Machine learning techniques for binary classification of microarray data with correlation-based gene selection

Abstract: Microarray analysis has made it possible to predict clinical outcomes or diagnosing patients with the help of biological data such as biomarkers or gene expressions. The data from microarrays are however characterized by high dimensionality and sparsity so that traditional statistical methods are difficult to use and machine learning algorithms are therefore applied for classification and prediction. In this thesis, five different machine learning algorithms were applied on four different microarray datasets from cancer studies and evaluated in terms of cross-validation performance and classification accuracy. A correlation-based gene selection method was also applied in order to reduce the amount of genes with the aim of improving accuracy of the algorithms. The findings of the thesis imply that the algorithm s elastic net and nearest shrunken centroid perform best on datasets with no gene selection, while support vector machine and random forest perform well on the reduced datasets with gene selection. However, no machine learning algorithm can be said to consistently outperform any of the other and the nature of the dataset seem to be a more important influence on the performance of the algorithm. The correlation-based gene selection method did however improve prediction accuracy of all the models by removing irrelevant genes.