k-SS: a sequential feature selection and prediction method in Microarray study

Abstract

Selection of the optimal feature subsets from high dimensional genomic data for predicting tumour conditions of tissue samples is the prime focus in many recent microarray studies. In this paper, we present a modified sequential features selection and classification procedure, the k-SS method, which ensures proper selection of relevant gene chips with correlated expression profiles to the existing cancer sub-groups in any binary-response microarray classification problem. The k-SS algorithm uses the misclassification error rate as its feature selection criteria. The procedure avoids being trapped in a local optimal feature selection step by allowing the level of significance level (of the k-SS features selection tests) at which the best combination of features is selected to be freely determined by cross-validation. This new method competes favourably with eight of the existing machine learning methods considered in this study. In many instances, the k-SS classification method performs excellently well in terms of better prediction accuracy relative to others. Nine published microarray data sets are used to demonstrate our results.