蘇秀媛臺灣大學：農藝學研究所金佩奇Chin, Pei-ChiPei-ChiChin2007-11-282018-07-112007-11-282018-07-112004http://ntur.lib.ntu.edu.tw//handle/246246/59117中 文 摘 要 ROC分析(receiver operating characteristic analysis)和存活函數(survival Analysis)都可以用來分析醫學診斷的資料，其中ROC分析為回溯性試驗，可以根據事件是否真正發生(D=0, D=1)及測驗所得的結果(T=0, T=1)將受試者分成四類；ROC曲線是由偽陽性率(T=1 , D=0)及真陽性率(T=1, D=1)所構成，不受盛行率的影響。而存活分析可以將有設限的資料也納入分析，不會遺漏任何一筆資料所提供的訊息，使分析結果能夠更符合實際的情況。 這兩種分析方法各有所長，也有部分相似之處，如：ROC曲線是一個非遞減函數，而存活分析中也有幾個常用的函數符合這項特徵，且兩種分析方式都能依事件的發生及測驗的結果作分析，所以可將ROC分析轉換成存活函數的形式，並作有意義的解釋。 用存活分析中的函數，能將盛行率的訊息帶入ROC分析中，當遇到兩疾病的ROC曲線相同，但盛行率不同的情況時，我們可以藉由這些函數來判斷那一條ROC曲線所造成的T=1誤診率較小。 另外，當ROC曲線下面積相等但來自不同的雙常態分布時，在ROC曲線中接近線性的區域，我們可以利用ROC分析中的參數b來判斷那一條ROC曲線較穩定，當b與測驗變異數Var的比值越小時，ROC曲線越穩定。Abstract Both of the ROC analysis (Receiver Operating Characteristic Analysis) and the survival analysis can be used to analyze data from the diagnostic medicine. In the ROC analysis, data will be summarized according to the true condition status (D=0, D=1) and the test results (T=0, T=1) of the patients. An ROC curve is a plot of TPR(T=1, D=1) versus its FPR(T=1 , D=0) of the test. It is not affected by the prevalence of the condition. The survival analysis can analyze the censoring data. Therefore, no information from each sample will be lost. That lets us get more matched results as the true condition. Each of the two analyses has its merits. And there are several similar points between two analyses. For example, an ROC curve is a non-decreasing curve, and several functions in survival analysis are also non-decreasing functions. Both two analyses are accordance with the true condition status and the test results. Therefore, the transformed functions have their meanings. The functions from survival analysis were used to transform the ROC data. In this way, the prevalence can be shown. When two ROC curves which come from different diseases with different prevalence are the same, the transformed functions can be used to distinguish which ROC curve which has a smaller error for T=1. If the areas of the ROC curves are the same, but two curves are came form different bionormail distributions. The parameter b can be used to determine the ROC curves’ stability at linear areas of the ROC curves. When the ratio of b/Var (the variances of the tests) is smaller, the ROC curve is more stable.目 錄 第一章 前言……………………………………………………..……..1 1.1 ROC分析的原理與方法…………....…………….……..2 1.2 存活分析的原理與方法…..……….….…………………5 1.3 判斷ROC分析準確度的指標…….…………………….9 1.3.1 ROC曲線下的面積……..……..…………………9 1.3.2 固定FPR值時的靈敏度………………………..10 1.3.3 部分ROC曲線下的面積……………………….11 第二章 前人研究……………………………………………………..12 2.1檢定兩ROC曲線是否完全相同……………………….13 2.2檢定兩ROC曲線在特定的FPR時是否有顯著差異 .................................................................................17 2.3檢定兩ROC曲線在特定的FPR區間是否有顯著差異 …………………………………………………………..19 第三章 資料轉換……………………………………………………..22 3.1以FPR代替存活分析中的t做資料轉換……….…….23 3.2將所有受試者都納入母群體……………….…………..25 3.3改變條件機率…………………………….……………..28 第四章 面積相同的ROC曲線………………………………………31 4.1兩ROC曲線完全疊合……………………………..31 4.2兩ROC曲線面積相同但來自不同函數…………. 35 第五章 討論與未來展望……………………………………………..41 l 參考文獻…………………………………………………….……431650332 bytesapplication/pdfen-USROC分析存活分析survival analysisROC analysisthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/59117/1/ntu-93-R90621203-1.pdf