A new sensitivity-preferred strategy to build prediction rules for therapy response of cancer patients using gene expression data

Abstract

Pre-therapeutic prediction of therapy response and clinical outcome is an important field in medicine and poses new challenges to statisticians. Statistical prediction rules for two-class problems are generally designated to maximize the overall correct classification rate and to reflect an optimal balance of sensitivity and specificity. In some clinical situations, however, correct prediction of one particular class is more important than of the other class. We therefore propose a new strategy of building prediction rules, which are designed to increase the sensitivity, while losing some specificity. This strategy is applied to artificial simulation data and to gene expression data from primary colon cancers. Our concept is generally applicable to most common classification methods.