Classification accuracy of structural and functional connectomes across different depressive phenotypes

Abstract

Phenotyping of major depressive disorder (MDD) in research can vary from study to study, which, together with heterogeneity of the disorder, may contribute to the inconsistent associations with various risk factors including neuroimaging features. These aspects also potentially underlie previous problems with machine learning methods using imaging data to inform predictive biomarkers. In this study we therefore aimed to examine the classification accuracy of structural and functional connectomes across different depressive phenotypes, including separating MDD subgroups into those with and without early childhood adversity (one of the largest risk factors for MDD associated with brain development). We applied logistic ridge regression to classify control and MDD participants defined according to six different MDD definitions in a large community-based sample (N = 14,507). We used brain connectomic data based on six structural and two functional network weightings and conducted a comprehensive analysis to (i) explore how well different connectome modalities predict different MDD phenotypes commonly used in research, (ii) investigate whether stratification of MDD based on the presence or absence of early childhood adversity (measured with the childhood trauma questionnaire) can improve prediction accuracies, and (iii) identify important predictive features that are consistent across MDD phenotypes. We find that functional connectomes consistently outperform structural connectomes as features for MDD classification across phenotypes. Highest accuracy of 61.06% (chance level 50.0%) was achieved when predicting the Currently Depressed phenotype (i.e. the phenotype defined by the presence of more than five symptoms of depression in the past two weeks) with features based on partial correlation functional connectomes. Accuracy of classifying Currently Depressed participants with added CTQ threshold criterion rose to 65.74%. Application of the Jaccard index to assess predictive feature overlap indicated that there were neurobiological differences between MDD patients with and without childhood adversity. Further to that, analysis of predictive features for different MDD phenotypes with binomial tests revealed sensorimotor and visual functional subnetworks as consistently important for prediction. Our results provide the basis for future research, and indicate that differences in sensorimotor and visual subnetworks may serve as important biomarkers of MDD.