Application of BP and RBF Neural Network in Classification Prognosis of Hepatitis B Virus Reactivation
Wu Guan-peng^{1}, Wang Shuai^{1}, Huang Wei^{2}, Liu Tong-hai^{2}, Yin Yong^{2}, Liu Yi-hui^{1}
^{1}School of Information, Qilu University of Technology, Jinan, China
^{2}Department of Radiation Oncology, Shandong Cancer Hospital, Shandong Academy of Medical Sciences, Jinan, China
Email address:
To cite this article:
Wu Guan-peng, Wang Shuai, Huang Wei, Liu Tong-hai, Yin Yong, Liu Yi-hui. Application of BP and RBF Neural Network in Classification Prognosis of Hepatitis B Virus Reactivation. Journal of Electrical and Electronic Engineering. Vol. 4, No. 2, 2016, pp. 35-39. doi: 10.11648/j.jeee.20160402.16
Received: March 23, 2016; Accepted: April 9, 2016; Published: April 13, 2016
Abstract: This study aims at finding the risk factors (the key feature subset) and building the classification prognosis model of hepatitis B virus (HBV) reactivation after precise radiotherapy (RT) in patients with primary liver carcinoma. We find out that the outer margin of RT, TNM of tumor stage and the HBV DNA levels are the risk factors (P<0.05) of HBV reactivation by feature extraction method of logistic regression analysis in this article. The feature extraction method reduced the dimension and improved the classification accuracy. Establish the classification prognosis model of BP and RBF neural network for original data set and the key feature subset. The experimental results show that BP and RBF neural network have good performance in classification of HBV reactivation.
Keywords: Primary Liver Carcinoma, HBV Reactivation, Feature Extraction, BP, RBF
1. Introduction
Primary liver carcinoma (PLC) is extremely prevalent malignant tumors in South China and mostly patients were infected with hepatitis B virus (HBV). In recent years, precise radiotherapy (RT) is a good method for treating primary liver carcinoma. Precise radiotherapy method can cause HBV [1] reactivation for PLC patients and the rate of HBV reactivation approximately reached to 25%. The mortality reached to 25% for patients after HBV reactivation. Finding the risk factors of HBV reactivation has important signification for PLC patients with HBV reactivation. In 2007, Kim [2] researched 32 PLC patients who underwent three-dimensional conformal radiotherapy (3D-CRT). Their study showed that the HBV reactivation may be associated with the HBV DNA levels. Wu [3] [4] speculated that HBV reactivation has closely associated with the HBV DNA replication. In 2013, Huang [5] studied 69 PLC patients who underwent RT. They adopted logistic regression analysis affection on HBV reactivation of clinical features. Their results showed that the HBV DNA levels were the independent risk factors which lead to HBV reactivation. In 2014, they found dosimetric parameters also were high risk factors which lead to HBV reactivation [6]. We still need to research the risk factors of HBV reactivation. So far, there is no the classification prognosis model of HBV reactivation. This study aims at finding the risk factors and building the classification prognosis model of HBV reactivation after precise radiotherapy in PLC patients.
Feature extraction has been widely used in complex data analysis in biomedical field. Such as Feature extraction [7,8] applied in mass spectrometry data analysis. Artificial neural network [9] [10] [11] [12] is deemed to important intelligent algorithm in pattern recognition. It has been widely used in a variety of biomedical respects. For example, used in magnetic resonance spectroscopy in hepatocellular carcinoma based on neural network [13] [14] [15]. We find out that the outer margin of RT, TNM of tumor stage and the HBV DNA levels are the risk factors (P<0.05) of HBV reactivation by feature extraction method of logistic regression analysis in this article. The feature extraction method reduced the dimension and improved the classification accuracy. Establish the classification prognosis model of BP and RBF neural network for original data set and the key feature subset. The experimental results show that BP and RBF neural network have good performance in classification of HBV reactivation.
2. Data and Feature Extraction Methods
2.1. Data
The research data comes from Shandong cancer hospital treated 90 PLC patients after precise radiotherapy. 20 PLC patients were occurred HBV reactivation, the rest of patients didn’t occurred HBV reactivation. Each patient is considered as a research sample. Each sample includes 30 features.
2.2. Feature Extraction Methods
We adopt two independent samples -test, test and rank sum test three methods to extract significant factors which are affecting on HBV reactivation. Put significant factors into logistic regression analysis for finding the risk factors. The risk factors are considered as the key feature subset that has good performance for distinguishing classification.
2.2.1. Two Independent Samples t-test Method
Two independent samples -test [16] is suitable for normal distribution or approximating normal distribution of sample with enumeration data. There is no connection between two samples. Mathematical formula is:
(1)
, represent the mean of sample 1 and sample 2 respectively. represent the standard error of mean differences for and .
2.2.2. x^{2} Test Method
test [17] is suitable for qualitative data analysis of sample. It can use to check the connection of feature. Mathematical formula is as follows.
(2)
The is the actual value, is theoretical value.
2.2.3. Rank Sum Test
Rank sum test [18] is nonparametric hypothesis test. It doesn’t consider the distribution of sample. It is suitable for the data with abnormal distribution, ranked data and heterogeneity of variance.
2.2.4. Logistic Regression Analysis
The outer margin of RT, TNM of tumor stage and the HBV DNA levels are the risk factors of HBV reactivation by feature extraction method with logistic regression analysis [19]. Mathematical formula as follows:
(3)
is partial regression coefficient, is independent variable, is true or false value.
3. BP and RBF Neural Network
3.1. BP Neural Network
BP neural network [20] is able to classify and recognize the complex data. There are three layer structure of classic BP neural network, including input layer, hidden layer output layer. As shown in figure 1.
Input layer: M represent the neurons of input. Hidden layer I consists of the weight between input layer and hidden layer, threshold value , accumulator and function . Output layer J consists of the weight between hidden layer and output layer, threshold value , accumulator and function . In this article, M=30, I=5, J=2 for original data; M=3, I=2, J=2 for feature subset. Function adopted Sigmoid. Training times 1000, learning rate 0.05, training precision 0.001.
3.2. RBF Neural Network
RBF neural network [21] has a good performance in dealing with local optimum. There are three layers structure similar to BP, including input layer, hidden layer and output layer. As shown in figure 2.
RBF neural network finished nonlinear mapping from input layer to hidden layer by connectional weight and linear mapping from hidden layer to output layer by connectional weight . Hidden layer is given by Gauss function. Gauss function is defined as follows:
(4)
Input layeris the number of feature. Output layer represent the number of output result that HBV reactivation and non-reactivation. Spread of radial basis functions adopted default: Spread = 1.0.
4. Experimental Results of Feature Extraction
4.1. Experimental Results of Two Independent Samples t-test
The enumeration data adopted two independent samples -test method. The experimental results show that the outer margin of RT is correlated with HBV reactivation (P<0.05), and HBV reactivation had no correlation with other parameter. As shown in table 1.
Parameter | Mean | Standard error | p-value |
Age, years | 56.14 | 10.63 | 0.866 |
AFP, ng/ml | 630.98 | 1022.58 | 0.644 |
Radiotherapy dose, Gy | 57.93 | 7.19 | 0.917 |
Equivalent biometric, Gy | 69.969 | 8.198 | 0.891 |
Radiotherapy times, times | 28.48 | 5.484 | 0.871 |
GTV, cm3 | 179.59 | 228.76 | 0.570 |
PTV, cm3 | 392.07 | 318.93 | 0.915 |
outer margin of RT(mm) | 11.04 | 2.76 | 0.012 |
V5(%) | 51.645 | 17.776 | 0.751 |
V10(%) | 16.357 | 1.724 | 0.722 |
V15(%) | 37.216 | 14.638 | 0.977 |
V20(%) | 31.299 | 13.262 | 0.862 |
V25(%) | 25.643 | 11.448 | 0.782 |
V30(%) | 21.205 | 10.304 | 0.635 |
V35(%) | 17.015 | 8.513 | 0.786 |
V40(%) | 13.352 | 7.099 | 0.982 |
V45(%) | 10.169 | 6.308 | 0.393 |
Dmax, Gy | 6902.6 | 1160.4 | 0.941 |
Dmean, Gy | 1597.1 | 623.8 | 0.733 |
AFP, -fetoprotein; GTV, gross tumor volume; PTV, planning target volume; Dmax, maximum dose; Dmean , Mean dose. |
4.2.Experimental Results of x^{2} Test
The measurement data adopted test method. The experimental results show that the measurement data had no correlation with HBV reactivation. As shown in table 2.
4.3. Experimental Results of Rank Sum Test
The HBV DNA levels and code of outer margin of RT had correlation with HBV reactivation. The P value of TNM tumor stage is so close to 0.05 that is deemed correlation with HBV reactivation. As shown in table 3.
4.4. Experimental Results of Logistic Regression Analysis
The outer margin of RT, TNM of tumor stage and the HBV DNA levels are the risk factors of HBV reactivation. As shown in table 4.
Parameter | (B) | SE | Exp(B) | P-value | 95% C.I for Exp(B) | |
Ceiling | Inferior | |||||
outer margin of RT | 0.558 | 0.160 | 1.747 | 0.001 | 2.392 | 1.276 |
TNM tumor stage | 1.566 | 0.592 | 4.785 | 0.008 | 15.281 | 1.498 |
HBV DNA levels | 1.630 | 0.463 | 5.103 | 0.000 | 12.663 | 2.061 |
code of outer margin of RT | -0.876 | 1.195 | 0.416 | 0.463 | 4.328 | 0.040 |
B, coefficient; SE, standard error; Exp(B), coefficient logarithm. |
5. Experimental Evaluation Methods and Results of BP and RBF Neural Network
5.1. Experimental Methods of K-fold Cross Validation and Performance Evaluation Standard of Classifier
Experiment took the k-fold cross validation in order to get the accuracy results. The k-fold cross validation defined as follows:
.
The is the mean value of sum of by k-fold cross validation. The k is the number of sample. The k value selected 3, 5, 10.
Define five formulas for evaluating the performance of classifier, including accuracy, sensitivity, specificity, balance accuracy (BACC), positive predictive value (PPV). The accuracy is the most important evaluation standard for classifier.
5.2. Experimental Results of BP and RBF Neural Network
Build the classification prognosis model of BP and RBF neural network for original data set (table 5) and the key feature subset (table 6) of HBV reactivation. Experimental results show that BP and RBF neural network have a good classification performance of HBV reactivation. As shown in following table.
Neural network | k value | Accuracy | Sensitivity | Specificity | BACC | PPV |
BP | 3 | 0.7333 | 0.7857 | 0.5500 | 0.6678 | 0.8286 |
5 | 0.6889 | 0.7429 | 0.5000 | 0.6214 | 0.8087 | |
10 | 0.7000 | 0.7857 | 0.4000 | 0.5927 | 0.8209 | |
RBF | 3 | 0.7556 | 0.8143 | 0.5500 | 0.6821 | 0.8636 |
5 | 0.6778 | 0.6571 | 0.7500 | 0.7036 | 0.9020 | |
10 | 0.6555 | 0.6714 | 0.6000 | 0.6357 | 0.8545 |
Neural network | k value | Accuracy | Sensitivity | Specificity | BACC | PPV |
BP | 3 | 0.7889 | 0.8143 | 0.7000 | 0.7571 | 0.9048 |
5 | 0.7444 | 0.7714 | 0.6500 | 0.7107 | 0.8852 | |
10 | 0.7111 | 0.7571 | 0.5500 | 0.6535 | 0.8548 | |
RBF | 3 | 0.8000 | 0.8857 | 0.5000 | 0.6929 | 0.8611 |
5 | 0.7556 | 0.8286 | 0.5000 | 0.6643 | 0.8529 | |
10 | 0.7222 | 0.7714 | 0.5500 | 0.6607 | 0.8571 |
The accuracy was worst when the k value adopted 10 whatever the model is BP or RBF neural network. These results due to samples of training were very less. The accuracy of results increased with the samples of training increasing. The obvious result showed by compared to table 5 and table 6. The classification performance was optimal when the experiment adopted 3 fold cross validation. The key feature subset has better classification performance than original data whatever the model is BP or RBF neural network. The classification accuracy of BP increased from 73.33% to 78.89%. Sensitivity: 81.43%; Specificity: 70.00%; BACC: 75.71%; PPV: 90.48%. The classification accuracy of RBF increased from 75.56% to 80.00%. Because of RBF has better learning ability and approximate consistency than BP. BP is easily fall into local optimum, especially for small sample data. So, the classification accuracy of RBF is superior to BP.
The experimental results show that BP and RBF neural network have good performance in classification of HBV reactivation. The classification performance of the key feature subset is superior to original data set. The feature extraction method reduced the dimension and improved the classification accuracy.
6. Conclusion
The classification prognosis model of BP and RBF has higher application value for HBV reactivation after precise radiotherapy in PLC patients. We can take antiviral treatment and liver protective measures for had infected HBV in PLC patients in order to improve their quality of survival and prolong survival time. In the future, we will continue to study other intelligent algorithms for application in HBV reactivation, in order to improve the classification performance.
Acknowledgements
This work was supported partly by National Natural Science Foundation of China (81402538), National Natural Science Foundation of China (61375013), Natural Science Foundation of Shandong Province, China (ZR2013FM020).
References