2009 International Association of Computer 2009 IACSIT Science Spring andConference Information Technology - Spring Conference
The Diagnosis of Hepatitis diseases by Support Vector Machines and Artificial Neural Networks Modjtaba Rouhani, IEEE Member, Islamic Azad University, Gonabad branch Gonabad, Iran
[email protected]
Mehdi motavalli haghighi Islamic Azad University, Gonabad branch Gonabad, Iran
[email protected] person affected by hepatitis B, 4-person who carries hepatitis C (no symptoms), 5-person affected by hepatitis C and 6-non viral hepatitis. We prepared a data set of 250 suspicious cases, from patients visited in two major hospitals in Mashhad, Iran, each of which has been carefully checked up by specialists and the diagnose is made. 58 of those patients was diagnosed as non hepatitis, 32 cases for person who carriers hepatitis B, 53 cases for person affected by hepatitis B, 32 specimens for person who carriers hepatitis C, 32 cases for person affected by hepatitis C and 43 specimens for person who affected by non-viral hepatitis. This paper organized as follow: in the second section a brief introduction to each ANN and SVM used is made. The third section presents the main results and in the forth section a combination of the networks is made which results the best accuracy. Section five concludes the paper.
Abstract—in this paper, we use Support Vector Machine (SVM) and artificial neural networks to diagnosis Hepatitis diseases. Furthermore, we use those networks to identify the type and the phase of disease. Considering the most important hepatitis cases leads us to six classes: hepatitis B (two phases), hepatitis C (two phases), non-viral hepatitis and no-hepatitis. For this purpose, we design various networks including RBF, GRNN, PNN, LVQ and SVM. The performance of each of them has studied and the best method is selected for each of classification tasks. The overall accuracy of diagnosis system is near 97%. Keywords- hepatitis disease, PNN, LVQ, SVM, RBF, GRNN)
I.
INTRODUCTION
Hepatitis disease is a fatal and deadly disease and is thought to be the fifth deadly disease worldwide. Hepatitis disease is the inflammation and damage to hepatocytes in the liver and can be caused by infections with viruses, bacteria, fungi, exposure to toxins, alcohol consumption and autoimmunity. Clinical symptoms of hepatitis are nausea, fever, general weakness, and jaundice. Five viruses have been identified and named hepatitis A through E [1,2]. In [1] and [2], the authors use different ANN structures to recognize the hepatitis disease. Reference [10] analyses a large database with hepatitis C virus infected patients. There are made a lot of statistical analyses on the records of this database in order to determine the evolution of biological parameters during the treatment. The results of the statistical analyses and the expert system predictions indicate the use of such a system to facilitate the physician work. In [8], authors use a feature selection (FS) and artificial immune recognition system (AIRS) with fuzzy resource allocation mechanism diagnosis of hepatitis disease with total accuracy of 92.59% on data set of UCI. In [9], ANN has been used to diagnosis chronic hepatitis disease. They report total accuracy of 93%. In fact, In addition to recognizing the hepatitis cases, it is important to identify the phase and the type of the hepatitis the person caused by, which is the main propose of this paper. There are fifteen parameters measured for each patient in order to diagnose hepatitis, as follow: 1-sex, 2-age, 3-ALK, 4-AST, SGOT, 5-ALT, SGPT, 6Bi, T, 7-Bi, D, 8-G.G.T, 9-HBSAg, 10-Alb, 11-LHD, 12-PT, 13-FBS, 14-CHO, and 15-HCVAb. These experimental data show six cases: 1-non hepatitis person, 2-person who carries hepatitis B (no symptoms), 3978-0-7695-3653-8/09 $25.00 © 2009 IEEE DOI 10.1109/IACSIT-SC.2009.25
II.
ANN CLASSIFIERS
The performance of five different networks in classification of hepatitis data sets are compared in this article. Those include radial basis functions (RBF), generalized regression neural network (GRNN), Probabilistic neural networks (PNN), LVQ network, and support vector machines (SVM). A brief description for them follows. A. RBF networks: Radial basis networks are powerful tools for classification. RBF has a two-layer structure. Neurons in the first layer compute the distance of the input vector to their centers, so the first layer determines which neuron is more similar (closer) to input vector. The second layer has linear activation function and its output is weighted sum of output of first layer. Learning of RBF is much easier and faster than BP networks as network parameters in first layer (centers) are set randomly and weights of the second layer are calculated to minimize the classification error. The activation function of hidden layer neurons is usually Gaussian function which is shown in equation (1). (1) Where ck is the center of radial basis function of neuron kth[4],[3]. B. GRNN networks: A generalized regression neural network (GRNN) is often used for function approximation. It has a radial basis 444 456
TABLE I.
layer and a special linear layer. The architecture for the GRNN is similar to the radial basis network, but has a slightly different second layer. The second layer also has as many neurons as input/target vectors. Suppose you have an input vector p close to pi, one of the input vectors among the input vector/target pairs used in designing layer 1 weights. This input p produces a layer 1 ai output close to 1. This leads to a layer 2 output close to ti, one of the targets used to form layer 2 weights [5]. C. PNN networks: Probabilistic neural networks (PNN) are a kind of radial basis network suitable for classification problems. The structure of PNN is like RBF networks. PNN suppose a probability distribution function for data (as Gaussian distribution) and for any input vector, the first layer estimate the conditional probability that the input vector belongs to each classes. The competitive transfer function in the second layer chooses the classes with the maximum probability as neuron with output one and all other neurons will have output zero [6], [7].
Disease
RBF
GRNN
PNN
No Hepatitis Carrier of HB Diseased by HB Carrier of HC Diseased by HC Non viral hepatitis
96.4 97.6
96.4 97.2
96.4 97.6
97.2
97.2
97.6
98.4
97.6
97.6
98.8
98.0
98.4
96.0
95.6
96.4
For RBF, GRNN and PNN networks, the spread (or radius) of radial basis functions plays an important rule in successful application of those networks. Increasing spread, generally, leads to smoother surface and better generalization property, but with more hidden layer neurons. We choose a maximum of 15 neurons at hidden layer as more neurons may lead to over-training of networks. The results for the best spread values are shown in table I. In each case 50% of data are used for training and all data used for testing.
D. LVQ networks: A Learning Vector Quantization Network (LVQ network) has a first competitive layer and a second linear layer. The competitive layer learns to classify input vectors in much the same way as the competitive layers of SelfOrganizing Nets. The linear layer transforms the competitive layer's classes into target classifications defined by the user. The classes learned by the competitive layer are referred to as subclasses and the classes of the linear layer as target classes. Both the competitive and linear layers have one neuron per (sub or target) class. LVQ learning in the competitive layer is based on a set of input/target pairs. The second layer needs no learning as the output classes (target classes) are known for each input pattern.
THE ACCURACY OF SVM FOR DIAGNOSIS OF HEPATITIS DISEASES BY VARIOUS KERNEL FUNCTIONS
TABLE II. disease
No Carrier Diseased Carrier Diseased Non viral Hepatit of HB by HB of HC by HC hepatitis is
Kernel RBF Linear Quadra tic
86.8 89.6 88.4
88.8 97.6 95.2
86.8 94.4 88.4
88.0 93.6 93.2
91.2 90.8 93.6
89.6 70.3 88.8
Polynomial
91.2
97.6
94.4
93.6
94.4
90.4
3
1
1
1
6
3
of Degree
For SVM, selection of proper Kernel function is important. So, we compare the performance of SVM classifiers for common kernel functions: RBF kernel, linear kernel, Quadratic kernel and polynomial kernel. In the case of polynomial kernel, polynomials with different degrees have been used. Table II summarizes the results.
E. SVM: Support vector machines (SVM) are basically linear classifiers or Linear Learning Machines (LLM). In SVM, a separator hyperplane between two classes is chosen to minimize the structural risk of misclassifying by maximizing the functional gap between two classes, the training data on the marginal sides of this optimal hyperplane called support vectors [4]. The learning process is the determination of those support vectors. For non linearly-separable data, SVM maps the input vector from input space to some normally higher dimensional feature space by introducing Kernel functions. Precisely selection of the Kernel function is an important step is successful design of a SVM in specific classification task. III.
THE ACCURACY OF THREE NEURAL NETWORKS FOR DIAGNOSIS OF HEPATITIS DISEASES
TABLE III.
MAIN RESULTS
In this section we design neural networks to diagnosis hepatitis disease. As maintained before, six classes have to be identified among learning data. As learning multi class problems are harder than two-class problems, we choose to design six classifiers, each of which for one of the classes.
THE ACCURACY OF LVQ FOR DIAGNOSIS OF HEPATITIS DISEASES
Diseases
Accuracy of train
Accuracy of test
average
No Hepatitis Carrier of HB Diseased by HB Carrier of HC Diseased by HC
98.4 87.2 98.4 99.2 99.2
92.8 84.0 92.0 96.8 94.4
95.6 85.6 95.2 98.0 96.8
Non viral hepatitis
98.4
91.2
94.8
For LVQ network, we considered a maximum number of 15 neurons in hidden layer and maximum epochs set to 150. As learning process of LVQ is stochastic, we repeat the learning for each set of parameters several times and choose
457 445
the best result for each case. Table III shows the accuracy of LVQ networks for train and test data. TABLE IV. Diseases
GRNN
PNN
SVM
LVQ
96.4
96.4
96.4
91.2
95.6
of
97.6
97.2
97.6
97.6
85.6
Diseased by HB
97.2
97.2
97.6
94.4
95.2
Carrier HC
of
98.4
97.6
97.6
93.6
98.0
Diseased by HC
98.8
98.0
98.4
94.4
96.8
Non viral hepatitis
96.0
95.6
96.0
90.4
94.8
Carrier HB
[3]
THE ACCURACY OF FOUR ANN AND SVM FOR DIAGNOSIS OF HEPATITIS DISEASES
RBF-rb
No Hepatitis
[2]
IV.
[4]
[5]
D. F .Specht, “A General Regression Neural Network”, IEEE Trans. Neural Networks, pp.568- 576,1991. [6] D.F.Specht, “Probabilistic Neural Networks”, Neural Networks, pp. 109-118, 1990. [7] F.Gorunescu and M. Gorunescu and E.El-Darzi and M. Ene and S. Gorunescu “Statistical Comparison of a Probabilistic Neural Networks Approach in Hepatic Cancer Diagnosis ” IEEE International conference on computer as a tool, 21-24, pp.237-240, 2005. [8] K. Polat and S. Gunes, “Hepatitis disease diagnosis using a new hybrid system based on feature selection (FS) and artificial immune recognition system with fuzzy resource allocation”, digital signal processing, Volume 16 , Issue 6 , (November 2006) [9] S. Shoimi et al, “Diagnosis of chronic liver disease from liver scintiscans by artificial neural networks,” Nuclear Medicine, Vol 11, No. 2, pp 75-80, 1997. [10] Andreea Drăgulescu et al, “Prognosis and Diagnosis in Hepatitis C using Expert Systems and Statistical Analyses”, bmf.hu/conferences/saci2006/Albu.pdf
PROPOSED DIAGNOSIS SYSTEM
The results of last section are summarized by table IV. In this table, the best accuracy for each class is printed in bold. In most cases the RBF network has the best accuracy; except for the third class (Hepatitis B) that the PNN performs better. So, as the best accuracy concerned, we may choose the PNN classifier for the third class and RBF networks for other five classes. The overall accuracy of diagnosis system is over 96.4% as indicated in table V. V.
TABLE V.
THE ACCURACY OF BEST NETWORKS FOR DIAGNOSIS OF HEPATITIS DISEASES
Diseases
CONCLUSIONS
In this paper, we train five different neural and SVM structures for diagnosis of hepatitis disease. The networks determine, based on clinical examinations, the person is affected by hepatitis or not and if so which type of hepatitis he or she is affected by. The data set consists of 250 cases carefully selected and examined by specialists. We found out that the RBF network outperforms other network including GRNN, LVQ, PNN and SVM. We obtain an overall accuracy of over 96.4% for train and test data. REFERENCES [1]
R.A.Vural and L.Ozyilmaz and T.Yildirim “A Comparative Study on Computerized Diagnostic Performance of Hepatitis Disease Using ANNs” ICIC, pp.1177-1182, 2006. R.J.Schalkoff, Artificial Neural Networks, McGraw-Hill Inc, Singapore ,1997. S.Haykin, Neural Networks: A Comprehensive Foundation, McMaster University,pp.278-300,1994.
L.Ozyilmaz and T.Yildirim , “Artificial Neural Network for Diagnosis of Hepatitis Disease,”. Proceedings of the International Joint Conference on Neural Networks, pp 586- 589, 2003.
458 446
Accuracy
No hepatitis
96.4
Carrier of HB
97.6
Diseased by HB
97.6
Carrier of HC
98.4
Diseased by HC
98.8
Non viral hepatitis
97.0
