Optimizing Specificity under Perfect Sensitivity for Computer-Aided Medical Data Analysis

• Main Authors: Cho-Yi Hsiao, 蕭卓毅
• Other Authors: 林守德
• Format: Others
• Language: en_US
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/64087082192842665924

碩士 === 國立臺灣大學 === 資訊工程學研究所 === 97 === In this thesis we purpose a novel evaluation criterion “specificity under perfect sensitivity” for medical data mining. This criterion aims at assessing the effectiveness of a classification model in confirming the perfection of the predicted negative data. We argue that this criterion could be useful for medical data mining when the penalty for false negative is extremely high so that no any false negative should be allowed. We further purpose two strategies to assist a classifier to obtain higher SUPS. The first strategy tries to loosen the criterion of positive by assigning negative instances closer to a positive one as suspicious, in order to enhance the confidence of predicted negative data. The second one tolerates the misclassified negative instances of positive patients to reduce the false negative rate of patients. The experiment results show that our methods can improve SUPS significantly comparing to the original classifiers.