Journals Proceedings

Abstract

Research in the health related field involves the use of high dimensional data where microarray gene expression data are used for the classifier based detection of diseases and abnormalities. Many machine learning tools and methods have been presented and proposed to detect diseases from microarray gene expression datasets where the overwhelming majority of work is for the detection of cancer. However, less attention is made to the detection of autism using such data. We experiment with autism detection using five gene expression data sets from five chromosomes. This data includes a low number of samples and a high number of features that reach tens of thousands. The task is difficult due to the large dimension of the data set and the high overlap in the class distributions. Therefore, a feature selection stage is necessary before the classifier and combiner design stages. We experiment with four feature selection methods, five classifier types and two existing combiner methods. Additionally, we propose six variants of a weighted fusion method, where this proposed method influences the classifier decision on a test sample based on its previous performance on the validation set. This is achieved by multiplying its decision by a predetermined weight. Results show that it outperforms or is equal to existing methods. This is achieved when the feature set size is very low reaching 50 or less.