Feature Selection from Gene Expression Data Using SVMRFE and Feed-Forward Neural Network Classifier

Abstract

Correct classification of tumors is an important problem in clinical oncology. Availability of gene expression profiles from DNA microarray experiments has made it possible to use computational techniques to analyze these profiles to identify the molecular biomarkers in samples and hence classify them according to various tumor types. Since gene expression data is very high dimensional and the number of samples is very small compared to the number of dimensions, selection of relevant genes from available gene expression dataset, called as feature selection, is very important to classify samples correctly and efficiently. Various approaches to feature selection have been applied by researchers in literature. In the current work, we have used a method called support vector machine recursive feature elimination for gene ranking with feed-forward neural network as classifier for evaluation of relevance of selected features for the problem of classification of three cancer types using three publicly available gene expression datasets. We compared the classification accuracy and f-score obtained using the proposed method with conventional feature selection-classification pairs—mutual information-random forest classifier, mutual information-support vector machines, particle swarm optimization-support vector machines, principal component analysis-naïve Bayes. The comparison on the basis of confusion matrix parameters—classification accuracy and f-score shows that the proposed method has a better performance as compared to other methods.