In this paper, we develop classification rules based on the consideration of measures of diagnostic accuracy. In particular, we are interested in finding gene expression profiles that can discriminate between two populations. A unique challenge is posed because of the large p, small n problem. Our solution is to combine the problems of variable selection and classification. We suggest an approach for classification using the LASSO approach (Tibshirani [11]). An advantage of this approach is that some of the effects of the variables in these models are estimated to be exactly zero. These will represent genes that have no discriminatory power between the two classes, while those with nonzero coefficients will represent genes that can separate classes of tumors successfully. Thus, a by-product of the approach is the generation of a gene list. We exploit an equivalence between LASSO and support vector machines (SVMs) in order to fit the proposed classifier. The structure of the paper is as follows. In “materials and methods,” we provide background on the data structures observed and the motivation based on biomarker combinations, which leads
