See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/270967090

Automatic recognition of cardiac arrhythmias based on the geometric

patterns of Poincaré plots

Article  in  Physiological Measurement · January 2015

DOI: 10.1088/0967-3334/36/2/283 · Source: PubMed

CITATIONS READS

16 656

7 authors, including:

Ying Wang
Xiamen University
68 PUBLICATIONS   1,103 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Obsessive compulsive disorder View project

All content following this page was uploaded by Ying Wang on 17 January 2016.

The user has requested enhancement of the downloaded file.

Home Search Collections Journals About Contact us My IOPscience

Automatic recognition of cardiac arrhythmias based on the geometric patterns of Poincaré

plots

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2015 Physiol. Meas. 36 283

(http://iopscience.iop.org/0967-3334/36/2/283)

View the table of contents for this issue, or go to the journal homepage for more

Download details:
This content was downloaded by: inging_wang
IP Address: 117.28.251.191
This content was downloaded on 14/01/2015 at 01:58

Please note that terms and conditions apply.

 Institute of Physics and Engineering in Medicine Physiological Measurement

Physiol. Meas. 36 (2015) 283–301 doi:10.1088/0967-3334/36/2/283

Automatic recognition of cardiac

arrhythmias based on the geometric
patterns of Poincaré plots
Lijuan Zhang2, Tianci Guo1,3, Bin Xi1, Yang Fan1, Kun Wang1,
Jiacheng Bi1, Ying Wang1
1
  Department of Automation, School of Information Science and Technology,
Xiamen University, Xiamen, Fujian, People’s Republic of China
2
  The First Affiliation Hospital of Xiamen University, Xiamen University, Xiamen,
Fujian, People’s Republic of China
3
  Institute of Industrial Process Control, Department of Control Science and
Engineering, Zhejiang University Hangzhou, Zhejiang, People’s Republic of China

E-mail: wangying@xmu.edu.cn

Received 9 July 2014, revised 25 November 2014

Accepted for publication 26 November 2014
Published 13 January 2015

Abstract
The Poincaré plot emerges as an effective tool for assessing cardiovascular
autonomic regulation. It displays nonlinear characteristics of heart rate variability
(HRV) from electrocardiographic (ECG) recordings and gives a global view of
the long range of ECG signals. In the telemedicine or computer-aided diagnosis
system, it would offer significant auxiliary information for diagnosis if the patterns
of the Poincaré plots can be automatically classified. Therefore, we developed
an automatic classification system to distinguish five geometric patterns of the
Poincaré plots from four types of cardiac arrhythmias. The statistics features are
designed on measurements and an ensemble classifier of three types of neural
networks is proposed. Aiming at the difficulty to set a proper threshold for
classifying the multiple categories, the threshold selection strategy is analyzed.
24 h ECG monitoring recordings from 674 patients, which have four types of
cardiac arrhythmias, are adopted for recognition. For comparison, Support
Vector Machine (SVM) classifiers with linear and Gaussian kernels are also
applied. The experiment results demonstrate the effectiveness of the extracted
features and the better performance of the designed classifier. Our study can be
applied to diagnose the corresponding sinus rhythm and arrhythmia substrates
disease automatically in the telemedicine and computer-aided diagnosis system.

Keywords: automatic recognition, cardiac arrhythmias, Poincaré plots, heart

rate variability (HRV), electrocardiographic (ECG), telemedicine

(Some figures may appear in colour only in the online journal)

0967-3334/15/020283+19$33.00  © 2015 Institute of Physics and Engineering in Medicine  Printed in the UK 283
Physiol. Meas. 36 (2015) 283 L Zhang et al

1. Introduction

Heart rate variability (HRV) from electrocardiographic (ECG) recordings has become an
important method for assessing cardiovascular autonomic regulation (Tulppo and Huikuri
2004). Time and frequency domain measures of HRV have been most commonly used, but
studies showed that the nonlinear dynamics characteristics of HRV may be more powerful in
terms of risk stratification (Huikuri et al 1999). The Poincaré plot emerges as an effective tool
to display nonlinear characteristics of HRV from ECG, such as the asymmetry in HRV signals
is novel defined using the Poincaré plot (Karmakar et al 2009b). Moreover, the Poincaré plot
gives a global view of the distribution and statistical features of long range of ECG signals.
Therefore, it is both a useful visual tool which is capable of summarizing an entire RR time
series derived from ECG in one picture, and a quantitative technique which gives nonlinear
information on the long- and short-term HRV. The Poincaré plot attracts more and more atten-
tion due to its proven clinical ability as a predictor of cardiac disease dysfunction. Moreover,
the Poincaré plots are also applied to analyze the differentiates between hypertensive preg-
nancy disorders with beat-to-beat interval and blood pressure signals (Seeck et al 2011) and
to diagnose sleep apnea hypopnea syndrome with overnight arterial oxygen saturation signals
(Morillo et al 2009).
In ECG signals, the Poincaré plot is constructed with successive RR intervals as X and Y
axis. Different geometries of the Poincaré plots reflect the global distribution of HRV and
the underlying information of corresponding cardiovascular status. The clinical and medical
researchers explored the physiological mechanisms behind different geometric features of
Poincaré plots, built the relationship between different shapes of Poincaré plots and heart fail-
ure diseases and, as a result, the Poincaré plots can assist to make the diagnosis of cardiovas-
cular status (Kamen and Tonkin 1995, Kamen et al 1996, Esperer et al 2001, 2008, Brennan
et al 2002, Ueda et al 2002, Mourot et al 2004). In the study of Esperer et al (Esperer et al
2008), the relationship between Poincaré plots and major types of cardiac arrhythmias was
systematically investigated. To depict the geometric features of the Poincaré plots, various
descriptor were studied: ellipse shape fitting (Marciano et al 1994, Huikuri et al 1996, Tulppo
et al 1996, D’Addio et al 1998), correlation coefficient calculation (Otzenberger et al 1998),
standard deviation of major and minor axis (Brennan et al 2001), complex correlation measure
(Karmakar et al 2009a) and intensity surface (Duong et al 2009). So the existing researches
focus on how to describe the Poincaré plots quantitatively and how to explore the relationship
between the patterns of Poincaré plots and specific cardiac arrhythmias substrates.
Poincaré plot gives a global and statistical view of long range of ECG signals, which is the
significant complement for the detailed but local information from original short-term signals.
Up to current applications, the Poincaré plots are examined and analyzed with doctor’s eyes
and assist the diagnosis. If the beat-to-beat-intervals can be detected with high precision in an
accurate way, the automatic analysis of such time series via Poincaré plot enables a computer
assisted diagnosis for the medical doctor. As far as we know, there is no publication about the
automatic recognition of the Poincaré plots, which implements the clinical ability as a predic-
tor of cardiac dysfunction with computational techniques.
Therefore, we develop an automatic pattern recognition system to distinguish the geometric
patterns of the Poincaré plot in our study. It helps to diagnose major cardiac arrhythmias based
on the relationship between five types of Poincaré plots and four cardiac arrhythmias explored
in the study (Esperer et al 2008). The spots distributions with sliding angular scan along the
regions of Poincaré plot and modified CCM (Complex Correlation Measure) (Jelinek et al
2011) are extracted as features to describe the characteristics of one Poincaré plot. An on-
line learning classifier is proposed as an ensemble of three types of neural networks. The

284
Physiol. Meas. 36 (2015) 283 L Zhang et al

Table 1. Five patterns of Poincaré plots and their numbers of various patterns
in datasets.
Number of Number of Number of
Sinus rhythm and arrhythmia substrates Patterns patients training testing
Sinus rhythm Comet pattern 97 63 34
Sinus rhythm Torpedo pattern 242 163 79
Atrial fibrillation (AF) and Multifocal atrial Fan pattern 152 119 33
tachycardia (MAT)
Sinus rhythm with atrial premature beats Double side lobe 34 21 13
(DSL) pattern
Sinus rhythm with ventricular premature Triple side lobe 149 102 47
beats (TSL)

three posterior probabilities from the three types of neural networks are integrated with the
optimized weights. Furthermore, aiming to the difficulty to set a proper threshold for multiple
categories, we analyze the effect of threshold on the performance of the classifier and get the
conclusion that judging a sample to be the pattern with highest output directly will obtain the
best classifying performance. To assess the performance of the classifier, the 24 h ECG moni-
toring recordings from 674 patients of the First Affiliation Hospital of Xiamen University are
adopted. The system is able to achieve an average good detection rate of 96.1% on the test-
ing data. By contrast, the SVM with linear and Gaussian kernels achieve 90.8% and 93.2%
detection rate respectively. The classification results demonstrate that the corresponding sinus
rhythm and arrhythmia substrates disease can be diagnosed automatically with the classifier
in the telemedicine or computer aided diagnosis system.

2. Methods

In this section, the framework of the classifier in a telemedicine system, the feature extraction
and the design steps for the classifier are described. The classifier can recognize the comet,
torpedo, fan, double side lobe and multiple side lobe patterns of the Poincaré plot, conse-
quently help to diagnose sinus rhythm, atrial fibrillation, atrial flutter and atrial tachycardia
automatically. The relationship between the five patterns and the major cardiac arrhythmias
were comprehensively studied in the study (Esperer et al 2008). The correspondence relation-
ships are summarized on table 1.

2.1.  Framework of the automatic pattern recognition system

The figure 1 shows the working mechanism of the automatic classifier for Poincaré patterns
of ECG signals in the telemedicine monitoring system. In the system, there are several sen-
sors for various medical detections. The sensors’ signals are transmitted to the data sever with
wireless communications, such as 3G, Wi-Fi or Bluetooth. The data server receives the sensor
signals and implements the functions of data storage and processing. There are different data
processing modules for specific medical signals. When the ECG signals are transmitted to
the data sever, RR intervals are detected with preprocessing module. Then the RR intervals
are input into the automatic classifier for Poincaré patterns proposed in this study. Aiming to
classify the five plot patterns, the classifier is an ensemble of classifiers with pre-trained three
types of neural networks and an online-learning module. The classifier outputs the recogni-
tion results and delivers them to the doctor via wireless communication. And the doctor will

285
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 1. The framework of automatic recognition for Poincaré patterns of ECG

signals in the telemedicine system.

feedback the final decision of the pattern, which is the reference knowledge to optimize the
weight of neural network classifier via online-learning module.

2.2.  Preprocessing of ECG signals

In the telemedicine system, the time series of ECG signals are transmitted to the data server
with wireless network. The RR intervals of ECG signals are detected by our previous devel-
oped R-wave detection tool (Feng et al 2013). The detected RR intervals are applied to obtain
the Poincaré plot and extract the features for the classifications.

2.3.  Features extractions

Based on the statistical analysis of RR-interval, the occurring frequency distributions of the
points along different regions of the Poincaré plot and the nonlinear index CCM (Complex
Correlation Measure) (Jelinek et al 2011) are extracted as features to describe the Poincaré
plot in our study.

2.3.1. The first- and second-order Poincaré plots.  The RR interval series are described as a
vector RR=[x1, x2, ⋯ , xn ]T . xi is the ith RR interval of ECG signals. The first order of Poincaré
plot is composed of the plots of xi on the x-axis versus xi + 1 on the y-axis, that is, (xi, xi + 1) . The
value of all RR intervals is nonnegative, so all the data are located on the first quadrant of the
coordinate system. After removing identical RR-interval points, an angular sliding window
scan is implemented from 0 to 90 degree with 0.5 degree changing step, as shown in figure 2.

286
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 2.  Scheme of angle-window sliding of the first-order of Poincaré plot.

In order to smooth the distribution, the width of the angular window is 1 degree. For each slid-
ing window, the number of points in this window was counted. The points counting numbers
of the 180 sliding windows constitute the histogram of the Poincaré plot. The first-order Poin-
caré plot and the corresponding histograms of the five patterns are shown in figures 3(a)–(e)
and (f)–(j) respectively.
It is easy to observe that these patterns exhibit different shapes, height of peaks, num-
ber of peaks, or distributions in the histogram. Traditional processing way is looking for
some representative measures to figure out the distinctive characteristics of each pattern as
the features of the classifier. In our study, instead, we use the 180 frequency counts in the
sliding window scan as the features to input into the designed ensemble of neural network
classifiers directly.
The second-order of Poincaré plots is a plot of xi − xi − 1 on the x-axis versus xi + 1 − xi on
the y-axis, that is, the (xi − xi − 1, xi + 1 − xi ) . Different with the first-order Poincaré plot, the
points are located over all four quadrants of coordinate system. Therefore, an angular sliding
window scan is implemented from 0 to 360 degree with 1 degree changing step after remov-
ing identical points. The window width is 4 degree for smoothing. The generation of histo-
gram is similar with that in the first-order Poincaré plot. For each sliding window, the number
of occurring points is counted to compose the histogram. The second-order of Poincaré plots
and the corresponding histograms of the five patterns are shown in figures 4(a)–(e) and (f)–(j)
respectively.
The histograms of the second-order shows more complicated characteristics than that in
the first-order Poincaré plots. Although each patterns exhibit more diverse outlines, it is more
hard to describe them with simple statistic measures, as shapes, height of peaks or number of
peaks. Therefore, we use the 360 points-counting numbers from the sliding window scan as
the features to input into the designed ensemble neural network classifier directly.

287
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 3.  The first-order of Poincaré plots and angle histograms of the patterns. (a)
The first-order Poincaré plots of Comet pattern. (b) The first-order Poincaré plots of
Torpedo pattern. (c) The first-order Poincaré plots of Fan pattern. (d) The first-order
Poincaré plots of DSL pattern. (e) The first-order Poincaré plots of TSL pattern. (f) The
histogram of first-order Poincaré plots of Comet pattern. (g) The histogram of first-
order Poincaré plots of Torpedo pattern. (h) The histogram of first-order Poincaré plots
of Fan pattern. (i) The histogram of first-order Poincaré plots of DSL pattern. (j) The
histogram of first-order Poincaré plots of TSL pattern.

288
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 4. The second-order of Poincaré plots and angle histograms of the patterns.
(a) The second-order Poincaré plots of Comet pattern. (b) The second-order Poincaré
plots of Torpedo pattern. (c) The second-order Poincaré plots of Fan pattern. (d) The
second-order Poincaré plots of DSL pattern. (e) The second-order Poincaré plots of
TSL pattern. (f) The histogram of second-order Poincaré plots of Comet pattern. (g)
The histogram of second-order Poincaré plots of Torpedo pattern. (h) The histogram of
second-order Poincaré plots of Fan pattern. (i) The histogram of second-order Poincaré
plots of DSL pattern. (j) The histogram of second-order Poincaré plots of TSL pattern.

289
Physiol. Meas. 36 (2015) 283 L Zhang et al

2.3.2.  Complex correlation measure of Poincaré plots.  Complex Correlation Measure (CCM)
(Karmakar et al 2009a) was proposed to measure the nonlinearity of HRV from Poincaré plot.
In the study (Karmakar et al 2009a), the CCM was calculated as:
N −m−1
1
CCM
 (m ) = ∑ A (i ) (1)
π * SD1 * SD2 * (N − 2) i=1

Where
xi xi + m 1
1
A (i ) = xi + 1 xi + m + 1 1 ,
2
xi + 2 xi + m + 2 1
and
SD12 = γRR (0) − γRR (m ) ,

SD22 = γRR (0) + γRR (m ) − 2RR 2,

γRR (0) and γRR (m ) is the autocorrelation function for lag-0 and lag-m RR interval and RR is
the mean of RR-intervals. The detail can be found in the paper (Karmakar et al 2009a).
During the CCM’s application in our study, we make the following two modifications:
(a) The original definition of CCM in paper (Karmakar et al 2009a) used SD1 * SD2 as a
shape factor to describe the ellipse shape of the first-order Poincaré plot, which gener-
ally happened to the patients with sinus rhythm. If more disease patterns are taken into
considerations, SD1 / SD2 can characterize different shapes more distinctively. Therefore,
in our study, we instead use the following formula (2) to define CCM:

N −m−1
SD1
CCM
 (m ) = ∑ A (i ) (2)
SD2 * (N − 2) i=1

(b) In paper (Karmakar et al 2009a), CCM was calculated with the following formula (for-
mula (11) in that paper):

⎡ N−m
CCM (m ) =
1 ⎢uN − muN − 1 + u2u1 + m − uN − 1 − muN − u1u2 + m + ∑ u iui − 2 + m
2Cn (N − 2) ⎢⎣ i=3
N−m N −1−m N −2−m ⎤
−2 ∑ uiui − 1 + m + 2 ∑ uiui + 1 + m − ∑ uiui + 2 + m ⎥
i=2 i=1 i=1 ⎥⎦

However, after double check, the formula was obtained with calculating the ∑iN=−1m − 1 A (i )
N −m−1
instead of ∑i = 1 A (i ) in the definition of CCM (m ) . Therefore, in our study, we use for-
mula (2) to calculate the CCM directly.
We calculate the distribution of CCM values for the patterns, as shown in table  2 and
­figure 5. It is easy to observe that with CCM, the comet and torpedo patterns have distinctive
difference from the other patterns, which can separate sinus rhythm out easily. In the boxplot
of figure 5, the sinus means the comet and torpedo patterns, and the non-sinus corresponding
to the rest four patterns. There is a big gap between the mean values of the two groups.

290
Physiol. Meas. 36 (2015) 283 L Zhang et al

Table 2. The mean and standard deviation about CCM of each patterns
(RR-interval data in second).
Patterns Mean of CCM Standard of CCM
Comet pattern 6.34  ×  10−4 8.58  ×  10−4
Torpedo pattern 1.56  ×  10−4 5.33  ×  10−4
Fan pattern 2.89  ×  10−2 2.55  ×  10−2
Double side lobe (DSL) pattern 1.24  ×  10−2 2.13  ×  10−2
Triple side lobe (TSL) pattern 1.58  ×  10−2 2.69  ×  10−2

Figure 5.  The boxplot of the CCM values of pure sinus and non-pure sinus patterns.

2.4.  An ensemble with on-line learning neural network classifiers

2.4.1. The classifier design.  Galina (Rogova 1994) proposed to combine the results of sev-
eral neural networks to improve the classification results in 1994. And it was applied to vari-
ous pattern classification targets (Giacinto 1997, Ghazikhani et al 2013). In this study, we
designed an ensemble classifier combining the Backpropagation neural network (BP), the
Multilayer Perceptron neural network (MLP), and the Radial Basis Functions neural network
(RBF) to classify the five patterns of Poincaré plots from the RR-Intervals of ECG directly.
The posterior probability from each neural network is combined with the optimized weight to
output the estimation of real posterior probability, which is shown in figure 6. There are 542
features as the inputs of the neural network. For the MLP neural network, architecture with
two hidden layers (450-20-10-5) was selected after several trials and testing. For the RBF

291
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 6.  Scheme of the ensemble classifier of neural networks.

neural network, architecture with one hidden layers (450-30-5) was selected. For the BP neu-
ral network, architecture with two hidden layers (450-20-10-5) was selected.
For multiple C classes (ω1, ω 2 … ωC ) , the real posterior probability of the input sample X is:
q (X ) = [p (ω1 X ) , p (ω 2 X ) … p (ωC X )]T .
For the ith classifier, the estimated posterior probability is:

qi ̂ (X ) = [pi (ω1 X ) , pi (ω 2 X ) , … , pi (ωC X )] , i = 1 ⋯ M

T

and M is the number of classifiers. If the weight of each classifier is wi, i = 1, 2 … M , the
final estimated posterior probability is q ̂ (X ) = ∑i = 1 wiqi ̂ (X ) . As discussed in paper (Wu et al
M

1998), the optimization of wi is:

 E q (X ) − q ̂ (X )
min 2
(3)
that is:
dE{ q ( X ) − q (̂ X ) }2
dwk
⎛ C ⎞
 d⎜⎜ ∑ E p ωj X − ∑ wipi ωj X ⎟
M

(( ) ( )) 2
⎟ (4)
⎝ j=1 i=1 ⎠
=
dwk
= 0.
From (4), we get the formula (5):
M C C
 wi ∑ E {pi (ωj
∑ ) (
X pk ωj X )} {(
= ∑ E p ωj X pk ωj X) ( ) }. (5)
i=1 j=1 j=1

C C

{ (
Let aik = ∑ E pi ωj X pk ωj X
j=1
) ( ) }, {(
bk = ∑ E p ωj X pk ωj X
j=1
) ( ) }, then the

formula (5) is written as:

AW = B (6)
where A = (aik ) M × M , B = (bi ) M × 1 , and W = (wk ) M × 1

292
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 7.  The scheme of the real time adjustment for the neural networks with the
doctor’s feedback.

The A and B can be obtained with statistical estimation:

N C
1
a ik̂ = ∑ ∑ pi ̂ (ωj ) (
Xn pk̂ ωj Xn ) (7)
CN n=1 j=1

N C
1
b i ̂ = ∑ ∑ p (ωj ) (
Xn pk̂ ωj Xn ) (8)
CN n=1 j=1

N is the number of training samples X1, X2, ..., XN . The solution of equation (6) is a typical
optimization problem, which can be solved with the classical optimization methods, such as
Least Square method.

2.4.2.  On-line learning of the neural network classifier.  One of the advantageous of telemedi-
cine system is the participation of doctors. As shown in figure  7, based on the transferred
sensor signals from patients, each data processing modules make the decisions with their
inside specific algorithms. The decision by the computer aided modules will be sent to the
doctor combined with the original signals via the communication channel of the telemedicine
system. Then, with the original signals, the doctor will feedback whether the automatic diag-
nosis results are correct or not to the data processing modules. For our study, if the automatic
decision for ECG signals is not correct, the wrong sample will resend to the neural network
classifier with proper punishment cost to retrain the neural networks and adjust their weights
to improve the performance of the classifier. The adjustment algorithm is very similar with the
general neural network training method.

3. Results

3.1.  Dataset description

The 24 h ECG signals from 674 patients of The First Affiliation Hospital of Xiamen University
are applied for the training and testing of the ensemble classifier of neural networks. The ECG
signals were recorded with Holter recorder DM S300-4A, which is the wearable 12 lead syn-
chronous ambulatory wearable electrocardiogram monitors with 256–4096 Hz frequency. All
patients’ signals were annotated by a specialist in cardiology with the aid of CardioScan12.
TOP ECG Holter analyzer system. The patients contained a variety of types of sinus rhythm

293
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 8.  The ROC curves of testing data for the ensemble classifier of neural
networks.

and arrhythmia substrates. The specific cardiac arrhythmias substrates are shown in the third
columns of table 1. Among all the 674 samples, there are more than 200 patients with sinus
rhythm of Torpedo pattern. Except the patients of sinus rhythm with atrial premature beats, the
samples of other types are more than or close to 100 cases. After the preprocessing process,
the R waves are detected and the corresponding RR interval series are produced. For each
sample, a 541-elements feature vector describes the ECG signals.

3.2.  Experiment results

70% of the samples, that is 468 patients are selected randomly to apply for the training of
the three neural networks. The training samples’ distribution for each pattern is shown in the
fourth column of table 1. Among the training samples, 70% of the samples are used for train-
ing the neural networks and 15% are used to validate the generalization of the network and to
stop training before overfitting. The last 15% are used as a completely independent test of the
generalization of the network.
The posterior possibility of each network is integrated with optimized weights to output the
final posterior possibility of each pattern. The remained 30% of samples, that is 206 patients,
are applied to test the performance of the ensemble classifier of the neural networks. The ROC

294
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 9.  The error rate analysis of the classifier. (a) The curves of error and rejecting
rate with the change of threshold for the first strategy. (b) The curves of error and
rejecting rate with the change of threshold for the second strategy. (c) The total error
rate curves for the first and second strategy.

curves and enlarged detail are shown in figure 8. For each pattern, the ROC curves are very
close to (0,1) point, which means the classifier shows significant sensitivity and specificity.
The threshold of the posterior possibility for pattern decisions has much effect on the per-
formance of the classifier. For two-category problem, the threshold is selected with the opti-
mal point of ROC curve. In our study, for the multi-category case, it is hard to obtain the
consistent optimal threshold for multiple ROC curves. For example, if the output posterior
possibility vector is [0.1 0.06 0.15 0.41 0.47]T and the threshold is set to 0.4, the output of 4th
and 5th pattern is higher than the threshold. Then we might have two decision strategies: we
can decide the sample is the 5th pattern because of the higher posterior possibility and we can
also reject to make a decision because of the closeness output value of the 4th and 5th patterns.
Then how to set the threshold and how to make a decision for multiple patterns with output
higher than the threshold?
Therefore, we analyze the effect of threshold on the performance of classifier under two
decision strategies: There are multiple categories with higher output than the threshold. The
first strategy is to reject making the decision; the second strategy is to judge the sample to be
the pattern with highest output.

295
Physiol. Meas. 36 (2015) 283 L Zhang et al

For the first decision strategy, the figure 9(a) shows the tendency of error rate and reject-
ing rate of the classifier with the change of threshold. When the threshold is in [0.25, 0.85],
the rejecting rate keeps about 10% and when the threshold is in [0, 0.25] and [0.85, 1], the
rejecting rate increase to 100% rapidly. The error rate shows an opposite tendency, but with
the threshold changing from 0 to 1, the classifying rate does not change much, on the range of
[0, 2.9%]. This means that if we do not take the rejecting rate into consideration, the designed
ensemble neural network classifier keeps robust to the threshold. And the rejecting error rate
affects the classifier performance greatly under the first strategy.
For the second decision strategy, the figure 9(b) shows the tendency of error rate and reject-
ing rate of the classifier with the change of the threshold. Compared with the first strategy,
the two curves present monotonically increments or decrements. The two rates still show the
opposite changing tendency with the increase of the threshold. With the threshold increasing
from 0 to 1, the error rate does not change much, dropping from 3.9% to 0%, and the rejecting
rate increases from 0% to 100%. When the threshold is in [0, 0.85], the rejecting rate does not
change much and is no higher than 10%. When the threshold is in [0.85, 1], the rejecting rate
increases to 100% rapidly.
The total error rates tendency of the two strategies are shown in figure 9(c). And the first
strategy shows the error rate with a depressed shape with peaks on both sides and trough on
the middle, while the second strategy shows a monotonically increment tendency of the total
error rates. For the first strategy, the minimum total error rate is 5.8% when the threshold is
0.45. For the second strategy, when the threshold is smaller than 0.85, the total error rate is
no higher than 10%. The minimum error rate is 3.9% when the threshold has a value between
[0, 4.5]. From the two total error rate curves of the two strategies, it is obvious to see that the
error rates of the first decision strategy is higher than those of the second one at almost each
threshold value. So the comparison shows that the second strategy is superior to the first one
in the total error rate. It means that there is no need to set the threshold and we can decide the
sample to be the pattern with maximum output directly. This conclusion is consistent with the
general processing way for the neural network classifier.
Therefore, in our study, we judge the sample to be the pattern with highest output directly,
that is, the second strategy with 0 as the threshold. For the threshold is 0.45 and 0 for the
first and second decision strategy respective, the corresponding confusion matrixes are shown
in figures  10(a) and (b). The accuracy rates for the two strategies are 94.21% and 96.12%
respectively. In our study, there are totally 674 patients. Among them, 468 patients are ran-
domly selected for classifier training and the remained 206 patients are applied for testing. As
shown in figure 10(b), the minimum misclassification rate for testing data are 3.88%, that is,
8 patients. The misclassification distribution can be found at the blocks out of the diagonal. 4
of the 34 Comet are misclassified as Torpedo; 1 of the 79 Torpedo is misclassified as Comet; 1
of the 13 DSL is misclassified as Comet; 2 of the 46 TSL is misclassified as Comet and DSL
respectively.
For the multiple-category classifications, Kappa statistics (Carletta 1996) are applied to
test the statistical significance of the classification results, as shown in table 3. The complete
agreement between the ground truth and classification results corresponds to Kappa = 1, and
the lack of agreement (i.e. purely random coincidences of rates) corresponds to Kappa = 0.
The closer to 1 the Kappa coefficient is, the more correct the classification is. For the 206 test-
ing samples, the Kappa coefficient is 0.92 and the p-value of the first strategy is 4.90  ×  10−65.
For the second strategy, the Kappa coefficient is 0.95 and the p-value is 5.76  ×  10−71. The
Kappa coefficient is highly close to 1 with almost 0 p-value, which means the classification
results are highly consistent with the ground truth with great statistical significance.

296
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 10. The confusion matrix of the neural network ensemble classifier. (a)
confusion matrix of the first threshold selection strategy. (b) The confusion matrix of
the second threshold selection strategy.

For comparison, with same selected features, the SVM (Support Vector Machine (Cortes
and Vapnik 1995)) classifiers with linear and Gaussian kernels are trained to classify the
same dataset respectively. This is a multi-category classification case and we use ‘one
versus the rest’ strategy to design the SVM classifiers for each pair of categories. For the

297
Physiol. Meas. 36 (2015) 283 L Zhang et al

Table 3.  The accuracy, Kappa coefficient and p-value of the classification
results.
Neural network ensemble classifier SVM classifier
Statistical First threshold Second threshold Linear Gaussian
significance strategy strategy kernel kernel
Accuracy 94.2% 96.1% 90.8% 93.2%
Kappa statistics 0.92 0.95 0.88 0.91
P-value 4.90  ×  10−65 5.76  ×  10−71 5.09  ×  10−56 2.62  ×  10−62

implementation of SVM, the toolbox LibSVM of Matlab was used in the study (Chang and
Lin 2011). For linear kernel SVM, the regularization parameter C was chosen from the set
{1, 10, 20, 50, 100} in internal cross validation. For the Guassian kernel, two parameters
need to be specified, i.e. the parameter γ that controls the width of the kernel and a regulari-
zation parameter C. To determine the optimal set of parameters γ and C, a grid search with
internal cross validation was perfromed (Chang and Lin 2011). SVM aims at minimizing an
upper bound of the generalization error through maximizing the margin between the separat-
ing hyperplane and the data. The maximum margin hyperplane gives the maximum separa-
tion between decision classes. The maximization of the separation is a quadratic program
which is a convex constrained optimization problem, and thus has a unique solution. In our
study, the margins for SVM classifiers of [Comet Torpedo Fan DSL TSL] are [0.97 1.04
3.66 2.34 3.80] and [0.0055 0.0090 0.0243 0.0049 0.0246] for Linear and Gaussian kernels
respectively. The margins between two kernels cannot be compared directly because of the
difference scaling spaces.
The corresponding confusion matrixes are shown in figures 11(a) and (b). The accuracy
rates for the linear and Gaussian kernels are 90.8% and 93.2% respectively. The Kappa statis-
tics results are also as shown in table 3. For linear kernel, Kappa coefficient is 0.88 and p-value
is 2.09  ×  10−56; for Gaussian kernel, Kappa coefficient is 0.91 and p-value is 2.62  ×  10−62.
From the comparison, it is easy to find that our extracted features are valid for other
classifier, and our proposed ensemble neural network classifier perform better than SVM
classifier. The experiment verifies that our proposed classifier achieves highly classifying
accuracy with great statistical significance. The different threshold decision strategies lead
to 2% difference in error rate. No matter what strategies it takes, the classifier keeps excel-
lent performance. It shows that the designed classifier is valid for automatic diagnosis of the
corresponding sinus rhythm and arrhythmia substrates disease, which can be integrated into
the telemedicine system.

4.  Conclusions and discussions

The Poincaré plot is both a useful visual tool which is capable of summarizing an entire
RR time series derived from ECG in one picture, and a quantitative technique which gives
nonlinear information on the long- and short-term HRV. Study shows the geometric patterns
of the Poincaré plots have the corresponding relationship with various cardiac arrhythmias.
The automatic recognition of geometric patterns of the Poincaré plots would be very useful
in the telemedicine system. As far as our literature surveys, there is no publication about the
automatic recognition of the geometric patterns of Poincaré plots. Therefore, in this study,
an ensemble classifier with neural networks is proposed combined with statistics features to
recognize the pattern of Poincaré plots. Based on the successive RR intervals, it can classify

298
Physiol. Meas. 36 (2015) 283 L Zhang et al

Figure 11.  The confusion matrix of the SVM classifier. (a) Confusion matrix of the
linear kernel. (b) The confusion matrix of the Gaussian kernel.

the five patterns of Poincaré plots automatically with satisfactory performance. As a result,
the corresponding sinus rhythm and arrhythmia substrates disease can be diagnosed automati-
cally. It will be a key module in the telemedicine system.

299
Physiol. Meas. 36 (2015) 283 L Zhang et al

In this study, our goal is to classify the five patterns to assist to diagnose four individual
cardiac arrhythmias. In clinical practice, a combination of hearth rhythms may occur, such as
paroxysmal AF or a mix of atrial and ventricular ectopic. We will attempt to study the cases
of multiple cardiac arrhythmias combination in the future study. And we plan to analyze this
problem through automatic segmentations and then identify the piecewise model with homo-
geneously statistical properties.

Acknowledgment

The research is partly supported by the National Natural Science Foundation of China
(61203282, 61202144), Information Platform of third stage of 985 in Xiamen University,
Telemedicine subproject, PI: Clarence de Silva; and National University students Innovation
Training Project of Xiamen University (2014, PI:Kun Wang)

References

Brennan M, Palaniswami M and Kamen P 2001 Do existing measures of Poincaré plot geometry reflect
nonlinear features of heart rate variability? IEEE Trans. Biomed. Eng. 48 1342–7
Brennan M, Palaniswami M and Kamen P 2002 Poincaré plot interpretation using a physiological model
of HRV based on a network of oscillators Am. J. Physiol. Heart Circ. Physiol. 283 H1873–86
Carletta  J 1996 Assessing agreement on classification tasks: the kappa statistic Comput. Linguist.
22 249–54
Chang C C and Lin C J 2011 LIBSVM: a library for support vector machines ACM Trans. Intell. Syst.
Technol. 2 27
Cortes C and Vapnik V N 1995 Support vector networks Mach. Learn. 20 273–97
D’addio G, Acanfora D, Pinna G D, Maestri R, Furgi G, Picone C and Rengo F 1998 Reproducibility of
short- and long-term poincare plot parameters compared with frequency-domain HRV indexes in
congestive heart failure Comput. Cardiol. 25 381–4
Duong N D, Jeong H, Youn C-H and Kim D 2009 Development of new cluster descriptors for image
analysis of poincaré plots IFMBE Proc. World Congress on Medical Physics and Biomedical
Engineering (Munich, Germany)
Esperer H, Esperer C and Cohen R 2008 Cardiac arrhythmias imprint specific signatures on Lorenz plots
Ann. Noninvasive Electrocardiol. 13 44–6
Esperer  H D, Esperer  M, Esperer  H C and Chernyak  Y B 2001 Heart rate variability assessed from
Lorenz plots predicts cardiovascular morbidity and risk in apparently healthy men Europace
pp 217–21
Feng K, Wang Y, Hong S and Al E 2013 A valid method of inverse R wave detection J. Comput. Inform.
Syst. 9 1027–34
Ghazikhani A, Monsefi R and Sadoghi Y H 2013 Ensemble of online neural networks for non-stationary
and imbalanced data streams Neurocomputing 122 535–44
Giacinto G and Roli F C 1997 Ensembles of neural networks for soft classification of remote sensing
images European Symp. on Intelligent Techniques (Bari, Italy)
Huikuri H V, MaKikallio T and Airaksinen K E J 1999 Measurement of heart rate variability: a clinical
tool or a research toy? J. Am. Coll. Cardiol. 34 1878–83
Huikuri H V, Seppanen T, Koistinen M, Juhani A, Ikaheimo M J, Castellanos A and Myerburg R J 1996
Abnormalities in beat-to-beat dynamics of heart rate before spontaneous onset of life-threatening
ventricular tachyarrhythmias in patients with prior myocardial infarction Circulation 93 1836–44
Jelinek  H F, Khandoker  A H, Quintana  D S, Imam  M H and Kemp  A H 2011 Complex correlation
measure as a sensitive indicator of risk for sudden cardiac death in patients with depression
Comput. Cardiol. 38 809–12 Hangzhou
Kamen  P W and Tonkin  A M 1995 Application of the Poincaré plot to heart rate variability: a new
measure of functional status in heart failure Aust. N Z J. Med. 25 18–26
Kamen P, Krum H and Tonkin A 1996 Poincar´e plot of heart rate variability allows quantitative display
of parasympathetic nervous activity in humans Clin. Sci. (Lond.) 91 201–8

300
 Physiol. Meas. 36 (2015) 283 L Zhang et al

Karmakar C K, Khandoker A H, Gubbi J and Palaniswami M 2009a Complex correlation measure: a

novel descriptor for Poincaré plot BioMed. Eng. Online 8 37–48
Karmakar C K, Khandoker A H, Gubbi J and Palaniswami M 2009b Defining asymmetry in heart rate
variability signals using a Poincar´e plot Physiol. Meas. 30 1227–40
Marciano F, Migaux M L, Acanfora D, Furgi G and Rengo F 1994 Quantification of Poincare maps for
the evaluation of heart rate variability Comput. Cardiol. 557–80
Morillo D S, Rojas J L, Crespo L F, Le´on A and Gross N 2009 Poincaré analysis of an overnight arterial
oxygen saturation signal applied to the diagnosis of sleep apnea hypopnea syndrome Physiol.
Meas. 30 405–20
Mourot L, Bouhaddi M and Perrey S 2004 Quantitative Poincaré plot analysis of heart rate variability:
effect of endurance training Eur. J. Appl. Physiol. 91 79–87
Otzenberger  H, Gronfier  C, Simon  C, Charloux  A, Ehrhart  J, Piquard  F and Brandenberger  G 1998
Dynamic heart rate variability: a tool for exploring sympathovagal balance continuously during
sleep in men Am. J. Physiol. 257 946–50
Rogova G 1994 Combining the results of several neural network classifiers Neural Netw. 7 777–81
Seeck  A, Baumert  M, Fischer  C, Khandoker  A, Faber  R and Voss  A 2011 Advanced Poincaré plot
analysis differentiates between hypertensive pregnancy disorders Physiol. Meas. 32 1611–22
Tulppo M and Huikuri H V 2004 Origin and significance of heart rate variability J. Am. College Cardiol.
43 2278–80
Tulppo M, Makikallio T H, Takala T E S, Seppanen T and Kuikuri H 1996 Quantitative beat-to-beat
analysis of heart rate dynamics during exercise Am. J. Physiol. 71 244 –52
Ueda T, Nakatsu T and Yamane S 2002 Correlation of Lorenz scatter plots with frequency-domain heart
rate variability Clin. Exp. Hypertens. 24 11–21
Wu Y, Fang Ch, Lin X and Ding X 1998 Knowledge Extraction Based Multilayer Feedforward Neural
Networks ICONIP’98. (Japan) 931–36

301

View publication stats