Classifying Gene Expression Data of Cancer Using Multistage Ensemble of Neural Networks

Abstract

Cancer classification based on the DNA microarray data is a difficult problem because of huge numbers of genes (features) relative to the number of training samples, and out of this huge number of genes most of are irrelevant. Therefore, it is a difficult task to create accurate classifiers for gene expression analysis that handles only a few samples with overwhelming number of noisy genes. Many contributions are already done for cancer classification of DNA microarray using neural network ensemble, which include, weighted averaging, majority voting, simple averaging and ranking. However, each method has some limitations, dependent on the application areas. For the proposed ensemble classification of gene expression data of cancer, first we extract the most relevant genes then apply a new multistage ensembles combination scheme, where the classified copy of results of training samples from first stage neural networks are used as an input features for second stage neural network. In this way we have checked the recognition accuracy for two benchmark gene expression data sets and got the encouraging results.