Machine Learning Framework for the Prediction of Alzheimer’s Disease Using Gene Expression Data Based on Efficient Gene Selection

Abstract

In recent years, much research has focused on using machine learning (ML) for disease prediction based on gene expression (GE) data. However, many diseases have received considerable attention, whereas some, including Alzheimer’s disease (AD), have not, perhaps due to data shortage. The present work is intended to fill this gap by introducing a symmetric framework to predict AD from GE data, with the aim to produce the most accurate prediction using the smallest number of genes. The framework works in four stages after it receives a training dataset: pre-processing, gene selection (GS), classification, and AD prediction. The symmetry of the model is manifested in all of its stages. In the pre-processing stage gene columns in the training dataset are pre-processed identically. In the GS stage, the same user-defined filter metrics are invoked on every gene individually, and so are the same user-defined wrapper metrics. In the classification stage, a number of user-defined ML models are applied identically using the minimal set of genes selected in the preceding stage. The core of the proposed framework is a meticulous GS algorithm which we have designed to nominate eight subsets of the original set of genes provided in the training dataset. Exploring the eight subsets, the algorithm selects the best one to describe AD, and also the best ML model to predict the disease using this subset. For credible results, the framework calculates performance metrics using repeated stratified k-fold cross validation. To evaluate the framework, we used an AD dataset of 1157 cases and 39,280 genes, obtained by combining a number of smaller public datasets. The cases were split in two partitions, 1000 for training/testing, using 10-fold CV repeated 30 times, and 157 for validation. From the testing/training phase, the framework identified only 1058 genes to be the most relevant and the support vector machine (SVM) model to be the most accurate with these genes. In the final validation, we used the 157 cases that were never seen by the SVM classifier. For credible performance evaluation, we evaluated the classifier via six metrics, for which we obtained impressive values. Specifically, we obtained 0.97, 0.97, 0.98, 0.945, 0.972, and 0.975 for the sensitivity (recall), specificity, precision, kappa index, AUC, and accuracy, respectively.