A Stacked Deep Neural Network for Cancer Classification and Ensemble Gene Selection Approach

Cancer is a genetic, incurable, survival disease characterized by various changes at the gene level. Microarray technology helps in capturing gene expression. This data can be computationally analyzed to find abnormal genes. Microarray helps in the easy diagnosis and classification of subtypes by analyzing gene expression datasets. Microarray data involves high dimensionality data. Only very few genes are associated with the incidence of this type of cancer. Mutual information, a feature selection technique identifies the major genes associated with the disease. The colon dataset contains 2000 genes with 62 samples, leukemia contains 7000 genes with 72 samples while lymphoma contains 70 samples with 7129 genes. Mutual information selects the top-ranking genes for further classification which also reduces high dimensionality and improves accuracy. Deep neural networks like multilayer perceptron and feed-forward neural networks with various epochs, dense layers, hidden layers, and batch sizes are trained and logistic regression is used as a meta learner for stacking which achieves an accuracy of 75% for colon, 76.5 for leukemia and 74% for lymphoma.