Functional Near-Infrared Spectroscopy-based Diagnosis Support System for Distinguishing between Mild and Severe Depression Using Machine Learning Approaches

Abstract

Abstract Due to the fast-paced lifestyle, individuals may experience varying degrees of depression attributed to stress, academic pursuits, and other causative factors. Hemodynamics and functional connectivity (correlation between channels) of the prefrontal lobe have been identified as crucial factors in assessing the severity of depression. As a non-invasive technique for monitoring cerebral blood flow, functional near-infrared spectroscopy (fNIRS) shows promising potential as a tool for objective auxiliary diagnosis of depression.This study aimed to develop prediction models for distinguishing patients with severe depression from those with mild depression based on the dataset collected by fNIRS.We collected the fNIRS data from 140 subjects, and used a complete ensemble empirical mode decomposition with adaptive noise-wavelet threshold combined denoising method (CEEMDAN-WPT) to remove the jitter and artefact noise generated during the verbal fluency task (VFT). The temporal and correlation features of 18 channels in the prefrontal lobe of the subjects were extracted as predictors. We screened out the optimal temporal features (TF) or correlation features (CF) using the RFECV, and investigated their role in distinguishing severe and mild depression, respectively. The fusion of TF and CF, as the input of the prediction model, yielded higher classification accuracy than using TF or CF alone as the prediction factor. Among the prediction models, the SVM-based predictive model performed well in nested cross-validation, with an accuracy rate of 92.8%.The proposed model effectively distinguishes mild depression from severe depression, provides an objective diagnostic method for mental health workers, and is significant in treating patients with depression.