Prognostic Accuracy of Biomarkers for Cure Status Prediction Under Diverse Censoring Patterns.
文献摘要
Cure survival data occur commonly in time-to-event studies, where a non-negligible proportion of individuals may never experience the event of interest and are regarded as cured. Distinguishing cured from uncured individuals using prognostic factors is clinically essential for advancing precision medicine. However, evaluating the accuracy of such a distinction is challenging as the cure statuses are usually partially observed due to censoring. In this article, we present statistical methods to estimate prognostic accuracy measures, including sensitivity, specificity, and positive and negative predictive values, for assessing biomarker discrimination of cure status under interval censoring. These measures are estimated in finite samples based on two classes of semiparametric mixture cure models, with logistic or more flexible single-index models for the incidence component and a semiparametric transformation model for the latency component. Our work complements existing studies by accommodating more diverse censoring mechanisms and offering a user-friendly estimation tool for these accuracy measures. Simulation studies and an application to the ADNI study demonstrate the satisfactory performance and flexibility of our proposed methods, compared to the approaches in conventional survival analysis.