Computers in Biology and Medicine 76 (2016) 113119

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/cbm

Discrimination of systolic and diastolic dysfunctions using multi-layer

perceptron in heart rate variability analysis
Yalcin Isler
Izmir Katip Celebi University, Department of Biomedical Engineering, Cigli, Izmir, Turkey

art ic l e i nf o

a b s t r a c t

Article history:
Received 29 January 2016
Received in revised form
26 June 2016
Accepted 28 June 2016

In this study, the heart rate variability (HRV) analysis is used to distinguish patients with systolic congestive heart failure (CHF) from patients with diastolic CHF. In the analysis performed, the best accuracy
performances of short-term HRV measures are compared. These measures are calculated in four different
ways with optional normalization methods of heart rate and data. The nearest neighbor and the multilayer perceptron (MLP) are used to evaluate the performances in discriminating these two groups. The
results point out that using both data and heart rate normalizations enhances the classier performance.
The maximum accuracy is obtained as 96.43% with MLP classier.
& 2016 Elsevier Ltd. All rights reserved.

Keywords:
Heart rate variability
Heart failure
Normalization
Multi-layer perceptron
Nearest neighbor

1. Introduction
When the pumping capacity of a heart reduces, this disorder is
called heart failure. As the accumulation of blood is so common in
patients with heart failure, it is also named congestive heart failure
(CHF). The failing heart does not manage either to pump or to ll
up, which are called systolic heart failure and diastolic heart failure, respectively [1]. Although the most essential element for the
success of the treatment is the reliable and exact diagnosis of CHF,
systolic dysfunction was determined in only half of all cases and
only 46% of physicians tried routinely distinguishing either systolic
or diastolic dysfunctionality in the patients with CHF [2]. The decreased pump function (i.e. systolic CHF) can be measured by
functional cardiac studies (e.g. ECG, BNP, and echocardiogram) [3
5]. On the other hand, the diastolic dysfunction is identiable by
only echocardiogram [6,7].
Almost 90% of physicians routinely performed ECG tests in
patients due to its simplicity [2]. Nonetheless, physicians accepted
that echocardiography is essential in diagnosis whether patient
with CHF is systolic or diastolic in addition to clinical sign and
symptoms [8]. Average waiting time for echocardiography was
1 month and that was 48 h for ECG [2]. Hence, simple and reliable
diagnostic procedures based on ECG may be helpful for primary
care physicians (PCPs) who are responsible for the early diagnosis
of CHF and implementation of adequate therapy.
Analysis of heart rate variability (HRV), derived from the ECG
E-mail address: islerya@yahoo.com
http://dx.doi.org/10.1016/j.compbiomed.2016.06.029
0010-4825/& 2016 Elsevier Ltd. All rights reserved.

analysis, is to analyze the variations between timings of heart

beats [9]. The sinoatrial node (SAN) is the pacemaker of the heart
and is responsible for the HRV. The cells of the SAN slowly but
automatically depolarize; when reaching a threshold, they rapidly
depolarize, followed by a repolarization, and the process repeats
itself constantly. The depolarization quickly propagates to the
surrounding cardiac muscle cells and the contraction of the heart
begins. In healthy subjects, the SAN cells generate depolarization
or action potentials at a frequency that is regulated through direct
innervation of both branches of ANS: sympathetic and parasympathetic (or vagal). The parasympathetic branch releases
acetylcholine (ACh) that slows the rate of SAN depolarization,
while the sympathetic branch releases norepineprine that increases the rate of SAN depolarization. The SAN effectively integrates both inputs from ANS, both temporally and spatially, and
this pacemaker activity is often modeled as an integrate pulse
frequency modulation [40]. This is the source of HRV [41].
In 1996, standards in HRV analysis was published by proposing
a number of time- and frequency-domain parameters calculated
from short-term (5-min) and long-term (24-h) HRV data [9]. In
order to achieve the results quickly, short-term (5-min long) HRV
measures have been preferred in many studies. HRV has been used
in a great number of clinical studies including prediction of sudden cardiac death after acute myocardial infarction [10], early diagnosis of diabetic neuropathy [11], dilated cardiomyopathy [12
14], CHF [1520], fetal distress conditions [21], and obstructive
sleep apnea [22].
In addition, Hallstrom et al. presented that rescaling RR

114

Y. Isler / Computers in Biology and Medicine 76 (2016) 113119

intervals to a xed HR reduces the variation in the measures and

increases their discrimination power [23]. If these measures are
calculated using heart rate normalized data, this analysis is called
as heart rate normalized heart rate variability (HRN) analysis
[20]. In a previous study, the discrimination of CHF patients who
are systolic or diastolic was reached by using kNN classier in only
12 patients as a case study [24].
In this study, it is aimed to determine whether systolic or
diastolic dysfunctionality exists in CHF patients using nearest
neighbor and multi-layer perceptron in both HRV and HRN measures. The following section covers brief information about the
methods used in the study including data, obtaining HRV measures, normalization techniques, classication and performance
assessment algorithms with the owchart of the implementation.
Section 3 gives the results. Finally, the results are discussed.

2. Methods
2.1. Experimental protocol
Patients came to the clinics after their breakfast at the morning,
which took approximately 2 h from their home to clinics because
of the morning trafc in Izmir, Turkey. Patients were requested to
rest at least 10 min just before connecting Holter device. After the
connection, patients waited for 10 min in resting position on a
comfortable chair. During this time, the required paper work was
also completed. Then, they left the clinics and continued their
daily life. They came back to the clinics again approximately 24 h
later. 30 patients from all patients were included into the study
after two experts analyzed the data.
The collected data were digitized at 128 samples per second.
There were 115,200 peaks in ECG by supposing 80 beats per
minute during 24 h for each patient. Ectopic beats are excluded
from HRV analysis in general [9]. Deciding such a big amount of
beats whether normal or ectopic is very time-consuming and irritable process for experts. Moreover, this process is costly for
health industry. That is the reason why the rst 5-min segments of
all recorded ECG data are included in this study. In addition, using
only 5-min (300-s) sections of data helps us achieving the discrimination faster and not disturbing the patients longer in a clinic
environment.
2.2. Data
The Holter ECG data used in this study were obtained from the
Faculty of Medicine in Dokuz Eyll University [24]. Two cardiologists decided whether patients would be included into the study
or not. The data included the study are as follows:

Age of the patient, time-domain, frequency-domain and non-linear methods are used in HRV studies. For further details, the studies given in the references could be examined [19,24,20].
Welch periodogram method, which uses fast Fourier transform
(FFT) method, is commonly used for the evaluation of frequency
domain [25]. Thanks to this method, power spectrum density
(PSD) over the data that is sampled at equal interval through the
time can be computed. In this study, HRV data is re-sampled by
using cubic interpolation method at 4 Hz before using FFT and
detrending of data is eliminated for providing stability analysis
[25]. Besides, evaluation of frequency domain obtained by Lomb
periodogram which is developed as an alternative to the classical
PSD method is also used. Owing to this method, PSD can be
computed from directly the HRV data without needing to resample on time domain [2628]. HRV evaluation of conventional
frequency domain for 5-min period is examined in detail by the
recommended standard [9].
There are three frequency bands commonly used in HRV analysis: VLF (00.033 Hz), LF (0.0330.15 Hz) and HF (0.150.4 Hz) [9].
In the evaluation of frequency domain, powers of these frequency
bands from calculated PSD were examined. Therefore, the following frequency domain measures were computed separately for
both methods of Welch periodogram and Lomb periodogram. The
frequency domain analysis contributed to the understanding of
autonomic background of RR interval uctuations in the HR record
[42,43].
On the other hand, wavelet analysis not only makes possible
the examination of one signal for the both time and scale domains
but also eliminates polynomial non-stability [29]. Wavelets are
reported as very useful for analysis of the RR intervals due to this
capability. Thus, Daubechies-4 main wavelet over HRV data resampled at 4 Hz with the 7-level wavelet transform method was
also used in this analysis to calculate the standard frequency domain features of HRV analysis [19].
Furthermore, non-linear measures derived from Poincare map
were used in this study. Poincare plot is a graph that each RR interval is plotted against the previous interval, which visualizes
detailed beating pattern of the heart. The Poincare plot is a popular technique thanks to its simplicity and its proved clinical
ability. Fitting an ellipse to the Poincare plot's shape and calculating measures from the plot is the most popular method among
nonlinear interpretation methods of HRV [3033].
As a result, 29 features were used. Those are obtained from
patient information, time domain evaluation, frequency spectrum
evaluation and non-linear methods with the numbers of 1, 6, 18,
and 4, respectively.
2.4. Normalization

age of 2066 years.

There are two optional normalization stages in the study:

(1) heart rate normalization and (2) data (or feature)
normalization.

the age of 3965 years.

 Heart rate normalization: Heart rate normalization process to x

 Systolic CHF data: 18 patients (11 men and 7 women) with the
 Diastolic CHF data: 12 patients (4 women and 8 unknown) with

the mean HR to its new value is applied as follows [23]:

Since the gender of 8 patients were not recorded, gender was
not used as a possible feature in the study.

HRN =

2.3. Heart rate variability measures

HRV data is collected after determining QRS wave structure in
the ECG signals because QRS wave structure is the component that
has the most distinctive amplitude values in the ECG signals. After
the determination of QRS using a Matlab implementation of Pan
Tompkins algorithm [38], the time difference between two consecutive R peaks is commonly called RR-interval (Tn = tn tn 1) [9].

60
HRV

NewHR
HRV

(1)

where HRN is HR normalized HRV data in seconds, HRV is raw

HRV data in seconds, NewHR is the new value of mean HR to be
xed in bpm, HRV is the mean value of whole HRV data in
seconds, and the constant 60 is used to make unit conversion
from bpm (beats-per-minute) to bps (beats-per-second). NewHR
is selected as 75 bpm since it is recommended and applied
value in [23,20].
Feature normalization: The used features must be equalized on

Y. Isler / Computers in Biology and Medicine 76 (2016) 113119

the same scale because units and amplitudes of features used in

classication and clustering algorithms differs highly compared
to each other [34]. In this study, MINMAX method is used to
make all features in the scale of 0, 1 using:

fi, N =

fi min(fi )
max(fi ) min(fi )

(2)

where fi, N is the normalized i-th feature, fi is the i-th feature,

min(fi ) and max(fi ) are the minimum and the maximum values
of the i-th feature, respectively.
2.5. Statistical analysis

115

2.7. Leave-one-out cross-validation method

In a cross-validation method, database is divided into two
groups. When the rst group (training database) is used for determination of model parameters of a classier, the other group
(test database) is used for testing the performance of the trained
classier [35]. In the leave-one-out cross-validation method, the
classier is trained with the n 1 samples (where n is the number
of samples in data) by leaving one sample for the test in each time.
Then the performances of the classier are calculated as a summation of these test results. In this study, the performance of
classier is determined by calculation of mean value for error of 30
times because there are 30 samples in the data [36].

The independent student t-test, which is a commonly used

method to show signicances of differences between measures of
two distinct groups, was used in this study [36]. The software
package of IBM SPSS Statistics version 22, which is one of the most
favorite programs in the literature related to the statistical analysis, is used to nd p-values that show statistical signicances. If the
p-value is less than or equal to 5%, the corresponding measure has
the statistical evidence on discriminating groups [39].

2.8. Performance measures

ACC =

TP + TN
TP + TN + FP + FN

(6)

2.6. Classication algorithms

SEN =

TP
TP + FN

(7)

SPE =

TN
TN + FP

(8)

2.6.1. Nearest neighbor

The main idea of the nearest-neighbor (NN) classier is to
predict the class of a test case based on its nearest neighbor. It
predicts the class of the test sample as that of its nearest neighbor.
Although there are many methods to calculate distances, the Euclidean distance is preferred in this paper [35]:
N

dist (j, k ) =

|fi j

fik |2

(3)

i=1
j

where N is the number of features, fi is the i-th feature of j-th

sample and fik is the i-th feature of k-th sample.
2.6.2. Multi-layer perceptron (MLP)
Multi-layer perceptron (MLP) is the most commonly used
structure among the applications based on articial neural networks due to its ability in using both linear and non-linear applications. MLP consists of an input layer, one or more hidden
layers, and an output layer [35]. The number of neurons in the
input layer is the number of features (or inputs) and the number of
neurons in the output layer is the number of outputs. The connection between the input-layer neurons and the hidden-layer
neurons which is similar to the connection between the hiddenlayer neurons and the output-layer neurons is calculated with
weights. The weight values are changed until the minimum error
value is found, which is called training. The general output function and the error function are dened as follows:

yi = f wjixi
i=1

E=

1
2

(4)

( di yi )
i

(5)

where xi is the input data, wji is weight values, f ( ) is an activation

function, yi is i-th network output, di is i-th expected output. The
technical implementation details of MLP can be found in the welldened literature [35].

There are a number of methods to compute performances of a

classier. Among them, the following one may be the easiest one
[34]:

If the instance is systolic and it is classied as systolic, it is counted

as true positive (TP); if it is misclassied as diastolic, it is counted
as false negative (FN). If the instance is diastolic and it is classied
as diastolic, it is counted as true negative (TN); if it is classied as
systolic, it is counted as false positive (FP). In the description of
these measures, the positive means the patient with systolic CHF
and the negative means the patient with diastolic CHF in this
study.
2.9. Implementation
The owchart of the implementation of the study is given in
Fig. 1. In this system, the heart rate and the data normalization
processes are applied optionally so these steps are drawn with
dashed line.
The rst stage of the proposed system is the loading of the
5-min HRV data from the hard disk. The next stage covers the
optional heart rate normalization process. The third stage is the
feature extraction. Time-domain, LS algorithm-based PSD and
Poincare plot measures are calculated directly from the HRV data.
Wavelet-based frequency domain measures require evenly sampled data. On the other hand, FFT-based frequency domain measures require both evenly sampled data and at least weakly stationary data. The cubic interpolation method with a rate of 4 Hz
(as offered in [9]) is used to make data evenly sampled and the
smoothness priors eye model with = 1000 (as offered in [37]) is
used to make data stationary. Both FFT- and LS-based frequency
domain PSD measures are calculated using 256 distinct frequency
points between 0 and 0.5 Hz . The wavelet entropy measures from
the resampled data are calculated using DB4 wavelet packet
transform method, which is probably the most preferred one in
the studies related to the HRV analysis [29]. The packets are calculated at the scale of 7 (i.e. 27 = 128 packets) in order to achieve
the limits of HRV frequency bands using 4-Hz resampled data. For
example, the rst packet is equivalent to the VLF band, the 2nd
10th packets are for the LF band, and the 11th32nd packets

116

Y. Isler / Computers in Biology and Medicine 76 (2016) 113119

Fig. 1. Flowchart of the system with optional heart rate and data normalization processes. The dashed lines show the optional steps.

correspond to the HF band. Four Poincare plot measures are then

calculated using the method of ellipse tting. At the end of this
stage, the HRV dataset with 29 features (1 for age, 6 time-domain
measures, 6 FFT-based PSD measures, 6 wavelet entropy measures,
6 LS-based PSD measures, and 4 Poincare plot measures) are
constructed. The last stage indicates the classication steps. This
also covers the optional data normalization step. MLP and NN
classiers are used to nd the best classier for distinguishing the
systolic CHF patients from the diastolic CHF patients. Using the
leave-one-out method, the performances of these classiers are
calculated.
If heart rate normalization process is applied, the system uses
the HRV data normalized to a x HR of 75 bpm [23]. If MINMAX
normalization is in use, the data to be applied to the classier is
limited to a range between 0 and 1. For each subject, all the
standard time-domain, frequency-domain (FFT-based, LS-based,
and wavelet entropies), and Poincare plot measures are calculated
by using Matlab 2015a. All the software implementations are
conducted using a portable workstation based on one Intel i7
processor with 32 GB RAM.

3. Results
All the features extracted from the HRV data and the HRN data
are given in Table 1. Statistical signicances (p-values) of HRN
measures are better than those of HRV measures in time-domain
and nonlinear parameters in general.
Statistical signicances (p-values) of measures are summarized
in Table 2 to show the effect of optional data normalization step.
After applying the HRN-based data normalization method, average
statistical signicance values are drastically reduced.
The application is run for eight different congurations:






No HR normalization and no MINMAX normalization,

HR normalization and no MINMAX normalization,
No HR normalization and MINMAX normalization,
HR normalization and MINMAX normalization.

with NN and MLP classiers. The performances of all the congurations are calculated based on their accuracy measures (see
Tables 3 and 4). The results show that using both MINMAX and
heart rate normalization procedures improve the performance of
the classication processes. The maximum accuracy is found to be
96.43% in the conguration of MINMAX and heart rate normalized

Y. Isler / Computers in Biology and Medicine 76 (2016) 113119

117

Table 1
Standard short-term HRV indices: patient information, time-domain, frequency-domain (estimated from periodogram techniques using both FFT and LS algorithms in
addition to wavelet entropy measures), and Poincare plots.
Feature

HRV

HRN

Systolic CHF

Diastolic CHF

Systolic CHF

Diastolic CHF

0.9727 0.38
1.521 7 1.79
1.698 7 2.16
1.0007 2.00
0.006 7 0.01
1.7317 2.13

0.542
0.313
0.379
0.349
0.349
0.360

0.800 7
0.282 7
0.3347
0.0007
0.0007
0.325 7

0.800 7
0.8917
0.996 7
0.250 7
0.0017
1.0097

1.000
0.214
0.257
0.150
0.150
0.236

Frequency-domain measures based on FFT

FFT_LFHF
9.317 14.79
FFT_VLF
401.617 1135.9
FFT_LF
90.346 7 255.53
FFT_HF
2.026 7 5.73
FFT_NLF
0.7347 0.22
FFT_NHF
0.2667 0.22

115.82 7 224.39
4.337 5.30
19.893 7 36.06
0.1807 0.26
0.598 7 0.41
0.402 7 0.41

0.187
0.468
0.705
0.559
0.424
0.414

9.317 14.79
84.658 7 239.45
19.0457 53.87
0.427 7 1.21
0.734 7 0.22
0.2667 0.22

115.82 7 224.39
1.602 7 1.87
10.363 7 19.72
0.059 7 0.07
0.598 7 0.41
0.402 7 0.41

0.187
0.468
0.705
0.559
0.424
0.414

Frequency-domain measures based on LS

LS_LFHF
2.7157 2.97
LS_VLF
0.229 7 0.19
LS_LF
0.1907 0.12
LS_HF
0.1297 0.08
LS_NLF
0.599 7 0.22
LS_NHF
0.401 7 0.22

0.454 7 0.14
0.220 7 0.19
0.073 7 0.04
0.1737 0.09
0.308 7 0.06
0.692 7 0.06

0.159
0.872
0.075
0.402
0.028
0.028

2.7157
0.229 7
0.1907
0.129 7
0.599 7
0.4017

0.454 7 0.14
0.220 7 0.19
0.0737 0.04
0.1737 0.09
0.308 7 0.06
0.692 7 0.06

0.159
0.872
0.075
0.402
0.028
0.028

Frequency-domain measures based on wavelet entropy

WS_LFHF
1.071 7 0.61
1.8077 1.90
WS_VLF
0.390 7 0.49
1.034 7 0.77
WS_LF
1.031 7 0.87
0.583 7 0.72
WS_HF
0.8477 0.49
0.265 7 0.13
WS_NLF
0.4827 0.14
0.543 7 0.20
WS_NHF
0.5187 0.14
0.457 7 0.20

0.315
0.101
0.366
0.039
0.518
0.507

1.0717 0.61
0.390 7 0.49
1.0317 0.87
0.8477 0.49
0.4827 0.14
0.5187 0.14

1.8077
1.034 7
0.583 7
0.265 7
0.543 7
0.4577

1.90
0.77
0.72
0.13
0.20
0.20

0.315
0.101
0.366
0.039
0.518
0.507

Poincare plot measures

SD1
0.4727
SD2
0.579 7
SD1SD2
1.846 7
RATIO
0.582 7

0.348
0.300
0.405
0.448

0.225 7
0.3067
0.398 7
0.581 7

0.7167
1.006 7
1.503 7
0.740 7

0.85
1.26
2.09
0.08

0.236
0.216
0.226
0.444

Time-domain measures
MEAN
0.874 7
SDNN
0.542 7
RMSSD
0.6827
NN20
0.250 7
PNN20
0.0027
SDSD
0.6787

0.34
1.38
1.73
0.71
0.00
1.85

1.26
1.44
5.16
0.38

1.2247
1.7327
4.859 7
0.7327

1.56
2.34
7.85
0.08

Table 2
Average statistical signicances of both HRV and HRN measures.
HRV

HRN

Time-domain measures
Frequency-domain measures (FFT periodogram)
Frequency-domain measures (LombScargle)
Frequency-domain measures (wavelet entropy)
Nonlinear measures (Poincare plot)

0.3820
0.4595
0.2607
0.3077
0.3753

0.2014
0.4595
0.2607
0.3077
0.2805

Mean values

0.3570

0.301

Feature set

Table 3
Classier results of the nearest neighbor classier, SEN is sensitivity (%), SPE is
specicity (%), and ACC is accuracy (%).

HRV
HRN

Before MINMAX

2.97
0.19
0.12
0.08
0.22
0.22

0.52
0.67
1.06
0.32

0.00
1.02
1.18
0.50
0.00
1.19

Table 4
Classier results of the MLP classier, SEN is sensitivity (%), SPE is specicity (%),
and ACC is accuracy (%).

Method

Feature set

0.00
0.62
0.76
0.00
0.00
0.76

After MINMAX

SEN

SPE

ACC

SEN

SPE

ACC

62.50
68.75

75.00
83.33

67.86
75.00

87.50
87.50

83.33
91.67

85.71
89.29

features using the MLP classier.

4. Conclusion
In this study, the effects of normalization methods of shortterm HRV feature sets are investigated for their discrimination in

HRV
HRN

Before MINMAX

After MINMAX

SEN

SPE

ACC

SEN

SPE

ACC

68.75
75.00

83.33
91.67

75.00
82.14

81.25
93.75

91.67
100.00

85.71
96.43

systolic and diastolic CHF patients. The class discrimination power

of different combinations of 29 distinct short-term HRV measures
are investigated using HRV data obtained from the Faculty of
Medicine in Dokuz Eyll University. In order to achieve the results
quickly, short-term (5-min long) HRV measures are preferred as
offered in the Task Force [9].
If there is a strong correlation among variables, then the accuracy of the classier is getting worse, which is the so-called
curse of dimensionality [35]. After applying the HRN-based data
normalization method, average statistical signicance values are
drastically reduced (Table 2). Heart rate normalization seems to
remove the correlation among HRV measures slightly. Nonetheless, an optimal or sub-optimal feature selection algorithm may
lead to more higher accuracies in classiers.
Only three measures were found to show statistically signicant differences ( p 5% ) among 29 measures. These measures
were normalized LF and normalized HF from LS-based frequency
analysis and HF from wavelet entropy based frequency analysis.

118

Y. Isler / Computers in Biology and Medicine 76 (2016) 113119

Since LF is affected by sympathetic activities and HF is affected by

vagal activities in the autonomous nervous system, this may
conclude that normalized calculations of these frequency components have signicant roles to discriminate whether CHF is systolic
or diastolic. Furthermore, LS and wavelet entropy methods may
give more statistically meaningful frequency-domain measures
because pre-processing steps such as re-sampling and detrending
may affect HRV measures [25].
The results, which are given in Tables 3 and 4, show that using
both MINMAX and HR normalization procedures improve the
performance of the classication. The maximum accuracy is found
to be 96.43% with both MINMAX and heart rate normalized features using a MLP classier. This shows that this classication
system can be used by physicians to detect systolic or diastolic CHF
patients from a ECG record easily. Nonetheless, there are only 30
participants in the study and this number is less than desired.
More patient data is needed to enhance the validity of this study.
In addition, HR normalization also improves the statistical
signicances in time-domain and non-linear HRV measures. This
indicates that the heart rate normalization methods improve the
discrimination capability of classiers.
In this study, only NN and MLP classiers are used to evaluate
the performance since these classication methods do not depend
on the data with any particular distribution. Using other classiers,
given in the literature related to machine learning, could further
improve the accuracy values.
Moreover, using long-term HRV measures in addition to the
short-term ones could also enhance the diagnostic ability of the
constructed classier systems. This is because the measures calculated from a 5-min HRV data do not reect the whole-day activity of the heart.

Conict of interest
None declared.

References
[1] Medical Encyclopedia, Heart Failure http://www.nlm.nih.gov/medlineplus/
ency/article/000158.htm (last visited in 2015).
[2] F. Hobbs, J. Korewicki, J. Cleland, et al., The diagnosis of heart failure in European primary care: the IMPROVEMENT programme survey of perception and
practice, Eur. J. Heart Fail. 7 (2005) 768779.
[3] S. Krger, K. Filzmaier, J. Graf, et al., QRS prolongation on surface ECG and
brain natriuretic peptide as indicators of left ventricular systolic dysfunction, J.
Int. Med. 255 (2004) 206212.
[4] V. Hombach, Electrocardiography of the failing heart, Cardiol. Clin. 24 (2006)
413426.
[5] G. Ashrith, N. Fida, A. Cordero-Reyes, A., et al., Echocardiographic Predictors of
Combined Pre and Post Capillary Pulmonary Hypertension (PH) in a Population of Systolic Heart Failure with WHO Group II Pulmonary Tension, American College of Cardiology (ACC.15), vol. 65(10S), March 17, 2015.
[6] M.R. Zile, C.F. Baicu, W.H. Gaasch, Diastolic heart failure-abnormalities in active relaxation and passive stiffness on the left ventricle, N. Engl. J. Med. 350
(19) (2004) 19531959.
[7] J. Wilbur, P. James, Diagnosis and management of heart failure in the outpatient setting, Prim. Care: Clin. Off. Pract. 32 (2005) 11151129.
[8] C. Gutierrez, D.G. Blanchard, Diastolic heart failure: challenges of diagnosis
and treatment, Am. Fam. Physician 69 (11) (2004) 26092616.
[9] Task Force of the European Society of Cardiology and the North American
Society of Pacing and Electrophysiology, Heart rate variability: standards of
measurement, physiological interpretation and clinical use, Circulation 93
(1996) 10431065.
[10] J.T. Bigger, J.L. Fleiss, R.C. Steinmann, et al., RR variability in healthy, middle
aged persons compared with patients with chronic coronary heart disease or
recent acute myocardial infarction, Circulation 91 (7) (1995) 19361943.
[11] M. Pagani, G. Malfatto, S. Pierini, et al., Spectral analysis of heart rate variability
in the assessment of autonomic diabetic neuropathy, J. Auton. Nerv. Syst. 23
(1988) 143153.
[12] L. Fauchier, D. Babuty, P. Cosnay, et al., Heart rate variability in idiopathic dilated cardiomyopathy: characteristics and prognostic value, J. Am. Coll.

Cardiol. 30 (4) (1997) 10091014.

[13] L. Fauchier, D. Babuty, M.L. Autret, et al., Temporal and spectral analysis of
heart rate variability in primary dilate cardiomyopathy: evaluation by case
control study, Arch. Mal. Coeur. Vaiss. 91 (2) (1998) 217224.
[14] G. Yi, J.H. Goldman, P.J. Keeling, et al., Heart rate variability in idiopathic dilated cardiomyopathy: relation to disease severity and prognosis, Heart 77 (2)
(1997) 108114.
[15] J.P. Saul, Y. Arai, R.D. Berger, et al., Assessment of autonomic regulation in
chronic congestive heart failure by heart rate spectral analysis, Am. J. Cardiol.
61 (1988) 12921299.
[16] G. Casolo, E. Balli, T. Taddei, et al., Decreased spontaneous heart rate variability
on congestive heart failure, Am. J. Cardiol. 64 (1989) 11621167.
[17] P.F. Blinkley, E. Nunziata, G.J. Haas, et al., Parasympathetic withdrawal is an
integral component of autonomic imbalance in congestive heart failure: demonstration in human subjects and verication in a paced canine model of
ventricular failure, J. Am. Coll. Cardiol. 18 (1991) 464472.
[18] M.H. Asyali, Discrimination power of long-term heart rate variability measures, in: Proceedings of the 25th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, Cancun, September 1721, 2003.
[19] Y. Isler, M. Kuntalp, Combining classical HRV indices with wavelet entropy
measures improves to performance in diagnosing congestive heart failure,
Comput. Biol. Med. 37 (10) (2007) 15021510.
[20] Y. Isler, M. Kuntalp, Heart rate normalization effects in the analysis of heart
rate variability analysis in congestive heart failure, Proc. IMechE Part H: J. Eng.
Med. 224(3) (2010) 453463.
[21] E.H. Hon, S.T. Lee, Electronic evaluations of the fetal heart rate patterns preceding fetal death: further observations, Am. J. Obstet. Gynecol. 87 (1965)
814826.
[22] A. Stys, T. Stys, Current clinical application of heart rate variability, Clin. Cardiol. 21 (10) (1998) 719724.
[23] A.P. Hallstrom, P.K. Stein, R. Schneider, et al., Structural relationships between
measures based on heart beat intervals: potential for improved risk assessment, IEEE Trans. Biomed. Eng. 51 (8) (2004) 14141420.
[24] Y. Isler, A detailed analysis of the effects of various combinations of heart rate
variability indices in congestive heart failure (Ph.D. dissertation), Dokuz Eylul
University, Izmir, Turkey, 2009.
[25] G.D. Clifford, L. Tarassenko, Quantifying errors in spectral estimates of HRV
due to beat replacement and resampling, IEEE Trans. Biomed. Eng. 52 (4)
(2005) 630638.
[26] N.R. Lomb, Least-squares frequency analysis of unequally spaced data, Astrophys. Space Sci. 39 (1976) 447462.
[27] J.D. Scargle, Studies in astronomical time series analysis (II): statistical aspects
of spectral analysis of unevenly spaced data, Astrophys. J. 263 (1982) 835853.
[28] G.B. Moody, Spectral analysis of heart rate without resampling, Comput.
Cardiol. 20 (1993) 715718.
[29] U. Wiklund, M. Akay, U. Niklasson, Short-term analysis of heart-rate variability
by adapted wavelet transforms, IEEE Eng. Med. Biol. 16 (5) (1997) 113118, and
138.
[30] P.W. Kamen, A.M. Tonkin, Application of the poincare plot to heart rate
variability: a new measure of functional status in heart failure, Aust. N. Z. J.
Med. 25 (1) (1995) 1826.
[31] P.W. Kamen, H. Krum, A.M. Tonkin, Poincare plot of heart rate variability allows quantitative display of parasympathetic nervous activity, Clin. Sci. 92
(1996) 201208.
[32] P.W. Kamen, Heart rate variability, Aust. Fam. Physician 25 (1996) 10871094.
[33] M. Brennan, M. Palaniswami, P. Kamen, Do existing measures of Poincare plot
geometry reect nonlinear features of heart rate variability? IEEE Trans.
Biomed. Eng. 48 (11) (2001) 13421347.
[34] R.C. Eberhart, R.W. Dobbins, Neural Network PC Tools, Academic Press, San
Diego, 1990.
[35] R.O. Duda, P.E. Hart, D.G. Stork, Pattern Classication, 2nd ed., Wiley, New
York, 2001.
[36] Y. Isler, A. Narin, M. Ozer, Comparison of the effects of cross-validation
methods on determining performances of classiers used in diagnosing congestive heart failure, Meas. Sci. Rev. 15 (4) (2015) 196201.
[37] M.P. Tarvainen, P.O. Ranta-aho, P.A. Karjalainen, An advanced detrending
method with application to HRV analysis, IEEE Trans. Biomed. Eng. 49 (2)
(2002) 172175.
[38] J. Pan, W. Tompkins, Real time algorithm detection for QRS, IEEE Trans.
Biomed. Eng. 32 (3) (1985) 230236.
[39] A. Narin, Y. Isler, M. Ozer, Investigating the performance improvement of HRV
indices in CHF using feature selection methods based on backward elimination and statistical signicance, Comput. Biol. Med. 45 (2014) 7279.
[40] K.L. Chang, K.J. Monahan, M.P. Grifn, et al., Comparison and clinical application of frequency domain methods in analysis of neonatal heart rate time
series, Ann. Biomed. Eng. 29 (2001) 764774.
[41] H. Cao, Linear and nonlinear approaches to heart rate analysis near the time of
birth (Ph.D. thesis), University of Virginia, Department of Biomedical Engineering, USA, 2004.
[42] M. Pagani, F. Lombardi, S. Guzzetti, et al., Power spectral analysis of heart rate
arterial pressure variabilities as a marker of sympathovagal interaction in man
and conscious dog, Circ. Res. 59 (1986) 178193.
[43] A. Malliani, M. Pagani, R. Furlan, et al., Individual recognition by heart rate
variability of two different autonomic proles related to posture, Circulation
96 (1997) 41434145.

Y. Isler / Computers in Biology and Medicine 76 (2016) 113119

Yalcin Isler (1971) received the B.Sc. degree in Electrical and Electronics Engineering from Anadolu University, Eskisehir, Turkey, in 1993, the M.Sc. degree in
Electronics and Communication Engineering, Suleyman Demirel University, Isparta,
Turkey, in 1996, and the Ph.D. degree in Electrical and Electronics Engineering,
Dokuz Eylul University, Izmir, Turkey, in 2009.
He worked as a lecturer in Burdur Vocational School of Suleyman Demirel
University (Burdur, 19932000), as a software engineer at the Department of
Research and Development in commercial companies (Izmir, 20002002), as a
research assistant in Bulent Ecevit University (Zonguldak, 20022003) and in

119

Dokuz Eylul University (Izmir, 20032010), as an assistant professor in Electrical

and Electronics Engineering of Bulent Ecevit University (Zonguldak, 20102012).
He has been working as an assistant professor with the Department of Biomedical Engineering in Izmir Katip Celebi University, Izmir, since 2012. His main
research interests are biomedical signal processing, medical device design,
computational neuroscience, pattern recognition, and embedded systems. He is
involved in several national projects on computer science in medicine and on
other elds that required designing electronic boards based on embedded
systems.