14th European Signal Processing Conference (EUSIPCO 2006), Florence, Italy, September 4-8, 2006, copyright by EURASIP

NEURAL NETWORK BASED ARRHYTHMIA CLASSIFICATION USING HEART

RATE VARIABILITY SIGNAL

Babak Mohammadzadeh-Asl, Seyed Kamaledin Setarehdan

Control and Intelligent Processing Center of Excellence, Electrical and Computer Engineering Department, Faculty of
Engineering, University of Tehran, Tehran, Iran
phone: + (98) 21 88020403 Ext 3377, fax: + (98) 21 88633029, email: b.mohammadzadeh@ece.ut.ac.ir

ABSTRACT (PVC), atrial fibrillation (AF), ventricular fibrillation (VF)

Heart Rate Variability (HRV) analysis is a non-invasive tool and and 2° heart block (BII).
for assessing the autonomic nervous system and specifically The HRV arrhythmia data, obtained using the ECG data
it is a measurement of the interaction between sympathetic from the MIT-BIH Arrhythmia Database which was digitized
and parasympathetic activity in autonomic functioning. In at a sampling rate of 360Hz. Moreover, due to the lack of the
recent years, HRV signal is mostly noted for automated VF data in the MIT-BIH arrhythmia database, the Creighton
arrhythmia detection and classification. In this paper, we University Ventricular Tachyarrhythmia Database was
have used a neural network classifier to automatic resampled at 360 Hz and then used for the VF arrhythmia
classification of cardiac arrhythmias into five classes. HRV class.
signal is used as the basic signal and linear and nonlinear Our analysis is carried out in three stages. First a
parameters extracted from it are used to train a neural preprocessing procedure is used to extract tachograms from
network classifier. The proposed approach is tested using the the ECGs. In this stage we have used Tompkins algorithm [2]
MIT-BIH arrhythmia database and satisfactory results were for detection of R peaks. The tachograms are segmented into
obtained with an accuracy level of 99.38%. small segments. Each segment contains 32 RR-intervals and
is characterized using the MIT-BIH arrhythmia database an-
1. INTRODUCTION notation. In the second stage, time and frequency domain and
nonlinear methods are applied to extract corresponding fea-
One of the methods which is mostly noted by specialists, for tures. In the third stage the extracted features are used to train
assessing the heart activity and discrimination of cardiac a neural network classifier.
abnormalities, is called Heart Rate Variability (HRV). HRV Next materials and methods are described. Then the
is a nonlinear and nonstationary signal that represents the different steps of the proposed algorithm are explained.
autonomic activity and its influence on the cardiovascular Finally results obtained on the MIT-BIH arrhythmia database
system. Hence, measurement of heart rate variations and are presented.
computerized analysis of it is a non-invasive tool for
assessing the autonomic nervous system and cardiovascular 3. FEATURE EXTRACTION
autonomic regulation . Furthermore, it could give us
information about heart deficiency at the present or in the The methods for HRV analysis can be divided into linear
future. (time and frequency domain) and nonlinear methods.
An automated method for the classification of cardiac In this work, we explored a combination of linear and
arrhythmias is proposed based on linear and non-linear nonlinear features.
analysis of HRV. Time and frequency domain measures in
3.1 Time domain analysis
heart rate variability analysis are less successful in the
classification of multiple rhythm changes. With the help of The time domain methods are the simplest to perform and
measures from non-linear dynamics we can quantify some various parameters can be extracted by time domain analysis
of the complex structures in heart rate time series [1]. of the segments [3]. We utilized five time domain parameters
Therefore, we have used a combination of linear and as follows:
non-linear parameters. These features are used as input in an Mean: Mean of all RR intervals in each segment
artificial neural network (ANN), which classifies each Rmssd: Root mean square successive difference of
segment into one of the arrhythmia classes. intervals in each segment
SDNN: Standard deviation of the RR intervals in each
2. MATERIALS AND METHODS segment
SDSD: Standard deviation of differences between
In this paper, we explore the HRV signal as the basic signal
adjacent RR intervals in each segment
to classify cardiac arrhythmias into five classes : normal
sinus rhythm (NSR), premature ventricular contraction
 14th European Signal Processing Conference (EUSIPCO 2006), Florence, Italy, September 4-8, 2006, copyright by EURASIP

pNN50: Number of successive difference of intervals entropy gives an estimate of the spectral entropy of the
which differ by more than 50 ms divided by the total number process, where entropy is given by
of all RR intervals in each segment
⎛ 1 ⎞
3.2 Frequency domain analysis H = ∑ p f log⎜ ⎟ (1)
⎜p ⎟
Time domain methods are computationally simple but lack f ⎝ f ⎠
the ability to discriminate between sympathetic and para-
sympathetic contributions of HRV [4]. Spectral analysis is where p f is the PDF (probability density function) value at
the most popular linear technique used in the analysis of
HRV signals [5]. Spectral power in the high frequency (HF) frequency f [5].
(0.15-0.4 Hz) band reflects respiratory sinus arrhythmia Heuristically, the entropy is enterpreted as a measure of
(RSA) and thus cardiac vagal activity. Low frequency (LF) uncertainty about the event at f . Thus entropy can be used
(0.04-0.15 Hz) power is related to baroreceptor control and
is mediated by both vagal and sympathetic systems [3,5]. as a measure of system complexity. The spectral entropy H
We used one frequency domain parameter which is describes the complexity of the HRV [5].
LF/HF: Ratio between LF and HF band powers
3.3.4. LLE
3.3 Nonlinear analysis Lyapunov exponent is simply a measure of how fast two
initially nearby points on a trajectory will diverge from each
A complex system like cardiovascular system can not be
other as the system evolves, thus giving information about
linear in nature and by considering it as a nonlinear system
the system’s dependence on initial conditions [12]. A
can lead to better understanding of the system dynamics. We
positive Lyapunov exponent is a strong indicator of chaos
utilized five nonlinear parameters in this work as follows:
[13,14]. Even though an m dimensional system has m
Lyapunov exponents, in most applications it is sufficient to
3.3.1. SD1/SD2
compose only largest Lyapunov exponent (LLE).
A relatively recent tool for HRV analysis is the Poincaré
The average largest Lyapunov exponent is calculated as
plot, which does not require the HRV signal to be stationary
follows. First, a starting point is selected in the reconstructed
[6]. Poincaré plot is a graphical representation of the
phase space and all the points which are closer to this point
correlation between successive RR intervals, i.e. plot of
than a predetermined distance, ε ,are found. Then the
RR(n + 1) as a function of RR(n) as described in Figure average value of the distances between the trajectory of the
1. The Poincaré plot may be analyzed quantitatively by initial point and the trajectories of the neighboring points are
calculating the standard deviations of the distances of the calculated as the system evolves. The slope of the line
RR(i ) to the lines y = x and y = − x + 2 * RRm , obtained by plotting the logarithms of these average values
versus time gives the LLE. To remove the dependence of
where RRm is the mean of all RR(i ) . The standard calculated values on the starting point, the procedure is
deviations are referred to as SD1 and SD2 , respectively. repeated for different starting points and the average is taken
as the average LLE [15].
SD1 related to the fast beat-to-beat variability in the data,
and SD2 described the longer-term variability of RR(i ) . 3.3.5. DFA
The detrended fluctuation analysis (DFA) is used to quantify
The ratio SD1 / SD2 may also be computed to describe the the fractal scaling properties of short interval R-R interval
relation between these components [7]. signals. This technique is a modification of the root-mean-
square analysis of random walks applied to nonstationary
3.3.2. ApEn signals [16].
Approximate entropy (ApEn) is a regularity statistic that The root-mean-square fluctuation of an integrated and
quantifies the unpredictability of fluctuations in a time detrended time series is measured at different observation
series. ApEn reflects the likelihood that similar patterns of windows and plotted against the size of the observation
observations will not be followed by additional similar window on a log-log scale [4].
observations. A time series containing many repetitive First, the R-R time series (of total length N) is integrated
patterns has a relatively small ApEn and a less predictable using the equation:
(i.e., more complex) process has a higher ApEn [8]. We have k
used the method proposed in [9] for calculating the ApEn
where m (pattern length) set to 2 and r (criterion of y(k) = ∑ ( RR ( i ) − RRavg ) (2)
i =1
similarity) set to 20% of the standard deviation of the
segment, as proposed in [10].
where y (k ) is the kth value of the integrated series, RR(i )
3.3.3. SpEn
is the ith inter beat interval and RRavg is the average inter
Spectral entropy (SpEn) quantifies the spectral complexity
of the time series [11]. Application of Shannon’s channel beat interval over the entire series.
 14th European Signal Processing Conference (EUSIPCO 2006), Florence, Italy, September 4-8, 2006, copyright by EURASIP

(a) (b)
Figure 1. Poincaré plot of (a) normal subject (b) heart rate with AF

Then, the integrated time series is devided into windows the appropriate class. The training of the neural network ends
of equal length, n. In each window of length n, a least- if the sum of the square errors for all segments is less than
squares line is fitted to the R-R interval data (representing the 0.01 or the maximum number of training epochs is reached
trend in that window). The y coordinate of the straight line (2000 epochs). The number of data set used for training and
segments are denoted by y n ( k ) . Next, we detrend the testing of the neural network classifier and the results for
each class are listed in Table2.
integrated time series, y n ( k ) , in each window. The root-
mean-square fluctuation of this integrated and detrended 5. RESULTS AND DISCUSSION
series is calculated using the equation: To evaluate the performance of the proposed classifier, three
measures are used and defined as follows:
1 N
F(n) = ∑ [ y(k) − yn (k)]2 . (3) TP
N k =1 Sensitivity (%) = × 100 (4)
TP + FN
This computation is repeated over all time scales TN
Specificity (%) = × 100 (5)
(window sizes) to obtain the relationship between F ( n) and TN + FP
window size n (i.e., the number of beats in a window that is (TP + TN )
the size of the window of the observation). Typically, F ( n) Accuracy = × 100 (6)
(TP + FN + TN + FP)
will increase with window size. The fluctuation in small
windows related to the fluctuations can be characterized by a
scaling exponent (selfsimilarity factor), α , representing the
where TP, TN, FP, and FN stand for true positive, true
negative, false positive and false negative, respectively. If
slope of the line relating log F ( n) to log n [4]. for example a segment of HRV with the VF arrhythmia is
classified as the VF, then it is said that the segment is
4. NEURAL NETWORK CLASSIFIER classified TP. On the other hand if a non-VF segment is clas-
Artificial neural networks (ANNs) are biologically inspired sified as non-VF, then it is said that the segment is classified
networks that are useful in application areas such as pattern TN. Any non-VF segment which is classified a VF segment
recognition, classification etc.. The decision making process by mistake will produce a FP, while any VF segment which
of the ANN is holistic, based on the features of input is classified a non-VF segment by mistake will produce a
toan dien FN result.
patterns, and is suitable for classification of biomedical data.
Typically, multilayer feed forward neural networks can be For the evaluation of proposed classifier, a total of 1317
trained as non-linear classifiers using the generalized back segments, which are obtained on the MIT-BIH arrhythmia
propagation algorithm [5,17]. database, were used and it consisted of 835 NSR segments,
The features extracted from linear and nonlinear analysis 57 PVC segments, 322 AF segments, 78 VF segments and 25
are used to train a back propagation neural network. The BII segments. Table1 shows the results of classification of
chosen architecture of the neural network contains: 11 inputs, test data for each class. The implementation was
one hidden layer with 20 neurons and 5 outputs, being a real experimented on a variety of datasets and results presented in
number in the interval [0,1]. The position of the maximum of Table1 and Table2 represent the average performances.
the outputs of neural network indicates the membership with
 14th European Signal Processing Conference (EUSIPCO 2006), Florence, Italy, September 4-8, 2006, copyright by EURASIP

Table 1. Results from the classification algorithm [3] Task force of the European society of cardiology and the
NSR PVC AF VF BII North American society of pacing and electrophysiology.
NSR 280 0 0 0 0 "Heart rate variability – standards of measurements
PVC 0 17.1 1.7 0 0.2 ,physiological interpretation,and clinical use," Eur.Hear
AF 0.1 3.6 104.2 0 0.1 J., vol.17, pp.354-381, 1996.
VF 0 0 1 24.8 0.2 [4] R.U. Acharya, N. Kannathal,and S.M. Krishnan
BII 0 0 0 0 10 "Comprehensive analysis of cardiac health using heart
rate signals," Physiol.Meas.vol. 25, pp. 1139-1151, 2004.
Table 2. Sensitivity, specificity, and accuracy for each class
[5] R.U. Acharya, A. Kumar, P.S. Bhat, C.M.Lim, S.S.
(units,%)
Arrhythmia # of Train Sensitivity Specificity Accuracy
Iyengar , N. Kannathal,and S.M. Krishnan,
Classes & Test "Classification of cardiac abnormalities using heart rate
segments signals," Med.Biol.Eng.Comp., vol.42, pp.288-293, 2004.
NSR 555,280 100 99.94 99.98 [6] J.L.A. Carvalho et.al, "Development of a Matlab
PVC 38,19 90 99.15 98.76 Software for Analysis of Heart Rate Variability,"
AF 214,108 96.48 99.19 98.53 Proc.ICSP02, pp. 1488-1491, 2002.
VF 52,26 95.38 100 99.73
BII 15,10 100 99.88 99.89 [7] M. Tulppo, T.H. Makikallio, T.E.S. Takala, and K.H.
Average 96.37 99.63 99.38 Seppanen, "Quantiative Beat-to-Beat Analysis of Heart
Rate Dynamics During Exercise," Am.J.Physiol.,vol.71,
6. DISCUSSION AND CONCLUSIONS pp.H244-52, 1996.
[8] KLL Ho, GB Moody, CK Peng, JE Mietus, MG Larson,
In this paper, the neural network classifier is presented as
D Levy,and AL. Goldberger, "Predicting Survival in
diagnostic tool to aid the physician in the analysis of heart
Heart Failure Case and Control Subjects by Use of Fully
diseases. The neural network classifier was fed by the
Automated Methods for Deriving Nonlinear and
combination of linear and non-linear parameters derived
Conventional Indices of Heart Rate Dynamics,"
from the HRV signal.
Circulation 1997(August),96(3),pp.842-48, 1997.
The proposed NN classifier showed satisfactory
performances in discriminating five types of arrhythmia. The [9] S.M. Pincus, "Approximate Entropy as a Measure of
accuracy of discrimination of NSR, PVC, AF, VF and BII System Complexity," Proc. Natl Acad Sci. USA,vol. 88,
were 99.98%, 98.76%, 98.53%, 99.73%, and 99.89%, pp. 2297–2301,1991.
respectively. It is noted that the percentage of NSR segments [10] S.M. Pincus,and A.L. Goldberger, "Physiological Time
in the dataset is high, but this is close to reality as ECG Series Analysis: What Does Regularity Quantify?,"
recordings have high percentages of normal beats. Am.J.Physiol.,vol. 266,pp.H1643-1656, 1994.
The results show that the proposed method is effective [11] I.A. Rezek, and S.J. Roberts, "Stochastic Complexity
for classification of cardiac arrhythmias, with an acceptable Measures for Physiological Signal Analysis," IEEE
high accuracy. It is evident that the combination of the linear Trans.Biomed.Eng., pp. 1186-1190,1993.
and nonlinear features together with the employed classifier
is very effective. [12] JP. Eckman, SO. Kamphorst, D. Ruelle,and S. Ciliberto,
The main advantage of the method, compared to other "Lyapunov Exponents From Time Series," Physical
approaches in the literature is that it is completely based on Review A, vol.34(6), pp. 4971-4979, 1986.
RR-interval signal which can be extracted with high accuracy [13] H. Kantz,and T. Schreiber, "Nonlinear Time Series
even for noisy or complicated ECG recordings, while the Analysis," Cambridge University Press, p.304, 1997.
extraction of all other ECG features or any other type of ECG [14] A. Wolf, JB. Swift, HL. Swinney, and JA. Vastano,
analysis is seriously affected by noise. "Determining Lyapunov Exponents from a Time Series,"
As a salient result, we can conclude that the HRV signal Physica D, vol.16, pp.285-317, 1985.
can be used as a reliable indicator of different kinds of heart [15] I.S. Uzun, M.H. Asyali, G. Selebi,and M. Pehlivan,
diseases. "Nonlinear Analysis of Heart Rate Variability,"IEEE, 23rd
EMBS conf., pp.1581-1584, 2001.
REFERENCES
[16] H.V. Huikuri, T.H. Makikallio, C.K. Peng, A.L.
[1] N. Wessel, A. Voss, J. Kurths, A. Witt, and KJ Osterziel, Goldberger, U. Hintze, and M. Moller, "Fractal
"24 Hour Heart Rate Variability Analysis Based on New Correlation Properties of R-R Interval Dynamics and
Methods of Non-Linear Dynamics," IEEE, Mortality in Patients with Depressed Left Ventricular
Comp.Cardio., pp. 693-696, 1995. Function after an Acute Myocardial Infraction,"
[2] J. Pan,and WJ. Tompkins, "A Real Time QRS Detection Circulation 101, pp.47-53, 2000.
Algorithm," IEEE.Trans.Biom.Eng., vol.32, pp. 230-236, [17] S. Haykin, "Neural Networks a Comprehensive Founda-
1985. tion," MacMillan College Publishing Company, New
York, 1995.