Dynamic Placement Values: A Basis for Evaluating Prognostic Potential

• Activity Number: 596
• Type: Topic Contributed
• Date/Time: Wednesday, August 12, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Medical Devices and Diagnostics
• Author(s): Aasthaa Bansal* and Patrick Heagerty and Paramita Saha-Chaudhuri and Jason Liang
• Companies: University of Washington and University of Washington and McGill University and University of Washington
• Keywords: time-dependent accuracy ; receive operating characteristic curve ; prognosis ; sensitivity ; specificity

Abstract

A primary goal associated with prognostic biomarkers and models for event-time outcomes is prediction of incident or future cases. Extensions of standard binary classification concepts have been developed in this context, namely time-dependent sensitivity, specificity, and receiver operating characteristic (ROC) curves (Zheng and Heagerty, 2005). We present a direct, non-parametric method to estimate the time-dependent area under the ROC curve (AUC). This approach builds on the binary classification setting, where the ROC curve can be interpreted as the probability distribution of placement values, which are standardized measurements of a raw marker calculated for each case, representing the proportion of controls with marker values larger than a given case (Hanley and Haijian-Tilaki, 1997). With event-time data, it is natural to consider controls dynamically over time, and additionally, for marker values to vary over time. We define the dynamic placement value for an observed case at time t compared to controls at time t and present graphical summaries of time-dependent prognostic accuracy. Application to a cystic fibrosis data set is presented to illustrate the methodology.