Design of ECG Processor for the prediction of

Ventricular Arrhythmia using RBF neural network

classifier
Mrs. Punitha .V
Mrs. Sherbin . A . K Assistant Professor
M Tech student Department of ECE
Department of ECE AWH Engineering College
AWH Engineering College Calicut, India
Calicut, India

Abstract— Ventricular arrhythmia is a severe heart efficiency. Although these methods have exhibited advantages
condition causing sudden cardiac arrest. Here the patients are in the detection of ventricular arrhythmia, there are so many
left a very few seconds when it occurs not enough to get an shortcomings while dealing with physical body. Most of them
immediate medical assistance. In this case, prediction of the
are very difficult to implement or possess high complexity in
ventricular arrhythmia is the best choice than detection. In this
thesis an ECG processor for the prediction of ventricular computation and some have low specificity in discriminating
arrhythmia using Radial basis function neural network classifier between normal and abnormal conditions, and all maintain late
is proposed and designed. To evaluate the performance database detection interval, which is usually not enough to take an
of the ECG recordings from the MIT PhysioNet is used. ECG action. Thus in this paper prediction is focussed than a mere
signals is filtered and real time and adaptive techniques were detection.
used to extract the required features. Seven features are .
extracted from the ECG delineation process. Sensitivity and
accuracy is maintained throughout the processing and is finally
classified using radial basis function neural network classifier. II. LITERATURE REVIEW
RBF network acts very advantageous and faster classification to
the corresponding ECG features extracted. Design is simulated A. Detection of Ventricular Arrhythmia
by using MATLAB and MODELSIM softwares and the given
ECG signals is classified to normal or abnormal signals. Shiu et.al [10] implemented an integrated
electrocardiogram signal processor (ESP) for the identification
Keywords—Electrocardiography (ECG), Ventricular of heart diseases using the 90-nm CMOS technology. The
arrhythmia, classifiers, radial basis function neural network. system employed an instrumentation amplifier and a low-pass
filter (LPF) to remove the baseline wander and the power line
I. INTRODUCTION interference form the ECG and employed a time-domain
Ventricular arrhythmia is an abnormal ECG rhythm and is morphological analysis for the feature extraction and
responsible for 75%–85% of sudden deaths in persons with classification based on the evaluation of the ST segment. The
heart problems unless treated within seconds [13] .Ventricular system was carried out in a field programmable gate array and
Arrhythmias are mostly caused by coronary heart disease, consumed a total of 40.3-μW power and achieved an accuracy
hypertension or cardiomyopathy and immediate death occurs if of 96.6%. The main disadvantage of the system is that it uses
it is not accurately diagnosed or treated [2]. VT is a fast rhythm fixed search window with predefined size to locate S and T
of more than three consecutive beats originating from the fiducial points, which is not suitable for real-time scenarios.
ventricles at a rate more than 100 beats/min [3]. Another ESP was proposed in [11].The system was fabricated
Electrocardiogram (ECG) is one of the most vital sign in on the 0.18-μm CMOS technology and executed different
diagnostics of Heart related diseases. It is a non-invasive functions for the three stages of pre-processing, feature
technique where the signal is measured on the surface of extraction, and classification. The algorithm behind these
human body, which is used in identification of the heart functions was based on the quad level vector [12] Moreover,
diseases. Any disorder of heart rate or rhythm, or change in the the functions were all pipelined to increase hardware utilization
morphological pattern, is an indication of cardiac arrhythmia, and reduce power consumption. Besides, the system employed
which could be detected by the analysis of the recorded ECG clock gating techniques to enable and disable each processing
waveform. unit individually according to the need and it applied voltage
scaling up to 0.7 V.
Recently, a wide range of methods were developed to
detect ventricular arrhythmia based on morphological [4], [6], B. Prediction over Detection
spectral [7], or mathematical [14] features extracted from the The ECG processor by Bayasi et.al [1] features to prediction
ECG signal. Machine learning techniques, such as neural over detection while compared to other arrhythmia analysing
networks [8] and support vector machine (SVM) [9] have also processors. Here the prediction is said to be expected 3 hours
been suggested as a useful tool to improve the detection before the onset. In the given design, the ECG data is
acquired, processed and analysed in a single chip. Then the it to 30% of the peak value (Thsb). As soon as the demarcation
extracted features are transmitted through the network and of the samples that exist in the QRS complex region is done,
data can be analysed by the physicians. the maximum value among them is set as a new R peak, and
the threshold is updated accordingly (50% of the last detected
R peak). The process repeats itself, and the threshold is
III. SYSTEM OVERVIEW adjusted according to the last detected R peak. The last step in
the QRS complex detection is to find the corresponding R
In the proposed system seven ECG features extracted peaks in filtered signal, which is done by subtracting the delay
is classified using radial basis function neural network encountered due to the filters.
classifier. Figure 1 shows the schematic diagram of the
proposed system which consists of three stages. The first stage
is ECG signal preprocessing and the second stage is feature
extraction and the third stage consist of classifier.

ECG Signal preprocessing

ECG preprocessing stage consist of three steps which

include ECG filtering, QRS detection and P and T wave
delineation. Fiducial points needed for the further stage is
delineated during this stage. They can be explained as follows:

A. ECG Filtering

Bandpass filtering of the raw ECG signal is the first step in Figure.1 Schematic of the proposed architecture
ECG preprocessing. Here the bandpass FIR filter is achieved
by cascading highpass and lowpass filter. In this section filter
coefficients are achieved using FDA tool and FIR filter is
simulated in Modelsim software.

B. QRS Delineation
After filtering, the PAT algorithm [5] is decomposed
into four steps. Differentiation of the filtered signal is used to
distinguish the QRS complex from other ECG waves by
finding high slopes. Then, a nonlinear transformation is
performed through point to point squaring of the filtered ECG
signal in which it is important to emphasize the higher
frequencies in the signal obtained from the previous step,
which are normally characteristic of QRS complex. After that,
integration is carried out by a moving time window to extract
additional features, such as the QRS width. Finally, adaptive
amplitude thresholds are applied to the averaged signal to
detect R peaks. Both the bandpass filtered signal and the
averaged signal are stored in separate SRAMs for further
analysis.
For real-time hardware implementation, we have
modified the peak detection technique, as shown in Figure 8.2.
Initially, the design reads the first 200 ECG samples from the
SRAM, which stores the samples of the averaged ECG signal.
The maximum value among these samples is set as an initial R
peak and used to compute the initial value of the threshold,
Th, which is set to 50% of the peak value. Then, the value of
every incoming sample is compared with Th, and only the
sample value, which is greater than the threshold, is used in
the next step. If none of the samples have a higher value than
Th, the algorithm redefined the value of the threshold and set
Figure 2 Flow chart of PAT peak detection technique
 basis function for the neural network of the classifier. The
RBF network is a three-layer (J1, J2, J3) feed forward neural
C. T and P Wave Delineation network, as shown in Figure 9.5. Each node in the hidden
layer uses a radial basis function (RBF), denoted φ(r), as its
The delineation of T and P waves is based on a novel nonlinear activation function. The hidden layer performs a
technique proposed in [15]. The method is based on adaptive nonlinear transform of the input, and the output layer is a
search windows along with adaptive thresholds to accurately linear combiner mapping the nonlinearity into a new space.
distinguish T and P peaks from noise peak. In each heartbeat, Usually, the same RBF is applied on all nodes; that is, the
the QRS complex is used as a reference for the detection of T RBF nodes have the nonlinearity φ1( ) = φ1( ), i=1,
and P waves in which two regions are demarcated with respect
…., J2, where _ci is the prototype or center of the ith node and
to R peaks.
φ( ) is an RBF. The biases of the output layer neurons can be
IV. FEATURE EXTRACTIN STAGE modeled by an additional neuron in the hidden layer, which
has a constant activation function φ0(r) = 1. The RBF network
achieves a global optimal solution to the adjustable weights in
The two main parameters that must be considered
the minimum mean square error (MSE) sense by using the
while developing a detection (or prediction) system are the
linear optimization method.
complexity and the accuracy of the feature extraction
The classification starts from the samples attained
technique in providing the best results. The result of such
from the preprocessing stage. The weight is initialized to the
analysis yielded in a unique set of ECG features, which were
samples and sets the initial learning rate. The winning neuron
found to be the most indicative characteristics of ventricular
is then selected from the input samples and the weight is
arrhythmia with a simple to realize system and high prediction
updated. If the variation of weight is equal to zero or
accuracy. The features include interval between P, Q, R, S, T
incredibly small, it is classified as center of radial basis
peak values. Figure.8.4 shows these intervals on ECG record.
function neural network.
It is worth mentioning that the features are extracted from two
consecutive heartbeats, unlike other methods that process each
heartbeat independently.

Figure. 3. ECG feature extraction from two consecutive

heartbeats.

V. CLASSIFICATION STAGE Figure 4. Flow chart of RBF classifier

The proposed classifier in this paper is Radial basis
function neural network classifier. RBF networks are used for VI. RESULTS AND DISCUSSION
function approximation and classification. They are trained to
approximate a nonlinear function, and the trained RBF
networks are then used to generalize. All applications of the ECG signal from MIT-BIH database is generated in
RBF network are based on its universal approximation MATLAB. The proposed design is stimulated using Modelsim
capability. Here the Gaussian function is taken as the radial
and Matlab R2014b. Figure 5 shows the ECG based processor function network as classifier the speed of the classifier is
simulated in Modelsim. much more increased which makes a fast computation process.
Increased speed in such a biomedical device is a productive
improvement in the working of the device.

VII. FUTURE SCOPE

Accuracy and power consumption is the significant
factors related to real time biomedical processors. Prediction
of human condition is becoming a must thing in today’s
human life style. When the life span of humans is coming
shorter the importance of life saviour devices is on demand.
We can induce a more accurate filtering and QRS
delineation technique by using DWT to the proposed design.
We can also employ some power reduction techniques to the
radial basis function network classifier. We can also include
other bio medical signals along with this and can develop a
Figure.10.4 ECG based processor simulated in Modelsim complete bio signal processor.

A. Synthesis Results
ACKNOWLEDGMENT
Low power ECG processor for the prediction of ventriculr This work has been done only because of constant
arrhythmia is designed by the proposed method and is guidance and efforts of my project guide Prof. Punitha. V. She
simulated in Modelsim. The classifier is synthesized in Xilinx not only showed me the way but molded my inexperience
14.2 Ise Design Suit and the coding is done in VHDL ideas and gave it practical outlook. I am thankful to her for
keen interest and supervision, which help in successful
language. The results obtained is shown in table no 11.1 and
completion of my work. I am also obliged to Prof. lekha
can be compared with the existing designs.
Pankaj, Head of Electronics and Communication Engineering
TABLE I. SYNTHESIS RESULTS OF RBFNN CLASSIFIER
Department for her constant encouragement.
REFERENCES

[1] Nourhan Bayasi and Temesghen Tekeste , Low-Power ECG Based

Temperature 29.2 0c Processor for Predicting Ventricular Arrhythmia”, IEEE Trans. On
VLSI systems 1063-8210, 2016.
Supply voltage 1.2 V [2] D. P. Zipes and H. J. J. Wellens, Sudden cardiac death, Circulation, vol.
98, no. 21, pp. 2334–2351, 1998.
[3] C. J. Garratt, Mechanisms and Management of Cardiac Arrhythmias.
Max. frequency 71.144 MHz London, U.K.: BMJ Books, 2001.
[4] 4. P. de Chazal, M. O’Dwyer, and R. B. Reilly, Automatic classification
of heartbeats using ECG morphology and heartbeat interval features,
Max. output time 4.591 ns IEEE Trans. Biomed. Eng., vol. 51, no. 7, pp. 1196–1206, Jul. 2004.
[5] J. Pan and W. J. Tompkins, A real-time QRS detection algorithm, IEEE
Trans. Biomed. Eng., vol. BME-32, no. 3, pp. 230–236, Mar. 1985.
No. of slices 4097 [6] A. Amann, R. Tratnig and K. Unterkofler, Detecting ventricular
fibrillation by time-delay methods, IEEE Trans. Biomed. Eng., vol. 54,
No. of slice flip flops 10 no. 1, pp. 174–177, Jan. 2007.
[7] O. Sayadi, M. B. Shamsollahi and G. D. Clifford, Robust detection of
premature ventricular contractions using a wave-based Bayesian
No. of 4 i/p LUTs 7331 framework, IEEE Trans. Biomed. Eng., vol. 57, no. 2, pp. 353–362, Feb.
2010.
VII CONCLUSION
[8] J. Pardey, Detection of ventricular fibrillation by sequential hypothes is
testing of binary sequences, in Proc. IEEE Comput. Cardiol, pp. 573–
576, Sep-Oct. 2007,
In this paper a digital ECG signal processor for the
[9] Q. Li, C. Rajagopalan and G. D. Clifford, Ventricular fibrillation and
prediction of ventricular arrhythmia with radial basis function tachycardia classification using a machine learning approach, vol. 61,
neural network classifier was proposed and simulated. Seven no. 3, pp. 1607–1613, Jun. 2013
features are extracted by the PAT algorithm and adaptive [10] B.-Y. Shiu, S.-W. Wang, Y.-S. Chu and T.-H. Tsai, Low-power low-noise
techniques. Features thus extracted is given to radial basis ECG acquisition system with dsp for heart disease identification,in
Proc. IEEE Biomed. Circuits Syst. Conf. (BioCAS), Oct./Nov. 2013, pp.
function neural network and the given ECG signals are 21–24.
classified to normal or abnormal. By using radial basis [11] H. Kim, R. F. Yazicioglu, T. Torfs, P. Merken, H.-J. Yoo and C. Van
Hoof, A low power ECG signal processor for ambulatory arrhythmia
 monitoring system, in Proc. IEEE Symp. VLSI Circuits (VLSIC), Jun. [14] X.-S. Zhang, Y.-S. Zhu, N. V. Thakor and Z.-Z. Wang, Detecting
2010, pp. 19–20. ventricular tachycardia and fibrillation by complexity measure, IEEE
[12] H. Kim, R. F. Yazicioglu, P. Merken, C. Van Hoof and H.-J. Yoo, ECG Trans. Biomed. Eng., vol. 46, no. 5, pp. 548–555, May 1999
signal compression and classification algorithm with quad level vector [15] N. Bayasi, T. Tekeste, H. Saleh, A. Khandoker, B. Mohammad and M.
for ECG holter system, IEEE Trans. Inf. Technol. Biomed., vol. 14, no. Ismail, Adaptive technique for P and T wave delineation in
1, pp. 93–100, Jan. 2010. electrocardiogram signals, in Proc. IEEE 36th Annu. Int. Conf. Eng.
[13] J. W. Schleifer and K. Srivathsan, Ventricular arrhythmias: State of the Med. Biol. Soc., pp. 90–93 , Aug. 2014
art, Cardiol. Clin., vol. 31, no. 4, pp. 595–605, 2013.