Predictive Accuracy of Cox Proportional Hazards Cure Models with Application in Cancer Studies
Han X, Zhang Y, and Shao Y. Predictive Accuracy of Cox Proportional Hazards Cure Models with Application in Cancer Studies. Cancer Res 2019; 79(13).
With recent development in cancer screening, diagnosis and treatment, many early-stage cancer patients will never experience recurrence, metastasis or death due to their primary cancer, whom can be considered as clinically cured. In the era of precision medicine, it is of interest to develop predictive models for cure status as well as for survival among uncured patients. The area under the ROC curve (AUC) is commonly used for assessing discriminative accuracy for dichotomous outcome (cured/uncured). Yet the conventional AUC cannot be directly used here due to unobserved cured status. In this study, we extend the conventional AUC to Cox proportional hazards (PH) cure models for evaluating prognostic utility in the presence of latent cured patients. We develop consistent and asymptotically normal estimates with explicit formulas for sensitivity, specificity and AUC. Furthermore, we reviewed the prognostic accuracy estimators for survival time among uncured patients for Cox PH cure models. Numerical studies show that our proposed estimators perform well for finite sample size. A melanoma data is used to illustrate the utility of our proposed prognostic metrics. We also have developed an R package called evacure to efficiently compute these estimates and their confidence intervals.