Classification of Obsessive-Compulsive Disorder Using Distance Correlation on Resting-State Functional MRI Images

Abstract

Both the Pearson correlation and partial correlation methods have been widely used in the resting-state functional MRI (rs-fMRI) studies. However, they can only measure linear relationship, although partial correlation excludes some indirect effects. Recent distance correlation can discover both the linear and non-linear dependencies. Our goal was to use the multivariate pattern analysis to compare the ability of such three correlation methods to distinguish between the patients with obsessive-compulsive disorder (OCD) and healthy control subjects (HCSs), so as to find optimal correlation method. The main process includes four steps. First, the regions of interest are defined by automated anatomical labeling (AAL). Second, functional connectivity (FC) matrices are constructed by the three correlation methods. Third, the best discriminative features are selected by support vector machine recursive feature elimination (SVM-RFE) with a stratified N-fold cross-validation strategy. Finally, these discriminative features are used to train a classifier. We had a total of 128 subjects out of which 61 subjects had OCD and 67 subjects were normal. All the three correlation methods with SVM have achieved good results, among which distance correlation is the best [accuracy = 93.01%, specificity = 89.71%, sensitivity = 95.08%, and area under the receiver-operating characteristic curve (AUC) = 0.94], followed by Pearson correlation and partial correlation is the last. The most discriminative regions of the brain for distance correlation are right dorsolateral superior frontal gyrus, orbital part of left superior frontal gyrus, orbital part of right middle frontal gyrus, right anterior cingulate and paracingulate gyri, left the supplementary motor area, and right precuneus, which are the promising biomarkers of OCD.

Reduced loss aversion in value-based decision-making and edge-centric functional connectivity in patients with internet gaming disorder

“…Behavioral measures included symptom severity of IGD, net-score and loss aversion acquired from the IGT. We calculated nFC ( Luo, Liu, Jin, Chang, & Peng, 2021 ; Zha, Li, et al., 2022 ), eFC and overlapping community features of eFC. And we used the generalized mediation analysis and dynamic causal modeling (DCM) analysis.…”

“…Following previous studies, we defined nodes and calculated the nFC Matrix ( Luo et al., 2021 ; Zha, Li, et al., 2022 ). Nodes were automatically divided by anatomical automatic labeling (AAL) parcellation ( Tzourio-Mazoyer et al., 2002 ).…”

Dissecting Psychiatric Heterogeneity and Comorbidity with Core Region-Based Machine Learning

“…Brain regions contributing to the discrimination consisted of the left superior temporal gyrus, the right middle temporal gyrus, the left supramarginal gyrus, and the superior parietal lobule. Jia et al [39] adopted the voxel-mirrored homotopic connectivity method to investigate interhemispheric coordination in [38] used distance correlation to construct the functional connectivity matrices (OCD = 61, HC = 67), and the best discriminative features were selected by SVM recursive feature elimination with a 10-fold CV strategy. The features from distance correlation achieved an accuracy of 93.0% (89.7% specificity and 95.1% sensitivity), superior to features from either Pearson correlation or partial correlation.…”

“…The SVM classifier with 10-fold CV successfully distinguished patients (83.6% accuracy, 80.8% sensitivity, and 86.3% specificity) from HCs with the most discriminative regions located in the inferior parietal lobule, dorsolateral PFC, middle occipital gyrus, and cuneus. Luo et al [ 38 ] used distance correlation to construct the functional connectivity matrices (OCD = 61, HC = 67), and the best discriminative features were selected by SVM recursive feature elimination with a 10-fold CV strategy. The features from distance correlation achieved an accuracy of 93.0% (89.7% specificity and 95.1% sensitivity), superior to features from either Pearson correlation or partial correlation.…”