Classification of Major Depressive Disorder using Machine Learning on brain structure and functional connectivity

Abstract

Timely identification of the risk of having Major Depressive Disorder (MDD) using the advanced machine learning (ML) approach has been popularized to generate neural indicators of MDD. However, past ML studies have usually employed a comprehensive whole-brain approach, resulting in high computational cost that is hardly affordable in practical and clinical contexts. In this study, we took an alternative approach by first conducting a meta-analysis to identify brain regions of interest (ROIs) where structural and/or functional alterations had been consistently reported in MDD patients. We then utilized a nonlinear multilayer deep neural network method to evaluate the classification accuracies of MDD patients (N = 41, 29 females) and healthy individuals (N = 20, 12 females) based on the structural volume and resting-state functional connectivity (rsFC) of the ROIs. We found that the ML model based on brain volumes, with the inferior frontal cortex (IFC), amygdala, and the pons receiving the highest weights, achieved higher classification accuracy and sensitivity relative to the model based on rsFC. For ethical reasons, people with MDD remained taking antidepressants at the time of the study. Our findings highlight the critical role of the pons-cortico-limbic network in MDD, which is a significant insight for early identification and diagnosis of the illness for timely intervention.