Latent Variable Dimensionality Reduction Using a Kullback-Leibler Criterion and Its Application to Predict Antidepressant Treatment Response

Abstract

In this paper, we propose a method for dimensionality reduction of high-dimensional input data in a binary classification problem. The method is based on selecting a few latent variables that maximize the Kullback-Leibler (KL) distance between the two class distributions, under the assumption that these distributions are multivariate Gaussian. Numerical performance is demonstrated by solving the challenging problem of classifying patients with major depressive disorder (MDD) into responders vs. non-responders to an anti-depressant treatment using pre-treatment resting electroencephalography (EEG) data. The extracted feature set measures consistent connectivity and includes the magnitude coherence features among all electrode pairs in a 3Hz to 30Hz bandwidth with 1Hz resolution. An overall 86% prediction performance indicates the effectiveness of the KLDR method in this application. This performance level was found to exceed that of other dimensionality reduction methods, namely the unsupervised principal component (PCA) and the supervised Fisher discriminant analysis (FDA) methods.