White matter abnormalities across the lifespan of schizophrenia: A harmonized multi-site diffusion MRI study

Abstract

Several prominent theories of schizophrenia suggest that structural white matter pathologies may follow a developmental, maturational, and/or degenerative process. However, a lack of lifespan studies has precluded verification of these theories. Here, we analyze the largest sample of carefully harmonized diffusion MRI data to comprehensively characterize age-related white matter trajectories, as measured by fractional anisotropy (FA) across the course of schizophrenia. Our analysis comprises diffusion scans of 600 schizophrenia patients and 492 healthy controls at different illness stages and ages (14–65 years), which were gathered from 13 sites. We determined the pattern of age-related FA changes by cross-sectionally assessing the timing of the structural neuropathology associated with schizophrenia. Quadratic curves were used to model between-group FA differences across whole-brain white matter and fiber tracts at each age; fiber tracts were then clustered according to both the effect-sizes and pattern of lifespan white matter FA differences. In whole-brain white matter, FA was significantly lower across the lifespan (up to 7%; p<0.0033) and reached peak maturation younger in patients (27 years) compared to controls (33 years). Additionally, three distinct patterns of neuropathology emerged when investigating white matter fiber tracts in patients: 1) developmental abnormalities in limbic fibers, 2) accelerated aging and abnormal maturation in long-range association fibers, 3) severe developmental abnormalities and accelerated aging in callosal fibers. Our findings strongly suggest that white matter in schizophrenia is affected across entire stages of the disease. Perhaps most strikingly, we show that white matter changes in schizophrenia involve dynamic interactions between neuropathological processes in a tract-specific manner.

Introduction

Structural brain abnormalities have been extensively studied in schizophrenia¹, especially since the emergence of neuroimaging technology^1,2,3. Early works have reported gray matter deficits localized mostly to frontal and temporal lobes, which are present prior to the onset of psychosis and tend to worsen within the first few years of illness^4,5,6. Recently, the field has shifted focus from gray matter regions to the cerebral white matter connections, propelled in part by the advent of diffusion magnetic resonance imaging (dMRI)^7–9. These studies have revealed a diffuse pattern of white matter alterations, as well as disruptions in structural connectivity, that have been found to be related to the manifestation of clinical symptoms and cognitive deficits¹⁰. However, despite the large number of imaging studies investigating white matter in schizophrenia, the field has not yet reached a consensus regarding the spatial pattern and lifespan trajectory of these changes.

Mapping white matter changes along the lifespan could provide unprecedented insight into the timing and the biological nature of the factors underlying white matter deficits that are observed in schizophrenia¹¹. According to disease simulations proposed by Kochunov et al.⁹, white matter abnormalities occurring prior to illness onset that subsequently remain stable with increasing age (i.e. lower but parallel to healthy aging), would indicate the contributions of genetic or early developmental risk factors to the observed white matter pathology. More specifically, a first possible trajectory, “neurodevelopmental models in schizophrenia”^9,12–14, postulate that schizophrenia is caused by environmental and/or genetic insults that occur during prenatal, perinatal, or early childhood/adolescence, leading to lower integrity in the cerebral white matter throughout the lifespan as compared to healthy controls. A second possible trajectory, “maturational model in schizophrenia”^15,16, suggests that disturbances during maturation would be reflected by different ascending slopes and a shift in peak maturation of white matter, indicating perturbed myelination. A third possible trajectory, “accelerated aging in schizophrenia”^{5,17,18,19,20,21}, supports that schizophrenia is marked by steeper descending slopes related to accelerated aging processes, such as myelin breakdown (Figure 1 depicts the three trajectory models). Longitudinal imaging studies focusing on schizophrenia patients have provided some evidence in support for all three models^5,6,19. However, these studies tend to have small samples of patients at one stage of illness, as well as limited age-windows, which likely obscured interpretations regarding the role development and aging in the white matter disruptions across the schizophrenia spectrum. A large cohort of schizophrenia patients at different stages of illness and ages could more accurately characterize white matter trajectories and in turn, by understanding the nature of these biological factors it might help develop more targeted treatment options.

Figure 1.

Trajectory models for early developmental, maturational or degenerative processes in schizophrenia.

World-wide collaborative efforts have led to the collection of large scale dMRI datasets that are aimed at boosting statistical power to allow increased sensitivity for detecting subtle white matter changes in schizophrenia. They have predominantly used statistical approaches^{22,23,24,25,26,27,28,29,5} to combine diffusion tensor-derived measures, such as fractional anisotropy (FA), collected from multiple, international studies. While these studies have paved the way toward replicating case-control differences in large sample populations, no previous studies have been successful in reconstructing the raw harmonized diffusion signal, which is critical for generating consistent microstructural, tractography and connectivity metrics across datasets. To address this challenge, we developed and validated a novel harmonization procedure³⁰ based on rotation invariant spherical harmonics features^31,32, which reconciles raw dMRI signals across disparate sites and acquisition parameters, while preserving inter-subject biological variability. This harmonization procedure can effectively leverage large, multi-site data to address more nuanced questions, including how white matter deficits progress with age in schizophrenia.

In the present study, we comprehensively characterize white matter changes across the course of schizophrenia, using the largest sample of carefully harmonized dMRI data to-date. Our multi-site sample is comprised 1092 dMRI scans from a combined sample of healthy controls and schizophrenia patients obtained at different ages (14–65 years of age). Using these harmonized data, we aim to identify and characterize global and tract-specific age-related trajectories of white matter pathology as measured by FA across the schizophrenia lifespan. Based on previously proposed trajectory models (Figure 1), we sought to determine whether structural white matter pathology in schizophrenia follows a developmental, maturational or degenerative pattern, and if this pattern varies between white matter fiber tracts across the brain.

Materials and methods

Participants and imaging acquisition.

This study included 600 patients diagnosed with a schizophrenia-spectrum disorder across multiple illness stages ¹ (mean [SD] age, 31.29 [12.02] years; 383 [63%] male) and 492 controls (mean [SD] age, 29.80 [13] years; 275 [56%] male), recruited from 13 different sites. All dMRI data were collected as part of separate, individual studies, and harmonized across the sites for subsequent joint analyses. Each site collected controls that were matched on age and sex to the site-specific patient population. Supplementary Table 1 and Supplementary Study Acquisitions summarize acquisition parameters, as well as basic demographic characteristics for each site.

All data (except for the Philadelphia Neurodevelopmental Cohort (PNC)^33,34) were provided by the principal investigators following procurement of institutional IRB approvals for sharing and analyzing de-identified data. PNC data were downloaded from the NIH database following NIH approval.

Image pre-harmonization procedure.

A standardized processing procedure was carried out for each subject across the 13 sites. This involved axis alignment, centering and eddy-current correction, using the Psychiatry Neuroimaging Laboratory pipeline: https://github.com/pnlbwh/pnlutil. In addition, brain masking was performed with the brain extraction tool (BET)³⁵. A total of 60 subjects (40 patients and 20 healthy controls) were excluded from our analysis, which resulted in 1092 subjects who were included in our data analysis. Exclusion was based on the presence of severe susceptibility distortion, cerebellar cropping, signal dropouts and artifacts upon careful visual/manual inspection. PNC data (8–17 years) were automatically quality checked as part of another study (Cetin Karayumak, Neuroimage, 2019): whole brain average DTI fit errors (residual value in the Figure below) were used to separate bad and good quality cases and a heuristic threshold (0.052) was set to separate the two clusters (refer to figure below). To preserve the even distribution of subjects between the two groups, only 58 healthy control good quality cases (ages 14–17 years) among 800 subjects were included in the present study from this PNC data set.

Harmonization.

Retrospective harmonization³⁰ was applied to raw dMRI data across the 13 study sites to remove the site related differences. To this end, a reference site was selected— Brigham and Women’s Hospital (BWH)-1 (site 1)—and raw dMRI data from the remaining 12 sites were mapped to this reference. A subset of age and sex-matched controls (of number of subjects >16; matched to BWH-1 were selected from each site as target for harmonization (refer to Supplementary Table 2 for the statistics of the matching and Supplementary Figure 1 for an overview of the harmonization pipeline)³⁰. This method was previously shown to successfully remove scanner-specific effects, while accounting for subtle acquisition discrepancies related to b-values, spatial resolution and number of gradient directions, as well as preserve anatomic variability across subjects³⁰. In particular, we previously showed that sex and age-related white matter differences were preserved after dMRI data harmonization, irrespective of the test sample size. To evaluate the performance of harmonization for the matched controls, whole-brain mean FA values were computed for the reference site and the target site before and after harmonization (Supplementary Figure 2) and unpaired t-tests assessed between-site difference. Statistical differences between the matched controls across sites (p=0.0077; t=3.52; Pearson’s r=0.0005) were removed after harmonization (p=0.3320; t=1.15; Pearson’s-r=0.93).

Image post-harmonization procedures.

Registration.

After dMRI data were harmonized to the reference BWH-1 site, FA was computed by fitting a single-tensor to the dMRI data using FSL’s DTIFIT³⁶. Resulting FA brain maps were registered to standard space (MNI space) following a two-step procedure to minimize miss-registration: first, a study specific FA template was generated for each site using antsMultivariateTemplateConstruction2.sh³⁷; next, site-specific templates were linearly and non-linearly registered to the Illinois Institute of Technology (IIT) Human Brain Atlas³⁸ FA map in MNI space using ANTs registration³⁹. The affine and non-linear transformations were subsequently applied to register each individual subject (in template space) to the FA map in MNI space.

White matter processing.

For each subject, a total of 14 probabilistic white matter fiber tracts ² were extracted from the IIT Human Brain probabilistic atlas³⁸, using a threshold to 0.25 (refer to Figure 3 for fiber tracts). The IIT atlas was chosen over the more popular and older white matter atlases, as it provides the highest spatial resolution. For each tract, mean FA was computed by averaging FA across all voxels traversing a probabilistic fiber bundle. In addition, whole-brain averaged FA was computed across voxels comprising a white matter skeleton using a template from IIT (IIT_WM_atlas_skeletonized.nii.gz³⁸).

Figure 3.

Age-related FA trajectories in schizophrenia patients and healthy controls across 14 white matter fibers tracts. Fibers tracts colors displayed in the surface rendering match the colors of curves for corresponding brain fiber tracts.

Statistical modeling.

The present study aimed to cross-sectionally examine white matter trajectories and group differences between those trajectories in a large, age-diverse and multi-site sample of patients and controls. Statistical inference was performed at the scale of the whole-brain as well as individual fiber tracts. To determine the best model for age-related change, mean FA was fitted separately to patients and controls using i) parametric models, such as quadratic, Poisson, gamma and linear; and ii) non-parametric models, such as smoothing-splines, local-polynomial-regression function and natural-splines, based on previous lifespan studies of white matter^{17,40,41–43}. The best fitted nonparametric model (optimal smoothness/regularization terms) was used as a ground truth and a parametric model with highest adjusted-r² was selected. This resulted in a quadratic curve providing the best parametric fit and was hence, applied herein. To estimate the peak age and confidence interval (CI) at 97.5% and 2.5% quantiles, quadratic curve fitting was repeated using a bootstrapping procedure with 5000 iterations.

To determine the influence of age on FA abnormalities in patients (i.e. between-group FA difference), a quadratic curve was subsequently applied, by independently modelling each year¹⁷:

$$\begin{gathered} FA=v_1+v_{2}dx+v_3(age-\alpha ) \\ +v_4(age-\alpha )dx+v_5{(age-\alpha )}^2 \\ +v_6{(age-\alpha )}^{2}dx+\kappa +ϵ \end{gathered}$$

where $v_1,v_2,\dots ,v_6$ are unknown parameters, $dx$ is the categorical variable for diagnosis: patients (0), controls (1), $\kappa$ represents other confounds, such as sex and $ϵ$ is the error term. The regression model was independently fitted with age centered between 14 and 65 years in yearly increments. This regression model yielded cross-sectional snapshots of FA change per year across the lifespan in schizophrenia and controls¹⁷. The between-group FA change was quantified as:

$$\frac{v_2}{v_1}x100\%$$

To demonstrate robustness of our results, we also computed the between-group effect-sizes (Cohen’s d) inside a sliding window within different age-strata for FA. A window size of 5 years was applied around each age, centered between 14 and 65 years along the sorted age-span of each group. After d was computed for the entire lifespan, k-means algorithm was used to cluster the fiber tracts, based on the characteristics of FA deficits in schizophrenia.

Results

We investigated age-related white matter alterations in patients compared to controls using a large cross-sectional and multi-site design with harmonized diffusion scans. FA, the most commonly used diffusion metric in clinical research, was selected as the dependent variable because of its ability to describe white matter coherence and axonal organization^2,44,45. To address the primary objective of this study, we examined whole-brain and tract-specific FA (across 14 fiber tracts) to understand whether white matter changes are consistent with abnormal neurodevelopment, maturation and/or accelerated aging processes in patients, relative to controls.

Modeling age-related white matter changes in patients and controls.

Mean FA extracted from whole-brain white matter was modeled over a 52-year period from 14 to 65 years (i.e. entire lifespan available throughout datasets) in patients and controls. Age-squared (the highest order term) was found to be the most significant term (${\upsilon }_5=4.5x{10}^{-10},t=4.7,p=6.1x{10}^{-4})$ in the quadratic model. Thus, the quadratic model was considered to provide the most parsimonious explanation of the relation between the FA and age.

Figure 2a shows trajectories of whole-brain FA in schizophrenia (brown-curve) and controls (orange-curve) for the entire cohort. Whole-brain FA showed a monotonic increase until reaching a peak at the age of 33 (median value after 5000 bootstrapping, SD=1.2) years in controls, and earlier, at the age of 27 (SD=2.7) years in patients. In both patients and controls, FA declined monotonically after reaching peak. Figure 2b depicts FA deviations in patients compared to controls at each age. Age epochs at which between-group FA significantly differed are denoted with a square (p<0.0033, Bonferroni-corrected ³; precise p-values have been included in Supplementary Table 4), as determined by the significance of the ${\upsilon }_2$ coefficient (Range of between-group FA% difference = [1.5, 7]). Additionally, the magnitude of d of whole-brain between-group FA differences, which were computed inside a sliding window within different age-strata, were significantly increased with age (Range of d = [0.55, 1.65]) in patients (Figure 2c).

Figure 2.

Whole-brain age-related FA difference/trajectories in schizophrenia patients and healthy controls: Panel a) Curves modeling the age-related trajectories in FA for healthy controls (orange curve) and schizophrenia subjects (brown curve) are shown within the confidence interval-CI (estimated with bootstrapping=5000 bootstrapped samples), dashed line. Standard deviation (shaded regions) is computed separately in schizophrenia and healthy controls centered at each age bin (5 years). Mean peak and standard deviation of each curve is depicted in black (estimated with bootstrapping=5000 bootstrapped samples); Panel b) Between-group FA difference as a function of age. Negative percentages indicate FA loss in schizophrenia patients; Panel c) Effect-size (d) difference between patients and controls (circles): Negative effect-size indicates a larger FA loss in schizophrenia patients compared to healthy controls. d was also modeled using smoothing splines for illustration (adjusted-r² = 0.89).

Since it is likely that whole-brain results are influenced by regional changes, we examined tract-specific changes to improve the spatial characterization of white matter trajectories in schizophrenia. To this end, individual white matter fiber tracts were modeled over the 52-year period. A quadratic curve (Figure 3 and 4) provided the best fit for cross-sectional, age-related trajectories (p<0.0033) in all but two white matter fiber tracts: the left ILF in controls and the right UF in patients, which were characterized by linear FA decline with increasing age (Figure 3).

Figure 4.

Age-related FA loss (%) in schizophrenia patients across 14 white matter fiber tracts. Significant reductions at each age in schizophrenia patients, relative to healthy controls are denoted by circles. Circle colors correspond to brain region colors displayed in the surface rendering (Figure 3). L: left hemisphere, R: right hemisphere.

Tract-specific white matter trajectories.

To better understand the regional patterns of white matter trajectories in schizophrenia, we examined whether differences in tract-specific trajectories could be clustered based on effect-sizes occurring along the lifespan. In doing so, we identified three distinct clusters of FA trajectories in patients compared to controls (Figure 5): i) small to medium effects that remained stable over the course of illness (Stable trajectory); ii) medium to strong effects, which gradually progressed with advancing age and reached peak maturation in FA younger (Declining trajectory) and, iii) strong effects from the outset of illness, which also deteriorated with increasing age (Deficit and Declining trajectory). As such, the findings are presented with reference to these three distinct clusters (Figure 5a).

Figure 5.

Effect-size trajectories across 14 fiber tracts, comparing schizophrenia patients and controls (Panel a). Tract-specific white matter pathology (i.e. based on between-group differences) could be organized into three clusters, according to effect-size occurring along trajectory, characterized by: Panel b) small to medium effects that remained stable over the course of illness (Stable trajectory), may signify abnormal developmental origins of white matter pathology in schizophrenia; Panel c) medium to strong effects, which gradually progressed with advancing age and short white matter maturational period (Declining trajectory), may provide evidence for maturational abnormality and accelerated aging; Panel d) strong effects from the outset of illness, which also progressed with increasing age (Deficit and declining trajectory), may indicate abnormal development and accelerated aging along white matter in schizophrenia.

Cluster 1: Stable trajectory cluster - abnormal early development.

White matter fiber tracts with small to medium effects that are stable across the entire lifespan are presented in Figure 5b. For example, CING2 displayed a stable FA abnormality, with d~0.5, for the entire age-range of schizophrenia. Consistent with this model, FA was significantly (p<0.0033) reduced by ~4% in patients relative to controls from youth to the fifth decade of life (Figure 4). Similar effect-sizes across the entire lifespan were also observed in other limbic white matter structures, such as CING1, UF, and ILF. Thus, FA deficits across limbic fiber tracts appear to emerge early and prior to the onset of schizophrenia and remain evident across the entire schizophrenia time-course, in accord with a developmental model of schizophrenia.

Cluster 2: Declining trajectory cluster - accelerated aging and abnormal maturation.

White matter fiber tracts showing medium to strong effects that increase with age, are presented in Figure 5c. These included long-range association and language-related connections, including the IFOF (left and right) and SLF (left and right), which showed a pattern of progressive FA decline with increasing age. Specifically, effect-sizes of FA differences in these tracts were small prior to their maturational peak in early adulthood, after which effect-sizes gradually increased with advancing age in late adulthood. After peak maturation, FA was significantly lower (p<0.0033) in schizophrenia, which reached a 7% reduction after the sixth decade (Figure 4-right SLF). Additionally, the most significant shifts from the healthy maturational peaks (p<0.0033) occurred in these fibers, e.g. the SLF (Figure 3), which reached a peak maturation earlier in schizophrenia subjects (at age 29) relative to controls (at age 38). These findings provide evidence for maturational abnormalities and accelerated aging in schizophrenia; the white matter maturational period is cut short and age-related degeneration/decline starts earlier.

Cluster 3: Deficit and Declining trajectory cluster - abnormal development and accelerated aging.

White matter fiber tracts displaying the strongest effects included the interhemispheric callosal fibers (forceps minor and forceps major), which were consistent with both abnormal development and accelerated aging in schizophrenia and are presented in Figure 5d. Significant between-group FA differences (p<0.0033) were observed throughout the entire lifespan in these fibers. They showed moderate to large effect-sizes for group differences presenting at early ages (14 years), as well as patterns of accelerated decline, characterized by steeper descending slopes in FA with advancing age. For example, in the forceps minor, effect-sizes of FA differences were stable (d~0.8) until age 38, after which d reached up to 1.8 in the sixth decade. Additionally, patients displayed a 10% FA reduction (Figure 4-forceps minor) after the sixth decade (Range of FA% difference = [3, 10]; Range of d = [0.8, 1.7]). These findings provide evidence for developmental abnormalities occurring within the largest interhemispheric connections of the brain.

Discussion

Schizophrenia is characterized by widespread white matter decline; however, the precise trajectories and whether these relate to early developmental risk factors, abnormal maturational development and/or degenerative processes (Figure 1) remains contentious. The large size (1092) and wide age-span (14–65 years) of our sample uniquely qualifies us to address the temporal evolution of white matter alterations across the time-course of schizophrenia illness.

Utilizing our unique, carefully harmonized dMRI dataset, we provide a breakthrough finding by identifying early-emerging white matter deficits, as well as early maturational shifts in the white matter of patients with schizophrenia, which varies in a tract-specific manner. Specifically, limbic connections appeared selectively vulnerable to early developmental anomalies that do not progress over time (cluster 1); whereas, long-range intra-hemispheric association tracts (including language tracts) displayed shorter maturational windows and faster declines (cluster 2), consistent with accelerated ageing processes in schizophrenia. Perhaps most strikingly, callosal-fibers exhibited severe deficits from the outset of illness, which became more pronounced with increasing age (reaching a 10% reduction after sixth decade in cluster 3). Reduced anisotropy of the corpus callosum reflects a well-replicated diffusion imaging finding in schizophrenia, with consistent reports across heterogeneous patient populations and maturational phases^46,47.

We observed patterns consistent with abnormal development (cluster 1, cluster 3), perturbed maturation (cluster 2) and accelerated aging (cluster 2, cluster 3). Specifically, limbic fibers displayed early-emerging deficits that did not worsen with advancing age. This preservation coheres with previous reports of static gray matter pathology in temporo-limbic fibers, connected by limbic circuitry⁴⁸. We note that loss of the significance in some portions of age range for limbic fibers may be due to variability in the data (refer to Supplementary Figure 3 for scattered data points, e.g., variability for cluster 1: left ILF increases after 30, which is not the case for whole brain white matter and the fiber tracts in cluster 2 and 3). In contrast, cortical association fibers, including the SLF and IFOF, were relatively preserved in the youngest patients, but were also characterized by premature peaks of white matter anisotropy, resulting in abnormally short maturational windows when compared to controls. In addition, the between-group FA difference gradually widened after reaching peak maturation in these fibers (reaching 8% loss in the IFOF and 9% in the forceps major), implying involvement of accelerated aging processes in schizophrenia. Consistent with previous reports, this finding suggests that white matter continues to deteriorate into the more chronic stages of illness. However, this deterioration appears localized to cortical and interhemispheric fibers in the brain. The tract-specific trajectories observed here could relate to spatial sequence of healthy white matter development, which reaches peak maturation in a nonlinear and tract-specific manner. In line with this notion, we observed the most extensive disruption across early maturing fibers, including the commissural fibers, which mature in adolescence to young-adulthood^6,40, which coincides with the timing of peak risk for developing schizophrenia.

This study represents the largest assessment of age-related white matter trajectories in schizophrenia to-date, and the first characterization of tract-specific trajectories. The power of the large sample size increased our sensitivity to detect FA deficits at the youngest ages, which was previously unavailable to MRI studies¹⁷. Despite these advantages, several limitations should be noted. First, our data is cross-sectional. Longitudinal designs offer direct examination of aging trajectories; however, the assessment from youth through late adulthood is more feasible with cross-sectional studies. Second, we were unable to examine the influence of disease onset age/duration, comorbid substance use/abuse or medication exposure/duration. It is of course possible that medication and other illness chronicity factors affect the integrity of white matter^49,50,51,52; however, a recent mega analysis⁵² reported that antipsychotic medication has minimal or no significant detrimental effect. Finally, our sample (starts at age 14), precludes us from distinguishing neurodevelopmental risk factors from early abnormal maturational processes, thus, our observed FA trajectories may be influenced by both developmental and early maturational changes in the younger subjects. In addition, it is possible that observations within adolescent-onset populations (ages 14 to 20 years) do not generalize to the wider schizophrenia population. To ensure that white matter trajectories were not influenced by demographic or clinical characteristics specific to the adolescent population, we compared sex, disease severity (PANSS) and medication dosage (mg) between patients that were younger and older than 20 years, which showed no significant differences (Supplementary Table 3). Thus, adolescent and adult patient groups were not significantly different in terms of basic demographic and clinical features.

This present work provides an initial benchmark for tract-specific trajectories of white matter abnormalities in schizophrenia. Our findings accord with a developmental perspective, suggesting that widely distributed white matter deficits emerge early or display perturbed maturation. In addition, callosal and long-range association (but not limbic) fibers undergo accelerated aging processes. This regional diversity could explain the heterogeneity encountered across previous dMRI studies and suggests that white matter pathology in schizophrenia dynamically interacts with neurodevelopment, abnormal maturation and aging processes, and manifests in a tract-specific manner at different ages and disease stages. In a future study, we plan to explore sex and diagnostic group interactions across a range of diffusion metrics (i.e., free-water, radial and axial diffusivity) and white matter clusters.