Abstract
Machine learning techniques have the potential to revolutionize medical diagnosis. Single Nucleotide Polymorphisms (SNPs) are one of the most important sources of human genome variability; thus, they have been implicated in several human diseases. To separate the affected samples from the normal ones, various techniques have been applied on SNPs. Achieving high classification accuracy in such a high-dimensional space is crucial for successful diagnosis and treatment. In this work, we propose an accurate hybrid feature selection method for detecting the most informative SNPs and selecting an optimal SNP subset. The proposed method is based on the fusion of a filter and a wrapper method, i.e., the Conditional Mutual Information Maximization (CMIM) method and the support vector machinerecursive feature elimination, respectively. The performance of the proposed method was evaluated against four state-of-the-art feature selection methods, minimum redundancy maximum relevancy, fast correlation-based feature selection, CMIM, and ReliefF, using four classifiers, support vector machine, naive Bayes, linear discriminant analysis, and k nearest neighbors on five different SNP data sets obtained from the National Center for Biotechnology Information gene expression omnibus genomics data repository. The experimental results demonstrate the efficiency of the adopted feature selection approach outperforming all of the compared feature selection algorithms and achieving up to 96% classification accuracy for the used data set. In general, from these results we conclude that SNPs of the whole genome can be efficiently employed to distinguish affected individuals with complex diseases from the healthy ones.