Constructing Classification Models Based on Mahalanobis-Taguchi System and Case-Based Reasoning

• Main Authors: Po-Chia Su, 蘇柏嘉
• Other Authors: Mei-Ling Huang
• Format: Others
• Language: zh-TW
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/41547156893842183643

碩士 === 國立勤益科技大學 === 工業工程與管理系 === 99 === One of the six emerging industries by the Executive Yuan to promote is Healthcare industry. Through the efforts of government and civil society, it has flourished. And the progress towards the target was established. Women were prone to breast cancer. This paper uses the database of breast cancer to construct classification models, and uses the database of SPECTF (Single Proton Emission Computed Tomography) to verify the efficiency of medical diagnostic classification models. Using data mining technology, we hope employing computer technology in medical diagnosis to reduce human error rate. So patients and their families do not suffer from another injury. This paper employs Mahalanobis-Taguchi System (MTS) and Case-Based Reasoning (CBR) with Mahalanobis distance in classification of breast cancer diagnosis of benign and malignant tumors, and in diagnostic of normal and abnormal of the SPECTF dataset. Then we calculate the area under the ROC curve and the accuracy of classification. We use T-test and Orthogonal Arrays with SN Ratio to filter characteristic variables. The variables will be filtered to the aforementioned diagnostic method to classify two categories forecast, calculated the area under the ROC curve and the accuracy of classification. Finally we compare the area under the ROC curve and accuracy of classification. We hope we can use more efficient and less characteristic variables to classify and predict. Before the feature selection, this paper performed classification model on Breast Cancer dataset, and the best performance of area under ROC curve was 0.998 with accuracy of classification 97.2%. After feature selection, the best performance was 0.998 with 96.9% accuracy of classification. For SPECTF dataset, before feature selection, the best performance was 1.0 with 100% accuracy of classification. After feature selection, the best performance was 1.0 with 99.6% accuracy of classification.