Resolving Heterogeneity in Posttraumatic Stress Disorder Using Individualized Structural Covariance Network Analysis

Abstract

The heterogeneity of posttraumatic stress disorder (PTSD) is an obstacle to both understanding and therapy, and this has prompted a search for internally homogeneous neuroradiological subgroups within the broad clinical diagnosis. We set out to do this using the individual differential structural covariance network (IDSCN). We constructed cortical thickness-based IDSCN using T1-weighted images of 89 individuals with PTSD (mean age 42.8 years, 60 female) and 89 demographically matched trauma-exposed non-PTSD (TENP) controls (mean age 43.1 years, 63 female). The IDSCN metric quantifies how the structural covariance edges in a patient differ from those in the controls. We examined the structural diversity of PTSD and variation among subtypes using a hierarchical clustering analysis. PTSD patients exhibited notable diversity in distinct structural covariance edges but mainly affecting three networks: default mode, ventral attention, and sensorimotor. These changes predicted individual PTSD symptom severity. We identified two neuroanatomical subtypes: the one with higher PTSD symptom severity showed lower structural covariance edges in the frontal cortex and between frontal, parietal, and occipital cortex-regions that are functionally implicated in selective attention, response selection, and learning tasks. Thus, deviations in structural covariance in large-scale networks are common in PTSD but fall into two subtypes. This work sheds light on the neurobiological mechanisms underlying the clinical heterogeneity and may aid in personalized diagnosis and therapeutic interventions.

Keywords: PTSD; heterogeneity; individual differential structural covariance network; psychoradiology; subtypes.

Figure 1

Flowchart for Individual Differential Structural Covariance Network (IDSCN) analysis. (a) Cortical thickness is calculated by surface-based morphometry and extracted based on Destrieux parcellation. (b) The reference network of structural covariance is constructed using Pearson correlation of cortical thickness between each pair of ROIs in the TENP control group, from which a perturbed network is created by adding PTSD patient k; the Z-score of the discrepancy between the perturbed and reference networks measures the IDSCN of patient k. (c) Heterogeneity comparison between the PTSD and TENP group uses the variability (Euclidean distance) between individualized SCN and group-level SCN. (d) The most differential edges are those different in at least 5% of the patients. (e) Predicted vs. observed CAPS scores show the performance of differential edges for symptom prediction using a random forest-based machine learning pipeline. (f) Subgroup clustering analysis uses a hierarchical tree; the height of each link in the dendrogram represents the distance between the clusters linked. (g) In the network distribution for differential edges shared by ≥5% of PTSD patients, the relative distribution (%RD) is the % of its overlapping voxels in a given network to the size of corresponding seed regions of differential edges. (h) In the functional annotation for differential edges, larger font size of functional terms corresponds to higher mean coactivation ratio (only terms surviving the permutation test [p  < 0.05] are shown). The scatterplots and word-clouds were created using R-based ggplot2 and wordcloud packages. For the circular ideograms of brain connectome, we utilized the circos toolbox in Perl, while the ball-and-stick charts were generated using the BrainNet Viewer toolbox in MATLAB. CAPS, Clinician-Administered PTSD Scale; PTSD, posttraumatic stress disorder; RD, relative distribution; ROI, region of interest; TENP, trauma-exposed non-PTSD.

Figure 2

PTSD patients cluster into two subgroups. (a) The top 32 differential structural covariance edges (i.e., differential in at least 5% of patients) (a1) and their Z-score distribution in the 2 subgroups (a2). (b) The five edges that show significant difference (FDR corrected) between the two PTSD subgroups (b1) and their Z-score distribution in the two subgroups (b2). ACC, anterior cingulate cortex; CAL, calcarine sulcus; FUS, fusiform gyrus; IFG, inferior frontal gyrus; INS, insula; IOG, inferior occipital gyrus; IPL, inferior parietal gyrus; L, left; MCC, middle cingulate cortex; MFG, middle frontal gyrus; OFC, orbitofrontal cortex; OLF, olfactory cortex; parCG, aracentral gyrus; PCC, posterior cingulate cortex; PCUN, precuneus; posCG, postcentral gyrus; preCG, recentral gyrus; R, right; SFG, superior frontal gyrus; SMG, supramarginal gyrus; SOG, superior occipital gyrus; STG, superior temporal gyrus.

Figure 3

Brain connectome patterns, their network distribution and function annotation. (a) The top 32 differential structural covariance edges (a1) were mapped onto 7 canonical brain networks based on functional coactivation; the relative distribution (%RD) is the % of its overlapping voxels in a given network to the size of corresponding seed regions of differential edges (a2); in the functional associations (a3), larger font in the word-cloud corresponds to higher mean coactivation ratio (only terms surviving the permutation test [p  < 0.05] are shown). (b) Shows the same for the five edges which are significantly different between the two PTSD subgroups. The 5 edges (b1) were also mapped onto 7 canonical brain networks with their network relative distribution (b2) and functional associations (b3). AFN, cortical affective network; CEN, central executive network; DAN, dorsal attention network; DMN, default mode network; RD, relative distribution; SMN, sensorimotor network; VAN, ventral attention network; VN, visual network.