Cerebral infarction classification using multiple support vector machine with information gain feature selection

Bulletin of Electrical Engineering and Informatics
Vol. 9, No. 4, August 2020, pp. 1578~1584
ISSN: 2302-9285, DOI: 10.11591/eei.v9i4.1997  1578
Journal homepage: http://beei.org
Cerebral infarction classification using multiple support vector
machine with information gain feature selection
Zuherman Rustam1
, Arfiani2
, Jacub Pandelaki3
1,2
Department of Mathematics, University of Indonesia, Indonesia
3
Department of Radiology, Cipto Mangunkusumo Hospital, Indonesia
Article Info ABSTRACT
Article history:
Received Aug 10, 2019
Revised Oct 14, 2019
Accepted Dec 4, 2019
Stroke ranks the third leading cause of death in the world after heart disease
and cancer. It also occupies the first position as a disease that causes both
mild and severe disability. The most common type of stroke is cerebral
infarction, which increases every year in Indonesia. This disease does not
only occur in the elderly, but in young and productive people which makes
early detection very important. Although there are varied of medical methods
used to classify cerebral infarction, this study uses a multiple support vector
machine with information gain feature selection (MSVM-IG). MSVM-IG
is a modification among IG Feature Selection and SVM, where SVM
conducted doubly in the process of classification which utilizes the support
vector as a new dataset. The data obtained from Cipto Mangunkusumo
Hospital, Jakarta. Based on the results, the proposed method was able to
achieve an accuracy value of 81%, therefore, this method can be considered
to use for better classification result.
Keywords:
Cerebral infarction
Information gain
Support vector machine
This is an open access article under the CC BY-SA license.
Corresponding Author:
Zuherman Rustam,
Department of Mathematics,
University of Indonesia,
Depok 16424, Indonesia.
Email: rustam@ui.ac.id
1. INTRODUCTION
Stroke is a leading cause of mortality and disability throughout the world [1, 2]. This far, ischemic
stroke is the most common type, which accounts for 70-90% of all stroke cases [3, 4]. Deaths that occur due to
ischemic stroke are still of foremost concern [5]. This disease becomes an important global health problem,
so that an effective way is needed to reduce mortality from this ischemic stroke. One way to diagnose whether
a patient has cerebral infarction, an examination from the radiology agency is needed, and one diagnostic
method often used to conduct these examinations is the computed tomography scanning (CT Scan). This
method is used to obtain a picture of the patient's head area. When some firmly demarcated dark areas are
visualized surrounding the brain tissue during the test, then that area is the chronic phase. As a result, a body
function regulated by the area tends to be permanently disrupted when early treatment isn’t provided.
Early medication helps to prevent diseases. Therefore, one important method used to prevent
chronic cerebral infarction is early identification to enable the patient to obtain the right treatment and care
immediately. One method used for this classification is machine learning such as the multiple support vector
machines with information gain feature selection (MSVM-IG) as proposed in this study. The cerebral
infarction data was obtained from RSCM hospital with as many as 206 patients who had undergone
the examination. Each patient was informed of the feature used to determine the severity of cerebral
infarction, and its data in this study consists of 10 features.

Bulletin of Electr Eng & Inf ISSN: 2302-9285 
Cerebral infarction classification using multiple support vector machine… (Zuherman Rustam)
1579
The previous researches on the classification of cerebral infarction had been carried out using
the Support Vector Machine method [6, 7] with great results. Similarly, the information gain feature selection
method has been used to detect Brain [8] and Lung Cancer [9]. In addition, the support vector machine
method has been used for the classification of schizophrenia data [10], to construct process maps for additive
manufacturing [11], often used for pattern recognition one of them in [12], for prediction of protein structural
classes [13], hyperspectral imagery [14], traffic incident detection [15], for image retrieval and image
process [16], fault interpretation, a study based on 3D seismic mapping of the Zhaozhuang coal mine in
the Qinshui Basin, China [17], intrusion detection system [18], pattern recognition to AVO classification [19],
for estimation of reservoir porosity and water saturation based on seismic attributes [20], elastic impedance
based facies classification [21], and the application of svm for prediction of coal and gas outburst [22].
2. RESEARCH METHOD
This research proposes a Multiple Support Vector Machine with Information Gain Feature Selection
(MSVM-IG) for early cerebral infarction classification. MSVM-IG is a method that uses support vector
obtained from SVM as an input in feature selection. Therefore, the amount of data processed by the IG
feature selection is not the same as the initial. The term multiple is used because after the feature selection
process with IG, SVM is re-evaluated. Due to the decrease in the amount of input data, IG selection features
are able to rank features more accurately with SVM producing better accuracy.
2.1. Data
The numeric data used in this study obtained from the results of the CT scan of Cipto
Mangunkusumo Hospital, Central Jakarta, which consists of 10 features, and they include: Gender, Age
of patient, Cerebral infarction area, Air normal cavity, Minimum value of area, Maximum value of area, Sum
of acute point, Length of area, Average of area, and Standar deviasi of area. The data include 206
observations with are 103 data labeled positive infarc and 103 data negative infarc.
2.2. Information gain feature selection
Information gain (IG) is one technique of filter type selection which works by sorting features based
on each value. Measurements from IG itself are based on the basic concept of entropy by determining
the difference between the entropy of all training data and the weighted sum of its subset of partition values
on a feature [23]. IG is also one of the easiest and fastest methods of sorting features. For example,
there is a training data set with -features and -classes, with is an attribute
consisting of different -classes. The value for the entropy of all training data is calculated based on
different -classes, therefore:
∑ (1)
with is the probability (relative frequency) of the class in the training data, with different
values used to calculate the weighed total sum of the entropy subset or partitioned values. Each value
contains an entropy value based on the class label in feature such that call acts as a subset, where
. Therefore, the weighted sum of the entropy subset of partition values on a feature is formulated
as follows:
∑
| |
| |
(2)
As previously explained, IG is obtained by looking at the difference between the entropy of all
training data and the weighted sum of the entropy subset of the partition values on a feature. Therefore,
the difference from equations (1) and (2) is the IG of a feature [23]:
∑
| |
| |
(3)
∑ ∑
| |
| |
(4)
2.3. Support vector machine
Support vector machine (SVM) which was introduced by Vapnik in the late 1990s, is a machine
learning algorithm used for classification and regression. SVM is related to structural risk minimization
(SRM) and was initially used for binary classification. It is currently used for multiclass classification

 ISSN: 2302-9285
Bulletin of Electr Eng & Inf, Vol. 9, No. 4, August 2020 : 1578 – 1584
1580
and takes the form of mapping input space into higher dimensional space to support nonlinear classification
problem where the maximum separation of the hyperplane is constructed. The hyperplane is a linear pattern
whose maximum margin provides separation between decision classes.
In the dataset { } , is the number of samples, is a feature vectors from sample- ,
with is the number of features (dimension), and is a class label. For the two-class classification problem
{ }, while in a multiclass { } with is the number of class. The main goal of SVM
is to determine the best hyperplane [24] and it illustrated in Figure 1:
(5)
Figure 1. SVM is trying to determine the best hyperplane to separate two classes
The problem of SVM optimization is summarized as follows:
‖ ‖ (6)
(7)
Objective function (6) to determine and subject to (7), with is the weights
and is bias. By completing the equation above, the formula and are obtained as follows:
∑ (8)
∑ ∑ (9)
and, the decision function as follows:
(10)
Below is the diagram flow of the proposed method, see Figure 2. First step is the data will be
processed by SVM so that the support vector is generated. Then, the IG feature selection will select
the selected features based on support vector. Lastly, SVM will be used again to get the measurement.
Figure 2. The flow diagram of MSVM-IG

1581
2.4. Kernel function
This research utilizes two kernel functions, namely radial basis function and polynomial kernel
functions with several parameters. The kernel function is given as follows:
( ) 〈 〉 (11)
with is a function that maps to the feature space . Every time 〈 〉 appears
in the classification algorithm, it is replaced with ( ) [25]. By using kernel functions, it is expected
that data is linearly separated linearly on higher dimensions. The formula of radial basis function (RBF)
and polynomial are shown below [26].
‒ RBF Kernel Function:
‖ ‖ (12)
‒ Polynomial kernel function:
[ ] (13)
2.5. Model performance evaluation
In this study, a performance evaluation model was conducted by measuring accuracy, precision,
sensitivity, specificity, and recall. Let TN, TP, FN, FP denote true negative, true positive, false negative,
and false positive, respectively. The following formulas below are used [27]:
(14)
(15)
(16)
(17)
(18)
3. RESULTS AND ANALYSIS
The support vector machine with information gain (IG-SVM) feature selection conventional
(without multiple SVM) is used to compare the proposed method. Two kernel functions were used, namely
radial basis function (RBF) and polynomial. Approximately 10 values of and are used in
the RBF and polynomial kernels respectively with the same parameter values; , -fold = 3,
and 5 main features.
3.1. Classification results with RBF kernel
For the RBF kernel we tried 10 different values that we determined randomly. The results are listed
in Tables 1 and 2.
Table 1. Results of cerebral infarction classification using MSVM-IG with RBF kernel
Accuracy (%) Precision (%) Sensitivity (%) Specificity (%) F1-score (%)
0.0001 81.863 81.553 83.333 81.372 81.951
0.001 81.127 79.812 83.333 78.922 81.535
0.05 80.882 79.439 83.333 78.431 81.34
0.1 80.76 79.254 83.333 78.186 81.243
1 80.686 79.143 83.333 78.039 81.184
10 80.637 79.07 83.333 77.941 81.146
50 80.602 79.017 83.333 77.871 81.118
100 80.576 78.978 83.333 77.819 81.097
1000 80.556 78.947 83.333 77.778 81.081
10000 80.539 78.923 82.353 77.745 81.068

 ISSN: 2302-9285
1582
Table 2. Results of cerebral infarction classification using IG-SVM with RBF kernel
0.0001 80.763 80.490 82.493 80.872 80.661
0.001 80.137 79.612 82.343 79.722 80.524
0.05 80.782 79.339 81.433 79.431 80.245
0.1 79.751 78.154 81.433 79.187 80.143
1 79.586 78.133 81.333 79.029 80.104
10 79.507 78.071 81.333 78.831 79.144
50 78.492 78.017 81.333 78.771 79.137
100 78.466 77.968 80.443 78.519 79.035
1000 78.446 77.847 80.443 77.769 79.033
10000 78.429 77.623 80.443 77.750 79.021
According to Tables 1 and 2, the smaller the value of the greater the classification results with
the highest accuracy, precision, sensitivity, specificity, and f1-score values obtained when the value of
for both methods. This is because the smaller the value of the faster the classification method
to learn data patterns and produce better results. The MSVM-IG produces better results than IG-SVM with
the highest accuracy, precision, sensitivity, specificity, and f1-score obtained by 81.863%, 81.553%,
83.333%, 81.372%, and 81.951% respectively. There was an approximate total difference of 1% between
the two methods, however, MSVM-IG is the method of choice for the classification of cerebral infarction.
3.2. Classification results with polynomial kernel
Also, for the polynomial kernel we tried 10 different values that we determined randomly.
The results are listed in Tables 3 and 4. The result shows that for experiment values from 1 to 10 produced
the same accuracy, precision, sensitivity, specificity, and F1-score.
Table 3. Results of cerebral infarction classification using MSVM-IG with polynomial kernel
1 80.392 78.704 83.333 77.451 80.952
2 80.392 78.704 83.333 77.451 80.952
3 80.392 78.704 83.333 77.451 80.952
4 80.392 78.704 83.333 77.451 80.952
5 80.392 78.704 83.333 77.451 80.952
6 80.392 78.704 83.333 77.451 80.952
7 80.392 78.704 83.333 77.451 80.952
8 80.392 78.704 83.333 77.451 80.952
9 80.392 78.704 83.333 77.451 80.952
10 80.392 78.704 83.333 77.451 80.952
Table 4. Results of cerebral infarction classification using IG-SVM with polynomial kernel
1 79.882 78.145 82.954 77.211 79.534
2 79.792 78.145 82.833 77.211 79.534
3 79.592 78.144 82.573 77.211 79.534
4 78.456 77.765 82.573 76.352 78.726
5 78.455 77.765 82.573 76.352 78.726
6 78.444 77.765 82.573 76.352 78.726
7 78.340 77.765 82.573 76.352 78.726
8 78.340 77.765 81.997 76.352 77.942
9 78.340 77.765 81.997 75.451 77.942
10 78.340 77.765 81.997 75.441 77.942
According to Tables 3 and 4, the smaller the value of the greater the classification results,
the higher the accuracy, precision, sensitivity, and specificity, with f1-score values are obtained when
for both methods. The smaller the value of the faster the classification method to quickly learn data
patterns and produce better results. The results of MSVM-IG is better than IG-SVM with the highest
accuracy, precision, sensitivity, specificity, and f1-score obtained by 80.392%, 78.704%, 83.333%, 77.451%,
and 80.952% respectively. The difference between the two methods is approximately 1%, however,
the MSVM-IG tends to be the method of choice for the classification of cerebral infarction.

1583
4. CONCLUSION
Stroke holds the second place of leading cause of death and the third the leading cause of disability.
Ischemic stroke is the most common type so we have to find the way to label stoke efficiently. This study
proposed a multiple support vector machine using the information gain feature selection (MSVM-IG) for
the classification of cerebral infarction. Additionally, the RBF and polynomial kernel functions are used
and based on the results as well as discussion, it was found that MSVM-IG tends to produce good accuracy,
sensitivity, specificity, and F1-score when using the RBF kernel ( ) with a high enough accuracy
of 81.863%. When compared with the conventional method, namely support vector machine with
information gain feature selection (IG-SVM), the difference was approximately 1% with MSVM-IG results
greater than IG-SVM. This indicated that MSVM-IG has a better result than the conventional method.
For future work, this modification could be improved again and the other kernel functions and techniques can
be used for comparison.
ACKNOWLEDGEMENTS
This research was financially supported by The Ministry of Research and Higher Education,
Republic of Indonesia (KEMENRISTEKDIKTI), with a PDUPT 2020 research grant scheme.
REFERENCES
[1] V. L. Feigin et al., “Global and regional burden of stroke during 1990-2010: Finfings from the global burden of
disease study 2010,” Lancet, vol. 838, pp. 245-255, 2014.
[2] D. Mukherjee and C. G. Patil, “Epidemiology and global burden of stroke,” World Neurosurgery, vol. 76; no. 6,
pp. 585-590, 2011.
[3] W. Wang et al., “Prevalence, incidence, and mortality of stroke in China: Results from a nationwide population-
based survey of 480 678 adults,” Circulation, vol. 135, no. 8, pp. 759-771, 2017.
[4] T. Xu et al., “Smoking, heart rate, and ischemic stroke: A population-based prospective cohort study among inner
Mongoloans in China,” Stroke, vol. 44, no. 9, pp. 2457-2461, 2013.
[5] K. B. Slot et al., “Impact of functional status at six months on long term survival in patients with ischaemic stroke:
Prospective cohort studies,” BMJ, vol. 336, no. 7640, pp. 376-379 2008.
[6] P. Bentley et al., “Prediction of stroke thrombolysis outcome using CT brain machine learning,” NeuroImage
Clinical, vol. 4, pp. 635-640, 2014.
[7] T. Kim et al, “Machine learning for detecting moyamoya disease in plain skull radiography using a convolutional
neural network,” EBioMedicine, vol. 40, pp. 636-642, 2018.
[8] C. M. Lai, W. C. Yeh, and C. Y. Chang, “Gene selection using information gain and improved simplified swarm
optimization,” Neurocomputing, vol. 218, pp. 331-338, Dec 2016.
[9] L. Gao, M. Ye, X. Lu, and D. Huang., “Hybrid method based on information gain and support vector machine for
gene selection in cancer classification,” Genomics Proteomics Bioinformatics, vol. 15, no. 6, pp. 389-395, 2017.
[10] T. Rampisela and Z. Rustam, “Classification of schizophrenia data using support vector machine,” Journal of
Physics: Conference Series, vol. 1108, no. 1, pp. 1-7, 2018.
[11] K. Aoyagi, H. Wang, H. Sudo, and A. Chiba, “Simple method to construct process maps for additive manufacturing
using a support vector machine,” Additive Manufacturing, vol. 27, pp. 353-362, 2019.
[12] C. J. C. Burges, “A tutorial on support vector machines for pattern recognition,” Data Mining and Knowledge
Discovery, vol. 2, pp. 121-167, 1998.
[13] Y-D. Cai, X-J. Liu, X-B. Xu, and K-C. Chou, “Prediction of protein structural classes by support vector machines,”
Computating & Chemistry, vol. 26, no. 3, pp. 293-296, 2002.
[14] Z. Chunhui, G. Bing, Z. Lejun, and W. Xiaoqing, “Classification of hyperspectral imagery based on spectral
gradient, SVM and spatial random forest,” Infrared Physics and Technology, vol. 95, pp. 61-69, 2018.
[15] J. Xiao, “SVM and KNN ensemble learning for traffic incident detection,” Physica A: Statistical Mechanics and its
Applications, vol. 517, pp. 29-35, 2019.
[16] L. Zhang, F. Lin, and B. Zhang, “Support vector machine learning for image retrieval,” Proceedings 2001
International Conference on Image Processing (Cat. No.01CH37205), vol. 2, pp. 721-724, 2001.
[17] G. Zhou, K. Ren, Z. Sun, S. Peng, and Y. Tang, “Fault interpretation using a support vectormachine: A study based
on 3D seismic mapping of the Zhaozhuang coal mine in the Qinshui Basin, China,” Journal of Applied Geophysics,
vol. 171, pp. 1-11, 2019.
[18] J. Maharani and Z. Rustam, “The application of multi-class support vector machines on intrusion detection system
with the feature selection using information gain,” 1st Annual International Conference on Mathematics, Science,
and Education (ICoMSE), pp. 1-4, 2017.
[19] J. Li and J. Castagna, “Support vector machine (SVM) pattern recognition to AVO classification,” Geophysical
Reseacrh Letter, vol. 31, no.2, pp. 475-496, 2004.
[20] S. R. Na’imi, S. R. Shadizadeh, M. A. Riahi, and M. Mirzakhanian, “Estimation of reservoir porosity and water
saturation based in seismic attributes using support vector regression approach,” Journal of Applied Geophysics,
vol. 107, pp. 93-101, 2014.

 ISSN: 2302-9285
1584
[21] Y. Nishitsuji and R. Exley, “Elastic impedance based facies classification using support vector machine and deep
learning,” Geophysics Prospect, vol. 67 no. 4, pp. 1040-1054, 2019.
[22] H. Zhang, L. U. Guangli, X. U. Lu, and Q. Yang, “Application of support vector machine in prediction of coal and
gas outburst,” Mining Safety & Environmental Protection, vol. 8, no. 2, pp. 2670-2681, 2013.
[23] A. Clare and R. D. King, “Knowledge discovery in multi-label phenotype data,” Proceedings 5th
European
Conference PKDD, vol. 2168, pp. 42-53, 2001.
[24] N. Cristianini and J. S. Taylor, “An introduction to support vector machines and other kernel-based learning
methods,” Cambridge University Press, 2000.
[25] I. H. Witten, E. Frank, and M. A. Hall, “Data mining: Practical machine learning tools and techniques-third edition,”
Elsevier Inc., Burlington, 2011.
[26] L. Liu, B. Shen, and X. Wang, “Research on kernel function of support vector machine,” In: Huang YM., Chao
HC., Deng DJ., Park J. (eds) Advanced Technologies, Embedded and Multimedia for Human-centric
Computing. Lecture Notes in Electrical Engineering, vol. 260, pp 827-834, 2014.
[27] M. Kohl, “Performance Measures in binary classification,” International Journal of Statistics in Medical Research
vol. 1, pp. 79-81, 2012.
BIOGRAPHIES OF AUTHORS
Zuherman Rustam is an Associate Professor and a lecturer of the intelligence computation at
the Department of Mathematics, University of Indonesia. He obtained his Master of Science in
1989 in informatics, Paris Diderot University, French, and completed his Ph.D. in 2006 from
computer science, University of Indonesia.
Assoc. Prof. Dr. Rustam is a member of IEEE who is actively researching machine learning,
pattern recognition, neural network, artificial intelligence.
Arfiani is a Bachelor of Science from the Department of Mathematics, University of Indonesia.
Her current research is machine learning and deep learning in various fields.
Jacub Pandelaki is an academic senate and lecturer at the Faculty of Medicine, University of
Indonesia.
He graduated from the medical doctor at the faculty of medicine, the University of Indonesia in
1989, and obtained his Ph.D. degree at the same university in 2010.
Dr. dr. Jacub Pandelaki, Sp. Rad works in the Department of Radiology, Cipto Mangunkusumo
Hospital, Indonesia as the doctor and the chair of the Interventional.

Cerebral infarction classification using multiple support vector machine with information gain feature selection

More Related Content

What's hot (19)

Similar to Cerebral infarction classification using multiple support vector machine with information gain feature selection (20)

More from journalBEEI (20)

Recently uploaded (20)

Cerebral infarction classification using multiple support vector machine with information gain feature selection