An Aggregated Cross-Validation Framework for Computational Discovery of Disease-Associative Genes

Numerous computational techniques have been applied to identify vital features of gene expression datasets that aim to increase efficiency of biomedical applications. Classification of samples is an important task to correctly recognize diseased people by identifying small but clinically meaningful genes. Conversely, it is a challenging issue for machine learning algorithms. In this paper, we apply a two-stage feature selection approach by using ensemble filter methods and Pareto Optimality. Although filter methods provide ranked lists of all features, they do not give any information about required (optimum) subset sizes of the features, namely, genes in this study. In order to address this issue, PO is incorporated with filter methods. The main aim of this study is therefore to develop a robust framework with PO, multiple feature selection methods and cross-validated subsets of the samples, which is also applicable to not only similar datasets but also different feature selection methods. The robustness of the framework has been successfully demonstrated over three well-known microarray gene expression data sets. The framework has been shown to yield equal or higher predictive accuracy with comparatively smaller feature sizes. Furthermore, the cross-validation and data variation approaches are considered in the framework. Consequently, the framework reduces the over-fitting problem and is observed to have made the gene selection more stable on different conditions.