| Summary: | 碩士 === 雲林科技大學 === 工業工程與管理研究所碩士班 === 96 === With the change of society, many people nowadays generally have overweight problems and some of them suffer from fatty liver symptoms. The symptoms often cause incorrect results from liver function checking, which leads to wrong prescription and delay in treatment.
Fatty liver is defined as that fat contents in liver cells are higher than 5 percent, and it can be generated from many causes - such as alcoholism, obesity, abnormal fat metabolism, rapid weight loss and malnutrition. In the past, fatty liver symptoms were mostly examined by performing a liver biopsy. With the prevalence of ultrasound technique, it gradually becomes a major tool for diagnosing fatty livers. Since ultrasound diagnostic methods require the combination of professional physicians and precise instruments, costs are high. Therefore, unnecessary examination may cause waste in medical resources. If we use objective information to filter out probable fatty liver patients from the results of physical examination, the method can become a supportive tool of medical decision-making in the prevention of fatty liver. Meanwhile, medical cost can be saved since only abdominal ultrasound examinations are performed on possible fatty liver patients
In this research, the data is collected from the 2007 physical examination of 1133 subjects from a medical center in the middle Taiwan. Those subjects’ situations match the conditions of normal fatty liver. The data are randomly separated into training data set (568 persons) and test data set (565 persons). Processed by data-mining methods, 28 characteristics are selected by Sequential Forward Selection (SFS) and 17 characteristics are clarified by Classification and Regression Tree（CART）. Then, those characteristics from SFS and CART are input to the General Regression Neural Network (GRNN) classifier to identify the performance by the Az value from the curve generated by Receiver Operating Characteristic (ROC) curve. Moreover, a GRNN classifier and a CART classification tree are constructed by classifiers, and some key characteristics of physical examination can be analyzed by the two forms to detect fatty liver symptoms.
The experimental results of the Az values of ROC curve show that the SFS classification method outperforms both the CART classification method and the method using all features. In terms of accurate rate, the GRNN classifier method is better than the one of CART classification tree method.
|
|---|
