Network-Based Spreading of Gray Matter Changes Across Different Stages of Psychosis

Abstract

Importance: Psychotic illness is associated with anatomically distributed gray matter reductions that can worsen with illness progression, but the mechanisms underlying the specific spatial patterning of these changes is unknown.

Objective: To test the hypothesis that brain network architecture constrains cross-sectional and longitudinal gray matter alterations across different stages of psychotic illness and to identify whether certain brain regions act as putative epicenters from which volume loss spreads.

Design, settings, and participants: This case-control study included 534 individuals from 4 cohorts, spanning early and late stages of psychotic illness. Early-stage cohorts included patients with antipsychotic-naive first-episode psychosis (n = 59) and a group of patients receiving medications within 3 years of psychosis onset (n = 121). Late-stage cohorts comprised 2 independent samples of people with established schizophrenia (n = 136). Each patient group had a corresponding matched control group (n = 218). A sample of healthy adults (n = 356) was used to derive representative structural and functional brain networks for modeling of network-based spreading processes. Longitudinal illness-related and antipsychotic-related gray matter changes over 3 and 12 months were examined using a triple-blind randomized placebo-control magnetic resonance imaging study of the antipsychotic-naive patients. All data were collected between April 29, 2008, and January 15, 2020, and analyses were performed between March 1, 2021, and January 14, 2023.

Main outcomes and measures: Coordinated deformation models were used to estimate the extent of gray matter volume (GMV) change in each of 332 parcellated areas by the volume changes observed in areas to which they were structurally or functionally coupled. To identify putative epicenters of volume loss, a network diffusion model was used to simulate the spread of pathology from different seed regions. Correlations between estimated and empirical spatial patterns of GMV alterations were used to quantify model performance.

Results: Of 534 included individuals, 354 (66.3%) were men, and the mean (SD) age was 28.4 (7.4) years. In both early and late stages of illness, spatial patterns of cross-sectional volume differences between patients and controls were more accurately estimated by coordinated deformation models constrained by structural, rather than functional, network architecture (r range, >0.46 to <0.57; P < .01). The same model also robustly estimated longitudinal volume changes related to illness (r ≥ 0.52; P < .001) and antipsychotic exposure (r ≥ 0.50; P < .004). Network diffusion modeling consistently identified, across all 4 data sets, the anterior hippocampus as a putative epicenter of pathological spread in psychosis. Epicenters of longitudinal GMV loss were apparent in posterior cortex early in the illness and shifted to the prefrontal cortex with illness progression.

Conclusion and relevance: These findings highlight a central role for white matter fibers as conduits for the spread of pathology across different stages of psychotic illness, mirroring findings reported in neurodegenerative conditions. The structural connectome thus represents a fundamental constraint on brain changes in psychosis, regardless of whether these changes are caused by illness or medication. Moreover, the anterior hippocampus represents a putative epicenter of early brain pathology from which dysfunction may spread to affect connected areas.

Figure 1.. Analysis Workflow for the Coordinated Deformation Model (CDM)

A, We derived voxelwise gray matter volume (GMV) estimates using deformation-based morphometry (DBM). Five separate contrasts were specified using a robust marginal model to infer baseline GMV differences and longitudinal GMV changes associated with illness and antipsychotic medication at 3 and 12 months. Cross-sectional contrasts were specified such that positive values (in red) in the resulting voxelwise t statistic maps indicate lower volume in patients compared with controls. For the illness-related longitudinal contrasts, positive values (in red) in the resulting voxelwise t statistic maps indicate greater longitudinal GMV decline in patients receiving placebo compared with controls. For the medication-related longitudinal contrasts, positive values in the resulting voxelwise t statistic maps indicate greater longitudinal GMV decline in patients receiving medication compared with both those receiving placebo and controls. Schizophrenia data sets included participants in the BrainGluSchi (BGS) and COBRE studies. Δ indicates change from baseline. B, The contrast statistics were mapped to a brain parcellation comprising 332 regions, and diffusion and functional magnetic resonance imaging data from an independent healthy sample were used to generate sample-averaged functional coupling (FC) and structural connectivity (SC) matrices. These matrices were used to model mean volume changes in structurally connected neighbors. Under the CDM, the estimated deformation of a node, d̂_i, is modeled as a weighted sum of the deformation values observed in its structurally connected neighbors, d_i (shown as light blue nodes in the example graphs). The weights are given by the adjacency matrix, A_ij. Three different matrices were used, yielding 3 CDM variants: (1) a model denoted as CDM-SC, in which A_ij = 1 if regions i and j share a connection and A_ij = 0 otherwise; (2) a model denoted as CDM-SCw in which the elements of A_ij correspond precisely to the weighted SC matrix, such that the contribution of each neighbor is weighted by the strength of its structural connectivity to the index node; and (3) a model denoted CDM-FCw, in which the elements of A_ij correspond precisely to the weighted FC matrix, such that the contribution of each neighbor is weighted by its FC with the index node. C, Model performance was evaluated using the product-moment correlation between regional estimates of observed and estimated GMV differences. D, We also compared model performance with 3 benchmark null models accounting for spatial autocorrelations in the deformation maps (Null-smash and Null-spin) and basic topological properties of the connectome (Null-rewire) (see Statistical Analysis subsection of the Methods section and eMethods 9 in Supplement 1).

Figure 2.. Baseline and Longitudinal Illness-Related Gray Matter Volume Changes Are Constrained by Connectome Anatomy

Top row, The contrast statistics for 4 cross-sectional contrasts (first-episode psychosis, early psychosis, schizophrenia-BrainGluSchi [BGS], and schizophernia-COBRE data sets) mapped to a brain parcellation comprising 332 regions. a Indicates anterior; AMY, amygdala; CAU, caudate nucleus; d, dorsal; DA, dorsoanterior; DMN, default mode network; DorsAttn, dorsal attention network; DP, dorsoposterior; FPN, frontoparietal network; GP, globus pallidus; HIP, hippocampus; l, lateral; Lim, cortical limbic network; m, medial; MTL, medial-temporal lobe (amygdala and hippocampus); NAc, nucleus accumbens; p, posterior; SomMot, somatomotor network; Stri, striatum; PUT, putamen; and THA, thalamus. Middle row, Performance of the equally weighted (SC), structural connectivity–weighted (SCw), and functional coupling–weighted (FCw) coordinated deformation models (CDMs) relative to the Null-smash, Null-spin, and Null-rewire benchmarks. Black circles indicate the observed product-moment correlations between estimated and actual regional deformation values for each model. Note that the observed value used for comparison against the Null-spin models is different because the subcortex was excluded. Bottom row, Scatterplots of the association between observed and estimated regional volume deformation values for the best-performing CDM-SCw model at each time point. ROI indicates region of interest. ^aP < .016.

Figure 3.. Longitudinal Illness-Related and Antipsychotic-Related Gray Matter Volume Changes Are Constrained by Connectome Anatomy

Top row, The contrast statistics for illness-related and antipsychotic-related contrasts mapped to a brain parcellation comprising 332 regions. a Indicates anterior; AMY, amygdala; CAU, caudate nucleus; d, dorsal; DA, dorsoanterior; DMN, default mode network; DorsAttn, dorsal attention network; DP, dorsoposterior; FPN, frontoparietal network; GP, globus pallidus; HIP, hippocampus; l, lateral; Lim, cortical limbic network; m, medial; MTL, medial-temporal lobe (amygdala and hippocampus); NAc, nucleus accumbens; p, posterior; SomMot, somatomotor network; Stri, striatum; PUT, putamen; and THA, thalamus. Middle row, Performance of the equally weighted (SC), structural connectivity–weighted (SCw), and functional coupling–weighted (FCw) coordinated deformation models (CDMs) relative to the Null-smash, Null-spin, and Null-rewire benchmarks. Black circles indicate the observed product-moment correlations between estimated and actual regional deformation values for each model at each time point. Note that the observed value used for comparison against the Null-spin models is different because the subcortex was excluded. Bottom row, Scatterplots of the association between observed and estimated regional deformation values for the best-performing CDM-SCw model at each time point. ROI indicates region of interest. ^aP < .004.

Figure 4.. Regional Epicenters of Gray Matter Volume (GMV) Loss

A, Epicenters were defined as potential sources of pathological volume loss from which GMV reductions spread (blue) to affect structurally connected areas. To identify such regions, we simulated a spreading process using a network diffusion model (NDM). Schizophrenia data sets included participants in the BrainGluSchi (BGS) and COBRE studies. B, Using each of the 332 parcellated regions as a seed, we retained the maximum correlation between the simulated and observed GMV abnormalities (maximum r). For each contrast, we then compared maximum r values for each region to a distribution of maximum r values from 2 benchmark null models accounting for spatial autocorrelations in the deformation maps (Null-smash) and basic topological properties of the connectome (see Statistical Analysis subsection of the Methods section [NDM]). Regional epicenters with significantly greater maximum r than a spatially constrained null model (orange indicates P < .05; red, familywise error [FWE] P < .05) are shown for cross-sectional (C-F) and longitudinal (G-J) effects. Results using Null-rewire benchmark models and scatterplots of observed and estimated volume abnormalities are provided in eFigure 3 in Supplement 1.