Improving classification accuracy of cancer types using parallel hybrid feature selection on microarray gene expression data.

Background: Data mining techniques are used to mine unknown knowledge from huge data. Microarray gene expression (MGE) data plays a major role in predicting type of cancer. But as MGE data is huge in volume, applying traditional data mining approaches is time consuming. Hence parallel programming frameworks like Hadoop, Spark and Mahout are necessary to ease the task of computation.

Objective: Not all the gene expressions are necessary in prediction, it is very essential to select important genes for improving classification accuracy. So feature selection algorithms are parallelized and executed on Spark framework to eliminate unnecessary genes and identify only predictive genes in very less time without affecting prediction accuracy.

Methods: Parallelized hybrid feature selection (HFS) method is proposed to serve the purpose. This method includes parallelized correlation feature subset selection followed by rank-based feature selection methods. The selected subset of genes is evaluated using parallel classification algorithms. The accuracy values obtained are compared with existing rank-weight feature selection, parallelized recursive feature selection methods and also with the values obtained by executing parallelized HFS on DistributedWekaSpark.

Results: The classification accuracy obtained with the proposed parallelized HFS method is 97% and 79% for gastric cancer and childhood leukemia respectively. The proposed parallelized HFS method produced ~ 4% to ~ 15% improvement in classification accuracy when compared with previous methods.

Conclusion: The results reveal the fact that the proposed parallelized feature selection algorithm is scalable to growing medical data and predicts cancer sub-types in lesser time with higher accuracy.