Depression Scale Prediction with Cross-Sample Entropy and Deep Learning

Abstract

A two-stage deep learning-based scheme is presented to predict the Hamilton Depression Scale (HAM-D) in this study. First, the cross-sample entropy (CSE) that allows assessing the degree of similarity of two data series are evaluated for the 90 brain regions of interest partitioned according to Automated Anatomical Labeling. The obtained CSE maps are then converted to 3D CSE volumes to serve as the inputs to the deep learning network models for the HAM-D scale level classification and prediction. The efficacy of the proposed scheme was illustrated by the resting-state functional magnetic resonance imaging data from 38 patients. From the results, the root mean square errors for the HAM-D scale prediction obtained during training, validation, and testing were 2.73, 2.66, and 2.18, which were less than those of a scheme having only a regression stage.