傳統上接受者作業特徵曲線(ROC)是針對二元的分類結果進行預測，然而存活資料為結合二元設限狀態及連續存活時間的資料型態，因此若將敏感度與特異度的定義經過適當的修改後，即可將接受者作業特徵曲線應用於存活資料上。在過去文獻中，已有學者將其推廣到時間獨立共變數下配適Cox比例風險模型，並結合時間相依敏感度與特異度預測精準度。然而在部分的醫學研究中，常有資料不符合比例風險假設，因此我們建議以參數化加速失效模型取代Cox比例風險模型，結合時間相依敏感度與特異度，並以接受者作業特徵曲線下面積(AUC)及一致性指標Concordance判斷生物指標對疾病的區別能力，亦進一步擴展此方法到含有長期追蹤共變數的資料上。;Traditionally, the receiver operating characteristic curve (ROC) are used to predict the binary classification results. However, survival data are data types that combine the binary censored status and continuous survival time. Therefore, if the definition of sensitivity and specificity have been slightly modified, the ROC curve can be applied to the survival data. In the past literature, some scholars have extended it when conditioned at time-independent covariate to fit Cox proportional hazard model, and combined with time-dependent sensitivity and specificity to predict model accuracy. However, in some medical studies, there are often data that violate the proportional hazard assumption. Therefore, we recommend to replace the Cox proportional hazard model as the parametric accelerated failure time model with combining time-dependent sensitivity, specificity, and AUC. And finally use AUC and Concordance to evaluate the ability of biomarkers to discriminate diseases and further extended this method to longitudinal covariates.
