Tobej 4 217
Tobej 4 217
Tobej 4 217
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/51177719
CITATIONS READS
10 222
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Uma Shanker Tiwary on 16 May 2014.
Open Access
Automated Risk Identification of Myocardial Infarction Using Relative
Frequency Band Coefficient (RFBC) Features from ECG
Gohel Bakul* and U.S. Tiwary
Abstract: Various structural and functional changes associated with ischemic (myocardial infarcted) heart cause
amplitude and spectral changes in signals obtained at different leads of ECG. In order to capture these changes, Relative
Frequency Band Coefficient (RFBC) features from 12-lead ECG have been proposed and used for automated
identification of myocardial infarction risk. RFBC features reduces the effect of subject variabilty in body composition on
the amplitude dependent features. The proposed method is evaluated on ECG data from PTB diagnostic database using
support vector machine as classifier. The promising result suggests that the proposed RFBC features may be used in the
screening and clinical decision support system for myocardial infarction.
Keywords: Coronary artery disease, Myocardial infarction (MI), Electrocardiogram (ECG), cardiac vector, Support Vector
Machine (SVM).
(a) (b)
Fig. (1). (a) Normal waveforms of ECG (b) 12-lead ECG signal.
(a) (b)
Fig. (2). (a) Cardiac vector (black arrow in centre) (b), electrode connectivity of I, II, III lead and generation of ECG wave in different lead(s)
through projection of cardiac vector.
composition variability across different persons gives impedance (Ri) exert on different leads are proportional to
undesirable effect on ECG amplitude and its dependent each other. So
features. To reduce the body composition effect on ECG
Ri
amplitude, different methods have been proposed in = kij , i j (1)
literature, which are mainly based on correlation of ECG Rj
amplitude with parameters that provide information
regarding body composition, like age, sex, Body Mass Index Where i, j are different leads of ECG and kij can be
(BMI), fat-free mass derived from bioelectric impedance considered as constant for selected pair of leads (i, j).
method, skin fold thickness etc. [9-11]. Methods that use It should be noted that proportion of body parts or
age, sex and BMI information are not very effective. The composition (anthropometric measures) varies from one
estimation of fat-free mass is a tedious job and requires a person to another and hence kij also varies. But the amount of
different instrument. In this paper, the relative amplitude variation is very small compared to absolute body
based approach is adopted which does not require any body composition variations. Thus the relative coefficients reduce
composition information.
the body composition effect on ECG considerably.
Body mass (mainly body fat), distance between the
Voltage (V) of ECG signal depends on electrical current
electrodes or distance of electrodes from the heart
through heart (C) and body impendence (R). So
characterizes the body impedance [9, 12]. Let us assume the
body as structurally proportional in all directions from the V = CR (2)
heart for each subject. So the body mass is approximately
Vi Ci Ri
Vj =
proportional to the distance between electrodes and distance (3)
C j Rj
of electrodes from heart for each subject. Thus the body
RFBC in Identification of Myocardial Infarction The Open Biomedical Engineering Journal, 2010, Volume 4 219
Where F represents the Fourier transform. Standard-12 leads ECG signal (x) from each subject was
used. To remove low frequency baseline drift, ECG signal xi
Furthermore, to account for the variability (person to (where i = 1 to 12 leads) was passed through high pass
person and male to female) in distribution of fat between forward-backward FIR filter with 2Hz cut off frequency
limb and chest area, we treated limb leads and chest leads (Fig. 4b). ECG waves like R and S waves as well as T and
separately. inverted T waves have opposite directions but same
2.2. Representation of ECG Signal in Terms of RFBC frequency spectra. To separate them ECG signal xi was
Features divided into positive peak signal xip (above the baseline) and
In our previous paper we had defined the relative negative peak signal xin (below the baseline) (Fig. 4c and
amplitude features [8] as follows: eqn. 7).
It is the relative amplitude between the R or S peak p xi, xi > 0 0, xi > 0
(whichever is maximum) of one lead and R or S peak xi = and xin = x , x < 0 (7)
0, xi < 0 i i
(whichever is maximum) of another lead of ECG signal (eqn.
5 and Fig. 3). Relative amplitude features (eqn. 6) [8] are Fourier spectrum X ip and X in was calculated by taking
calculated for different combination of limb leads (I,II, III,
aVR, aVL, AVF) and different combination of chest leads fast discrete fourier transform of xip and xin respectively
(V1, V2, V3, V4, V5, V6 ) as follow (eqn. 8). As most of the energy of ECG waveforms is
concentrated in frequency up to 40 Hz only, the frequency
(ai a j ) spectrum up to 40 Hz has been considered in the analysis
Ai j = i j (5)
(ai + a j ) (Fig. 4d). This also removes high frequency noise and power
line noise (50 Hz), Frequency spectrum was broken into S
ai = argmax{ Ri , Si } equally spaced bins (frequency bands, sized (40-2)/S Hz))
and the cumulative amplitude of each bin bm was calculated
a j = argmax{ R j , S j } (Fig. 4e and eqn. 9)
N 1 j 2lk
here i, j = 1 to 6 for limb leads and i, j = 7 to 12 for chest Xip [k] = xip [l]. e N
(8)
leads l=0
1.4
1.2
0.8
mV
0.6
0.4
0.2
-0.2
Time
(a)
1.2
0.8
0.6
mV
0.4
0.2
-0.2
Time
0. 9
(b)
0.8
0.7
0.6
0.5
mV
0.4
0.3
0.2
0.1
0
Time
(c)
700
600
500
400
300
200
100
0
2 5 10 15 20 25 30 35 40
Frequency (Hz)
(d)
90
80
70
60
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10 11 12
Bins
(e)
p p
Fig. (4). (a) ECG signal (b) baseline corrected signal (c) positive peak signal xi (d) DFT spectrum of xi (2 Hz to 40 Hz) (e) value of 12
p
equally spaced bins of DFT spectrum of xi .
RFBC in Identification of Myocardial Infarction The Open Biomedical Engineering Journal, 2010, Volume 4 221
Wi p [bm ] W jp [bm ] In the non-linear SVM, first the input vector xi is mapped
RFBC[bm ]ijp = (10) to higher dimension feature space F using mapping function
Wi [bm ] + W j [bm ]
p p
(x) and then a linear classification is performed in this
here, for limb leads (L) i = 1, j = 2 to 6 (5 pair(i, j)) feature space F. In practice kernel function K(xi, xj) is used,
as it offers implicit mapping into inner product feature space
For chest leads (C) i = 7, j = 8 to 12 (5 pair(i, j))
(reproducing kernel Hilbert spaces), K(xi, xj) = ( (xi), (xj))
Calculation of RFBC features of xin is also done in the = (xi, xj)D. This means explicit mapping (x) is not required,
same way. For one subject, total number of RFBC features as the inner product provides sufficient mapping. Mostly
with S number of bins is S*2*10. We have taken S = 12 in used kernels are polynomial, gaussian radial basis function
this analysis. (RBF), exponential radial basis function, sigmoid kernel
(tanh) etc. In this study, a sigmoid kernel was used.
3. CLASSIFICATION BASED ON RFBC FEATURES
4. RESULTS AND DISCUSSION
Various machine learning techniques have been
intensively employed in clinical decision support system ECG signal of healthy (HC) subjects and subjects with
particularly artificial neural network [5, 6] and support MI was downloaded from online PTB diagnostic database
vector machine [13]. SVM is preferred over ANN as it yields [15]. This dataset has 12 standard leads for each subject,
sampled at 1000 Hz sampling frequency. RFBC features
global and unique solution, have a simple geometrical
were calculated on 52 HC and 104 MI subjects and
interpretation, use structural risk minimization, and it is less
classification was done using support vector machine with
prone to over fitting etc. Thus we have used SVM for one leave out cross validation. Value of S is selected as 12,
classification. providing 3.16 Hz wide frequency band bin. By manually
SVM is the novel classification technique based on VC optimizing the parameters and kernel of SVM the following
dimension theory [13, 14]. The central idea of SVM is to results (Table 1) are obtained.
find hyperplanes (H1, H2) with largest possible margin M To compare the accuracy of RFBC features, we have also
that separates the data points X = {x1, x2 …xi } Rd in two tabulated our previous results using relative amplitude
classes yi {1,-1} by finding a weight vector w Rd and features [8]. The relative amplitude features yields 71.07 %
offset b. The hyperplanes (H1, H2) are pushed apart until they and 74.35% accuracy with Back-propagation neural network
meet corresponding class of data points which are called and Support vector machine based classifications
support vectors. Since respectively, while RFBC yields accuracy of 85.23 %,
sensitivity 85.57 % and Specificity: 83.97 % with support
2 vector machine (Table 1). Thus RFBC features outperform
M= (11)
w the relative amplitude features. Increase in accuracy with
RFBC is due to the fact that it captures information
((x, w) + b) +1 for yi = +1 (12) regarding all characteristic waves in terms of frequency
spectrum while the relative amplitude features capture
((x, w) + b) 1 for yi = 1 (13) information based on only R and S. However, it is worth
mentioning that the measurement of classification accuracy
Maximization of margin M requires minimization of
of acute MI is only possible when ECG data is labeled
w . So simultaneously with confirmatory test for acute MI e.g.
blood troponin level. The PTB diagnostic database is not
min (w, w)
w R d labeled with confirmatory test. Further, features may
represent chronic structural changes associated with
subject to yi ((x, w) + b) 1 0 i ischemic heart instead of ischemic event directly. So the
results mentioned above may not represent sensitivity and
One of the variants of this algorithm is soft margin SVM
specificity of acute MI in true sense. It is better to infer that
which allows minimum amount of misclassification or error the proposed features represent the risk of having MI rather
for dataset that are not linearly separable by hyper planes and than true sensitivity and specificity of acute ischemic event.
lead to following optimization problem
Different ECG features e.g. prominent Q-wave and
n
min (w, w) + C ventricular hypertrophy changes which are used presently to
2
i (14)
w R d i=1 identify risk of MI have low sensitivity [3, 9, 16]. Prominent
Q-wave in ECG also denotes old myocardial infarction.
subject to yi ((x, w) + b) 1 i and i 0 i Methods representing ventricular hypertrophic (VH) changes
have low sensitivity because of the effect of body
Where parameter are called slack variables that composition variation across the subjects [9, 16]. Exercise
represent total training error. The tradeoff between training Tolerance Test [1] is also useful in identifying risk of MI but
error and maximal margin M is controlled by the regulation cannot be performed in emergency. Hence our proposed
parameter C. If C is set too large, classifier is over fitted and features, as it has high accuracy, may be beneficial in risk
if it is set too small, classifier is under fitted, which means identification of MI at the time of emergency as well as for
large training error. So proper value of C is important for the regular screening of MI.
performance of the classifier.
222 The Open Biomedical Engineering Journal, 2010, Volume 4 Bakul and Tiwary
Table 1. Result of MI and HC ECG Signal Classification Based on Relative Amplitude Features and RFBC Features
Furthermore, this method also reduces the body [6] B. Hedén, H. Öhlin, R. Rittner and L. Edenbrandt “Acute
composition effect on features without using any body myocardial infarction detected in the 12-lead ECG by artificial
neural networks”, Circulation, vol. 96, pp. 1798-1802, 1997.
specific information like age, sex, body weight, BMI and [7] G. Cliford, F. Azuaje and P. McSharry. Advanced Methods and
body impedance. Thus this method may be helpful in the Tools For ECG Data Analysis, Artech House Publishers, Oct.
diagnosis of ventricular hypertrophies through ECG. 2006.
[8] B. Gohel, U.S.Tiwary and T.Lahiri, “Relative Amplitude based
5. CONCLUSION AND FUTURE WORK Features of characteristic ECG-Peaks for Identification of Coronary
Artery Disease”, in proceeding of International Conference on
Here, we have presented the new features, RFBC from Intelligent Human Computer Interaction, Jan 20-23, 2009
12-lead ECG that discriminate MI subject from healthy Allahabad, India. Springer, 2009.
subject automatically with good accuracy. The main [9] O. Tochikubo, E. Miyajima, T. Shigemasa and M. Ishii “Relation
between body fat–corrected ECG voltage and ambulatory blood
characteristic of this method is the reduced effect of inter- pressure in patients with essential hypertension”, Hypertension,
subject body composition variation. It might be used for the vol. 33, pp. 1159-1163, 1999.
identification of the risk of MI or in clinical decision support [10] P. M. Okin, S. Jern, R. B. Devereux, S. E. Kjeldsen and B. Dahlöf,
system for MI effectively. In future, we would like to “Effect of Obesity on Electrocardiographic left ventricular
evaluate the proposed features in identification of ventricular hypertrophy in hypertensive patients”, Hypertension, vol. 35, pp.
13-18, 2000.
hypertrophy. [11] H. C. Lukaski, W. W. Bolonchuk, C. B. Hall and W. A. Siders
“Validation of tetrapolar bioelectrical impedance method to assess
REFERENCES human body composition”, J. Appl. Physiol., vol. 60, pp. 1327-
1332, 1986.
[1] N. Herring and D.J. Paterson, “ECG diagnosis of acute ischemia [12] F. Morris, J. Edhouse, W. J. Brady and J. Camm, ABC of clinical
and infarction: past, present and future”, QJM, vol. 99, no. 4, pp. ectrocardiography, BMJ book, 2003.
219-230, 2006. [13] D. Confortia and R. Guido, “Kernel-based Support Vector Machine
[2] P. H. Ginn and Barbara, “How accurate is the use of ECGs in the classifiers for early detection of myocardial infarction”, Optim.
diagnosis of myocardial infarct?”, J. Fam. Pract., vol. 55, no. 6, Methods Softw., vol. 20, pp. 401-413, April 2005.
June, 2006. [14] S. Sanei and J.A. Chambers, EEG signal processing, John Wiley &
[3] F.M. Asch, Shah S, C. Rattin, S. Swaminathan, A. Fuisz and J. Sons Ltd, 2007.
Lindsay, “Lack of sensitivity of the electrocardiogram for detection [15] A. L. Goldberger, A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch.
of old myocardial infarction: a cardiac magnetic resonance imaging Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C. Peng and H. E.
study”, Am. Heart J., vol. 152, no. 4, pp. 742-748, Oct. 2006. Stanley, “PhysioBank, PhysioToolkit, and PhysioNet: Components
[4] J. Zimmerman, R. Fromm, D. Meyer, A. Boudreaux, C. Wun, R. of a new research resource for complex physiologic signal”,
Smalling, B. Davis, G. Habib and R. Roberts, “Diagnostic Marker Circulation, vol. 101, no. 23, pp. 215-220, June 2000.
Cooperative Study for the Diagnosis of Myocardial Infarction”, [16] D. Levy, S. B. Labib, K. M. Anderson, J. C. Christiansen, W. B.
Circulation, vol. 99, pp. 1671-1677, 1999. Kannel and W. P. Castelli, “Determinants of sensitivity and
[5] H. Haraldsson, L. Edenbrandt and M. Ohlsson “Detecting acute specificity of electrocardiographic criteria for left ventricular
myocardial infarction in the 12-lead ECG using Hermite hypertrophy”, Circulation, vol. 81, pp. 815-820, 1990.
expansions and neural networks”, Artif. Intell. Med., vol. 32, pp.
127-136, 2004.
Received: March 14, 2010 Revised: July 03, 2010 Accepted: July 09, 2010