Training feedforward neural network using genetic algorithm to diagnose left ventricular hypertrophy

TELKOMNIKA Telecommunication, Computing, Electronics and Control
Vol. 18, No. 3, June 2020, pp. 1285~1291
ISSN: 1693-6930, accredited First Grade by Kemenristekdikti, Decree No: 21/E/KPT/2018
DOI: 10.12928/TELKOMNIKA.v18i3.15225  1285
Journal homepage: http://journal.uad.ac.id/index.php/TELKOMNIKA
Training feedforward neural network using genetic algorithm
to diagnose left ventricular hypertrophy
J. Revathi1
, J. Anitha2
, D. Jude Hemanth3
1
Department of Biomedical Engineering, KIT-Kalaignarkarunanidhi Institute of Technology, India
2,3
Department of Electronics and Communication Engineering, Karunya Institute of Technology and Sciences, India
Article Info ABSTRACT
Article history:
Received Jan 8, 2020
Revised Feb 6, 2020
Accepted Feb 24, 2020
In this research work, a new technique was proposed for the diagnosis of left
ventricular hypertrophy (LVH) from the ECG signal. The advanced imaging
techniques can be used to diagnose left ventricular hypertrophy, but it leads
to time-consuming and more expensive. This proposed technique overcomes
thesef issues and may serve as an efficient tool to diagnose the LVH disease.
The LVH causes changes in the patterns of ECG signal which includes
R wave, QRS and T wave. This proposed approach identifies the changes
in the pattern and extracts the temporal, spatial and statistical features
of the ECG signal using windowed filtering technique. These features
were applied to the conventional classifier and also to the neural network
classifier with the modified weights using a genetic algorithm. The weights
were modified by combining the crossover operators such as crossover
arithmetic and crossover two-point operator. The results were compared
with the various classifiers and the performance of the neural network
with the modified weights using a genetic algorithm is outperformed.
The accuracy of the weights modified feedforward neural network is 97.5%.
Keywords:
ECG signal
Feedforward neural network
Genetic algorithm
Left ventricular hypertrophy
This is an open access article under the CC BY-SA license.
Corresponding Author:
D. Jude Hemanth,
Department of Electronics and Communication Engineering,
Karunya Institute of Technology and Sciences,
Karunya Nagar, Coimbatore, Tamil Nadu, India.
Email: judehemanth@karunya.edu
1. INTRODUCTION
The patients who suffered from left ventricular hypertrophy may lead to the risk of mortality.
Sudden cardiac death may also causes because of left ventricular hypertrophy [1]. Left ventricular
hypertrophy is arising due to prolonged high blood pressure. The high blood pressure is also called
as cardiovascular syndrome [2]. It makes the chamber walls to be thick and it enlarges the chamber. It is
also stated that the increase in left ventricular mass leads to changes in the structure of myocardium [3].
The factors influences the LVH disease are age, increase in blood pressure, glucose intolerance and
obesity [4]. It alters ventricle repolarization and depolarization which leads to the changes in QRS and
T patterns in the ECG signal [5]. The various LVH-ECG criteria were reported which includes
Cornell Voltages, Romhillt Estees and Sokolov Lyon (SL) [6]. The ECG is a common method on the account
of wide availability [7]. Thus, the electrocardiography can be used as a tool to diagnose LVH in an early
stage. The LVH can be diagnosed using imaging modalities with advanced techniques. This method is high
priced and the process is delayed. The shortage of financial resources to acquire the echocardiographic
facility with skilled personnel is a challenging issue in the hospitals present in the rural area [8]. To overcome

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 18, No. 3, June 2020: 1285 - 1291
1286
these issues, a suitable algorithm is proposed to diagnose left ventricular hypertrophy using ECG signal.
The various Neural Network techniques are widely used to diagnose cardiovascular events [9].
In this research work, the classification of LVH disease is implemented using feedforward neural
network with modified weights. The weights were modified by using a crossover operator in a genetic
algorithm. Several researchers have utilized a genetic algorithm and a feedforward neural network for signal
processing applications. Zeinab Arabasadi et al. proposed a hybrid neural network and genetic algorithm for
the diagnosis of heart diseases. The neural network performance was increased by 10% [10]. Arpit Bhardwaj
et al proposed a network model for multiclass classification. This algorithm provides a 5% to 8% better
result [11]. Kinjal Jadav et al reported that the neural network training may lead to local minima and slow
convergence. To overcome this issue author applied a genetic algorithm for global optimization [12].
Zhen-Guo Che et al compared the neural network and genetic algorithm performance and found that the BPN
performs better [13].
Janati Idrissi et al. optimized a neural network based on metaheuristic algorithm by formulating
multi-objective mathematical function [14]. Ze Li et al. reported a Genetic Algorithm backpropagation
network which increases the convergence speed [15]. Hongqiang Li et al. reported the GA-BPNN network
for the signal classification it produced very good results [16]. Arpit Bhardwaj et al. developed a new
crossover and mutation operator to solve a classification problem [17]. V. S. R. Kumari et al reported that
the genetic algorithm optimizes the learning rate to classify arrhythmia disease [18]. Ali Bahadorinia et al.
developed a PSO and genetic algorithm-based neural network for the classification of arrhythmias [19].
Mansoura Sekkal et al developed a neuro-genetic algorithm to classify cardiac arrhythmia disease [20].
Manab Kumar Das et al used S transform to extract the features and optimized using a genetic algorithm to
obtain a better result [21]. Section 2 explains the dataset, Methodology, ECG pre-processing and ECG signal
analyzing and Classification techniques. Further, this article results are discussed and concluded in section 3
and section 4.
2. RESEARCH METHOD
This section is described with the dataset utilized in this study, methodology, signal pre-processing,
and feature extraction and classification methods.
2.1. ECG signal dataset
The PTB Diagnostic ECG database is utilized in this work to acquire 12 lead ECG signals of healthy
persons and the patients who are affected by left ventricular hypertrophy. This database also contains
the ECG signal with bundle branch block, myocarditis, and cardiomyopathy [22]. The sampling frequency
of the signal is 1000 Hz.
2.2. Methodology
The Figure 1 shows the flow of proposed work to diagnose left ventricular hypertrophy. It includes
feature extraction and classification. The fast Fourier transform was implemented to remove the low
frequencies and inverse fast Fourier transform was applied to restore the ECG signal. This was followed by
a windowed filter to determine the local maxima. The size of the filter was adjusted and filtering is repeated
to identify the peaks present in the signal. The temporal, spectral and statistical features were obtained.
These features were fed to the various classifiers to diagnose the LVH disease.
Figure 1. Diagnosis of Left Ventricular Hypertrophy from ECG signal

TELKOMNIKA Telecommun Comput El Control 
Training feedforward neural network using genetic algorithm… (J. Revathi)
1287
2.3. ECG signal pre-processing
The ECG signals are corrupted with the artifacts due to other monitoring devices, cables [23].
The elimination of these artifacts is essential to diagnose the disease more accurately. The FFT
is implemented to remove the low-frequency noise present in the signal.
2.4. Feature extraction
The essential features that are extracted to diagnose LVH are variance, kurtosis, form factor, average
Heart rate, mean RR interval, RR interval root mean square distance, standard deviation, heart rate
variability, and power spectral Entropy. In this study, the ECG leads such as Lead I, II and III, aVL, V1 to
V6 obtained from the PTB database. The low-frequency signals removed by using a fast Fourier transform.
This is followed by an inverse fast Fourier transform to restore the ECG signal as shown in Figure 2.
Figure 2. ECG signal after applying FFT
This is followed by a windowed filtering technique to find the local maxima. This helps to find
the peaks of the signal. The window sizes were adjusted filtering is applied to the signal to extract features
such as kurtosis, form factor, average Heart rate, mean RR interval, RR interval root mean square distance,
standard deviation, heart rate variability, and power spectral Entropy. This is further applied to the various
classifier to diagnose the disease.
2.5. Classification
In this research work, classifier includes K-nearest neighbor (KNN), support vector machine
(SVM), feedforward network and modified weight feedforward network was employed for diagnosis
and classification of LVH.
2.5.1. Support vector machine
This classifier segregates the feature by plotting the hyperplane that separates all the features of one
class to another. SVM employs mapping to convert nonlinear features to high dimensions [24]. The SVM
is analyzed using the Gaussian kernel function in this work [25]. The advantage is that it is flexible.
It supports complex models. This function is utilized by the equation is;
𝑘( 𝑥, 𝑦) =
𝑒−‖𝑥−𝑦‖2
22
(1)
2.5.2. K-Nearest neighbor classifier
The classifier K-nearest neighbor is an instant based learning method [26]. It is considered to be
effective, to decrease the misclassification of error, when there is a large number of components in
the training set [27]. During classification, a group of 'k' features from the training set are proposed that are
close to the features that are to be tested [28]. The classifier performance is based on the value K and distance
metrics. In this study, these values are implemented by cross-validation for the assessment of classification
accuracy [27]. For testing, a single data sample is employed for validation and remaining K-1 data samples
are used for training [27]. This is repeated a K time with each of the k data samples.

 ISSN: 1693-6930
1288
2.5.3. Feedforward neural network with genetic algorithm
This work proposed that the genetic algorithm is employed to train the feed-forward neural network.
The classification was employed using the Levenberg-Marquart network, a scaled conjugate gradient
backpropagation neural network. The total number of neurons presented in the hidden layer is 12, 15, and 20.
The trial and error method is implemented for the selection of neurons in the hidden layer. In the output
layer, one neuron was employed to diagnose the LVH based ECG signal. 70% of the dataset is implemented
for training and 15% of the dataset is used for validation and the last 15% of data is used for testing.
This performance of these classifiers was compared with the neural network classifier
with the modified weights. In this research work, the weights of the feedforward neural networks
were updated based on the genetic algorithm. In a genetic algorithm, a population is considered as a weight.
The chromosomes are prescribed by using an objective function. This is followed by a selection of two
chromosomes to perform the genetic operation. The offspring is generated which alters the initial population.
This is repeated until the desired output is achieved. The objective function is stated as a normalized
mean square error.
e(x)= (Oi-ti)2 (2)
𝐹(𝑥) =
𝑚𝑒𝑎𝑛(𝑒(𝑥))
𝑚𝑒𝑎𝑛(𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒(𝑂𝑖)
(3)
Where Oi represents the desired output, ti represents output obtained from net.
ALGORITHM
Step 1: The size of the weight matrix is defined and considered as a population.
Step 2: Population set is produced. Each component present in this set represents weights.
This weight depends on the number of layers in the network [29].
Step 3: Every matrix of weights is applied to the fitness function.
Step 4: The weight matrix with the best solution is chosen for the modification.
Step 5: Genetic operations mainly based on crossover and mutation. This process creates new
offspring.
Step 6: Crossover operator:
In this work crossover arithmetic and the crossover, two-point is implemented to generate
new offspring. The obtained results were taken and the new offspring is generated which is
followed by mutation.
The crossover arithmetic is a linear combination of two parents who are selected randomly.
The crossover two-point copied values based on the two points. In-between these two points
the genes get swapped. The weights are further computed with the formula (X+X1)/3.
Where X represents the weights obtained from the crossover arithmetic and X1 represents
the weights obtained from the crossover two-point.
Step 7: Stopping criteria is based on the number of generations and the convergence rate
the network reaches the objective.
Thus feedforward neural network was implemented by adjusting the weights of neural networks using a
genetic algorithm.
3. RESULTS AND ANALYSIS
In this research work, the fast Fourier transform was applied to the signals to eliminate lower
frequencies. This is followed by inverse fast Fourier transform to restore the ECG signal. Further,
the windowed filter was implemented to find local maxima. The peaks of the signals were identified.
The window size is adjusted and again filter applied to find the peaks. It improves the quality of the filtering.
The features extracted from this study includes variance, kurtosis, form factor, average Heart rate, mean
RR interval, RR interval root mean square distance, standard deviation, heart rate variability, and power
spectral entropy. This acquired features fed to the various classifiers to determine LVH based ECG signal
from the healthy individual ECG signal.
The 432 ECG signals acquired from the database which includes both LVH and healthy ECG
signals. The total features obtained from this study were (432x12=5184 features). These features were
applied to the classifiers such as SVM, KNN, Levenberg-Marquardt neural network (LM NN), scaled
conjugate gradient neural network (SCG NN), and weights modified feedforward neural networks using GA.
This experiment shows the potency of modified GA based feedforward neural network. In GA, the crossover
operator was implemented which includes arithmetic crossover and two-point crossover to create a new
offspring. Further new weights obtained by combining the weights from these crossover operators.

1289
To determine the constructiveness of extracted features, a receiver-operating characteristic analysis
was implemented. The area under the curve (AUC) denotes the goodness of fit [30]. The allocation of feature
vectors is similar in two classes for the AUC value is 0.5 [30]. The allocation of features of two classes does
not overlap with each other for the AUC value is 1 [30]. This value is more than 0.5 then the guessing model
is good. In this work, the fivefold cross-validation scheme is employed for the training and testing
of classifiers. The parameters such as sensitivity, specificity, and accuracy are evaluated to identify
the performance of the classifier. These parameters are judged by identifying true positives, true negatives,
false positives, and false negative values. The TP is true positive which indicates the number of events
matched and FN is false negative which indicates the number of events that are not recognized [31]. The FP
is false positive which represents the number of events that are not matched and true negative (TN)
represents the events that are exactly recognized as not defectives [31]. Table 1 shows the classifier
performance evaluated in terms of sensitivity, specificity, accuracy and area under the curve.
Table 1. Performance of the classifiers
Classifiers Number of neurons Sensitivity (%) Specificity (%) Accuracy (%) AUC
SVM - 86.1 92 90.7 0.90
KNN - 85.1 93.3 91.4 0.91
LM NN 12 89.8 98.8 96.5 0.96
LM NN 15 79.6 96.9 92.6 0.92
LM NN 20 69.4 91.4 85.9 0.86
SC NN 12 43.5 96.6 83.3 0.83
SC NN 15 49.1 96.9 85.0 0.85
SC NN 20 72.2 96.6 90.5 0.90
LM NN weights optimized using GA 12 85.2 98.5 95.1 0.95
SCG NN weights optimized using GA 12 37 98.1 82.9 0.83
SCG NN weights optimized using GA 15 62 98.1 83.1 0.83
SCG NN weights optimized using GA 20 60.2 97.8 83.3 0.83
The feedforward NNs were trained with 12, 15 and 20 hidden neurons. The results show that
the accuracy of LM NN with modified weights is increased than the LM NN without modified weights.
The accuracy of the Levenberg-Marquardt neural network (LM NN) shows better results than scaled
conjugate gradient neural network (SCG NN). The accuracy of SVM and KNN is low as compared to LM
NN feedforward neural network with modified weights.
4. CONCLUSION
In the current scenario, various imaging modality techniques used to diagnose LVH. But this leads
to a delay in treatment and high cost. Hence, the diagnosis of LVH using an ECG signal is a necessary part
of effective treatment. This study involves the analysis of temporal, spectral and statistical features
of ECG signal. The classification of disease with LVH compared with SVM, KNN and various feedforward
neural networks with and without modified weights. In this work, Genetic algorithm was implemented to
optimize the weights of feedforward NN. The weights modified based on the crossover operator. Thus,
it proves that the Levenberg-Marquardt NN weights optimized using GA perform better than another
classifier to diagnose LVH.
REFERENCES
[1] K. Porthan et al., "ECG left ventricular hypertrophy as a risk predictor of sudden cardiac death," International
Journal of Cardiology, vol. 276, pp. 125-129, 2019.
[2] X. Jiang et al., "Electrocardiographic criteria for the diagnosis of abnormal hypertensive cardiac phenotypes,"
The Journal of Clinical Hypertension, vol. 21, no. 3, pp. 372-378, 2019.
[3] A. Linhart and F. Cecchi, "Common presentation of rare diseases: Left ventricular hypertrophy and diastolic
dysfunction," International Journal of Cardiology, vol. 257, pp. 344-350, 2018.
[4] H. Lee et al., "Clinical characteristics associated with electrocardiographic left ventricular hypertrophy in clinical
normotensives without a history of hypertension: A cross-sectional study," Korean Journal of Family Medicine,
vol. 40, no. 2, pp. 106-115, 2019.
[5] G. Schillaci et al., "Improved electrocardiographic diagnosis of left ventricular hypertrophy," American Journal
of Cardiology, vol. 7, no. 7, pp. 714-719, 1994.
[6] W. B. Nurdin, H. Januari, S. Suryani, and S. Dewang,"Best LVH-ECG criteria for Indonesian hypertensives,"
Journal of Physics: Conference Series, vol. 1242, no. 1, pp. 1-4, 2019.

 ISSN: 1693-6930
1290
[7] V. Nomsawadi and R. Krittayaphong, "Diagnostic performance of electrocardiographic criteria for left ventricular
hypertrophy among various body mass index groups compared to diagnosis by cardiac magnetic resonance
imaging," Annals of Noninvasive Electrocardiology, vol. 24, no. 2, pp. 1-9, 2019.
[8] J. Pinto, P. George, and N. Hegde, "Study in Southern India among hypertensive patients using ECG to screen left
ventricular hypertrophy-can we do it in rural health centres?," Journal of Clinical and Diagnostic Research: JCDR,
vol. 8, no. 3, pp. 59-62, 2014.
[9] Z. N. Alimbayeva et al, "Neural network analysis of electrocardiogram signal," International Conference of Young
Specialists on Micro/Nanotechnologies and Electron Devices, pp. 606-611, 2019.
[10] Z. Arabasad et al., "Computer-aided decision making for heart disease detection using a hybrid neural
network-Genetic algorithm," Computer Methods and Programs in Biomedicine, vol. 141, pp. 19-26, 2017.
[11] A. Bhardwaj, A. Tiwari, H. Bhardwaj, and A. Bhardwaj, "A genetically optimized neural network model for
multi-class classification," Expert Systems with Applications, vol. 60, pp. 211-221, 2016.
[12] K. Jadav and M. Panchal, "Optimizing weights of artificial neural networks using genetic algorithm," International
Journal of Advanced Research in Computer Science and Electronics Engineering, vol. 1, no. 10, pp. 47-51, 2012.
[13] Z. G. Che, T. A. Chiang, and Z. H. Che, "Feed-forward neural networks training: a comparison between genetic
algorithm and back-propagation learning algorithm," International Journal of Innovative Computing, Information,
and Control, vol. 10, pp. 5839-5850, 2011.
[14] M. A. J. Idrissi, H. Ramchoun, Y. Ghanou, and M. Ettaouil, "Genetic algorithm for neural network architecture
optimization," 3rd International Conference on Logistics Operations Management, pp. 1-4, 2016.
[15] Z. Li, D. Y. Sun, J. Li, and Z. F. Li, "Social network change detection using a genetic algorithm based back
propagation neural network model," IEEE/ACM International Conference on Advances in Social Networks
Analysis and Mining, pp. 1386-1387, 2016.
[16] H. Li, D. Yuan, X. Ma, D. Cui, and L. Cao, "Genetic algorithm for the optimization of features and neural networks
in ECG signals classification," Scientific Reports, vol. 7, pp. 1-12, 2017.
[17] A. Bhardwaj and A. Tiwari, "Breast cancer diagnosis using genetically optimized neural network model," Expert
Systems with Applications, vol. 42, no. 10, pp. 4611-4620, 2015.
[18] V. S. R. Kumari, P. R. Kumar, and P. Rajesh, "Optimization of multi-layer perceptron neural network using genetic
algorithm for arrhythmia classification,” Communications, vol. 3, no. 5, pp. 150-157, 2015.
[19] A. Bahadorinia, A. Dolatabadi, and A. Hajipour, "A hybridized artificial neural network and optimization
algorithms for the diagnosis of cardiac arrhythmias," Advances in Computer Science: an International Journal,
vol. 3, no. 4, pp. 51-58, 2014.
[20] M. Sekkal and M. A. Chikh, “Neuro-genetic approach to classification of cardiac arrhythmias” Journal
of Mechanics in Medicine and Biology, vol. 12, no. 1, 2012.
[21] M. K. Das, D. K. Ghosh, and S. Ari, "Electrocardiogram (ECG) signal classification using s-transform, genetic
algorithm, and neural network." 1st International Conference on Condition Assessment Techniques in Electrical
Systems, pp. 353-357. 2013.
[22] L. N. Sharma, R. K. Tripathy, and S. Dandapat, "Multiscale energy and eigenspace approach to detection
and localization of myocardial infarction," IEEE Transactions on Biomedical Engineering, vol. 62, no. 7,
pp. 1827-1837,2015.
[23] K. S. Kumar, B. Yazdanpanah, and P. R. Kumar, "Removal of noise from electrocardiogram using digital FIR
and IIR filters with various methods," International Conference on Communications and Signal Processing,
pp. 0157-0162, 2015.
[24] M. Turhan, D. Şengur, S. Karabatak, Y. Guo, and . Smarandache, "Neutrosophic weighted support vector machines
for the determination of school administrators who attended an action learning course based on their conflict-
handling styles," Symmetry, vol. 10, no. 5, pp. 1-11, 2018.
[25] M. Hofmann, "Support vector machines-Kernels and the kernel trick," Notes, pp. 1-16, 2006.
[26] R. He, K. Wang, Q. Li, Y. Yuan, N. Zhao, Y. Liu, and H. Zhang, “A novel method for the detection of R-peaks in
ECG based on K-nearest neighbors and particle swarm optimization,” EURASIP Journal on Advances in Signal
Processing, vol. 2017, no. 82, pp. 1-14, 2017.
[27] I. Saini, D. Singh, and A. Khosla, "QRS detection using K-Nearest Neighbor algorithm (KNN) and evaluation on
standard ECG databases," Journal of advanced research., vol. 4, no. 4, pp. 331-344, 2013.
[28] M. Suvarna and Venkategowda N., “Performance measure and efficiency of chemical skin burnclassification using
KNN method,” Procedia Computer Science, pp. 48-54, 2015.
[29] N. Mohankumar, B. Bhuvan, N. Devi, and S. Arumugam, "A modified genetic algorithm for evolution of neural
network in designing an evolutionary neuro-hardwar," International Conference on Genetic and Evolutionary
Methods, pp. 108-111, 2008.
[30] M. H. Imam, C. Karmakar, H. F. Jelinek, M. Palaniswami, and A. H. Khandoker, "Detecting subclinical diabetic
cardiac autonomic neuropathy by analyzing ventricular repolarization dynamics," J. Biomedical and Health
Informatics, vol. 20, no. 1, pp. 64-72, 2016.
[31] P. C. Chang, J. J. Lin, J. C. Hsieh, and J. Weng, "Myocardial infarction classification with multi-lead
ECG using hidden Markov models and Gaussian mixture models," Applied Soft Computing, vol. 12, no. 10,
pp. 3165-3175, 2012.

1291
BIOGRAPHIES OF AUTHORS
Mrs. Revathi Jothiramalingam is an Assistant Professor in the Department of Biomedical
Engineering at KIT-Kalaignarkarunanidthi Institute of Technology. She has more twelve years
of professional experience in Biomedical field. Her research topics include Biomedical
Instrumentation, Biomedical Image and Signal Processing and Nanotechnology. She has
published more than 10 research articles in journals and conference proceedings.
Dr. Anitha Jude Hemanth is an Associate Professor in the Department of Electronics and
Communication Engineering at Karunya Institute of Technology and Sciences. She completed
her Ph. D at the Karunya Institute of Technology, Coimbatore. Her research interest includes
Machine learning, Classification, Neural Networks and Artificial Intelligence, Image Analysis,
Medical Image and signal Processing. She has presented papers at conferences both home and
abroad, published articles and papers in various journals, and contributed a chapter to the
books. She published more than 92 research papers in these areas.
Dr. D. Jude Hemanth received his B.E degree in ECE from Bharathiar University in 2002,
M.E degree in communication systems from Anna University in 2006 and Ph.D. from Karunya
University in 2013. His research areas include Computational Intelligence and Image
processing. He has authored more than 100 research papers in reputed SCIE indexed
International Journals and International Conferences with leading publishers such as Elsevier,
Springer, IEEE, etc. His Cumulative Impact Factor is more than 100. He has authored 1 book
with (VDM-Verlag, Germany) and 21 edited books with reputed publishers such as Elsevier,
Springer, IET and IOS Press.

Training feedforward neural network using genetic algorithm to diagnose left ventricular hypertrophy

More Related Content

What's hot (20)

Similar to Training feedforward neural network using genetic algorithm to diagnose left ventricular hypertrophy (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

Training feedforward neural network using genetic algorithm to diagnose left ventricular hypertrophy