Statistical Issues in the Development of Clinically Useful Gen(omic) Tests for Prognosis and Therapy Selection
View Presentation View Presentation
*Lisa M. McShane, National Cancer Institute 

Keywords: Genomics, biomarker, prognostic, predictive, risk model, classifier

A better understanding of disease through identification of biological characteristics of the disease process and host that predict disease course and responsiveness to therapy will be essential for the discovery of new therapies and for optimization of clinical care for individual patients. Many attempts have been made to develop omics-based tests that could provide these clinically informative biological characterizations, but only a few omics-based tests have been moved successfully into clinical use. Several pitfalls in the development of such tests have been identified. Many studies aiming to develop omics-based tests are poorly designed because they fail to define an appropriate patient population, they do not include relevant control groups to enable distinction between prognostic and predictive tests, and they do not include a sufficient number of patients to draw reliable conclusions. Problems in model development and validation are also common. Models based on high-dimensional omics data are frequently overfit. Often the overfitting is not detected because researchers use resubstitution estimates to assess model performance, or they attempt internal validation methods such as cross-validation but do not perform these procedures correctly. Several published examples of omics-based tests will be discussed to illustrate the concepts and demonstrate the impact of poor design and analysis approaches.