An Ensemble Feature Selection Framework for the Early Non-invasive Prediction of Parkinson’s Disease from Imbalanced Microarray Data

Abstract

Early identification of Parkinson’s disease (PD) using a non-invasive method is essential to slow down the disease progression with appropriate therapy. This can be accomplished by analysing microarray gene expression data from blood samples. This study proposes a computational framework for predicting PD from blood-based microarray gene expression data. Pre-processing, data balancing and feature reduction, and prediction are the three stages of the proposed system. In the pre-processing stage, annotation, cross-platform normalisation, and integration were performed. Balanced subsets were created using k-means clustering on majority samples and random undersampling. The ANOVA filter extracted critical features from balanced subsets in the feature reduction stage, and various cost-sensitive classification models and an ensemble model were built in the prediction stage. The method could achieve an AUC of 82.6% using the cost-sensitive Logistic regression classifier and 83.2% using the ensemble model on independent test data. The experimental results indicate that the suggested framework could effectively diagnose PD at the early stages.