الفهرس 
Chapter 1 IntroductionOnly 14 pages are availabe for public view
 |  | from | 92 |  |  |
| | | from | 92 | | |
Abstract
The assessment of liver fibrosis stage in chronic hepatitis B virus (HBV) infected patients is very important. Liver biopsy is the gold standard and reference style to evaluate different fibrosis stages but with several drawbacks. So, using alternative non-invasive methods is necessary and better to avoid these drawbacks. In this thesis, seven clinical laboratory parameters of 235 chronic HBV Egyptian patients were collected from HBV clinic at National Liver Institute that belongs to Menoufia University in Egypt. The first aim of this study is to apply multiple machine-learning techniques based on clinical parameters to build efficient classification models that predict two liver related issues: the first is fibrosis stage prediction by differentiating between moderate fibrosis stage and advanced fibrosis stage, and the second is cirrhosis of liver prediction by differentiating between non-cirrhosis liver and cirrhosis liver in chronic HBV Egyptian patients. Also, two attribute selection methods were applied to reduce the dataset dimensionality and obtain the most relevant parameters. For fibrosis stage prediction, a classification model based on Logistic Regression using all seven parameters achieved accuracy of 93.61% and AUROC of 0.99. Besides, using only four parameters selected as the most relevant, accuracy of 95.74% and AUROC of 0.971 were achieved. For cirrhosis of liver prediction, a classification model based on Logistic Regression and cost sensitive with penalty of 2 using all seven parameters achieved accuracy of 91.49% and AUROC of 0.936. Besides, using only three parameters selected as the most relevant, accuracy of 85.11% and AUROC of 0.92 were achieved. The classification models outperformed noninvasive index-based method, FIB-4 that based on four clinical parameters, in both fibrosis stage and liver cirrhosis prediction in chronic HBV Egyptian patients.
Also, knowing the response of patient to treatment is a very important issue. Therefore, it is necessary to predict the treatment response to understand the efficiency of drugs. Recently, several machine learning techniques were experimented for treatment response prediction. The second aim of this study is to apply multiple machine learning techniques to predict the antiviral treatment response of Hepatitis B Egyptian patients, which are Decision Tree, Random Forest, k-Nearest Neighbor, Gradient Boosting. The input parameters include clinical laboratory parameters plus fibrosis stage and the antiviral drug type (Tenofovir, Entecavir and Lamivudine). Also, two over-sampling methods were applied to solve the data imbalance issues, which are Adaptive Synthetic Sampling (ADASYN) and Synthetic Minority Over-sampling Technique (SMOTE). Also, attribute selection method was used to reduce the dimensionality of features. The highest accuracies achieved by Decision Tree, Random Forest, k-Nearest Neighbor, Gradient Boosting were 85.7%, 85.7%, 92.9% and 85.7% respectively. The best classification model for antiviral treatment response on HBV Egyptian patients was kNN classifier model that achieved an accuracy of 92.9%, precision of 0.91 for response class and 1.0 for non-response class and recall of 1.0 for response class and 0.75 for non-response class.