Extraction of qrs complexes using automated bayesian regularization neural network

International Journal of Advanced Research in Engineering and Technology (IJARET), IN 0976
INTERNATIONAL JOURNAL OF ADVANCED RESEARCH ISSN
– 6480(Print), ISSN 0976 – 6499(Online) Volume 3, Number 2, July-December (2012), © IAEME
ENGINEERING AND TECHNOLOGY (IJARET)
ISSN 0976 - 6480 (Print)
ISSN 0976 - 6499 (Online) IJARET
Volume 3, Issue 2, July-December (2012), pp. 37-42
© IAEME: www.iaeme.com/ijaret.html
Journal Impact Factor (2012): 2.7078 (Calculated by GISI)
©IAEME
www.jifactor.com

EXTRACTION OF QRS COMPLEXES USING AUTOMATED BAYESIAN
REGULARIZATION NEURAL NETWORK

Nilesh Parihar, Ph. D*, Scholar, Department of ECE, J.N.U., Jodhpur, Rajasthan,
Dr. V. S. Chouhan, Department of Electronics and Communication Engineering, M. B. M. Engineering
College, Jodhpur, Rajasthan, India

ABSTRACT

An efficient algorithm for the detection of QRS complexes in 12 lead ECG is presented in this
paper. The algorithm is developed in MATLAB with standard CSE – ECG data base.
Preprocessing is done by using Kaiser-window for minimizing the noise interference and
differentiator for baseline drift removal. Bayesian regularization neural network is used to learn
the characteristics of QRS complex to detect R peak. This algorithm yields high detection
performance with detection rate of 98.5% sensitivity is 98.41% and positive predictivity of
98.6%.
1. INTRODUCTION

ECG is a tool that is widely used to understand the condition of the heart. ECG signal is the
electrical representation of the heart activity. In ECG different type of noise commonly
encountered artifacts included as a power line interference, electrode contact noise, motion
artifacts, base line drift, instrumentation, electrosurgical noise generated by electric devices [1].
Baseline drift is another important parameter to be suppressed for correct detection of QRS
complex. Many researchers have worked on development of methods for reduction of baseline
drift by using Kalman filter, cubic-spline, moving average algorithm and Chebyshave filters.

37

International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976

V.S. Chouhan and S.S. Mehta developed an effective algorithm for baseline drift removal using
least squares error correction and median based correction. [2].
For automatic ECG data monitoring various implementations have been done previously with
Multi level Perceptron and backpropagation training. It is a supervised learning algorithm, in
which a sum square error function is defined, and the learning process aims to reduce the overall
system error to a minimum [3, 4]. The network has been trained with moderate values of learning
rate and momentum. The weights are updated for every training vector, and the training function
is terminated when the sum square error reaches a minimum value. Also few problems generally
occur during neural network training that is, over fitting, early stopping and slow processing. For
effective training, it is desirable that the training data set be uniformly spread throughout the
class domains [5, 6]. In this algorithm automated regularization training function is used to
implement the optimal regularization in an automated fashion for the detection of R peaks. The
network with input layer consists of nodes and subsequent hidden layers [7, 8]. The neurons are
processed with the standard sigmoid activation function in this paper.

2. METHODOLOGY
2.1 Filter Design
In order to attenuate noise and remove the baseline drift we design and implement a FIR band
pass filter with Kaiser Window. The cut off frequency of filter is 0.5 – 40 Hz and order is 7.
ଶ଴ ୪୭୥൫௦௤௥௧ሺ௡∗௥ೞ ሻ൯ ିଵଷ
݇ܽ݅‫ = ݓ݋݀݊݅ݓ ݎ݁ݏ‬ భర.ల൫೑ೞ – ೑೛ ൯
೑

The resulting ECG signals are stable.
2.2 Bayesian Regularization Neural Network
In this algorithm we use a training algorithm which consistently produces networks with good
generalization. This method for improving generalization contains the size of the network
weights and is referred to as regularization.

38


Fig. 1 feed forward Neural Network
In this technique the data is divided into three subsets. The first subset is the training set, which
is used for computing the gradient and updating the network weights and biases. The second
subset is the validation set. The error on the validation set is monitored during the training
process. The validation error normally decreases during the initial phase of training, as does the
training set error. However, when the network begins to over fit the data, the error on the
validation set typically begins to rise. When the validation error increases for a specified number
of iterations, the training is stopped, and the weights and biases at the minimum of the validation
error are returned. The test set error is not used during training, but it is used to compare different
models. It is also useful to plot the test set error during the training process. If the error in the test
set reaches a minimum at a significantly different iteration number than the validation set error,
this might indicate a poor division of the data set [3, 8].
In that once the network weights and biases are initialized, the network is ready for training with
proper network inputs p and target outputs t.
In Automated Regularization (trainbr) it is desirable to determine the optimal regularization
parameters in an automated fashion. A possible step towards this process is the Bayesian
framework. The use of Bayesian regularization function is a combination with Levenberg-
Marquardt training process [9].
The trainbr algorithm generally works best when the network inputs and targets are scaled so that
they fall approximately in the range [-1, 1]. If the inputs and targets do not fall in this range, we
can use the function mapminmax to perform the scaling.
The algorithm is said to be converged if the sum squared error (SSE) and sum squared weights
(SSW) are relatively constant over several iterations.
ே
1
‫ = ݁ݏ݉ = ܨ‬෍ሺ݁௜ ሻଶ
ܰ
௜ୀଵ

39


3. ALGORITHM FOR DETECTION
1. Load ECG database files case by case as shown in fig (2a)
2. Implement FIR bandpass filter with Kaiser Window as shown in fig (2b).
3. Differentiate the output of step 2. These outputs pass through a moving average integrator
for form a proper shape and desired level as shown in fig (2c).
4. The output of the above step is passed through the neural network, which is having P as
an input. The training function is to train feed forward neural network with Bayesian
regularization function having target T with a defined learning rate µ=0.05 and number of
epoch=4. Further, in this step mapminmax function is used.
5. Then the output of mapminmax is trained and simulated with the input. After
simulation we get the stable and desired output.
6. We apply the threshold condition to detect and mark the R–wave as shown in fig (2d).
4. Graphical Results of R-peaks

Input ECG signal (MO1122A VF) Input ECG signal (MO1122A VF)
2000 2000
0 0
-2000 -2000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Band Pass Filter output Band Pass Filter output
2000 2000
0 0
-2000 -2000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
" - Neural Network After Training - " " - Neural Network After Training - "
-0.8192 -0.9404
-0.8194 -0.9406
-0.8196 -0.9408
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
==== R - Peak ==== ==== R - Peak ====
2000 2000

0 0

-2000 -2000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

40


Input ECG signal (MO1122AVF)
Input ECG signal (MO1122A VF) 2000
2000
0
0
-2000
-2000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Band Pass Filter output
Band Pass Filter output 2000
2000
0
0
-2000
-2000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 " - Neural Network After Training - "
" - Neural Network After Training - "
-0.899
-0.816
-0.817
-0.9

-0.818 -0.901
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
==== R - Peak ==== ==== R - Peak ====
2000 2000

0 0

-2000 -2000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

Fig 2 a. Input ECG, b. Filtered output, c. Neural Network output, d. Detection of R peaks
5. TESTING RESULTS
In this paper Bayesian regularization neural network is used to learn the characteristics of QRS
complex to detect R peak on the standard CSE database. Bayesian regularization gives very good
results. Table shows the actual number of QRS complexes (R peaks), number of R peaks
detected, true positive (TP), false negative (FN), and false positive (FP) detection for entire CSE-
ECG library dataset-3. Each ECG record of the dataset is of 10 sec duration sampled at 500
samples per second, thus giving 5000 samples. The table also shows the detection rate (DR),
positive predictivity (+P) and sensitivity (Se): -

Table.1. QRS-detection results
Actual no. True False False Detection Positive
Sensitivity
of QRS Positive Negative Positive Rate Predictivity
Se
complexes TP FN FP DR +P

17760 17477 283 32 98.5 % 98.6 % 98.41%

41


6. CONCLUSION
The algorithm developed in MATLAB is implemented, with ANN based on Bayesian
regularization function, for detection of QRS-complexes. The algorithm is rigorously
tested on entire CSE-ECG dataset-3, which includes an exhaustive range of morphologies
and vast variety of cases. The resultant values are shown in Table 1 with significantly low
values of FP and FN and excellent values of DR, +P and Se. The suitability of the
algorithm for this purpose is conspicuously evident.
7. REFERENCES
1. V.S. Chouhan, S.S. Mehta, “Total Removal of Baseline Drift from ECG Signal”, IEEE
Proceedings of International Conference on Computing: Theory and Applications,
ICCTA–2007, ISI, Kolkata, India, 0-7695-2770-1/07, pp. 512-515, March 5-7, 2007.
2. Manpreet kaur, Birmohan Singh, “Comparisons of different approaches for
removal of baseline wonder from ECG signal”, 2nd international journal of
computer aplication 2011.
3. F. Dan Foresee and martin t. Hagan, gauss-Newton approximation to bayesian
learning, School of Electrical and Computer Engineering Oklahoma State University
Stillwater.
4. Leong Chio ln Wan Feng, Vai Mang, Mak Peng Un, “QRS Complex detector using
artificial neural network”, university of Macau, China.
5. Yu Hen Hu, J. Thompkins, Joes L. Urrusti, valtino X. afonso, “Applications of
artificial neural network for ECG signal detection and classification”, journal of
Electro cardiology voll. 26 supplement.
6. S. Issac Niwas, R. Shantha Selva kumara, Dr. V. sadasivam, “Artificial neural
network based automatic cardiac abnormalities classification”, Proceedings of the sixth
international conference on computational intelligence and multimedia
application, ICCIMA05, 2005, IEEE.
7. Qiuzhen Xue, Yu Hen Hu, J Tompkins, “Neural network based adaptive matched
filtering for QRS detection”, IEEE transaction on biomedical engineering, vol.39,
no.04, April 1992.
8. Matlab the language of technical computing 7.7.0 (R2008b), September 17 2008.
9. Kuryati kipli, Mohd Saufee Muhammad, Masniah Wan Masr, “Performance of
Levenberg – Marquardt backpropagation for full reference hybrid image quality matrix”,
proceeding of the international multi conference of Engineering and computer scientists
2012, vol 1, IMECS 2012, Hang Kong.
10. Alireza behrad and karim faez, “New method for QRS – wave reorganization in
ECG using MART neural network”, Seventh Australian and newzeland intelligent
information systems conference, 18-21 Nov. 2001, Perth, west. Australia.

42

Extraction of qrs complexes using automated bayesian regularization neural network

More Related Content

Similar to Extraction of qrs complexes using automated bayesian regularization neural network (20)

More from iaemedu (20)

Extraction of qrs complexes using automated bayesian regularization neural network