Wavelet based detection of ventricular arrhythmias with neural network classifier

doi:10.4236/jbise.2009.26064

Paper Menu >>

Journal Menu >>

Vol.2, No.6, 439-444 (2009)

doi:10.4236/jbise.2009.26064

JBiSE

Wavelet based detection of ventricular arrhythmias with

neural network classifier

Sankara Subramanian Arumugam1, Gurusamy Gurusamy2, Selvakumar Gopalasamy3

1Department of Electrical and Electronics Engineering, NPA Centenary Polytechnic College, Kothagiri, Tamilnadu, India;

2Department of Electrical and Electronics Engineering, Bannariamman Institute of Technology, Sathyamangalam, Tamilnadu, India;

3Department of Electrical Sciences, Vinayaka Mission’s Kirupananda Variyar Engineering College, Salem, Tamilnadu, India.

Email: [email protected]; [email protected]

Received 29 August 2009; revised 10 September 2009; accepted 15 September 2009.

ABSTRACT

This paper presents an algorithm based on the

wavelet decomposition, for feature extraction

from the Electrocardiogram (ECG) signal and

recognition of three types of Ventricular Ar-

rhythmias using neural networks. A set of Dis-

crete Wavelet Transform (DWT) coefficients,

which contain the maximum information about

the arrhythmias, is selected from the wavelet

decompositio n. Th ese co eff icients are fed to t h e

feed forward neural network which classifies

the arrhythmias. The algorithm is applied on the

ECG registrations from the MIT-BIH arrhythmia

and malignant ventricular arrhythmia databases.

We applied Daubechies 4 wavelet in our algo-

rithm. The wavelet decomposition enabled us to

perform the task efficiently and produced reli-

able result s.

Keywords: Daubechies 4 Wavelet; ECG; Feed

Forward Neural Network; Ventricular Arrhythmias;

Wavelet Decomposition

1. INTRODUCTION

The cardiac disorders which are life threatening are the

ventricular arrhythmias such as Ventricular Tachycardia

(VT), Ventricular Fibrillation (VFIB) and Ventricular

Flutter (VFL). The classification of ECG into these dif-

ferent pathological disease categories is a complex task.

Successful classification is achieved by finding the

characteristic shapes of the ECG that discriminate effec-

tively between the required diagnostic categories. Con-

ventionally, a typical heart beat is identified from the

ECG and the component waves of the QRS, T and P

waves are characterized using measurements such as

magnitude, duration and area (Figure 1).

In an arrhythmia monitoring system or a defibrillator,

it is important that the algorithm for detecting ECG ab-

normalities should be reliable. The patient will be loos-

ing a chance of treatment if the system is not able to de-

tect the arrhythmia. Also a false positive detection will

initiate a defibrillator to give improper therapeutic inter-

vention. Both situations are linked with the patient’s life.

An algorithm based on Wavelet Transform is quite ef-

ficient for the detection of ventricular arrhythmias com-

pared to the Discrete Fourier Transform (DFT) since it

permits analyzing a discrete time signal using frequency

components. The DWT has been used for analyzing,

decomposing and compressing the ECG signals [1].

The wavelet transforms make possible, the decompo-

sition of a signal into a set of different signals of restrict-

ed frequency bands. Wavelet processing can be consid-

ered as a set of band pass filters [2]. Moreover, the dis-

crete wavelet transform corresponds to a multiresolution

analysis [3] which can reduce the redundancy of each

filtered signal so that the processing algorithm can be

applied effectively to a small data subset of the original

signal [4,5,6,7,8,9].

A classification scheme is developed in which a feed

forward neural network is used as a classification tool

depending on the distinctive frequency bands of each

arrhythmia. The Back Propagation (BP) algorithm al-

lows experiential acquisition of input/output mapping

Figure 1. Normal ECG waveform.

A. S. Subramanian et al. / J. Biomedical Science and En gineering 2 (2009) 439-444

440

knowledge within multi-layer networks [10]. BP per-

forms the gradient descent search to reduce the mean

square error between the actual output of the network

and the desired output through the adjustment of the

weights. It is highly accurate for most classification

problems because of the property of the generalized data

rule. In the traditional BP training, the weights are

adapted using a recursive algorithm starting at the output

nodes and working back to the first hidden layer.

In our study, an algorithm is implemented to detect

and classify Ventricular Tachycardia (VT), Ventricular

Fibrillation (VFIB) and Ventricular Flutter (VFL) using

wavelet decomposition and neural classification.

The ECG registrations from MIT-BIH arrhythmia and

malignant ventricular arrhythmia databases are used here.

The analysis is done using Daubechies 4 wavelet.

2. VENTRICULAR ARRHYTHMIA

Arrhythmias are the abnormal rhythms of the heart.

They cause the heart to pump the blood less effectively.

Most cardiac arrhythmias are temporary and benign. The

ventricular arrhythmias are life threatening and need

treatment. Such ventricular arrhythmias are Ventricular

Tachycardia, Supraventricular Tachycardia, Ventricular

Fibrillation and Ventricular Flutter.

VT is a difficult clinical problem for the physicians

(Figure 2). Its evaluation and treatment are compli-

cated because it often occurs in life-threatening situa-

tions that dictate rapid diagnosis and treatment. Ven-

tricular tachycardia is defined as three or more beats of

ventricular origin in succession at a rate greater than 100

beats/minute. There are no normal-looking QRS com-

plexes. Because ventricular tachycardia originates in the

ventricle, the QRS complexes on the electrocardiogram

are widened (>0.12 seconds). Ventricular tachycardia has

the potential of degrading to the more serious ventricular

fibrillation.

VFIB is a condition in which the heart's electrical ac-

tivity becomes disordered (Figure 3). During Ventricular

Figure 2. ECG waveform with VT.

Figure 3. Ventricular fibrillation.

Figure 4. Ventricular flutter.

fibrillation, the heart's ventricles contract in a rapid and

unsynchronized way. It is a medical emergency. If this

condition continues for more than a few seconds, blood

circulation will cease, as evidenced by lack of pulse,

blood pressure and respiration, and death will occur [11].

VFL is a tachyarrhythmia characterized by a high

ventricular rate with a regular rhythm (Figure 4). The

ECG shows large sine wave-like complexes that oscil-

late in a regular pattern. There is no visible P wave. QRS

complex and T wave are merged in regularly occurring

undulatory waves with a frequency between 180 and 250

beats per minute. In severe cardiac or systemic disease

states, ventricular tachycardia can progress to ventricular

flutter, then to ventricular fibrillation.

3. COMPARISION OF EXISTING

ALGORITHMS

Computer based detection and classification algorithms

can achieve good reliability, high degree of accuracy and

offer the potential of affordable mass screening for car-

diac abnormalities.

Till now, many linear techniques for the detection of

ventricular arrhythmias have been developed, such as the

probability density function method [12], rate and ir-

regularity analysis, analysis of peaks in the short-term

A. S. Subramanian et al. / J. Biomedical Science and En gineering 2 (2009) 439-444

441

autocorrelation function [13], sequential hypothesis tes-

ting algorithm [14], correlation waveform analysis, four

fast template matching algorithms, VF-filter method [15,

16], spectral analysis [17], and time-frequency analysis

[18]. However, these methods exhibit disadvantages,

some being too difficult to implement and compute for

automated external defibrillators (AED’s) and implant-

able cardioverter defibrillators (ICD’s), and some only

successful in limited cases. For example, the linear tech-

niques [13,18] using the features of amplitude or fre-

quency have shown their limits, since the amplitude of

ECG signal decreases as the VFIB duration increases,

and the frequency distribution changes with prolonged

VFIB duration.

Openly accessible at

The VF filter method relies on approximating the

VFIB signal as a sinusoidal waveform. The method is

equivalent to using a bandpass filter, the central fre-

quency of which is the mean signal frequency. The spec-

tral analysis technique relies on the fact that the VFIB

frequency contents are concentrated in the bandwidth

4-7Hz [13]. The increased power in this band of fre-

quencies is the major indication of the presence of VFIB.

The above spectral analysis technique is applied to sta-

tionary signals. However, abrupt changes in the non-

stationary ECG signal are spread over the whole fre-

quency range. Important time varying statistical charac-

teristics are lost once the signal has been Fourier trans-

formed.

In recent years time-frequency analysis techniques

have proved to be useful in experimental and clinical

cardiology. These techniques are needed to fully de-

scribe and characterize the various ventricular arrhyth-

mias and facilitate the development of new detection

schemes with high correct detection rate, or equivalently,

with low false-positive and false-negative performance

statistics. The wavelet transforms make possible, the

decomposition of a signal into a set of different signals

of restricted frequency bands. Wavelet processing can be

considered as a set of band pass filters [15]. Moreover,

the discrete wavelet transform corresponds to a multi-

resolution analysis [12] which can reduce the redun-

dancy of each filtered signal so that the processing algo-

rithm can be applied effectively to a small data subset of

the original signal.

A classification scheme is developed in which a feed

forward neural network is used as a classification tool

depending on the distinctive frequency bands of each

arrhythmia. The algorithm is applied to ECG signals

with ventricular arrhythmia disorders.

4. WAVELET DECOMPOSITION

The wavelet transform is a mathematical tool for de-

composing a signal into a set of orthogonal waveforms

localized both in time and frequency domains. The de-

composition produces coefficients, which are functions

of the scale (of the wavelet function) and position (shift

across the signal).

A wavelet which is limited in time and frequency is

called “mother wavelet”. Scaling and translation of the

mother wavelet gives a family of basis functions called

“daughter wavelets”.

The wavelet transform of a time signal at any scale is

the convolution of the signal and a time-scaled daughter

wavelet. Scaling and translation of the mother wavelet is

a mechanism by which the transform adapts to the spec-

tral and temporal changes in the signal being analyzed.

The continuous wavelet transform for the signal x(t) is

defined as:



ba 











 





)(



(1)

where a and b are the dilation (scaling) and translation

parameters, respectively. A wide variety of functions can

be chosen as mother wavelet ψ, provided ψ(t) Є L2 and

0)( 







dtt



(2)

The DWT makes possible the decomposition of ECG

at various scales into its time-frequency components. In

DWT two filters, a Low Pass Filter (LPF) and a High

Pass Filter (HPF) are used for the decomposition of ECG

at different scales. Each filtered signal is down sampled

to reduce the length of the component signals by a factor

of two. The output coefficients of LPF are called the

Approximation while the output coefficients of HPF are

called the Detail. The approximation signal can be sent

again to the LPF and HPF of the next level for second

level of decomposition; thus we can decompose the sig-

nal into its different components at different scale levels.

Figure 5 shows the decomposition process of a signal

into many levels. The details of all levels and the ap-

proximation of the last level are saved so that the origi-

nal signal could be reconstructed by the complementary

filters. The reconstructed signals at each level are repre-

sented by the notations D1, D2, D3 and so on.

Figure 5. Wavelet decomposition and reconstruction.

A. S. Subramanian et al. / J. Biomedical Science and En gineering 2 (2009) 439-444

442

Figure 6. Ideal frequency bands for the various details.

Figure 6 shows the ideal frequency bands for a sam-

pling frequency of 360 samples/second. Depending on

the scaling function and the mother wavelet, the actual

frequency bands and consequent frequency selectivity of

the details are slightly different.

5. METHODOLOGY

In this section we present the algorithm which efficiently

detects and classifies the various ventricular arrhythmias

using wavelet decomposition and neural classification.

5.1. Description of the Algorithm

The algorithm first decomposes the ECG signals using

wavelet transform. The decomposed signals are fed to

the feed forward neural network. The wavelet theory is

used as a time-frequency representation technique to

provide a method for enhancing the detection of life

threatening arrhythmias. It reveals some interesting

characteristic features such as low frequency band (0-5

Hz) for VFL, two distinct frequency bands (2-5, 6-8 Hz)

for VT, and a broad band (2-10 Hz) for VFIB.

A classification scheme is developed in which a neural

network is used as a classification tool depending on the

above distinctive frequency bands of each arrhythmia.

The algorithm is applied to ECG signals obtained from

patients suffering from the arrhythmias, mentioned above.

5.2. ECG Analysis Using Wavelet

Decomposition

To quantify the differences between the various ar-

rhythmias with the help of the wavelet transform, the

densities for different frequency bands are compared.

The wavelet transform is performed using Daubechies 4

wavelet since it provides better sensitivity [19].

The algorithm computes the volume underneath the

3D plots of the square modulus of the wavelet transform

for several regions of the time-frequency plane. The

time-frequency plane is divided into seven bands rang-

ing from 0 to 15 Hz. For sinus rhythm the energy is cal-

culated within the time intervals T1 and T2 integrated

over the whole frequency axis. The time interval T1 is

determined by the region of QRS complex, and the time

interval T2 is determined by the region of the T-wave.

As the wavelet transform is very sensitive to abrupt

changes in the time direction, the energy parameter over

the given time intervals attains relatively large values for

normal subjects. This parameter is referred to as Tv and

it defines the sum of the energy parameters computed

within the intervals T1 and T2. Although the signals of

VFL and VT exhibit a QRS-complex, the parameter

value T2 for these signals remains relatively small, ow-

ing to the absence of abrupt changes in the region of the

T-wave. Therefore, the value of Tv will still be smaller

than that of the normal subjects.

The 2D and 3D wavelet transform contours of Normal

Sinus Rhythm (NSR), VT and VFL are shown in Figure 7.

Figure 7. (a) 2D wavelet of NSR, (b) 2D wavelet of VT, (c) 2D wavelet of VFL, (d) 3D wavelet of NSR, (e) 3D wavelet of VT

and (f) 3D wavelet of VFL.

A. S. Subramanian et al. / J. Biomedical Science and En gineering 2 (2009) 439-444

443

ful

learni

(3)

The weights and biases of the A

work (ANN)

in the direction of the gradient, but in the

3. Classification of Ventricular

Arrhythmias Using Neural Network

There are quite lot of network topologies with power

ng strategies which exist to solve nonlinear prob-

lems [20,21,22,23]. For the classification of Ventricular

Arrhythmias, back propagation with momentum is used

to train the feed forward neural network [24].

A multilayer feed forward neural network with one

layer of hidden (Z) units is used for this problem. The

tput (Y) units have weights wjk and the hidden units

have weights vjk. During the training phase, each output

neuron compares its computed activation yk with its tar-

get value dk to determine the associated Error (E) by

using (3) for the pattern with that neuron.



)(





kk ydE

rtificial Neural Net-

are adjusted to minimize the least-square

ror. The minimization problem is solved by gradient

technique, the partial derivatives of E with respect to

weights and biases are calculated using the generalized

delta rule. This is achieved by the back propagation of

the error.

When using momentum, the neural network is pro-

ceeding not

direction of the combination of the current gradient and

the previous direction of weight correction. Convergence

is sometimes faster if a momentum term is added to the

weight update formula. In the back propagation with

momentum, the weights for the training step t+1 are

based on the weights at training steps t and t-1. The

weight update formulae for back propagation with mo-

mentum are given by (4)

)]1()([)()1(  twtwztwtw kjkjk

)]1()([)()1(  tvtvxtvtv ijijijijij

jkjkj







(4)

where the learning factor α and momentum param

eural net-

d for the analysis [25]. The

rom the records 200, 203, 205, 207, 208,

14, 215, 217, 221, 223, 233, 106 etc were

re used for training. The learning

onverge

of VT,

) and (6).

eter µ

are constrained to be in the range 0 to 1, exclusive of the

end points. The weights and biases are initialized to

some random values and updated in each iteration until

the network has settled down to a minimum.

The classification of the VT, VFIB and VFL arrhyth-

mia is carried out using a Back-propagation n

work whose input is the energy level calculated by the

wavelet transform as described above, the output is a

three bit pattern, in which 100 corresponds to VFL, and

010 corresponds to VT, and 001 corresponds to VFIB.

The network has three layers: The input, the output

and one hidden layer. The input layer has seven nodes

representing the seven different frequency bands from

0-15 Hz. The output layer has three nodes that represent

the three different types of arrhythmia signals VFL, VT

effective size of the network and acceptable efficiency.

The learning rate and the momentum term are chosen to

be 0.7 and 0.3 respectively.

5.4. Data

The MIT-BIH arrhythmia and malignant ventricular ar-

rhythmia databases were use

and VFIB. The hidden layer consists of six nodes fo

VT episodes f

210, 213, 2

taken from the MIT-BIH arrhythmia database. VFIB and

VFL episodes were from MIT-BIH malignant ventricular

arrhythmia database from the records 418 – 430, 602,

605, 607, 609, 610, 611, 612, 614, 615 and cu01 – cu35.

The data files from arrhythmia database and malignant

ventricular arrhythmia database were sampled at 360

samples / second.

6. RESULTS AND CONCLUSIONS

Twenty-five signals of VT, ten signals of VFL and five

signals of VFIB a

process took approximately 1000 iterations to c

with classification error of 0.001. Thirty signals

fifteen signals of VFL and four signals of VFIB are se-

lected as test set. For decomposition of ECG signal

Daubechies 4 wavelet is used.

In medical statistics, few parameters are important to

evaluate the performance of the algorithm. These pa-

rameters are sensitivity (Se) and positive predictivity (+P)

which can be computed using (5

FNTP

Se

 (5)

FPTP

P

 (6)

where TP is True Positive, FP is False Positive and

False Negative.

Classification results of testi

Forward Neural Network with BP are shown in Table 1.

ias are shown in Table 2. The results

Detected Ventricular Arrhythmia Total

FN is

ng data sets using Feed

Sensitivity and Positive predictivity for the various ven-

tricular arrhythm

ow that the BP algorithm is reliable in the classifica-

tion of all the three types of ventricular arrhythmias.

Table 1. A comparison between the actual and detected ven-

tricular arrhythmias in terms of the number of patterns.

Actual Ventricular

Arrhythmia VT VFL VFIB 49

VT 30 0 0 30

VFL 1 13 1 15

VFIB 0 04 4

A. S. Subramanian et al. / J. Biomedical Science and En gineering 2 (2009) 439-444

444

Table 2sitivity and positive predictivity r the testing s

Ven Ar-TP FP FN sitive Pr

ctivity (%Sensitivity (%)

. Senfoet.

Openly accessible at

tricular

rhythmia

Po e-

di)

VT 30 1 0 96.77 100

VFL

VFIB

100

86.67

100

Tabl shows that te BP network misclashe

properythmi mcasee VFL

classified as VFIB, this is ause energy level in the

equency bands is high and common between VFL

ents, IEEE Signal

2(7).

Wavelets and filter banks: Theory and

rs,

ibrillation, IEEE Eng. Boil., 152–159.

matic im-

detection by a regression test on the autocor-

ion algorithm for cardiac arrhythmia classification,

of ventricular fibrillation using neural net-

–349.

lter banks, IEEE

Wavelet decomposition for detection and classifi-

nal of Elec-

l networks, Journal on Medical

s, Journal of

analysis, IEEE Trans. on Sig-

nd radial basis neural networks for

tection, Proceedings of World Congress on Medical

Physics and Biomedical Engineering, Chicago, USA.

[10] D. E. Rumelhart, G. E. Hinton, and R. J Williams, (1986)

Learning representations by back-propagation erro

Nature.

[11] V. X. Afonso and W. J. Tompkins, (1995) Detecting ven-

tricular f

[12] A. Langer, M. S. Heilman, and M. M. Mower, (1976)

Considerations in the development of the auto

e 1

arrh

sified t

case was a ine s, on

becthe

and relation function, Med. Biol. Eng. Comput., 25(3), 241–

249.

[14] S. W. Chen, P. W. Clarkson, and Q. Fan, (1996) A robust

detect

plantable defibrillator, Medical Instrumentation, 10(3),

163–167.

[13] S. Chen, N. V. Thakor, and M. M. Mover, Ventricular

fibrillation

FIB. The algorithm is able to classify VT and VFIB

with 100% sensitivity. The positive predictivity for VFL

episodes is 100%. The algorithm is reliable by providing

the overall sensitivity of 95.56% and the overall positive

predictivity of 92.26%. The algorithm can be validated

using more number of ECG samples.

REFERENCES

[1] K. Anant, F. Dowla, and G. Rodrigue, (1995) Vector

quantization of ECG wavelet coeffici

IEEE Transactions on Biomedical Engineering, 43,

1120–1125.

[15] R. H. Clayton, A. Murray, and R.W. F. Campbell, (1994)

Recognition

works, Med. Bio. Eng. Comp., 32, 217–220.

[16] S. Kuo and R. Dillman, (1978) Computer detection of

ventricular fibrillation, Computer Cardiology, 347

Processing Letters,

[2] M. Vetterli, (1992)

design, IEEE Transactions on Signal Processing, 2207–

2232.

[3] R. M. Rao and A. S. Bopardikar, (1998) Wavelet trans-

forms: Introduction to theory and applications, Addison

Wesley Longman.

[4] L. Khadra, A. S. Al-Fahoum, and H. Al-Nashash, (1997)

Detection of life threatening cardiac arrhythmia using the

wavelet transformation, Med. Biol. Eng. Comput., 35,

626–632.

[5] P. S. Addison, J. N. Watson, G. R. Clegg, M. Holzer, F.

Sterz, and C. E. Robertson, (2000) Evaluating arrhyth-

mias in ECG signals using wavelet transforms, IEEE En-

gineering in Medicine and Biology Magazine, 19, 104–

109.

[6] H. A. N. Dinh, D. K. Kumar, N. D. Pah, and P. Burton,

(2001) Wavelets for QRS detection, Proceedings of the

23rd Annual Conference, IEEE EMS, Istanbul, Turkey,

35–38.

[7] S. Kadambe, R. Murray, and G. F. Boudreaux-Bartels,

(1999) Wavelet transform based QRS complex detector,

IEEE Transaction on Biomedical Engineering, 46(7),

838–848.

[8] I. Romero, L. Serrano, and Ayesta, (2001) ECG fre-

quency domain features extraction: A new characteristic

for arrhythmias classification, Conference of the IEEE

Engineering in Medicine and Biology Society.

[9] S. M. Szilagyi and L. Szilagyi, (2000) Wavelet transform

and neural network based adaptive filtering for QRS de-

[17] S. Barro, R. Ruiz, D. Cabello, and J. Mira, (1989) Algo-

rithmic sequential decision-making in the frequency do

main for life threatening ventricular arrhythmias and

imitative artifacts: A diagnostic system, Journal on Bio-

medical Engineering, 11(4), 320–328.

[18] V. X. Afonso, W. J. Tompkins, T. Q. Nguyen, and S. Luo,

(1999) ECG beat detection using fi

Transactions on Biomedical Engineering, 46(2), 192–

202,.

[19] G. Selvakumar, B. K. Bhoopathy, and R. B. Chidhambara,

(2007)

cation of critical ECG arrhythmias, Proc. of the 8th

WSEAS Int. Conf. on Mathematics And Computers in

Biology and Chemistry, Vancouver, Canada.

[20] G. Bortalan and J. L. Willems, (1993) Diagnostic ECG

classification based on neural networks, Jour

trocardiology, 26, 75–79.

[21] Z. Dokur, T. Olmez, and E. Yazgan, (1997) Detection of

ECG waveforms by neura

Engineering and Physics, 19(8), 738–741.

[22] L. Edenbrandt, B. Heden, and O. Pahlm, (1993) Neural

networks for analysis of ECG complexe

Electrocardiology, 26, 74.

[23] R. Silipo and C. Marchesi, (1998) Artificial neural net-

works for automatic ECG

nal Processing, 46.

[24] A. S. Al-Fahoum and I. Howitt, (1999) Combined wave-

let transformation a

classifying life threatening cardiac arrhythmias, Med.

Biol. Eng. Comput., 37, 566–573.

[25] MIT-BIH (http://www.physionet.org)