Biomedical Signal Processing and Control 34 (2017) 1–8

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

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

Technical Note

Evaluation of effect of unsupervised dimensionality reduction

techniques on automated arrhythmia classification
Rekha Rajagopal ∗ , Vidhyapriya Ranganathan
Department of Information Technology, PSG College of Technology, Coimbatore 641 004, India

a r t i c l e i n f o a b s t r a c t

Article history: Automation in cardiac arrhythmia classification helps medical professionals to make accurate decisions
Received 9 December 2015 upon the patient’s health. The aim of this work is to evaluate the performance of five different linear and
Received in revised form nonlinear unsupervised dimensionality reduction (DR) techniques namely principal component analysis
16 September 2016
(PCA), fast independent component analysis (fastICA) with tangential, kurtosis and Gaussian contrast
Accepted 20 December 2016
functions, kernel PCA (KPCA) with polynomial kernel, hierarchical nonlinear PCA (hNLPCA) and principal
polynomial analysis (PPA) on classification of cardiac arrhythmias using probabilistic neural network
Keywords:
classifier (PNN). The design phase of the classification model comprises of the following stages: pre-
Biomedical signal processing
Decision support systems
processing of the cardiac signal by eliminating detail coefficients that contain noise, feature extraction
Feature extraction through Daubechies wavelet transform, dimensionality reduction through unsupervised DR techniques,
Supervised learning and arrhythmia classification using PNN. PCA is a widely used DR technique for mapping high dimen-
sional data to its low dimensional representation. But real world data like electrocardiogram (ECG) signals
are complex and nonlinear in nature. This work concentrates on performance analysis of four nonlinear
DR techniques and conventional linear PCA technique on classification of cardiac arrhythmias. Entire
MIT-BIH arrhythmia database is used for experimentation. The experimental results demonstrate that
the combination of PNN classifier (at spread parameter, ␴ = 0.4) and fastICA DR technique with tangential
contrast function exhibit highest F score of 99.83% with a minimum of 10 dimensions. hNLPCA and KPCA
requires more computation time for low dimensional mapping. PPA performs about 10% better than PCA
and serves intermediate between linear and nonlinear techniques.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction Dimensionality reduction is the mapping of data in higher

Cardiovascular disease is a leading cause of global mortality.
Hence, there is a need to develop automation strategies for the
management of sudden cardiac death [1]. Abnormality in the nor-
mal rhythm of heartbeat causes arrhythmia. ANSI/AAMI EC57: 1998
classification standard categorize arrhythmias into five classes,
namely: non ectopic beat (N), supra ventricular ectopic beat (S),
ventricular ectopic beat (V), fusion beat (F) and unknown beat (Q).
The diagnosis of a specific class of arrhythmia is done by careful
monitoring of long term electrocardiograph (ECG) signal. Automa-
tion in ECG arrhythmia classification is very essential in order to
make fast and accurate decision on arrhythmia class. The important
requirements of an automated system are reduced complexity, fast
decision making and less memory. This can be accomplished with
the use of dimensionality reduction techniques.
∗ Corresponding author.
E-mail addresses: rekha.psgtech@gmail.com, jp.psgtech@gmail.com
(R. Rajagopal), rvidhyapriya@gmail.com (V. Ranganathan).
dimensional space to meaningful representation in lower dimen-
sional space such that the dimensionality of data is reduced to
manageable size. Dimensionality reduction is very important in a
classification system in order to mitigate the curse of dimension-
ality, computational complexity and memory requirement. Curse
of dimensionality means that for a given sample size, there is a
maximum number of features above which the performance of
a classifier gets degraded. The objective of this work is to evalu-
ate the performance of five different unsupervised dimensionality
reduction techniques in the classification of ECG arrhythmias.
2. Related works
Several research works are carried out for automation in
arrhythmia classification. In general, the algorithm used for
automated classification includes (i) preprocessing, (ii) feature
extraction, (iii) dimensionality reduction and (iv) classification. The
preprocessing of recorded ECG signals is done in order to elimi-
nate the important noises that degrade the classifier performance,
such as baseline wandering, motion artifact, power line interfer-

http://dx.doi.org/10.1016/j.bspc.2016.12.017
1746-8094/© 2016 Elsevier Ltd. All rights reserved.
2 R. Rajagopal, V. Ranganathan / Biomedical Signal Processing and Control 34 (2017) 1–8

Table 1
Number of heartbeats in each class. Connuous ECG recordings from MIT_BIH
Heartbeat type N S V F Q arrhythmia database
Full database 87643 2646 6792 794 15

ECG beat segmentaon

ence and high frequency noise. Currently, researchers use many
filtering techniques like morphological filtering [2], integral coef-
ficient band stop filtering [3], finite impulse response filtering [4],
Wavelet based feature extracon
5–20 Hz band pass filtering [5,6], median filtering [7] and wavelet
based denoising [7–9] for preprocessing.
Commonly extracted ECG features include (i) Temporal features
of heartbeat like P–Q interval, QRS interval, S–T interval, Q–R inter-
val, R–S interval and R–R interval between adjacent heartbeats. (ii) Dimensionality reducon
Amplitude based features like P peak amplitude, Q peak ampli- (Linear – PCA)
tude, R peak amplitude, S peak amplitude and T peak amplitude. (Non-linear - fast ICA, Kernel PCA, hNLPCA,
(iii) Wavelet transform based features that include Haar wavelet, PPA)
Daubechies wavelet and discrete Meyer wavelet at various decom-
position levels of 4, 6 and 8. (iv) Stockwell transform based features
that include statistical features taken from a complex matrix of
Stockwell transform, time–frequency contour and time–maximum Divide enre samples into 10 equal sets
amplitude contour.
Dimensionality reduction is applied in order to remove redun-
dancy and extract useful information from captured features. A
training set with class label information is required by supervised
dimensionality reduction techniques to learn the lower dimen- Set Set Set ……... Set
sional representation. Commonly used supervised techniques are
1 2 3 10
linear discriminant analysis (LDA) [10], generalized discriminant
analysis (GDA) [5], neighborhood component analysis and metric
learners. Unsupervised techniques do not require class label infor-
mation and few examples of unsupervised techniques are principal
component analysis, independent component analysis, canonical 90 % Training 10% Tesng
correlation analysis, partial least squares, isomap, kernel PCA, max-
imum variance unfolding, diffusion maps, local linear embedding,
Laplacian eigenmaps, Hessian LLE, Local tangent space analysis,
Sammon mapping and multilayer auto encoders [11]. PNN Classifier
Support vector machine (SVM) [5,12–17], probabilistic neural
network (PNN) [9,15,18,19], multilayer perceptron neural network
(MLPNN) [6,15,20], linear discriminant classifier [7], mixture of
experts [15] and unsupervised clustering [21,22] are commonly Ten fold cross validaon
used by researchers for classification of ECG arrhythmia. The
parameters like accuracy, sensitivity and specificity are used in
literature for evaluating the performance of a classifier.
Most of the research works have compared the performance Fig. 1. Architecture of the proposed work.
of classifiers using PCA [2,9], LDA [9,10] and ICA [9,23,24] for
dimensionality reduction. Research works that use nonlinear DR
techniques in arrhythmia classification are very minimal. This work
concentrates on evaluating the performance of an automated clas-
sification system when DR techniques such as fast ICA, Kernel PCA, 4. Methodology
hNLPCA, PPA and PCA are used. The next section discusses the mate-
rials and methods used in this work. Fig. 1 shows the architecture of the proposed work. The entire
experiments are carried out using Matlab R2014a. The details of the
methodology carried out are summarized below.

3. Materials used

The MIT BIH arrhythmia database [25] is used in this work. It
contains 48 half–hour excerpts of two channel ambulatory ECG
recordings, obtained from 47 subjects studied by the BIH arrhyth-
mia laboratory. The recordings were digitized at 360 samples per
second per channel with 11 bit resolution over 10 mV range. The
reference annotations for each beat were included in the database.
Four records containing paced beats (102, 104, 107 and 217) were
removed from analysis as specified by AAMI. The total number of
heart beats in each class is given in Table 1.
4.1. Data preprocessing
The records contain continuous ECG recording for duration
of 30 min. The raw ECG signals include baseline wander, motion
artifact and power line interference noise. The discrete wavelet
transform (DWT) is used for denoising the ECG signal and also for
extracting the important features from original ECG signal [26,27].
DWT captures both temporal and frequency information. The DWT
of the original ECG signal is computed by successive high pass and
 R. Rajagopal, V. Ranganathan / Biomedical Signal Processing and Control 34 (2017) 1–8
Fig. 2. Segment of continuous ECG waveform before preprocessing.
Fig. 3. Segment of continuous ECG waveform after preprocessing.
low pass filtering of that signal. This can be mathematically repre-
sented in the following Eqs. (1) and (2),
∞
yhigh [k] = x [n] g [2k − n]
n=−∞
∞
ylow [k] = x [n] h [2k − n]
n=−∞
where, x[n] is the original ECG signal samples, g and h are the
impulse response of high pass and low pass filters respectively,
and yhigh [k],ylow [k] are the outputs of high pass and low pass fil-
ters after sub sampling by 2. This procedure is repeated until the
required decomposition level is reached. The low frequency com-
ponent is called approximation and high frequency component is
called detail.
In this work, the raw ECG signals are decomposed into approxi-
mation and detail sub bands up to level 9 using Daubechies (‘db8 )
wavelet basis function [28]. The detail coefficients contain most of
the noise information. Hence, soft thresholding is applied to the
detail coefficient at each level. Denoised ECG signal is computed
based on the original approximation coefficient at 9th level and
modified detail coefficient of levels from 1 to 9. After denoising,
the continuous ECG waveform is segmented into individual heart-
beats. This segmentation is done by identifying the R peaks using
Pan Tompkins algorithm [29] and by considering 99 samples before
R peak and 100 samples after R Peak. This choice of 200 samples,
including R peak for segmentation is taken because it constitutes
one cardiac activity with P, QRS and T waves. Figs. 2 and 3 show a
segment of the recorded ECG waveform of patient identifier: 209
before and after preprocessing respectively.
3
4.2. Feature extraction
The entire database (97,890 heartbeats) is divided into ten sets
each containing 9789 heartbeats. Nine sets are used for train-
(1) ing (88,101 heartbeats) and one set for testing (9789 heartbeats).
From each heartbeat, wavelet based features are extracted by
using Daubechies wavelet (‘db4 ). Daubechies wavelet with level
(2) 4 decomposition is selected in this work after making perfor-
mance comparisons with discrete Meyer wavelet and other levels
of Daubechies wavelet including ‘db2 , ‘db6 . The power spectral
density of ECG beats showed that most of the ECG signal variations
occur in the fourth sub band frequency range of (0–11.25 Hz) and
(11.25–22.5 Hz). A total of 107 features are produced by the 4th
level approximation sub band coefficients and another 107 fea-
tures by the 4th level detail sub band coefficients. Fig. 4 shows
the decomposition of heart beat signal (sampled at 360 Hz) into
approximation and detail sub bands up to level 2. This procedure
can be repeated for further decomposition.
4.3. Dimensionality reduction
The coefficients of approximation and detail sub bands at level
4 decomposition of Daubechies wavelet transform are used as
features to represent each heart beat. Hence each heartbeat is rep-
resented by 214 features (107 coefficients from approximation sub
band summed with 107 coefficients from detail sub band pro-
vide 214 features). In order to extract useful information from the
features and to remove redundancy the following dimensionality
reduction techniques are applied.
4 R. Rajagopal, V. Ranganathan / Biomedical Signal Processing and Control 34 (2017) 1–8

ECG signal samples x [n]

4.3.2.1. Fast independent component analysis (Fast ICA). ICA makes a
(Sampled at 360 Hz) data representation such that transformed components are statisti-
cally independent from each other. In order to have computational
simplicity and little memory space, fixed point ICA algorithm is
h[n] g[n] used in this work [30]. The steps involved are,

1 Data centering
2 2
This step produces a data set whose mean (x̄) is zero and is
Level 1 detail coefficients shown in (6).
h[n] g[n] (90 to 180 Hz)
XC = X − E{X} (6)

2 2 where, X is the input data matrix.

2 Whitening
Level 2 approximation Level 2 detail coefficients
coefficients (45 to 90 Hz)
Eigenvector V and Eigen value E are computed from the covari-
(0 to 45 Hz)
ance matrix of centered data, and whitening is performed as in
Fig. 4. Two level wavelet filter bank. (7),

Z = E−1/2 V ∗ XC (7)
4.3.1. Linear technique
Linear mapping of high dimensional data to low dimensional 3 Fixed point iteration for one unit
data is done with this technique. The important advantages of linear
technique are its simple geometric interpretations and attractive It estimates one row of the demixing matrix w as a vector wT .
computational properties [11]. Estimation of w proceeds iteratively as in (8) until convergence is
achieved.
4.3.1.1. Principal component analysis (PCA). PCA that maximizes the
amount of variance in the data is used in this work. The steps fol- • Selection of initial random vector w
lowed for dimensionality reduction of data matrix X consisting of
n ECG samples xi (i &1028; {1, 2, .,n}) with D features (The value of wp = E{x g(wT x)} − E{xg (wT x)}w (8)
D is 214 in the proposed work due to 214 features extracted from
each heartbeat using wavelet transform) are given below, where, wp ∗ = wp /wp  and wp  is the norm of w and g is the contrast
function.
1 Subtraction of the mean
4 Evaluation of next independent components
This step produces a data set whose mean x̄ is zero.
To estimate the other independent components, step 3 is
2 Calculation of covariance matrix C is shown in (3). repeated for getting weight vectors wi , i = 2, 3,.n. To prevent dif-
ferent vectors from converging to the same optimum, the weight
C = (x − x̄)(x − x̄)T (3) vectors are decorrelated using Gram-Schmidt orthogonalization
p T w )w and renormalize
and are given by wp+1 = wp+1 − j=1 (wp+1
 j j
3 Calculation of eigenvectors V and eigenvalues E is shown in (4) wp+1 as wp+1 = wp+1 / T w
wp+1 p+1 .

V−1 CV = E (4)
4.3.2.2. Kernel principal component analysis (Kernel PCA). Using
4 Sorting of eigenvectors Kernel PCA [31], the linear operations of PCA are done with non-
linear mapping. Kernel PCA computes the principal eigenvectors
of the kernel matrix instead of computation from the covariance
Eigen vector with highest eigenvalue is the principal compo-
matrix. The steps involved are,
nent of the data. Eigen vectors are ordered by eigenvalue, highest
to lowest.
1 Computation of kernel matrix K
5 Projection of data is shown in (5)
For the data samples {xi }, kernel matrix is computed. The entries
 T T
 in the kernel matrix are given by ki,j = k(xi , xj ), where k is a kernel
Projected data = V (x − x̄) T (5)
function such as Gaussian or polynomial.
Arrhythmia classification is done by varying the selection of
components from 1 to 10 in the projected data and the results are 2 Double centering of K
analyzed.
It means subtracting the mean of the data in the feature space
4.3.2. Nonlinear techniques defined by the kernel function and is shown in (9).
1 1 
Nonlinear techniques have the ability to deal with complex
1 1
nonlinear real world data. The following nonlinear techniques are ki,j = − (ki,j − kil − l kjl + 2 klm ) (9)
2 n l n n lm
experimented to analyze its performance in comparison with tra-
ditional PCA. Where, n is the number of data samples.
 R. Rajagopal, V. Ranganathan / Biomedical Signal Processing and Control 34 (2017) 1–8
3 Computation of eigenvectors vi of centered kernel matrix.
4 Computation of eigenvectors of covariance matrix ai in the fea-
ture space constructed by k.
5 Projection of data is shown in (10)
n  n 
(j) (j)
yi = a1 k xj , xi , . . ..., ad k x j , x i (10)
j=1 j=1
(j)
where, a1 indicates the jth value in the vector a1 and k is the kernel
function. Arrhythmia classification is done by varying the selection
of components from 1 to 10 in the projected data and the results
are analyzed.
4.3.2.3. Hierarchical nonlinear PCA (hNLPCA). hNLPCA [32] is an
extension of PCA in which both the principal component values and
the mapping function is provided by the neural network approach.
Layers involved are input, hidden and output layer. Hidden layer
performs nonlinear mapping of high dimensional data to lower
dimensions. Input and output layers have D nodes and hidden layer
has d nodes, where, D is the actual dimension and d is the reduced
dimension. hNLPCA means that the same hierarchical order as
the linear components of standard PCA is maintained. The steps
involved are,
1 Feed forward neural network is trained such that the mean
squared error between the input and output is minimized.
hNLPCA tries to minimize the hierarchical error as well.
The hierarchical error function is given in (11),
EH = E1 + E 1,2 + E 1,2,3 + . . .. . .. + E 1,2,3...d (11)
and E = 12 x̂ − x2 , where x is the data input with Dimension D and
x̂ is its reconstructed value.
2 Optimal network weights are found using conjugate gradient
descent approach.
3 The addition of a weight decay term in order to penalize large
network weights w.
 uated using performance metrics such as accuracy, sensitivity,
Etotal = E + v w2 , w is the network weight and v is the weight
i i
decay coefficient.
3 At each iteration, single error terms such as E1, E1,2 . . . are cal-
culated separately and d-dimensional space with minimal mean
square error is found.
4.3.2.4. Principal polynomial analysis (PPA). PPA models the direc-
tions of maximal variance by means of curves instead of straight
lines as modeled by PCA [33]. PPA performs simple univariate
regression such that it becomes computationally feasible. The steps
involved are,
1 Data centering:
This step produces a data set whose mean (x̄) is zero and is
shown in (12).
XC = X − E{X} (12)
2 The best vector for data projection is located by finding the lead-
ing eigenvector of conventional PCA.
3 Conditional mean m̂ of D dimensional data sample x is estimated
from the projection by a polynomial fitted to minimize the resid-
ual |x − m̂|.
4 Conditional mean is subtracted to every data sample
5
Table 2
Properties of DR techniques.
Technique Mixing model Parameters Computational complexity
PCA Linear None O(D3 )
FastICA Non linear g, ε O(2(D+1)ni)
Kernel PCA Non linear k (., .) O(n3 )
NLPCA Non linear Net size O(inw)
PPA Non linear  O(D3 +(D−1)(+1)3 ))
D–Dimensionality of input sample.
g–Contrast function.
ε–Convergence parameter.
n–Number of data points.
i–Number of iterations.
k–Kernel function.
w–Number of weights in a neural network.
–Polynomial degree.
5 Steps 3 and 4 are repeated for the remaining (D-1) dimensions.
Data of reduced dimension will be the input for the next step in
the sequence.
Properties of above discussed DR techniques are shown in
Table 2.
4.4. Probabilistic neural network classifier
Probabilistic neural network (PNN) is used for classification of
ECG beats. It is a feed forward network with input, hidden, sum-
mation and output layer. When an input is given, the hidden layer
computes the distances from the inputs and training input vectors
to produce a vector whose elements indicates how close the input is
for a training input. Summation layer sums these contributions for
each class of inputs to produce as its net output a vector of proba-
bilities. Output layer picks the maximum of these probabilities, and
produces a 1 for that class and a 0 for the other classes. Radial basis
function (RBF) is used as the transfer function. PNN is trained with
88101 ECG beat samples which includes training examples from all
five classes. Training and testing matrix is computed such that each
row represents an ECG heartbeat and features occupy columns.
The performance of the arrhythmia classifier system is eval-
specificity, false positive rate and F score. These metrics are com-
puted by calculating TP (True Positive), TN (True Negative), FP
(False Positive) and FN (False Negative) count and are defined as
follows: Sensitivity = TP/(TP + FN), Specificity = TN/(TN + FP), False
positive rate = FP/(TN + FP), F score = 2TP/(2TP + FP + FN) and Accu-
racy = (TP + TN)/(TP + FP + FN + TN). The process is repeated ten
times such that each set is used once for testing. The overall per-
formance of the classifier is computed by taking the average of all
ten folds.
5. Results and discussion
The goodness of a classifier in accurately classifying the test
heartbeat class is measured mainly by the sensitivity and F score.
The reason for not considering accuracy is that even a poor classi-
fier can show good accuracy in favoring class with more training
examples. Fig. 5 shows the results of average accuracy of PNN clas-
sifier taken from 10 fold cross validation for different DR techniques
and dimensions. The result shows that kernel PCA with polynomial
nonlinearity is able to get the highest accuracy of 99.88% even at
lower dimensions. But the time taken for dimensionality reduction
is higher than PCA, fastICA and PPA because KPCA computes the
eigenvectors of the kernel matrix instead of the covariance matrix.
Sensitivity is an important classification parameter which indi-
cates the number of correctly identified arrhythmias from patients.
Fig. 6 shows the results of average sensitivity taken from 10 fold
6 R. Rajagopal, V. Ranganathan / Biomedical Signal Processing and Control 34 (2017) 1–8

100 PCA 105

Accuracy in %
98 95 PCA
KPCA-POLY

F score in %
96
85 KPCA-POLY
94 PPA
fastICA-pow3 75 PPA
92
90 fastICA-tanh 65 fastICA-pow3
1 2 3 4 5 6 7 8 9 10 fastICA-tanh
fastICA-gauss 55
Number of dimensions
hNLPCA 1 2 3 4 5 6 7 8 9 10 fastICA-gauss
Number of dimensions hNLPCA
Fig. 5. Results of average accuracy of PNN classifier for different DR techniques and
dimensions. Fig. 9. Results of average F score of PNN classifier for different DR techniques and
dimensions.
100
95 PCA Table 3
Sensivity in %

90
85 KPCA-POLY Variation of average F scores with respect to different spread values during PNN
80 classification using fastICA and Kernel PCA DR techniques.
75 PPA
70 Spread parameter Average F Score
65 fastICA-pow3
60 FastICA–tanh KernelPCA–poly
fastICA-tanh
55
fastICA-gauss 0.1 78.11 92.92
1 2 3 4 5 6 7 8 9 10
0.2 99.62 93.05
Number of dimensions hNLPCA 0.3 99.75 93.05
0.4 99.83 92.97
Fig. 6. Results of average sensitivity of PNN classifier for different DR techniques 0.5 99.83 92.97
and dimensions. 0.6 95.87 92.97
0.7 95.87 92.9
0.8 95.87 92.74
100 PCA 0.9 95.87 92.52
Specificity in %

98 1 95.72 91.66
KPCA-POLY
96 PPA
94 fastICA-pow3 Table 4
Variation of average F scores with respect to different spread values during PNN
92 fastICA-tanh classification using PCA, hNLPCA and PPA DR techniques.
1 2 3 4 5 6 7 8 9 10
fastICA-gauss
Number of dimensions Spread parameter Average F Score
hNLPCA
PCA hNLPCA PPA

Fig. 7. Results of average specificity of PNN classifier for different DR techniques 0.01 61.77 79.13 61.21
and dimensions. 0.02 76.74 92.53 89.19
0.03 78.19 90.51 91.2
0.04 78.6 85.91 91.49
7 0.05 78.74 80.98 91.49
False posive rate in %

6 PCA 0.06 78.74 73.5 91.49

5 KPCA-POLY
4 0.08
PPA
3
fastICA-pow3
2
1 fastICA-tanh
0
fastICA-gauss
1 2 3 4 5 6 7 8 9 10
Number of dimensions hNLPCA
Fig. 8. Results of average false positive rate of PNN classifier for different DR tech-
niques and dimensions.
cross validation of a PNN classifier for different DR techniques and
dimensions. Experimental results show that fast ICA with Gaus-
sian non linearity is able to achieve 99.54% sensitivity with only
5 dimensions. PCA, which performs a linear mapping, is able to
produce only a maximum sensitivity of 78.74% even when 10
dimensions are used.
Specificity indicates the true negative rate and is shown in Fig. 7.
The result shows that kernel PCA with polynomial nonlinearity is
able to get the highest specificity of 99.9% even at lower dimensions.
Average false positive rate for different DR techniques and dimen-
sions taken from 10 fold cross validation is shown in Fig. 8. FastICA
with Gaussian non linearity produces the least false positive rate of
0.05% even with two dimensions. F score is an important parameter
that combines both precision and recall. The average F score result
0.07 78.74 66.63 91.51
78.82 59 91.28
0.09 78.6 51.34 91.28
0.1 78.5 53.22 91.43
of various DR techniques and different dimensions are shown in
Fig. 9. Results show that fastICA with Gaussian nonlinearity gives
highest F-score of 99.67% with 5 dimensions itself.
Table 3 shows the variation of average F scores with respect to
different spread values during PNN classification using fastICA and
Kernel PCA DR techniques. PNN with fastICA (tangential contrast
function) coefficients yielded highest F-score of 99.83% on spread
value 0.4. PNN with Kernel PCA (polynomial kernel) coefficients
yielded highest F-score of 93.05% on spread value 0.2. Table 4 shows
the variation of average F scores with respect to different spread
values during PNN classification using PCA, hNLPCA and PPA DR
techniques. PNN with PCA coefficients yielded highest F-score of
78.82% on spread value 0.08. PNN with hNLPCA coefficients yielded
highest F-score of 92.53% on spread value 0.02 and PNN with PPA
coefficients yielded highest F-score of 91.51% on spread value 0.07.
Time taken by various dimensionality reduction techniques for
making a mapping from high dimensional representation of ECG
samples (97,890 × 214) with 97,980 heartbeats and 214 features to
low dimensional representation (97,890 × 10) reduced to 10 fea-
tures is shown in Table 5.
 R. Rajagopal, V. Ranganathan / Biomedical Signal Processing and Control 34 (2017) 1–8 7

Table 5 evaluated by using various linear and nonlinear dimensionality

Computation time required by various DR techniques for the input ECG data samples.
reduction techniques with different dimensions of data represen-
DR Technique Computation time (Seconds) tations. Experimental results revealed that the conventional linear
PCA 0.6 DR technique like PCA is easy to apply, but is not able to capture
fastICA–gauss 2 important information required for class discrimination from very
KernelPCA–poly 170 lower dimensional data representation. Nonlinear DR techniques
hNLPCA 4576 are able to capture significant information from the lower dimen-
PPA 0.8
sional representation itself. At the same time, nonlinear techniques
face the important drawback of tuning the parameters. The time
Table 6 required for computation is also high compared to linear DR tech-
Comparison of arrhythmia classifiers classifying classes N, S, V and F. niques. hNLPCA requires comparatively very huge time for low
F Score of each arrhythmia class
dimensional mapping than other techniques because of the tedious
training procedure and slow convergence. PPA is computationally
Class N Class S Class V Class F
feasible, but not able to achieve high sensitivity and F score as
Chazal et al. [10] 92.60 51.09 79.76 0.15 other nonlinear techniques since it is just a nonlinear generaliza-
Llamedo et al. [34] 87.15 53.66 85.38 8.10 tion of PCA. FastICA in combination with PNN classifier provides
Ye et al. [24] 92.80 56.25 70.01 4.43
Llamedo et al. [35] 97.46 58.43 88.58 5.33
the highest sensitivity and F score with reduced computation time.
Zhang et al. [36] 93.69 49.45 88.96 23.95
Herry et al. [37] 95.18 12.19 75.15 0
Proposed 99.91 80.97 91.54 86.76 References

[1] R. Mehra, Global public health problem of sudden cardiac death, J.

Based on experimental results obtained, it is found that fast
ICA dimensionality reduction technique with ‘tanh’ non-linearity
produced the best average F-score of 99.83% when tenfold cross
validation was done by dividing the entire dataset into ten equal
sets. The problem with this cross validation approach is that there
is a chance for training and testing matrix to get ECG beats from
same patient. Hence experiments were repeated by dividing entire
patient records into ten instead of dividing entire ECG beats into
ten partitions, so that 10% of patient records are tested against 90%
of patient records. The results obtained by using fast ICA (tanh
non-linearity) DR technique are compared with six arrhythmia
classifiers and results are shown in Table 6. Best F Score obtained
for each class by arrhythmia classifiers are shown bolded.
Automatic heartbeat classification system was proposed by [10].
Morphological and temporal features were used by them and
achieved 75.9% sensitivity, 38.5% positive predictivity and 4.7% false
positive rate for S class. Two ECG leads were used by them for
feature extraction and final decision was obtained by combining
the outputs from two linear discriminant classifiers. The proposed
method used ECG signal from only one lead. Random selection
of training samples was done by [24]. Cluster based classification
was done by [35]. Classifier performance was improved by them
with expert assistance. The proposed method does not require any
expert assistance. Feature selection was made by one-versus-one
feature ranking stage by [36] and they achieved 88.94%, 79.06%,
85.48% and 93.81% sensitivities for classes N, S, V and F respec-
tively using two ECG leads. Synchrosqueezing transform was used
to enhance R peak detection by [37] and achieved 91.02% over-
all accuracy. But classification performance of classes other than N
and V were very poor. The proposed model used only DWT features
reduced using fast ICA. The results show that F Score of classifier
in classifying class S is 80.97% and class F is 86.76%. The reason for
low F Score is because of heavy overlap of class S with class N and
few training examples available for certain classes. Future work is
to overcome class overlap problem and class imbalance problem in
arrhythmia classification. Comparison of proposed system is done
with existing systems that followed AAMI recommended practice
of arrhythmia classification in which training and testing data do
not contain ECG beats from same patient.
6. Conclusion
Classification of cardiac arrhythmias into five different classes
as per AAMI standard is performed. Classifier performance is
Electrocardiol. (2007) 118–122, http://dx.doi.org/10.1016/j.jelectrocard.2007.
06.023.
[2] J. Kim, S. Shin H, K. Shin, M. Lee, Robust algorithm for arrhythmia
classification in ECG using extreme learning machine, Biomed. Eng. Online 8
(1) (2009) 1–12, http://dx.doi.org/10.1186/1475-925x-8-31.
[3] W. Liang, Y. Zhang, J. Tan, Y. Li, A novel approach to ECG classification based
upon two-layered HMMs in body sensor networks, Sensors (2014)
5994–6011, http://dx.doi.org/10.3390/s140405994.
[4] H. Kim, R.F. Yazicioglu, P. Merken, C. Van hoof, H.J. Yoo, ECG signal
compression and classification algorithm with quad level vector for ECG
holter system, IEEE Trans. Inf. Technol. Biomed. (2010) 93–100, http://dx.doi.
org/10.1109/TITB.2009.2031638.
[5] B.M. Asl, S.K. Setarehdan, M. Mohebbi, Support vector machine—based
arrhythmia classification using reduced features of heart rate variability
signal, Artif. Intell. Med. 44 (2008) 51–64, http://dx.doi.org/10.1016/j.artmed.
2008.04.007.
[6] J.J. Oresko, Z. Jin, S. Huang, Y. Sun, Heather Duschland Allen Cheng C. A
wearable smart phone based platform for real time cardiovascular disease
detection via electrocardiogram processing, IEEE Trans. Inf. Technol. Biomed.
14 (2010) 734–740, http://dx.doi.org/10.1109/TITB.2010.2047865.
[7] Kai Huang, Liqing Zhang, Cardiology knowledge free ECG feature extraction
using generalized tensor rank one discriminant analysis, EURASIP J. Appl.
Signal Process. 2014 (1) (2014) 1–15, http://dx.doi.org/10.1186/1687-6180-
2014-2.
[8] B. Zhu, Y. Ding, K. Hao, A novel automatic detection for ECG arrhythmias using
maximum margin clustering with immune evolutionary algorithm, Comput.
Math. Methods Med. 2013 (April) (2013) 1–8, http://dx.doi.org/10.1155/2013/
453402.
[9] R.J. Martis, U. Rajendra Acharya, Min Lim Choo, ECG beat classification using
PCA, LDA, ICA and discrete wavelet transform, Biomed. Signal Process. Control
8 (2013) 437–448, http://dx.doi.org/10.1016/j.bspc.2013.01.005.
[10] P. Chazal, M.O. Dwyer, Automatic classification of heartbeats using ECG
morphology and heartbeat interval features, IEEE Trans. Biomed. Eng. 51
(2004), http://dx.doi.org/10.1109/tbme.2004.827359.
[11] Laurens van der Maaten, Eric Postma, Dimensionality Reduction: A
Comparative Review, TiCC, Tilburg University, 2009.
[12] M.H. Song, Jeon Lee, Sungpilcho, Kyoung Joung Lee, Sun Kook Yoo, Support
vector machine based arrhythmia classification using reduced features, Int. J.
Control Automat. Syst. 3 (2005) 571–579.
[13] Z. Zidelmal, A. Amirou, D. Ould Abdeslamand, J. Merckle, ECG beat
classification using a cost sensitive classifier, Comput. Method Prog. Biomed.
111 (2013) 570–577, http://dx.doi.org/10.1016/j.cmpb.2013.05.011.
[14] A. Daamouche, L. Hamami, N.A. Farid Melgani, A wavelet optimization
approach for ECG signal classification, Biomed. Signal Process. Control 7
(2012) 342–349, http://dx.doi.org/10.1016/j.bspc.2011.07.001.
[15] E.D. Ubeyli, Usage of Eigen vector methods in implementation of automated
diagnostic systems for ECG beats, Digit. Signal Process. 18 (2008) 33–48,
http://dx.doi.org/10.1016/j.dsp.2007.05.005.
[16] F. Melgani, Y. Bazi, Classification of electrocardiogram signals with support
vector machines and particle swarm optimization, IEEE Trans. Inf. Technol.
Biomed. 12 (2008) 667–677, http://dx.doi.org/10.1109/TITB.2008.923147.
[17] J. Wiensand, J.V. Guttag, Patient adaptive ectopic beat classification using
active learning, Proceedings of the 2010 Computing in Cardiology, IEEE (2010)
109–112 http://hdl.handle.net/1721.1/73888.
[18] Manab Kumar Das, Samit Ari, ECG beats classification using mixture of
features, Int. Sch. Res. Notices 2014 (2014).
[19] Manab Kumar Das, Samit Ari, Electrocardiogram beat classification using
S-transform based feature set, J. Mech. Med. Biol. (2014) 14.
8 R. Rajagopal, V. Ranganathan / Biomedical Signal Processing and Control 34 (2017) 1–8
[20] Hari Mohan Rai, Anurag Trivedi, Shailja Shukla, ECG signal processing for
abnormalities detection using multi-resolution wavelet transform and
artificial neural network classifier, Measurement 46 (2013) 3238–3246,
http://dx.doi.org/10.1016/j.measurement.2013.05.021.
[21] Behnaz Ghoraani, Sridhar Krishnan, Discriminant non-stationary signal
features clustering using hard and fuzzy cluster labeling, EURASIP J. Adv.
Signal Process. 2012 (1) (2012) 1–20, http://dx.doi.org/10.1186/1687-6180-
2012-250.
[22] Fahim Sufi, Ibrahim Khalil, Abdun Naser Mahmood, A clustering based system
for instant detection of cardiac abnormalities from compressed ECG, Expert
Syst. Appl. 38 (2011) 4705–4713, http://dx.doi.org/10.1016/j.eswa.2010.08.
149.
[23] Sung Nien Yu, Combining independent component analysis and back
propagation neural network for ECG beat classification, Proc. Conf. IEEE Eng.
Med. Biol. Soc. (2006) 3090–3093.
[24] Can Ye, B.V.K. Vijayakumar, M.T. Coimbra, Heartbeat classification using
morphological and dynamic features of ECG signals, IEEE Trans. Biomed. Eng.
(2012) 59.
[25] G.B. Moody, R.G. Mark, The impact of the MIT BIH arrhythmia database, IEEE
Eng. Med. Biol. 20 (2001) 45–50, http://dx.doi.org/10.1109/51.932724.
[26] E.B. Mazomenos, D. Biswas, A. Acharyya, Taihai Chen A low-complexity ECG
feature extraction algorithm for mobile healthcare applications, IEEE J.
Biomed. Health Inform. 17 (2013) 459–569, http://dx.doi.org/10.1109/TITB.
2012.2231312.
[27] Li Deqiang, Pedrycz Witold, J. Nicolino Pizzi, Fuzzy wavelet packet based
feature extraction method and its application to biomedical signal
classification, IEEE Trans. Biomed. Eng. 52 (2005) 1132–1139, http://dx.doi.
org/10.1109/TBME.2005.848377.
[28] B.N. Singh, A.K. Tiwari, Optimal selection of wavelet basis function applied to
ECG signal denoising, Digit. Signal Process. 16 (2006) 275–287, http://dx.doi.
org/10.1016/j.dsp.2005.12.003.
[29] J. Pan, J.W. Tompkins, A real time QRS detection algorithm, IEEE Trans.
Biomed. Eng. (1985) 32, http://dx.doi.org/10.1109/tbme.1985.325532.
[30] Aapo Hyvarinen, Fast and robust fixed point algorithms for independent
component analysis, IEEE Trans. Neural Netw. (1999) 626–634, http://dx.doi.
org/10.1109/72.761722.
[31] J. Shawe Taylor, N. Christianini, Kernel Methods for Pattern Analysis,
Cambridge University Press, Cambridge, UK, 2004, http://dx.doi.org/10.1017/
CBO9780511809682.
[32] M. Scholz, R. Vigario, P.C.A. Nonlinear, A new hierarchical approach, Proc.
ESANN (2002) 439–444.
[33] V. Laparra, S. Jimenez, D. Tuia, G. Camps-Valls, J. Malo, Principal polynomial
analysis, Int. J. Neural Syst. (2014) 26.
[34] M. Llamedo, Martinez Juan Pablo, Heartbeat classification using feature
selection driven by database generalization criteria, IEEE Trans. Biomed. Eng.
58 (2011) 616–625, http://dx.doi.org/10.1109/TBME.2010.2068048.
[35] M. Llamedo, J.P. Martinez, An automatic patient adapted ECG heartbeat
classifier allowing expert assistance, IEEE Trans. Biomed. Eng. 59 (2012)
2312–2320, http://dx.doi.org/10.1109/TBME.2012.2202662.
[36] Z. Zhang, J. Dong, X. Luo, K. Choi, X. Wu, Heartbeat classification using disease
specific feature selection, Comput. Biol. Med. 46 (2014) 79–89, http://dx.doi.
org/10.1016/j.compbiomed.2013.11.019.
[37] C.L. Herry, M. Frasch, H. Wu, Characterization of ECG patterns with the
synchrosqueezing transform. (2015). arXiv:1510.02541.