Classification rules in brain tumor gene expression data

In this paper a gene expression study is presented aiming at distinguishing between brain tumors of different grades. Several prediction methods have been employed (diagonal linear discriminant analysis, k nearest neighbor, support vector machines), in combination with two different gene filtering methods, and compared to wrapper methods (nearest shrunken centroids and random forest). An alternative strategy for gene selection has also been considered. Error rates have been also compared with respect to the stability of the solutions obtained, since prediction was not the only objective of the study but also biological interpretability of the results was relevant. In fact, before wondering about the identity of the selected genes (in order to understand molecular pathways ot to find targets for drug development), it is important to evaluate how stable the list of the selected genes is.