://doi.org/10.4258/
hir.2016.22.3.196.
[5] J. Kranjec, S. Begus, G. Gersak, J. Drnovsek, Non-contact heart rate and heart
rate variability measurements: a review, Biomed. Signal Process. Control 13
(2014) 102–112, https://doi.org/10.1016/j.bspc.2014.03.004.
[6] M. Elgendi, On the analysis of fingertip photoplethysmogram signals, Current
Cardio. Rev. 8 (1) (2012) 14–25, https://doi.org/10.2174/
157340312801215782.
[7] Q. Fan, K. Li, Non-contact remote estimation of cardiovascular parameters,
Biomed. Signal Process. Cont. 40 (2018) 192–203, https://doi.org/10.1016/j.
bspc.2017.09.022.
[8] A.S. Al-Fahoum, A. Al-Zaben, A multiple signal classification approach for
photoplethysmography signals in healthy and athletic subjects, Int. J. Biomed.
Eng. Tech 17 (1) (2015) 1–23, https://doi.org/10.1504/IJBET.2015.066962.
[9] J. Abdul Sukor, S.J. Redmond, N.H. Lovell, Signal quality measures for pulse
oximetry through waveform morphology analysis, Physio. Meas. 32 (3) (2011)
369–384, https://doi.org/10.1088/0967-3334/32/3/008.
[10] E. Gill, P.J. Laguna, P. Martinez, O.B. Perez, A.G. Alberola, L. Sornmo, Heart rate
turbulence analysis based on photoplethysmography, IEEE Trans. Biomed. Eng.
60 (11) (2013) 3149–3155, https://doi.org/10.1109/TBME.2013.2270083.
[11] J. Bhattacharya, P.P. Kanjilal, V. Muralidhar, Analysis and characterization of
photo-plethysmographic signal, IEEE Trans. Biomed. Eng. 48 (1) (2001) 5–11,
https://doi.org/10.1109/10.900243.
[12] J.L. Moraes, M.X. Rocha, G.G. Vasconcelos, J.E. Vasconcelos Filho, de
Albuquerque V.H.C. de Albuquerque, A.R. Alexandria, Advances in
photopletysmography signal analysis for biomedical applications, Sensors 18
(6) (2018) 1–26, https://doi.org/10.1007/s00521-018-3801-x.
[13] J. Allen, Photoplethysmography and its application in clinical physiological
measurement, Physiol. Meas. 28 (3) (2007) 1–39, https://doi.org/10.1088/
0967-3334/28/3/R01.
[14] E. Gil, M. Orini, R. Baiĺon, J.M. Vergara, L. Mainardi, P. Laguna,
Photoplethysmography pulse rate variability as a surrogate measurement of
 D. Ramachandran et al. / Measurement 150 (2020) 107048
heart rate variability during non-stationary conditions, Physiol. Meas. 31 (9)
(2010) 1271–1290, https://doi.org/10.1088/0967-3334/31/9/015.
[15] K. Takazawa, N. Tanaka, M. Fujita, O. Matsuoka, T. Saiki, M. Aikawa, S. Tamura,
C. Ibukiyama, Assessment of vasoactive agents and vascular aging by the
second derivative of photoplethysmogram waveform, Hypertension 32 (2)
(1998) 365–370.
[16] R. Gonzalez, A. Delgado, A.J.- Padilla, B. Trenor, J. Ferrero, J. Saiz,
Photoplethysmographic augmentation index as a non invasive indicator for
vascular assessments, Proceed. Int. Federat. Med. Biol. Eng. Springer 22 (2008)
1167–1170.
[17] M. Soltane, M. Ismail, Z.D. Abdul Rashid, Artificial Neural Networks (ANN)
approach to PPG signals classification, Int. J. Comp. Inform. Sci. 2 (1) (2004)
58–65.
[18] R. Baid, N. Krupa, A.M. Muhammad, Classification of photoplethysmographic
signals using support vector machines for vascular risk assessment, Proceed.
IASTED Biomed. Eng. Conf. 791 (2013) 183–187.
[19] L.F. Polanıa, L.K. Mestha, D.T. Huang, J.P. Couderc, Method for classifying
cardiac arrhythmias using photoplethysmography, Proceed. IEEE 37 (2015)
6574–6577.
[20] S. Shobitha, R. Sandhya, B. Niranjana Krupa, M.A. Mohd Ali, Recognizing
cardiovascular risk from photoplethysmogram signals using ELM, Proceed.
IEEE 2 (2016) 1–5.
[21] Z.S. Hosseini, E. Zahedi, A.H. Movahedian, H. Fakhrzadeh, M.H. Parsafar,
Discrimination between normal-to-mild and severe coronary artery disease
using time-domain features of the finger photoplethysmogram response to
reactive hyperemia, Biomed Sig. Proc. Cont. 18 (2) (2015) 282–292, https://doi.
org/10.1016/j.bspc.2014.12.011.
[22] R. Harikumar, B. Vinoth Kumar, Performance analysis of neural networks for
classification of medical images with wavelets as a feature extractor, Int. J.
11
Imaging Syst. and Tech. 25 (1) (2015) 33–40, https://doi.org/10.1002/
ima.22118.
[23] K.S. Maedeh, S. Asadollah, Classification of ECG arrhythmias using discrete
wavelet transform and neural networks, Int. J. Compt. Sci. Eng. Appl. 2 (1)
(2012) 1–13, https://doi.org/10.5121/ijcsea.2012.2101.
[24] S.P. Singh, S. Urooj, Wavelets: biomedical applications, Int. J. Biomed. Eng.
Tech. 19 (1) (2015) 1–25, https://doi.org/10.1504/IJBET.2015.071405.
[25] I. Daubechies, Orthonormal bases of compactly supported wavelets, Comm.
Pure Appl. Math. 41 (7) (1988) 909–996, https://doi.org/10.1002/
cpa.3160410705.
[26] Jalal Karam, Theory and practice of wavelets in signal processing, IJMISSP 1 (2)
(2014) 174, https://doi.org/10.1504/IJMISSP.2014.065696.
[27] F. Zang, F.J.S. Zhang, Softmax discriminant classifier, Proceed. IEEE 3 (2011)
16–20.
[28] R. Li, R. Perneczky, I. Yakushev, S. Förster, A. Kurz, A. Drzezga, S. Kramer,
Gaussian mixture models and model selection for [18f] fluorodeoxyglucose
positron emission tomography classification in alzheimer’s disease, PLoS ONE
10 (4) (2015) 1–22, https://doi.org/10.1371/journal.pone.0122731.
[29] W. Karlen, M. Turner, E. Cook, G.A. Dumont, J.M. Ansermino, CapnoBase: Signal
database and tools to collect, share and annotate respiratory signals, Annual
Meeting of the Society for Technology in Anesthesia (STA) Conference, 2010),
25.
[30] W. Karlen, S. Raman, J.M. Ansermino, G.A. Dumont, Multiparameter
respiratory rate estimation from the photoplethysmogram, IEEE Trans.
Biomed Eng. 60 (7) (2013) 1946–1953, https://doi.org/10.1109/
TBME.2013.2246160.
[31] A.P. Bradley, The use of the area under the roc curve in the evaluation of
machine learning algorithms, Pattern Recog. 30 (7) (1997) 1145–1159.