Evaluating Predictive Accuracy of Prognostic Model for Censored Time-to-Event Data Analysis in Clinical Trials

In clinical trial time-to-event analysis, prognostic model, which is a class of clinical prediction models, is useful tool to identify factors (biomarkers) that are associated with clinical trial health outcomes of interest. The risk of an individual developing a particular state of health or experiencing a future outcome can be estimated using this tool. This plays an increasingly important role in assisting health professionals in making decisions and improving the health outcomes of patients. However, the quality of the decision made by health professionals depends on the accuracy of the estimated future risk. Therefore, it is imperative to evaluate the predictive performance of a given prognostic model, such as the Cox proportional hazards model, before using its risk estimates as the basis for important clinical decisions. The statistical performance of a prognostic model can be evaluated in terms of its ability to discriminate subjects with and without events of interest. To this end, the receiver characteristics curve (ROC) and its associated summary measures play an important role. This chapter provides review of current estimation approaches introduced for the time-dependent ROC curve and its associated summary measures in the context of censored data. The Youden index-based optimal cutoff value selection method is also presented. These approaches are illustrated by publicly available datasets using the R software.