Vol.2, No.6, 405-411 (2009)
doi:10.4236/jbise.2009.26058
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
JBiSE
ECG arrhythmia classification based on logistic model
tree
V. Mahesh1, A. Kandaswamy2, C. Vimal2, B. Sathish2
1Department of Information Technology, PSG College of Technology, Coimbatore, India; 2Department of Bio-Medical Engineering,
PSG College of Technology, Coimbatore, India.
Email: vm@ity.psgtech.ac.in; hod@bme.psgtech.ac.in; vimalfu@gmail.com; satfrd@gmail.com
Received 6 May 2008; revised 25 May 2009; accepted 9 June 2009.
ABSTRACT
This paper presents a diagnostic system for
classification of cardiac arrhythmia from ECG
data, using Logistic Model Tree (LMT) classifier.
Clinically useful information in the EC G is found
in the intervals and amplitudes of the charac-
teristic waves. Any abnormality in the wave
shape and duration of the wave features of the
ECG is considered as arrhythmia. The ampli-
tude and duration of the characteristic waves of
the ECG can be more accurately obtained using
Discrete Wavelet Transform (DWT) analysis.
Further, the non-linear behavior of the cardiac
system is well characterized by Heart Rate
Variability (HRV). Hence, DWT and HRV tech-
niques have been employed to extract a set of
linear (time and frequency domain) and non-
linear characteristic features from the ECG
signals. These features are used as input to the
LMT classifier to classify 11 different arrhyth-
mias. The results obtained indicate an impres-
sive prediction accuracy of 98%, validating the
choice and combined use of the current popular
techniques (DWT and HRV) for cardiac ar-
rhythmia classification. The system can be de-
ployed for practical use after validation by ex-
perts.
Keywords: ECG; Arrhythmia; Wavelet Transform;
HRV Analysis; Feature Extraction
1. INTRODUCTION
Electrocardiography is a commonly used, non-invasive
procedure for recording electrical changes in the heart.
The record, which is called an electrocardiogram (ECG
or EKG), shows the series of waves that relate to the
electrical impulses which occur during each beat of the
heart. The information present in the ECG characteristic
wave peaks and time intervals between them are impor-
tant. The waves in a normal record are named P, Q, R, S,
and T and follow in alphabetical order. Any abnormal
change in the shape and variation of time intervals is
conside red as arrhythmia.
Detection of abnormal ECG signals is a critical step in
administering aid to patients. Arrhythmias can occur in a
healthy heart and be of minimal consequence. They may
also indicate a serious problem and lead to heart disease,
stroke or sudden cardiac death. Cardiac arrhythmia is
one of the major causes of sudden death. To detect the
presence of arrhythmia, patients are hooked to cardiac
monitors in hospitals. This requires continuous monitor-
ing by the physicians. Visual inspection is tedious and
physician dependent. Computer programs have been
developed to help in this visual analysis by providing
condensed printouts. This again requires meticulous
study by the physician to identify arrh ythmia. To cater to
large number of patients, to eliminate subjective inaccu-
racies and to aid the physician in the diagnosis several
methods for automated arrhythmia detection have been
developed in the past few decades to attempt simplify
the monitoring task and improve diagnostic efficiencies.
In pursuit of arrhythmia detection and classification
work, many computer techniques have been developed.
Notably, Palreddy et al. employed a multiple-classifier
architecture composed of Self Organizing Maps (SOM)
and Learning Vector Quantization (LMQ) to classify
premature ventricular contraction (PVC) beats and the
non-PVC beats [1]. Babak Mohammadzadeh-Asl et al
used both lin ear and non-linear p arameter extracted from
heart rate signals with multilayer feed forward neural
networks to classify only five types of arrhythmias [2]. J.
Lee et al. proposed a wavelet based approach along with
Linear Discriminant Analysis (LDA) for classifying only
five types of arrhythmias using multilayer perceptron
classifier [3]. Chazal et al. has proposed a method for
automatic classification of heartbeats using ECG mor-
phology, heartbeat interval features and RR intervals to
discriminate only five different beat types [4]. Dingfie et
al. classified only six arrhythmias using autoregressive
V. Mahesh et al. / J. Biomedical Science and Engineering 2 (2009) 405-411
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
406
modeling and Generalized Linear Model (GLM) [5].
Linh et al. selected the Hermite Function Expansion as
the feature extraction method to represent the QRS com-
plex. They proposed a fuzzy neural network where Her-
mite coefficients served as the features to classify only
seven different types of arrhythmias [6]. Kannathal et al.
used three non-linear parameters as inputs to the pro-
posed ANF classifier for classification of only ten dif-
ferent types of arrhythmias [7]. Kadbi et al. used wavelet
parameters along with RR interval and Form Factor as
inputs to an ANN classifier to discriminate only ten dif-
ferent arrhythmias [8].
In clinical domains, one has to face the problem of
developing classifiers that are able to deal with nonlinear
discrimination between classes, incomplete or ambigu-
ous input patterns, and suppression of false alarms. It is
necessary to develop new detection schemes with a high
level of accuracy, or equivalently, low false-positive and
false-negative statistics, for them to be useful in practical
applications. In this direction a new approach based on
Logistic Model Tree classifier is presented in this paper.
LMT is a recent addition to decision trees that replace
the terminal nodes of a decision tree with logistic regres-
sion functions. This has the advantage of producing de-
cision trees that are more comprehensible, have higher
accuracy, and have higher fidelity than previous decision
tree extraction algorithms [9].
2. DATA SOURCE AND CONTENT
ECG data for use in this classification work has been
collected from the MIT-BIH arrhythmia database as
published in Physionet, a site dedicated to data for vari-
ous diseases and their study [10]. The database contains
48 recordings, each containing two 30-min ECG lead
signals (denoted A and B). In 45 recordings, lead A is
modified-lead II and for the other three is lead V5. Lead
B is lead V1 for 40 recordings and is either lead II, V2,
V4 or V5 for the other recordings. Twenty-three records,
numbered from 100 to 124 with some numbers missing,
serve as a representative sample of routine clinical re-
cordings and the remaining twenty-five records, num-
bered from 200 to 234 again with some numbers missing ,
contain complex ventricular and supraventricular ar-
rhythmias. In this work, ECG signals from Modified
Lead II (MLII) leads are chosen. Prior to recording, the
ECG signals in these records have been sampled at a
frequency of 360Hz and preprocessed to remove noise
due to power-line interference, muscle tremors, spikes
etc. This database was selected because it contains a
variety of beat types. Another reason for considering this
database was its use in other studies and thus compari-
son of results can be performed. One minute segments of
each beat type were extracted from the records for fur-
ther processing. This work focused on several important
arrhythmia types such as Paced beat (P), Atrial prema-
ture beat (A), Right bundle branch block beat (R), Left
bundle branch block beat (L), Ventricular escape beat (E),
Ventricular flutter wave (!), premature ventricular con-
traction (V), Fusion of ventricular and normal beat (F),
Fusion of paced (f), Blocked Atrial Premature Beat (x)
and the Normal beat segment (Normal). The number of
segments extracted for each type from the database re-
cords is given in Table 1.
3. FEATURE EXTRACTION
The main objective of the feature extraction process is to
derive a set of parameters that best characterize the sig-
nal. These parameters, in other words, should contain
maximum information abou t the signal. Hence the selec-
tion of these parameters is an important criterion to be
considered for proper classification. Arrhythmia classi-
fication, therefore, involves determination of several
characteristic features of the ECG signal. This work ex-
plores a combination of linear (time and frequency do-
main) and non-linear characteristic features of the ECG
signal. The Discrete Wavelet Transform has been used to
obtain the amplitude and duration of the characteristic
waves of the ECG from which a set of time-domain pa-
rameters are derived. The DWT is also used to obtain the
RR interval time series. Heart Rate Variability (HRV)
helps in understanding the non linear behavior of the
cardiac system. Using the RR series a set of non linear
paramete rs are al so derived.
3.1. Time-Domain Analysis
For each of the segments extracted from the records, the
characteristic points P, Q, R, S and T are obtained using
Discrete Wavelet T ransform.
3.1.1. Discrete Wavelet Transform (DWT)
The wavelet transform is a convolution of the wavelet
function (t) with the signal x(t). Orthonormal dyadic
discrete wavelets are associated with scaling functions
Table 1. Arrhythmia types classified in proposed method.
Type of Arrhythmia No of Segments Ex-
tracted
Normal 459
P 105
A 123
R 99
L 108
E 18
! 24
V 290
F 16
f 27
x 12
V. Mahesh et al. / J. Biomedical Science and Engineering 2 (2009) 405-411
SciRes Copyright © 2009 http://www.scirp.org/journal/JBISE/
407
tion A1 respectively. The first approximation A1 is de-
composed again and this process is continued. The de-
composition of the signal into different frequency bands
is simply obtained by successive highpass and lowpass
filtering of the time domain signal. The signal decompo-
sition can be mathematically expressed as follows:
(t). The scaling function can be convolved with the
signal to produce approximation coefficients A. The
wavelet transform of the signal x(t) can be written as:
,().
,
()
mn tdt
mn
Txt


(1)
By choosing an orthonormal wavelet basis, m, n (t),
one can reconstruct the original signal [11]. The ap-
proximation coefficients of the signal at scale m and
location n can be rep resented by:
hi
y[k] =x[n].g[2k -n]
(3)
lo
y[k]=x[n].h[2k -n]
(4)
The characteristic points P, Q, R, S and T are obtained
at different decomposition levels as shown in Figure 2.
,().
,
()
mn tdt
mn
Axt


(2) Segment selection
8-level wavelet decomposition using Daube-
chies 6 wavelet functions
3.1.2. DWT Decomposition
Discrete Wavelet Tr ansform involves decomposition of a
signal by wavelet filter banks. DWT uses two filters, a
low pass filter (LPF) and a high pass filter (HPF) to de-
compose the signal into different scales. The output co-
efficients of the LPF are called approximations while the
output coefficients of the HPF are called details. The
approximations of the signal are what define its identity
while the details only imparts nuance [12].
Detection of R peak at level 4 using adaptive
threshold value (related to the maximum and
mean values of the signal)
Determination of R-R interval using R-R dis-
tance
Detection of Q, S points as local minimum
points at level 0, before and after R wave
Elimination of the QRS complex from the signal
to obtain other parameters
The DWT decomposition of an input signal x[n] is
schematically shown in the Figure 1 below. Each stage
consists of two digital filters and two downsamplers to
produce the digitized signal. The first filter, g[n] is a
high-pass filter, and the second, h[n] is a low-pass filter.
The downsampled outputs of the first high pass filter and
low-pass filter provide the detail D1 and the approxima-
Detection of T wave at level 6 and 7 for finding
QT distance
Detection of P wave at level 6 and 7 for finding
P-R and P-P distance
From the values obtained the following five time-
domain parameters have been calculated:
Feature Meaning Formula
P-P Mean of P-P interval duratio n s. TPP = Pi+1 –Pi , i=1…N – 1
R-R Mean of R-R interval durations. TRR=Ri+1 –Ri , i=1…N – 1
P-R The time duration between successive P and R waves in each beat. TPR=R – Pon-set
QRS Duration The time duration from the beginning of the Q wave to the end o
f
the S wave. TQRS=TS –TQ
QT Inte rval D uration It is the time from the beginning of the Q-wave to the end of the
T-wave TQT =Toff-set–Q
Figure 1. DWT decomposition.
Openly accessible at
V. Mahesh et al. / J. Biomedical Science and Engineering 2 (2009) 405-411
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
408
n
3.2. Frequency Domain Analysis
Time-domain methods are computationally simple but
lack the ability to discriminate between sympathetic and
parasympathetic contributions of HRV. Spectral analysis
is the most popular linear technique used in the analysis
of HRV signals [13]. Spectral power in the high fre-
quency (HF) (0.15–0.4 Hz) band reflects respiratory
sinus arrhythmia (RSA) and thus cardiac vagal activity.
Low frequency (LF) (0.04–0.15Hz) power is related to
baroreceptor control and is mediated by both vagal and
sympathetic systems [14]. Hence, the frequency domain
parameter LF/HF, which is the ratio between LF and HF
band powers, is obtained for each segment.
3.3. Non-Linear Analysis
The cardiovascular system is a complex non-linear sys-
tem and is characterized by many complex estimators. In
this classification work the following parameters have
been derived from the RR-interval time series obtained
using DWT.
3.3.1. Spectr al Entropy
The power spectral density (PSD) of a signal is the dis-
tribution of power as a function of frequency. This PSD
can be obtained using Fourier transform. The normaliza-
tion of this PSD yields the probability density function
(PDF) [15]. This PDF has a value in the range
011, 2,...,
f
pf  (5)
1
1
n
f
f
p
(6)
The spectral entropy H which describes the complex-
ity of the heart rate variability (HRV) signal is obtained
using Eq.(7).
1
f
ff
Hp
p



(7)
Here pf is the probability density function at f. The spec
tral entropy H calculated for each segment is used as one
of the classifying parameters [16].
3.3.2. Detrended Fluctuation Analy sis (DF A )
The Detrended Fluctuation Analysis (DFA) is used to
quantify the fractal scaling properties of short time R-R
interval signals. This technique is a modification of the
root-mean square analysis of random walks applied to
nonstationary signals [17]. The root-mean-square fluc-
tuation of an integrated and detrended time series is
measured at different observation windows and plotted
against the size of the observation window on a log-log
scale. First, the R-R time series (of total length N) is
integrated using the equation:
1
()( ()))
k
avg
i
ykRRi RR

(8)
where y(k) is the kth value of the integrated series, RR(i)
is the ith inter beat interval and RRavg is the average
inter beat interval over the entire series [18]. Then, the
integrated time series is divided into windows of equal
length, n. In each window of length n, a least squares
line is fitted to the R-R interval data (representing the
trend in that window). The ‘y’ coordinate of the straight
line segments are denoted by yn(k) . Next, we detrend
the integrated time series, yn(k) in each window. The
root mean-square fluctuation of this integrated and de-
trended series is calculated using Eq.(9) for each seg-
ment.
2
1
1
()[() ()]
N
n
k
F
nyky
N

k (9)
4. LOGISTIC MODEL TREES CLASSIFIER
Logistic Model Trees are a combination of a tree struc-
ture and logistic regression functions to produce a single
decision tree [19,20,21,22]. The decision tree structure
has the logistic regression functions at the leaves. The
leaf node has two child nodes which is branched right
Figure 2. Characteristic points extraction from ECG signal at various decomposition levels.
V. Mahesh et al. / J. Biomedical Science and Engineering 2 (2009) 405-411
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
409
k
and left depending on the threshold. If the value of the
attribute is smaller than the threshold it is sorted to left
branch and value of attribute greater than the threshold it
is sorted to right branch as shown in Fi gure 3.
The threshold is usually fixed by Logit Boost method
[19]. Logit Boost uses a ensemble of functions FK to
predict classes 1, . . . , K using M “weak learners”.
1
() ()
K
km
m
F
xf
x
(10)
Steps followed for developing the LMT classifier:
The linear regression function is fitted using the
Logitboost method to build a logistic model.
The Logitboost method uses 5 examples for the
cross validation to determine the best number of
iterations to run, when fitting the logistic regres-
sion function at a node of the decision tree
The logistic model is built using all data.
The split of the data at the root is constructed
using the threshold.
This splitting is continued till some stopping
criterion is met. Here the stopping criterion is 5
examples, since it helps in cross validation for
logitboost method.
Once the tree has been build it is pruned using
CART-based pruning [19].
Reasons for choosing the Logistic Model Tree classifier:
Logistic Regression is very good at detecting
linear relationships and then combining those
relationships into an equation that provides the
odds of the dependent variable reaching a par-
ticular outcome, when the various independent
variables are fed into the resulting equation.
Logistic Regression models are widely used and
they are considered robust and not prone to over
fitting the data.
These models can be built with high level of
accuracy using little data preparation.
Logistic Model Trees give explicit class prob-
ability estimates rather than just a classification.
The classification task, depicted in Figure 4, involves
the following steps:
Figure 3. Tree structure of logistic model tree (LMT).
Figure 4. Block diagram of the proposed method.
V. Mahesh et al. / J. Biomedical Science and Engineering 2 (2009) 405-411
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
410
One minute segments of each beat type are ex-
tracted from ECG records in the database.
Each segment is then decomposed using DWT
into various levels for extracting linear time-
domain parameters.
The different nonlinear parameters are calcu-
lated using their respective formula.
Both linear and non-linear parameters for all the
segments are combined and a dataset is formed.
75% of the dataset, called training set, is used
for training the classifier.
The remaining 25% of the dataset, called test set,
is used for testing the classifier.
5. RESULTS AND DIS CUSSIONS
The objective of any clinical research is to find the rela-
tionship between results and presence of any disease.
For the evaluation of the proposed classifier, a total of
1281 segments, extracted from the MIT BIH arrhythmia
database records were used. Five time-domain, one fre-
quency domain and two non-linear parameters were de-
rived from these segments. These eight parameters along
with the corresponding output class (type of arrhythmia)
forms a feature vector. Thus 1281 feature vectors com-
prise the dataset. 75% of each type from this dataset was
used as the train dataset and the remaining 25% as the
test dataset. The output obtained from the Logistic
Model Tree was used to calculate the accuracy of each
type of beat using Eq.(11)).
Number ofbeatscorrectly classified
Accuracy Totalnumber ofbeats
(11)
The experimental results are presented in Table 2.
Table 2. Performance of the proposed method.
Type of
Arrhythmia No of Segments
Extracted No of Segments for
Training No of Segments for
Testing Correctly
Classified Accuracy %
Normal 459 353 106 102 96.22
P 105 74 31 29 93.54
A 123 95 28 26 92.85
R 99 74 25 25 100
L 108 77 31 31 100
E 18 13 5 5 100
! 24 18 6 6 100
V 290 219 71 70 98.6
F 16 12 4 4 100
F 27 19 8 8 100
X 12 7 5 5 100
Average=98.29
Table 3. Performance comparison of different ECG arrhythmia classifiers.
Work Reference Types Accuracy (%) Feature Extraction Method Classifier
Palreddy [1] 2 98.58% LVQ SOM
Babak [2] 5 99.38% HRV NN
Lee [3] 5 99.48% WT LDA/MLP
Chazal[4] 5 96.87% ECG Morphology/ Interval LDA
Dingfie [5] 6 93.2% AR Modeling GLM
Linh [6] 7 96% HER FNN
Kannathal [7] 10 94.64% HRV ANN
Kadbi [8] 10 90% WT Cascade ANN
Proposed Method 11 98.29% DWT/HRV LMT
V. Mahesh et al. / J. Biomedical Science and Engineering 2 (2009) 405-411
SciRes Copyright © 2009 Openly accessible at http://www.scirp.org/journal/JBISE/
411
Table 3 shows the performance comparison of the
different ECG arrhythmia classifiers. The proposed me-
thod shows comparable performance even when 11 dif-
ferent types of arrhythmias have been considered.
6. CONCLUSIONS
In this paper, the effectiveness of the Logistic Model
Tree classifier for arrhythmia classification has been
demonstrated. The Logistic Model Tree classifier was
fed by the combination of linear and non-linear parame-
ters derived from ECG data using DWT and HRV. The
results indicate that the proposed method employing the
LMT classifier with linear and nonlinear parameters is
effective for classification of cardiac arrhythmias with an
acceptably high accuracy. Compared to other approaches
in the literature cited, the proposed method exploits the
power of HRV and DWT techniques in discriminating 11
different arrhythmia types. Parameters derived from
ECG features and HRV analysis can therefore be used as
a reliable indicator of different types of arrhythmias. The
proposed system, after validation by experts, can serve
as a diagnostic tool and aid the physician in the detection
and classification of cardiac arrhythmias.
REFERENCES
[1] S. Palreddy, Y. H. Hu, V. Mani, and W. J. T ompkins
(1997) A multipleclassifier architecture for ECG beat
classification, IEEE Workshop on Neural Network for
Signal Processing, 172–181.
[2] B. Mohammadzadeh-Asl and S. K. Setarehdan, (2006)
Neural network based arrhythmia classification using
heart rate variability signal, Proceedings of the EUSI-
PCO.
[3] J. Lee, K. L. Park, M. H. Song, and K. J. Lee, (2005)
Arrhythmia classification with reduced features by linear
discriminant analysis, IEEE EMBS 2005, 27th Annual
International Conference, Engineering in Medicine and
Biology Society, 1142–1144.
[4] P. Chazal, M. O’Dwyer, and R. B. Reilly, (2004) Auto-
matic classification of heartbeats using ECG morphology
and heartbeat interval features, IEEE Trans. Biomedical
Engineering, 51(7), 1196–1206.
[5] G. Dingfei, N. Srinivasan, and S. Krishnan, (2002) Car-
diac arrhythmia classification using autoregressive mod-
eling, BioMedic a l E n g i n e e r i n g OnLine, 1–5.
[6] T. H. Linh, S. Osowki, and M. Stodolski, (2003) On-line
heart beat recognition using hermite polynomials and
neuro-fuzzy network, IEEE Trans. Instrumentation and
Measurement, 52(4), 1224–1231.
[7] N. Kannathal, S. K. Puthusserypady, C. M. Lim, U. R.
Acharya, and S. Laxminarayan, (2005) Cardiac state di-
agnosis using adaptive neuro-fuzzy technique, Proceed-
ings of the IEEE Engineering in Medicine and Biology
27th Annual Conference Shanghai, China.
[8] M. H. Kadbi, J. Hashemi, H. R. Mohseni and A. Magh-
soudi, (2006) Classification of ECG arrhythmias based
on statistical and time-frequency features, 3rd Interna-
tional Conference on Advances in Medical, Signal and
Information Processing, MEDSIP.
[9] D. Dancey, Z. A. Bandar, and D. McLean, (2007) Logis-
tic model tree extraction from artificial neural networks,
IEEE Trans on Systems, Man and Cybernetics, 37(4),
794–802.
[10] http://www.physionet.org/physiobank/database/mitdb/.
[11] S. Z. Mahmoodabadi, A. Ahmadian, M. D. Abolhasani,
M. Eslami, and J. H. Bidgoli, (2005) ECG feature extrac-
tion based on multiresolution wavelet transform, in Proc.
IEEE Engineering in Medicine and Biology, 3902–3905.
[12] E. M. Tamil, N. H. Kamarudin, R. Salleh, M. Yamani
Idna Idris, M. N. Noorzaily, and A. M. Tamil, (2008)
Heartbeat electrocardiogram (ECG) signal feature extrac-
tion using discrete wavelet transforms (DWT), in Pro-
ceedings of CSPA, 1112–1117.
[13] D. N. Dutt and S. M. Krishnan, (2000) Computer proc-
essing of heart rate variability signals for detection of pa-
tient status in cardiac care units, Current Science, 78(7),
864–868.
[14] Task force of the European society of cardiology and the
North American society of pacing and electrophysiology
(1996), Heart rate variability–standards of measurements,
physiological interpretation, and clinical use, European
Heart Journal, 17, 354–381.
[15] S. I. Niwas, R. S. S. Kumari, and V. Sadasivam, (2005)
Artificial neural network based automatic cardiac ab-
normalities classification, Proceeding of the Sixth Inter-
national Conference on Computational Intelligence and
Multimedia App lication.
[16] U. R. Acharya, A. Kumar, P. S. Bhat, C. M. Lim, S. S.
Iyengar, N. Kannathal, and S. Krishnan, (2004) Classifi-
cation of cardiac abnormalities using heart rate signals,
medical, Biological Engineering and Computing, 42,
288–293.
[17] H. Kun, C. I. Plamen, C. Zhi, C. Pedro, and H. E. Stanley,
(2001) Effect of trends on detrended fluctuation analysis,
Physical Review E, 64, 011114.
[18] R. U. Acharya, C. M. Lim, and P. Joseph, (2002) Heart
rate variability analysis using correlation dimension and
detrended fluctuation analysis, ITBM–RBM, 23(6), 333–
339.
[19] N. Landwehr, M. Hall, and E. Frank, (2005) Logistic
model trees, Machine Learning, 59(1/2), 161–205.
[20] D. W. Hosmer and S. Lemeshow, (2000) Applied logistic
regression, 2nd Edition, Wiley-Interscience, New York.
[21] J. Friedman, T. Hastie, and R. Tibshirani, (2000) Additive
logistic regression: A statistical view of boosting, Annals
of Staistics, 32(2), 337–374.
[22] Z. H. Zhou, Y. Jiang, and S. F. Chen, (2003) Extracting
symbolic rules from trained neural network ensembles,
Artificial Intelligence Communication, 16(1), 2–15.