Neural Networks and Classi cation & Regression Trees Are Able to Distinguish Female with Major Depression from Healhy Controls Using Neuroimaging Data

Abstract

Recently, magnetic resonance imaging and proton magnetic resonance spectroscopy studies of major depression identified structural and neurochemical alterations in several brain regions, including the hippocampus and prefrontal cortex. However, many contradictory endings exist. Most previous studies used a few cases and features, and conventional statistics. Therefore, we decided to use computational intelligence tools -neural networks and decision trees. Our data mining approach was applied to a neuroimaging dataset from 23 depressed women and 25 female controls. The goal was to identify complex relationships between the spectroscopic and volumetric inputs and diagnostic output. We tried to overcome data-related problems by using neural networks combined with genetic algorithms, ensemble methods, resampling, and extensive data preprocessing. The approach seems very promising. Some neural networks and two classification & regression trees - one involving hippocampal subregional volume and the other related to limbic-cortical neuronal integrity - successfully classified the subjects with 100% accuracy.