Comparing Different Settings of Parameters Needed for Pre-processing
of ECG Signals used for Blood Pressure Classification
Monika Simjanoska
1
, Gregor Papa
2
, Barbara Korou
ˇ
si
´
c Seljak
2
and Tome Eftimov
2
1
Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University,
Rugjer Boshkovikj 16, 1000 Skopje, Macedonia
2
Computer Systems Department, Jo
ˇ
zef Stefan Institute, Jamova cesta 39, 1000 Ljubljana, Slovenia
Keywords:
Biomedical Signal Analysis, ECG, Signal Length, Baseline Removal, Blood Pressure Classification.
Abstract:
Because a real-time monitoring using electrocardiogram (ECG) signals is a challenging task, the pre-
processing techniques used for ECG signal analysis are crucial for obtaining information that is further used
for some more complex analysis, such as predictive analyses. We compared different settings of parameters
needed for pre-processing of ECG signals in order to estimate the valuable information that can be further
used for blood pressure classification. Two parameters were involved in the comparison: i) the signal length
used for ECG segmentation; and ii) the cut-off frequency used for baseline removal. The first parameter is the
parameter used for obtaining ECG segments that are further used, and the second one is the frequency used for
baseline removal. Thirty different combinations, each a combination of a signal length and a cut-off frequency,
were evaluated using a dataset that contains data from five commercially available ECG sensors. For signal
lengths: 10 s, 20 s, and 30 s, were used for data segmentation, while the cut-off frequency for baseline removal
starts from 0.05 Hz, till 0.50 Hz, with a step length of 0.05 Hz. The evaluation of these combinations was done
in combination with complexity analysis used for features extraction that are further used for blood pressure
classification. Experimental results, obtained using a data-driven approach by comparing the combinations
using the results obtained from the classification for 17 performance measures, showed that a signal length
of 30 s carries the most information in a combination with cut-off frequency between 0.10 Hz and 0.20 Hz.
Results contribute to the arguments published in the literature discussing the optimal ECG sample lengths
needed for building predictive models, as well as the lower frequencies where the ECG components overlap
with the baseline wander noise.
1 INTRODUCTION
A biomedical processing system includes a biological
system of interest, sensors used to follow the physio-
logical conditions, and a methodology for biomedical
signal analysis in order to extract the useful informa-
tion from the system. In this paper, the biological sys-
tem addressed is the heart system, and its electrical
activity is followed using electrocardiogram (ECG)
signals. Through ECG we want to follow blood pres-
sure. Two phases for managing blood flow exist: i) a
diastole phase known as the filling phase and ii) a sys-
tole phase known as the pumping phase. The blood
pressure (BP) is defined as the force of the blood
pushing against the walls of the arteries as the heart
pumps blood (Shriram et al., 2010).
Nowadays, commercially available wearable bio-
sensors (e.g., ECG sensors, sweating-rate sensors,
respiration-rate body sensors) provide an opportunity
for real-time monitoring of human vital signs, which
further helps the process of preventive, timely notifi-
cation and real-time diagnosis (Cosoli et al., 2015).
Recently, many systems for non-invasive BP moni-
toring have been developed, such as: i) the Superfi-
cial Temporal Artery Tonometry-based device (Sackl-
Pietsch, 2010), ii) the PPG optical sensor (Canning
et al., 2016), iii) ARTSENS (ARTerial Stiffness Eval-
uation for Non-invasive Screening) for brachial ar-
terial pressure (Mouradian et al., 2015), iv) an elec-
tronic system based on the oscillometric method (Sa-
hani et al., 2014), v) a BP estimation device based
on the principle of volume compensation (Marani
and Perri, 2012), vi) a Modulated Magnetic Signa-
ture of Blood mechanism (Tanaka et al., 2007), and
vii) portable equipment that includes a cuff-based BP
sensing system (Li et al., 2013). However, they are
specialized only for BP measurements, and exclude
62
Simjanoska, M., Papa, G., Seljak, B. and Eftimov, T.
Comparing Different Settings of Parameters Needed for Pre-processing of ECG Signals used for Blood Pressure Classification.
DOI: 10.5220/0007390100620072
In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 62-72
ISBN: 978-989-758-353-7
Copyright
c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
the other vital signs.
The data collected from sensors that measure ECG
signal as physiological signal can be further used
for developing smart phone applications for real-
time diagnosis (Cosoli et al., 2015). In (Simjanoska
et al., 2018), the relation between the ECG signal
and BP using supervised machine learning (ML) ap-
proaches has been investigated. For this reason, a pre-
processing that involves ECG signal complexity anal-
ysis was applied in order to extract features that are
further used for training a meta-model used for blood
pressure classification. The ECG signal analysis was
performed using a signal length of 30 s and a band-
pass Butterworth filter with a frequency of 0.30 Hz for
removing the baseline wander without deforming the
ECG signal. These parameters (30 s; 0.03 Hz) were
chosen following the state-of-the-art literature. Using
such setting, promising results were achieved that in-
dicate the relation between the complexity features of
the ECG signals and the BP.
However, in real-life scenarios (e.g., civil and mil-
itary emergency situations) there is no guaranty that
there will be enough time to measure long ECG sam-
ples. For this reason, in this study we focus on com-
parison of different combinations of parameters for
signal length and frequency used for baseline removal
(i.e. different ECG signal analysis) and their influence
on the end result of the methodology used for BP clas-
sification. We consider signal lengths of 10 s, 20 s,
and 30 s, and frequency starting from 0.05 Hz, till
0.50 Hz, with a step length of 0.05 Hz. The novelty
is also in the evaluation of the influence, in which we
did not use only one performance measure, but a set
of performance measures with a data-driven approach
in order to make a general conclusion.
In the remainder of the paper, we first present an
overview of the related work. Then we present the
methodology used for comparison of the parameters
values used for pre-processing of ECG signals is ex-
plained in detail, followed by the experimental results
and discussion. Finally, the conclusions of the paper
are presented followed by directions for future work.
2 RELATED WORK
Most of the recently published studies on BP esti-
mation use a combination of ECG and photoplethys-
mogram (PPG) sensors (Sahoo et al., 2011; Ilango
and Sridhar, 2014; Thomas et al., 2016). The com-
mon techniques for obtaining measurements on BP
mainly rely on Pulse Wave Velocity (PWV), Pulse Ar-
rival Time (PAT), and Pulse Transit Time (PTT) (Choi
et al., 2013; Goli and Jayanthi, 2014; Nye et al., 2015;
Seo et al., 2015), all of which require an accurate
PPG measurement, which is not simple to get unob-
trusively yet, and no clear proof on the PPG measure-
ment’s relation to the BP has been provided (Payne
et al., 2006; Wong et al., 2009; i Car
´
os, 2011). The
relationship between the ECG signal and BP estima-
tion has been discussed in (Chan et al., 2001; Ah-
mad et al., 2012). Both methods used an additional
PPG sensor with an ECG sensor. The results from
several studies that address the ECG-BP relationship
(Schroeder et al., 2003; Hassan et al., 2008) confirm
that no strong relationship exists between hyperten-
sion occurrence and morphological changes in ECG.
For this reason, in (Simjanoska et al., 2018), instead
of using morphological changes in ECG, complexity
analysis was used to detect bio-system complexity.
For this reason, the ECG signals were segmented us-
ing signal length of 30 s, for which cut-off frequency
of 0.30 Hz was applied for baseline removal.
3 METHODOLOGY
To compare different combinations of parameters
used for pre-processing of ECG signals, we propose
an approach, which goes through four different lev-
els: data segmentation and baseline removal, feature
extraction, classification, and classifiers ranking.
Level 1: The key part is the data segmentation
and baseline removal, because the end result
is comparing different parameters that are used
for obtaining ECG segments that are further used
for blood pressure classification. For this reason,
30 different combinations were included, a sig-
nal length of 10 s, 20 s, and 30 s, and frequency
for baseline removal starting from 0.05 Hz, till
0.50 Hz, with a step length of 0.05 Hz.
Level 2: To compare the different combinations,
for each of them the same complexity analysis
was applied for feature extraction.
Level 3: The extracted features for each combi-
nation are further used to learn a classification
model for blood pressure following the stacking
procedure.
Level 4: Finally, by performing classifiers rank-
ing using a set of different performance measures,
the influence of each combination can be esti-
mated.
The flowchart of the methodology is presented in Fig-
ure 1. Further, each level is explained in more detail.
Comparing Different Settings of Parameters Needed for Pre-processing of ECG Signals used for Blood Pressure Classification
63
Raw ECG signal
Data
Segmentation
Combination 1 Combination n
. . .
Feature
Extraction
Data set 1 Data set n
Classification
. . .
Classifier 1 Classifier n
Classifiers
Ranking
Best
Combination
Level 2
Level 1
Level 3
Level 4
Figure 1: The flowchart of the methodology.
3.1 Data Segmentation and Baseline
Removal
The raw ECG signals are segmented using signal
lengths of 10 s, 20 s, and 30 s. In the process of seg-
mentation, the segments were also labeled with the
appropriate BP class. The result of this are three data
sets, the first consists of segments of 10 s, the sec-
ond consists of segments of 20 s, and the third of seg-
ments of 30 s. Further, a band-pass filter is applied on
each of these datasets in order to preserve only valid
ECG information. The cut-off frequency for base-
line removal starts from 0.05 Hz, till 0.50 Hz, with
a step of 0.05 Hz. Performing 10 cut-off frequencies
on each dataset, it results in 30 data sets, which were
further called “combinations” because each of them
is described with a combination of signal length and
cut-off frequency.
3.2 Feature Extraction
After creating 30 datasets (combinations), each of
them is used for feature extraction, in which the prob-
lem is addressed from the perspective of complex-
ity analysis (Eke et al., 2002; Morabito et al., 2012;
McBride et al., 2014), and not as other traditional
approaches in which morphogical features are used
(Wong et al., 2009; Zhang and Zhang, 2006; Nitzan,
2011). For this reason, for each dataset, five com-
plexity metrics are calculated: signal mobility, signal
complexity, fractal dimension, autocorrelation, and
entropy.
3.2.1 Signal Mobility
Given x
i
, i = 1,. . . ,N is the ECG signal of length N
and d
j
= x
j+1
x
j
is the first-order variation in the
signal, then the first-order factors, S
0
and S
1
, are cal-
culated as:
S
0
=
s
N
i=1
x
2
i
N
(1)
S
1
=
s
N1
j=2
d
2
j
N 1
(2)
The signal mobility quantitatively measures the
level of variation in the signal. It is calculated as a
ratio between the factors S
1
and S
0
:
Mobility =
S
1
S
0
, (3)
3.2.2 Signal Complexity
Given the first-order variation of the ECG signal d
j
,
j = 1, . . . ,N 1, the second-order variation of the sig-
nal is presented by g
k
= d
k+1
d
k
. Then, the second-
order factor is calculated as:
S
2
=
s
N2
k=3
g
2
k
N 2
, (4)
Both the signal mobility and signal complex-
ity were computed by using the Hjorth parameters
method (Kugiumtzis and Tsimpiris, 2010).
BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing
64
3.2.3 Fractal Dimension
Fractal dimension measures the self-similarity of the
signal. By zooming and comparing different portions,
it describes fundamental patterns hidden in the signal.
Higuchi algorithm (Monge-
´
Alvarez, 2015) has been
used to calculate the fractal dimension. The method
works with a set of k subseries with different resolu-
tions, creating a new time series X
k
, for m = 1,. . . , k:
X
m
k
: x(m),x(m + k),x(m + 2k),...,x(m + b
N m
k
ck)
(5)
The length of the curve X
m
k
, l(k) is calculated as:
l(k) =
(
bNm/kc
i=1
|x(m + ik) x(m + (i 1)k)|(N 1))
(b
Nm
k
c)k
(6)
Then, for each k in range 1 to k
max
, the average
length is calculated as the mean of the k lengths l(k)
for m = 1, ...k. The fractal dimension is the estimation
of the slope of the plot ln(l(k)) vs. ln(1/k).
3.2.4 Autocorrelation
The similarity between the signal and its shifted ver-
sion is measured via autocorrelation. Let τ be the
amount of shift, and then the autocorrelation is cal-
culated as:
r
xx
(τ) =
Z
+inf
inf
x(t)x(t τ)p
xx
(x(t),x(t τ))dt, (7)
where p
xx
(x(t),x(t τ)) presents the joint probability
density function of x(t) and x(t τ).
3.2.5 Entropy
Entropy expresses the randomness of the signal. The
decrease of entropy often indicates a disease (Zhou
et al., 2017). The amount of information is expressed
through the concept of probability. Let p
i
denote the
probability of each outcome x
i
within the ECG signal
X for i = 1,. . . ,N 1. Then, entropy is calculated as:
Entropy =
N1
i=0
p
i
log(
1
p
i
) (8)
3.3 Classification
Performing the feature extraction for each of the 30
datasets (combinations), each of them is used to train
a meta-model that will be used to predict the BP class.
The performance of the meta-model on each data set
is used to compare different signal lengths and cut-
off frequencies. The methodology that is applied is
taken from a previous study (Simjanoska et al., 2018),
where the meta-model was learned for a signal length
of 30 s and cut-off frequency of 0.30 Hz. Using this
combination, it was shown that promising results can
be achieved.
The meta-model follows a stacking design, in
which seven different classifiers are used: Bagging
(Breiman, 1996), Boosting (Freund et al., 1996),
SVM (Hearst et al., 1998), K-means (Liao and Ve-
muri, 2002), Random Forest (Liaw et al., 2002),
Naive Bayes (Rish et al., 2001), and J48 (Patil and
Sherekar, 2013). The probabilities from each of them
for all the feature vectors to belong in each class, com-
prise a new feature vector upon which a META clas-
sifier (Random Forest) is trained. More details about
the classification methodology are provided in (Sim-
janoska et al., 2018).
3.4 Ranking of Classifiers
For each dataset (combination), a meta-model is
trained using a training set and it is evaluated on a
corresponding validation set. For this reason, each
dataset is m-times split into training and validation
set, since we have multiple independent measure-
ment for each subject (i.e. participant) and as the
number of measurements varies for each subject.
We follow the rule that if a subject is included in one
set, none of its measurements may occur in another
set. So for each combination, m meta-models are
trained.
To evaluate the performance of the meta-models
trained for each combination of signal length and cut-
off frequency, we used an ensemble (a fusion) of a set
of 17 performance metrics:
Accuracy (ACC) - represents the fraction be-
tween the number of correct predictions by the to-
tal number of predictions made.
Cohen’s Kappa - compares the observed accu-
racy (the number of correctly classified instances)
with the expected accuracy (taking into account
the number of instances in each class, along with
the number of instances that the classifier agreed
with the ground truth label).
Precision (PR) - captures the effect of the large
number of negative examples on the classifier’s
performance, by comparing false positives (FP)
to true positives (TP) rather than true negatives
(TN), i.e. it measures how many of the predicted
instances were TP.
Recall (RC) - to measure how many of the TP
instances were predicted.
Comparing Different Settings of Parameters Needed for Pre-processing of ECG Signals used for Blood Pressure Classification
65
F-Measure - measures the trade-off between re-
call and precision giving equal importance to re-
call and precision.
Area Under PRC - precision-recall curve does
not account for TN, since TN is not a component
of either Precision or Recall. Given there are more
negatives (normal) than positives (hypertension),
PRC might be a very informative evaluation met-
ric of the performance of the classifier. Having
a model at the upper right corner, means that the
classifier gets only the TP with no FP and no false
negatives (FN) at all and thus is a perfect classi-
fier.
Area Under ROC - the ROC curve plots the
True Positive Rate (TPR) vs. False Positive Rate
(FPR). Therefore achieving a model at the upper
left corner, means that the classifiers is getting no
FP at all and thus is a perfect classifier.
Matthews Correlation (COR) - takes the advan-
tage from all four metrics TP, FP, TN and FN to
calculate the correlation coefficient between the
observed and predicted classifications (Vihinen,
2012).
Relative Absolute Error (RAE) - is an error
measure relative to a simple predictor, meaning
it takes the total absolute error and normalizes it
by dividing it by the one of the simple predictor.
Root Relative Squared Error (RRSE) - takes
the total squared error and normalizes it by divid-
ing by the total squared error of the simple predic-
tor. The square root reduces the error to the same
dimensions as the problem at hand.
Root Mean Squared Error (RMSE) - explains
the standard deviation of the prediction error.
Informedness (INF) - quantifies how informed a
classifier is for the specified class. It provides in-
sight into how consistently the classifier predicts
the class (Powers, 2011).
Markedness (MAR) - quantifies how marked a
class is for the specified classifier, i.e. how con-
sistently the class has the classifier as a marker by
combining measures about correct classifications
(Powers, 2011).
Micro F-measure (MF) - aggregates the contri-
butions of all classes to compute the average F-
measure by considering the total TPs, FNs and
FPs.
Table 1: Decision matrix.
q
1
q
2
. . . q
17
C
1
q
1
(C
1
) q
2
(C
1
) . . . q
17
(C
1
)
C
2
q
1
(C
2
) q
2
(C
2
) . . . q
17
(C
2
)
.
.
.
.
.
.
.
.
.
.
.
.
C
30
q
1
(C
30
) q
2
(C
30
) . . . q
17
(C
30
)
Log Likelihood (LL) - presents the probability of
the observed prediction given the real class.
Mutual Information (MI) - measures whether
the real and the predicted labels are statistically
depended.
Pearson’s Chi-squared Test (PRS) - to measure
whether the observed and expected frequencies
are the same by comparing the predicted contin-
gency table to an expected table.
To compare the results of the meta-models ob-
tained for each combination using a set of perfor-
mance measures, their results are organized into a
decision matrix (Table 1). The rows of the deci-
sion matrix correspond to different meta-models and
the columns correspond to the values obtained for
the performance measures. First, a generalized pref-
erence function should be selected for each perfor-
mance measure. In our case, the V -shape general-
ized preference function is used for each performance
measure, where the threshold of strict preference is
set to the maximum difference that exists for each
preference measure from all pairwise comparisons
according to that performance measure (Brans and
Mareschal, 2005). After that, the average preference
index for each pair of meta-models should be calcu-
lated, which gives information of global comparison
between them using all performance measures. To
rank the meta-models, a net flow for each one needs
to be calculated. It is a difference between a posi-
tive preference flow and a negative preference flow of
the meta-model. The positive preference flow gives
information about how a given meta-model is glob-
ally better than the other meta-models, while the neg-
ative preference flow gives the information about how
a given meta-model is outranked by all the other meta-
models. This approach has already been used for
evaluation of multi-objective meta-heuristic stochas-
tic optimization algorithms regarding a set of perfor-
mance measures. More details about the ranking ap-
proach and the equations for the net, positive, and
negative flow, can be found in (Eftimov et al., 2018).
BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing
66
4 EVALUATION
4.1 Data Collection from Multiple
Sensors
The data was collected using five ECG sensors, from
which four are commercially available ECG sensors
(Hacks, 2015; Biosignals, 2016; Technology, 2017;
Trobec et al., 2018a) and one is from the online avail-
able Physionet database (Goldberger et al., 2000). For
each ECG signal, reference SBP and DBP values are
attached, which are measured in parallel using elec-
tronic sphygmomanometer. The information about
each sensor, its reliability, the number of participants
measured by each sensor, and their age, is presented
in Table 2. All the human subjects involved in the re-
search signed an agreement allowing their ECG data
to be used for research goals. The involved partici-
pants come from different age and health status, and
the measurements were performed in moving and sit-
ting positions. The patients included from the Phy-
sionet database are explicitly selected to suffer brain
injuries.
As we mentioned before, each ECG signal is ac-
companied with SBP and DBP values, we also added
the BP class, which is the target in the classification.
To obtain the BP class for each ECG measurement, a
publicly available scheme presented in Table 3 (Pro-
gram et al., 2004; AHA, 2016) was used. The BP
classes seem not to be mutually exclusive, however
this was solved by giving priority to the more se-
vere BP conditions by checking the AND” condi-
tions at the end of the ECG samples mapping proce-
dure. To solve the imbalanced-class data, BP classes
were grouped into three main categories: hypotension
(HPTN) and normal (N) as Normal class (denoted as
0); prehypertension (PHTN) as Prehypertension class
(denoted as 1); and stage 1 hypertension (S1HTN),
stage 2 hypertension (S2HTN), isolated systolic hy-
pertension (ISHTN), and hypertensive crisis (HTNC)
as Hypertension class (denoted as 2).
4.2 Data Segmentation and Baseline
Removal
After the data from all sensors was collected and la-
beled with an appropriate BP class, data segmentation
was applied on each ECG signal. The raw ECG sig-
nals were segmented using signal lengths of 10 s, 20 s,
and 30 s, and each of the signal length was analyzed
with different cut-off frequency for baseline removal
starting from 0.05 Hz, till 0.50 Hz, with a step of
0.05 Hz. The segments were labeled with the BP class
from the corresponding ECG signal from which they
are obtained. Finally, we create 30 different combina-
tions that should be analyzed in the case of BP clas-
sification in order to provide information about which
pre-processing combination gives the most promising
results.
4.3 Feature Extraction
For each of 30 datasets (i.e. combinations), we ex-
tracted the five complexity features that will be further
used for learning a meta-model used for BP classifica-
tion. Each of the datasets is described with the same
five features, however the features values are different
due to different signal length and cut-off frequency for
baseline removal that are used.
4.4 Classification
Having the 30 datasets, for each one the classifica-
tion was done using a previously published methodol-
ogy that follows a stacking design (Simjanoska et al.,
2018). Output probabilities from seven different clas-
sifiers, already mentioned above, were used as fea-
tures to train a meta-model for multi-class BP classi-
fication using Random Forest. In our case, the number
of classes is three.
4.5 Ranking of Classifiers
To evaluate the performance of each meta-model ob-
tained for each dataset (combinations), we follow the
idea of splitting the datatsets into training and valida-
tion set. The splitting is made in a way that if a subject
is included in the training set, none of its measure-
ments may occur in the validation set, and vice versa.
For this reason and in order to build a robust model
for each combination, each dataset was split 30 times
into training and validation sets. We followed the rule
of 75% of the subjects were used for training and 25%
of the subjects were used for validation.
Using this approach, for each combination we ob-
tained 30 different meta-models and their evaluation
on the corresponding validation sets described by the
17 performance measures. The evaluation was per-
formed in two scenarios.
First, within each configuration, we used the rank-
ing scheme based on PROMOTHEE method in or-
der to find the best split for each combination, which
results in the best meta-model for each combina-
tion. Further, the best meta-models for all combi-
nations were again evaluated with the same ranking
scheme and the set of 17 performance measures in
order to obtain the most promising combination of
Comparing Different Settings of Parameters Needed for Pre-processing of ECG Signals used for Blood Pressure Classification
67
Table 2: Sensors summary information.
Dataset Reliability Participants Age
Cooking hacks (Hacks, 2015) (Winderbank-Scott and Barnaghi, 2017) 16 16–72
180° eMotion FAROS (Biosignals, 2016) (Ahonen et al., 2016) 3 25–27
Zephyr Bioharness module (Technology,
2017)
(Johnstone et al., 2012) 14 20–73
Savvy sensor platform (Trobec et al.,
2018a)
(Trobec et al., 2018b) 21 15–54
Charis Physionet database (Goldberger
et al., 2000)
(Kim et al., 2016) 7 20–74
Table 3: Rules and categorization.
Joined class Class SBP (mmHg) Logical DBP (mmHg)
Normal
HPTN <=90 OR <=60
N 90–119 AND 60–79
Prehypertension PHTN 120–139 OR 80–89
Hypertension
S1HTN 140–159 OR 90–99
S2HTN >=160 OR >=100
ISHTN >=140 AND <90
HTNC >=180 OR >=110
pre-processing that was applied. The performance
of the best meta-model within each configuration are
presented in Table 4. The first column presents the
parameters settings for each combination. The last
column reported the ranking for each combination,
which is obtained by the ranking scheme when the
best meta-models from all combinations are com-
pared.
Using the information reported in Table 4, we can
see that the best results are obtained using the meta-
model for the combination with the signal length
of 10 s and cut-off frequency for baseline removal
0.30 Hz, the second best is obtained for the signal
length of 20 s and the cut-off frequency of 0.45 Hz,
and the third with the signal length of 30 s and the
cut-off frequency of 0.15 Hz. From here, there is no
general conclusion for which data segmentation pro-
vides the most valuable information for BP classifi-
cation; however there is a question of the robustness
of the results, or if selecting the best meta-model (or
best splitting) for each combination corresponds to a
real-life application.
Second, in order to have more robust results, the
results from 30 splits for each combination were ag-
gregated by averaging the results for each perfor-
mance measure separately, and the average value for
each performance measure was further used for each
combination. This was made with the purpose of fol-
lowing the cross-validation idea. Then, the 30 com-
binations were compared using the ranking scheme
and the set of 17 performance measures. The aver-
aged performance for each configuration is presented
in Table 5. The first column presents the parameters
settings for each combination. The last column re-
ported the ranking for each combination, which is ob-
tained by the ranking scheme when all combinations
are compared using the averaged results.
Using the results reported in Table 5, it follows
that the best meta-model can be learned using the sig-
nal length of 30 s and cut-off frequency of 0.10 Hz,
the second best can be learned using the signal length
of 30 s and the cut-off frequency of 0.15 Hz, and the
third with the signal length of 30 s and the cut-off fre-
quency of 0.20 Hz. We can also concluded that the 10
best combinations are obtained when the signal length
is set at 30 s, so it carries the most valuable informa-
tion. Depending on the cut-off frequency that is used,
best models can be learned when it is between 0.10 Hz
and 0.20 Hz. As we mentioned before, in real-life sce-
narios (e.g., civil and military emergency situations),
maybe there is no time to obtain a signal of 30 s, so if
we have a signal length of 10 s, using the results re-
ported in Table 5, we can select the best combination
when the signal length is 10 s, and this combination
is obtained with a cut-off frequency of 0.25 Hz.
We need to mention that the focus here is made
on different settings of parameters for pre-processing
of ECG signals used for blood pressure classification.
Regarding the classification methodology, it could be
also tested with some other methodologies; we se-
lected one that was recently published. The novelty
in the comparison is that we used a set of 17 per-
formance measures instead of focusing of compar-
ing them separately to the most commonly used per-
formance measures in which a general conclusion is
made. The performance measures were selected man-
ually.
5 CONCLUSIONS
We investigated different parameter values that can be
used for pre-processing of ECG signals in order to
estimate the valuable information that can be further
used for blood pressure classification. One parameter
is related to the signal length used for data segmen-
tation of the raw ECG signals and the other is related
to the cut-off frequency that is applied on each seg-
BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing
68
Table 4: Performance results for best meta-model trained for each combination.
Comb. ACC KAPPA PRC ROC F COR PR RC RAE RRSE RMSE INF MAR MF LL MI PRS Ranking
10/0.05 55.16 0.33 0.54 0.71 0.53 0.32 0.54 0.55 68.23 108.14 0.51 0.31 0.37 0.55 56.59 0.18 96.75 14
10/0.10 46.99 0.24 0.53 0.68 0.45 0.25 0.54 0.47 77.09 109.25 0.54 0.27 0.29 0.47 32.67 0.12 58.12 27
10/0.15 55.30 0.30 0.53 0.70 0.55 0.33 0.56 0.55 66.71 104.72 0.50 0.32 0.34 0.55 40.72 0.19 61.14 16
10/0.20 51.46 0.22 0.51 0.67 0.49 0.23 0.50 0.51 73.99 110.49 0.52 0.24 0.28 0.51 19.38 0.09 29.71 28
10/0.25 54.01 0.31 0.52 0.69 0.54 0.31 0.55 0.54 71.47 110.71 0.52 0.32 0.30 0.54 33.26 0.14 57.64 22
10/0.30 63.14 0.44 0.62 0.78 0.64 0.47 0.68 0.63 58.84 95.76 0.46 0.47 0.45 0.63 52.12 0.22 93.10 1
10/0.35 56.38 0.28 0.60 0.72 0.56 0.32 0.56 0.56 67.75 101.45 0.49 0.32 0.32 0.56 25.18 0.13 40.82 18
10/0.40 50.19 0.24 0.48 0.64 0.50 0.24 0.51 0.50 80.58 116.81 0.55 0.23 0.27 0.50 26.72 0.10 45.47 30
10/0.45 52.68 0.30 0.48 0.66 0.51 0.28 0.52 0.53 75.96 112.93 0.53 0.27 0.34 0.53 32.58 0.15 54.57 24
10/0.50 52.68 0.29 0.49 0.67 0.50 0.28 0.50 0.53 73.16 112.33 0.53 0.27 0.34 0.53 29.21 0.13 49.54 25
20/0.05 56.88 0.35 0.55 0.73 0.54 0.34 0.53 0.57 63.21 106.86 0.51 0.33 0.42 0.57 45.39 0.21 73.83 11
20/0.10 56.07 0.34 0.55 0.72 0.52 0.33 0.53 0.56 69.25 106.40 0.50 0.33 0.42 0.56 48.09 0.20 77.33 12
20/0.15 57.24 0.24 0.57 0.65 0.57 0.23 0.58 0.57 66.19 103.87 0.50 0.23 0.24 0.57 11.33 0.07 20.14 23
20/0.20 57.14 0.32 0.55 0.70 0.57 0.33 0.57 0.57 67.37 104.82 0.50 0.34 0.34 0.57 29.84 0.15 47.90 15
20/0.25 51.64 0.25 0.51 0.66 0.51 0.25 0.52 0.52 75.33 109.22 0.51 0.26 0.23 0.52 26.49 0.11 44.91 26
20/0.30 55.56 0.28 0.58 0.70 0.56 0.30 0.57 0.56 66.80 101.72 0.48 0.30 0.29 0.56 12.89 0.08 20.42 21
20/0.35 65.56 0.37 0.63 0.73 0.65 0.39 0.65 0.66 56.85 94.44 0.46 0.39 0.39 0.66 19.70 0.13 34.84 5
20/0.40 51.03 0.24 0.45 0.61 0.51 0.23 0.51 0.51 78.26 115.49 0.55 0.24 0.22 0.51 39.74 0.10 68.30 29
20/0.45 61.03 0.41 0.58 0.76 0.59 0.41 0.59 0.61 65.11 93.24 0.44 0.41 0.45 0.61 53.41 0.25 86.30 2
20/0.50 54.87 0.33 0.51 0.67 0.55 0.33 0.56 0.55 77.03 103.46 0.49 0.34 0.35 0.55 35.22 0.16 60.30 20
30/0.05 58.64 0.36 0.58 0.75 0.57 0.37 0.56 0.59 61.19 101.95 0.49 0.36 0.40 0.59 70.00 0.24 115.57 7
30/0.10 54.15 0.32 0.53 0.68 0.55 0.34 0.59 0.54 72.85 109.31 0.52 0.35 0.30 0.54 68.93 0.23 115.65 13
30/0.15 60.96 0.41 0.60 0.75 0.62 0.43 0.64 0.61 62.35 97.53 0.47 0.43 0.42 0.61 45.74 0.20 80.81 3
30/0.20 58.48 0.38 0.53 0.69 0.57 0.37 0.58 0.58 67.72 105.90 0.50 0.37 0.41 0.58 47.47 0.21 77.22 10
30/0.25 52.73 0.30 0.54 0.69 0.52 0.34 0.59 0.53 74.38 108.59 0.52 0.36 0.32 0.53 52.02 0.19 88.54 19
30/0.30 56.07 0.33 0.52 0.68 0.56 0.33 0.56 0.56 70.54 105.38 0.50 0.32 0.33 0.56 43.67 0.16 79.52 17
30/0.35 58.37 0.37 0.58 0.74 0.57 0.37 0.57 0.58 64.60 101.66 0.48 0.37 0.40 0.58 49.26 0.22 79.46 9
30/0.40 62.73 0.36 0.59 0.74 0.60 0.43 0.65 0.63 58.09 100.81 0.48 0.48 0.42 0.63 27.00 0.25 40.09 4
30/0.45 59.41 0.38 0.59 0.73 0.59 0.38 0.59 0.59 68.63 100.51 0.47 0.39 0.38 0.59 45.21 0.15 88.38 8
30/0.50 60.48 0.40 0.59 0.74 0.60 0.41 0.60 0.60 68.74 95.42 0.45 0.41 0.41 0.60 37.36 0.18 67.44 6
Table 5: Averaged performance results from all meta-models trained for each combination.
Comb. ACC KAPPA PRC ROC F COR PR RC RAE RRSE RMSE INF MAR MF LL MI PRS Ranking
10/0.05 44.32 0.14 0.45 0.60 0.43 0.14 0.46 0.44 83.80 117.68 0.56 0.14 0.16 0.44 14.64 0.06 24.40 16
10/0.10 43.05 0.12 0.46 0.58 0.43 0.13 0.48 0.43 85.39 119.59 0.57 0.14 0.12 0.43 15.59 0.06 26.19 21
10/0.15 44.02 0.13 0.45 0.59 0.43 0.13 0.46 0.44 84.55 119.00 0.57 0.14 0.14 0.44 14.43 0.06 23.65 19
10/0.20 44.28 0.15 0.44 0.59 0.43 0.15 0.47 0.44 85.16 118.83 0.57 0.16 0.16 0.44 18.22 0.07 30.27 14
10/0.25 45.89 0.15 0.47 0.60 0.45 0.15 0.48 0.46 81.93 116.21 0.55 0.16 0.15 0.46 14.88 0.06 24.46 12
10/0.30 43.78 0.14 0.46 0.60 0.43 0.14 0.47 0.44 83.49 117.40 0.56 0.15 0.14 0.44 15.84 0.06 26.07 15
10/0.35 44.07 0.13 0.46 0.59 0.43 0.14 0.46 0.44 84.54 117.58 0.56 0.13 0.14 0.44 11.94 0.05 18.98 18
10/0.40 42.43 0.11 0.43 0.58 0.41 0.11 0.44 0.42 86.35 120.03 0.57 0.11 0.11 0.42 12.52 0.04 21.03 26
10/0.45 44.51 0.13 0.46 0.59 0.43 0.13 0.45 0.45 83.31 116.29 0.55 0.13 0.15 0.45 10.94 0.05 17.06 17
10/0.50 41.80 0.10 0.44 0.58 0.40 0.10 0.43 0.42 86.91 118.43 0.57 0.10 0.12 0.42 10.41 0.04 16.55 29
20/0.05 42.80 0.13 0.45 0.59 0.42 0.13 0.46 0.43 85.39 118.85 0.57 0.13 0.15 0.43 16.57 0.06 28.14 20
20/0.10 46.08 0.16 0.46 0.59 0.45 0.16 0.49 0.46 81.75 117.28 0.56 0.17 0.17 0.46 18.36 0.06 31.33 11
20/0.15 41.65 0.10 0.44 0.57 0.41 0.11 0.45 0.42 87.05 121.09 0.58 0.12 0.11 0.42 11.59 0.05 18.65 28
20/0.20 42.90 0.12 0.45 0.58 0.42 0.13 0.47 0.43 86.06 118.77 0.57 0.13 0.13 0.43 15.25 0.06 25.31 22
20/0.25 41.97 0.09 0.44 0.57 0.41 0.10 0.45 0.42 86.71 120.46 0.58 0.10 0.10 0.42 11.93 0.05 19.48 30
20/0.30 42.65 0.12 0.44 0.57 0.41 0.11 0.44 0.43 86.93 119.17 0.57 0.12 0.13 0.43 14.81 0.05 24.20 23
20/0.35 41.61 0.10 0.44 0.57 0.40 0.10 0.44 0.42 87.53 120.54 0.58 0.10 0.11 0.42 14.66 0.06 23.60 27
20/0.40 41.76 0.11 0.44 0.57 0.40 0.11 0.45 0.42 87.62 120.57 0.58 0.12 0.12 0.42 15.25 0.06 23.65 25
20/0.45 44.69 0.13 0.47 0.60 0.44 0.14 0.47 0.45 83.58 116.39 0.56 0.15 0.14 0.45 15.72 0.07 24.58 13
20/0.50 42.97 0.10 0.45 0.58 0.42 0.11 0.44 0.43 86.02 118.99 0.57 0.11 0.11 0.43 11.88 0.05 18.69 24
30/0.05 47.96 0.19 0.49 0.62 0.47 0.19 0.50 0.48 78.77 115.69 0.55 0.20 0.21 0.48 23.61 0.09 40.29 5
30/0.10 49.21 0.19 0.49 0.62 0.49 0.20 0.52 0.49 78.07 114.34 0.54 0.21 0.20 0.49 21.45 0.08 36.08 1
30/0.15 47.88 0.20 0.49 0.63 0.47 0.21 0.52 0.48 78.95 114.58 0.55 0.22 0.21 0.48 21.63 0.09 37.32 2
30/0.20 48.21 0.20 0.48 0.62 0.47 0.21 0.51 0.48 79.81 115.27 0.55 0.22 0.22 0.48 24.05 0.10 40.22 3
30/0.25 46.82 0.18 0.48 0.62 0.46 0.18 0.49 0.47 80.50 115.71 0.55 0.19 0.19 0.47 24.06 0.09 40.49 7
30/0.30 45.69 0.16 0.46 0.61 0.44 0.16 0.48 0.46 82.11 116.52 0.56 0.17 0.17 0.46 19.97 0.08 33.25 10
30/0.35 47.53 0.16 0.48 0.60 0.46 0.17 0.50 0.48 79.27 114.75 0.55 0.18 0.16 0.48 19.43 0.08 32.22 8
30/0.40 48.41 0.18 0.48 0.60 0.47 0.18 0.50 0.48 78.94 114.78 0.55 0.19 0.19 0.48 19.29 0.07 32.74 6
30/0.45 48.31 0.18 0.47 0.62 0.47 0.19 0.49 0.48 79.17 114.26 0.54 0.19 0.19 0.48 25.97 0.10 43.66 4
30/0.50 45.79 0.18 0.48 0.63 0.44 0.19 0.50 0.46 82.13 115.70 0.56 0.20 0.21 0.46 19.59 0.07 32.44 9
ment for baseline removal. For this reason, 30 com-
binations were experimentally evaluated; each one is
a combination of different signal length used for data
segmentation and cut-off frequency for baseline re-
moval. For the signal lengths: 10 s, 20 s, and 30 s,
were used; while the cut-off frequency starts from
0.05 Hz, till 0.50 Hz, with a step of 0.05 Hz.
The evaluation of these combinations was per-
formed in a combination with complexity analysis
used for feature extraction that are further used for
blood pressure classification. The evaluation was
done using a dataset that contains data from five com-
mercially available ECG sensors. The dataset was
pre-processed for each combination separately. The
same classification methodology was used for each
pre-processed dataset, which results are only used to
estimate the valuable information that these combina-
tions convey. A data-driven methodology from multi-
criteria decision analysis was used to rank these com-
binations regarding a set of 17 performance measures.
Evaluation results showed that the best results can
be achieved using a signal length of 30 s, so it fol-
lows that it carries the most valuable information. Re-
garding the cut-off frequency, it follows that good
models are achieved when it is between 0.10 Hz and
0.20 Hz. This is a contribution to the arguments pub-
Comparing Different Settings of Parameters Needed for Pre-processing of ECG Signals used for Blood Pressure Classification
69
lished in the literature discussing the optimal ECG
sample lengths needed for building predictive models
(Takahashi et al., 2017; Shdefat et al., 2018), as well
as the lower frequencies where the ECG components
overlap with the baseline wander noise (Xu et al.,
2017). Also, this analysis can further give additional
directions for research, where the pre-processing will
not be made regarding a fix signal length and cut-off
frequency for baseline removal, but it can produce in-
formation that can be used for information fusion ap-
proaches.
ACKNOWLEDGEMENTS
This work was supported by the Slovenian Research
Agency (research core funding No. P2-0098, and
project J1-8155), and by SIARS, NATO multi-year
project NATO.EAP.SFPP 984753.
REFERENCES
AHA (2016). Understanding blood pressure readings.
Ahmad, S., Chen, S., Soueidan, K., Batkin, I., Bolic, M.,
Dajani, H., and Groza, V. (2012). Electrocardiogram-
assisted blood pressure estimation. IEEE Transactions
on Biomedical Engineering, 59(3):608–618.
Ahonen, L., Cowley, B., Torniainen, J., Ukkonen, A., Vi-
havainen, A., and Puolam
¨
aki, K. (2016). Cognitive
collaboration found in cardiac physiology: Study in
classroom environment. PloS one, 11(7):e0159178.
Biosignals, B. (2016). Emotion faros (2016).
Brans, J.-P. and Mareschal, B. (2005). Promethee methods.
In Multiple criteria decision analysis: state of the art
surveys, pages 163–186. Springer.
Breiman, L. (1996). Bagging predictors. Machine learning,
24(2):123–140.
Canning, J., Helbert, K., Iashin, G., Matthews, J., Yang, J.,
Delano, M. K., Sodini, C. G., and Zhang, Q. (2016).
Noninvasive and continuous blood pressure measure-
ment via superficial temporal artery tonometry. In En-
gineering in Medicine and Biology Society (EMBC),
2016 IEEE 38th Annual International Conference of
the, pages 3382–3385. IEEE.
Chan, K., Hung, K., and Zhang, Y. (2001). Noninva-
sive and cuffless measurements of blood pressure for
telemedicine. In Engineering in Medicine and Biol-
ogy Society, 2001. Proceedings of the 23rd Annual In-
ternational Conference of the IEEE, volume 4, pages
3592–3593. IEEE.
Choi, Y., Zhang, Q., and Ko, S. (2013). Noninvasive cuf-
fless blood pressure estimation using pulse transit time
and hilbert–huang transform. Computers & Electrical
Engineering, 39(1):103–111.
Cosoli, G., Casacanditella, L., Pietroni, F., Calvaresi, A.,
Revel, G. M., and Scalise, L. (2015). A novel ap-
proach for features extraction in physiological sig-
nals. In Medical Measurements and Applications
(MeMeA), 2015 IEEE International Symposium on,
pages 380–385. IEEE.
Eftimov, T., Koro
ˇ
sec, P., and Seljak, B. K. (2018). Data-
driven preference-based deep statistical ranking for
comparing multi-objective optimization algorithms.
In International Conference on Bioinspired Methods
and Their Applications, pages 138–150. Springer.
Eke, A., Herman, P., Kocsis, L., and Kozak, L. (2002). Frac-
tal characterization of complexity in temporal physio-
logical signals. Physiological measurement, 23(1):R1.
Freund, Y., Schapire, R. E., et al. (1996). Experiments with
a new boosting algorithm. In Icml, volume 96, pages
148–156. Citeseer.
Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff,
J. M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody,
G. B., Peng, C.-K., and Stanley, H. E. (2000). Phys-
iobank, physiotoolkit, and physionet. Circulation,
101(23):e215–e220.
Goli, S. and Jayanthi, T. (2014). Cuff less continuous non-
invasive blood pressure measurement using pulse tran-
sit time measurement. International Journal of Recent
Development in Engineering and Technology, 2(1):87.
Hacks, C. (2015). e-health sensor platform v2. 0 for arduino
and raspberry pi.
Hassan, M. K. B. A., Mashor, M., Nasir, N. M., and Mo-
hamed, S. (2008). Measuring of systolic blood pres-
sure based on heart rate. In 4th Kuala Lumpur Interna-
tional Conference on Biomedical Engineering 2008,
pages 595–598. Springer.
Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J., and
Scholkopf, B. (1998). Support vector machines. IEEE
Intelligent Systems and their applications, 13(4):18–
28.
i Car
´
os, J. M. S. (2011). Continuous non-invasive blood
pressure estimation. PhD thesis, ETH.
Ilango, S. and Sridhar, P. (2014). A non-invasive blood pres-
sure measurement using android smart phones. IOSR
Journal of Dental and Medical Sciences, 13(1):28–31.
Johnstone, J. A., Ford, P. A., Hughes, G., Watson, T., and
Garrett, A. T. (2012). Bioharness™ multivariable
monitoring device: part. i: validity. Journal of sports
science & medicine, 11(3):400.
Kim, N., Krasner, A., Kosinski, C., Wininger, M., Qadri,
M., Kappus, Z., Danish, S., and Craelius, W. (2016).
Trending autoregulatory indices during treatment for
traumatic brain injury. Journal of clinical monitoring
and computing, 30(6):821–831.
Kugiumtzis, D. and Tsimpiris, A. (2010). Measures of anal-
ysis of time series (mats): A matlab toolkit for compu-
tation of multiple measures on time series data bases.
arXiv preprint arXiv:1002.1940.
Li, Y., Gao, Y., and Deng, N. (2013). Mechanism of cuff-
less blood pressure measurement using mmsb. Engi-
neering, 5(10):123.
Liao, Y. and Vemuri, V. R. (2002). Use of k-nearest neigh-
BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing
70
bor classifier for intrusion detection1. Computers &
security, 21(5):439–448.
Liaw, A., Wiener, M., et al. (2002). Classification and re-
gression by randomforest. R news, 2(3):18–22.
Marani, R. and Perri, A. G. (2012). An intelligent sys-
tem for continuous blood pressure monitoring on re-
mote multi-patients in real time. arXiv preprint
arXiv:1212.0651.
McBride, J. C., Zhao, X., Munro, N. B., Smith, C. D.,
Jicha, G. A., Hively, L., Broster, L. S., Schmitt, F. A.,
Kryscio, R. J., and Jiang, Y. (2014). Spectral and com-
plexity analysis of scalp eeg characteristics for mild
cognitive impairment and early alzheimer’s disease.
Computer methods and programs in biomedicine,
114(2):153–163.
Monge-
´
Alvarez, J. (2015). Higuchi and katz fractal dimen-
sion measures. https://www.mathworks.com/matlab
central/fileexchange/50290-higuchi-and-katz-fractal-
dimension-measures/content/Fractal dimension
measures/Higuchi FD.m.
Morabito, F. C., Labate, D., La Foresta, F., Bramanti, A.,
Morabito, G., and Palamara, I. (2012). Multivariate
multi-scale permutation entropy for complexity anal-
ysis of alzheimer’s disease eeg. Entropy, 14(7):1186–
1202.
Mouradian, V., Poghosyan, A., and Hovhannisyan, L.
(2015). Noninvasive continuous mobile blood pres-
sure monitoring using novel ppg optical sensor. In
Biomedical Wireless Technologies, Networks, and
Sensing Systems (BioWireleSS), 2015 IEEE Topical
Conference on, pages 1–3. IEEE.
Nitzan, M. (2011). Automatic noninvasive measurement
of arterial blood pressure. IEEE instrumentation &
measurement magazine, 14(1).
Nye, R., Zhang, Z., and Fang, Q. (2015). Continuous
non-invasive blood pressure monitoring using pho-
toplethysmography: A review. In Bioelectronics
and Bioinformatics (ISBB), 2015 International Sym-
posium on, pages 176–179. IEEE.
Patil, T. R. and Sherekar, S. (2013). Performance analy-
sis of naive bayes and j48 classification algorithm for
data classification. International journal of computer
science and applications, 6(2):256–261.
Payne, R., Symeonides, C., Webb, D., and Maxwell, S.
(2006). Pulse transit time measured from the ECG:
an unreliable marker of beat-to-beat blood pressure.
Journal of Applied Physiology, 100(1):136–141.
Powers, D. M. (2011). Evaluation: from precision, recall
and f-measure to roc, informedness, markedness and
correlation.
Program, N. H. B. P. E. et al. (2004). The seventh report of
the joint national committee on prevention, detection,
evaluation, and treatment of high blood pressure.
Rish, I. et al. (2001). An empirical study of the naive bayes
classifier. In IJCAI 2001 workshop on empirical meth-
ods in artificial intelligence, volume 3, pages 41–46.
IBM New York.
Sackl-Pietsch, E. (2010). Continuous non-invasive arterial
pressure shows high accuracy in comparison to inva-
sive intra-arterial blood pressure measurement. Un-
published manuscript.
Sahani, A. K., Ravi, V., and Sivaprakasam, M. (2014). Au-
tomatic estimation of carotid arterial pressure in art-
sens. In India Conference (INDICON), 2014 Annual
IEEE, pages 1–6. IEEE.
Sahoo, A., Manimegalai, P., and Thanushkodi, K. (2011).
Wavelet based pulse rate and blood pressure esti-
mation system from ECG and PPG signals. In
Computer, Communication and Electrical Technology
(ICCCET), 2011 International Conference on, pages
285–289. IEEE.
Schroeder, E. B., Liao, D., Chambless, L. E., Prineas, R. J.,
Evans, G. W., and Heiss, G. (2003). Hypertension,
blood pressure, and heart rate variability. Hyperten-
sion, 42(6):1106–1111.
Seo, J., Pietrangelo, S. J., Lee, H.-S., and Sodini, C. G.
(2015). Noninvasive arterial blood pressure wave-
form monitoring using two-element ultrasound sys-
tem. IEEE transactions on ultrasonics, ferroelectrics,
and frequency control, 62(4):776–784.
Shdefat, A. Y., Joo, M.-I., Choi, S.-H., and Kim, H.-C.
(2018). Utilizing ECG waveform features as new
biometric authentication method. International Jour-
nal of Electrical and Computer Engineering (IJECE),
8(2).
Shriram, R., Wakankar, A., Daimiwal, N., and Ramdasi, D.
(2010). Continuous cuffless blood pressure monitor-
ing based on ptt. In Bioinformatics and Biomedical
Technology (ICBBT), 2010 International Conference
on, pages 51–55. IEEE.
Simjanoska, M., Gjoreski, M., Gams, M., and Made-
vska Bogdanova, A. (2018). Non-invasive blood pres-
sure estimation from ECG using machine learning
techniques. Sensors, 18(4):1160.
Takahashi, N., Kuriyama, A., Kanazawa, H., Takahashi, Y.,
and Nakayama, T. (2017). Validity of spectral analysis
based on heart rate variability from 1-minute or less
ECG recordings. Pacing and Clinical Electrophysiol-
ogy.
Tanaka, S., Nogawa, M., Yamakoshi, T., and Yamakoshi,
K.-i. (2007). Accuracy assessment of a noninvasive
device for monitoring beat-by-beat blood pressure
in the radial artery using the volume-compensation
method. IEEE Transactions on Biomedical engineer-
ing, 54(10):1892–1895.
Technology, Z. (2017). Zephyr bioharness 3.0 user manual.
https://www.zephyranywhere.com/media/download/
bioharness3-user-manual.pdf.
Thomas, S. S., Nathan, V., Zong, C., Soundarapandian, K.,
Shi, X., and Jafari, R. (2016). Biowatch: A noninva-
sive wrist-based blood pressure monitor that incorpo-
rates training techniques for posture and subject vari-
ability. IEEE journal of biomedical and health infor-
matics, 20(5):1291–1300.
Trobec, R., Toma
ˇ
si
´
c, I., Rashkovska, A., Depolli, M., and
Avbelj, V. (2018a). Commercial ecg systems. In
Body Sensors and Electrocardiography, pages 101–
114. Springer.
Comparing Different Settings of Parameters Needed for Pre-processing of ECG Signals used for Blood Pressure Classification
71
Trobec, R., Toma
ˇ
si
´
c, I., Rashkovska, A., Depolli, M., and
Avbelj, V. (2018b). Commercial ECG systems. In
Body Sensors and Electrocardiography, pages 101–
114. Springer.
Vihinen, M. (2012). How to evaluate performance of pre-
diction methods? measures and their interpretation
in variation effect analysis. In BMC genomics, vol-
ume 13, page S2. BioMed Central.
Winderbank-Scott, P. and Barnaghi, P. (2017). A non-
invasive wireless monitoring device for children and
infants in pre-hospital and acute hospital environ-
ments.
Wong, M. Y.-M., Poon, C. C.-Y., and Zhang, Y.-T. (2009).
An evaluation of the cuffless blood pressure estima-
tion based on pulse transit time technique: a half year
study on normotensive subjects. Cardiovascular En-
gineering, 9(1):32–38.
Xu, Y., Luo, M., Li, T., and Song, G. (2017). Ecg
signal de-noising and baseline wander correction
based on ceemdan and wavelet threshold. Sensors,
17(12):2754.
Zhang, X.-Y. and Zhang, Y.-T. (2006). A model-based
study of relationship between timing of second heart
sound and systolic blood pressure. In Engineering in
Medicine and Biology Society, 2006. EMBS’06. 28th
Annual International Conference of the IEEE, pages
1387–1390. IEEE.
Zhou, P., Zong-Xia, M., Chun-Lan, H., and Huang, Y.-X.
(2017). Power spectral entropy in the ECG of patients
suffered from nocturnal frontal lobe epilepsy. Journal
of Pharmaceutical and Biomedical Sciences, 7(3).
BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing
72