Classification improvement of gene expression for bipolar disorder using weighted sparse logistic regression

Abstract

The computer-aided diagnosis system plays an important role in the classification of diseases and genes such as psychological or other diseases. Bipolar disorder (BD) is a commond psychological disease nowdys. Genes that describe this type of disease may include irrelative values to bipolar disorder disease. These values may adversely impact the classification performance. Logistic regression (LR) and recently sparse logistic regression (SLR) were used as a common technique to solve such binary classification problems. Gene selection has been applied to be a successful technique to get better classification output by excluding the irrelative values of genes. In this work we go further in improving the classification accuracy by restoring to incorporating the weight of these genes utilizing integrating the standardization of T-test with the sparse logistic regression, aiming to accomplish high classification accuracy. A bipolar dataset of gene expressions measured for 22283 genes using Affymetrix technology was used. Two performance indicators; classification accuracy, and geometric-mean of specificity and sensitivity are considered in evaluating the proposed method. Experimental results show an improvement over the two competitor methods; SLR-smoothly clipped absolute deviation (SCAD) and SLR-lasso in three indicators: classification accuracy, geo-means, and area under the curve. Therefore, our technique is beneficial to predict and classify BD psychopaths.