Elucidating the relationship between white matter structure, demographic, and clinical variables in schizophrenia - a multicenter harmonized diffusion tensor imaging study

Abstract

White matter (WM) abnormalities are repeatedly demonstrated across the schizophrenia time-course. However, our understanding of how demographic and clinical variables interact, influence, or are dependent on WM pathologies is limited. The most well-known barriers to progress are heterogeneous findings due to small sample sizes and the confounding influence of age on WM.

The present study leverages access to the harmonized diffusion magnetic-resonanceimaging data and standardized clinical data from 13 international sites (597 schizophrenia patients (SCZ)). Fractional anisotropy (FA) values for all major WM structures in patients were predicted based on FA models estimated from a healthy population (n=492). We utilized the deviations between predicted and real FA values to answer three essential questions. 1) “Which clinical variables explain WM abnormalities?”. 2) “Does the degree of WM abnormalities predict symptom severity?”. 3) “Does Sex influence any of those relationships?”.

Regression and mediator analyses revealed that a longer duration-of-illness is associated with more severe WM abnormalities in several tracts. Additionally, they demonstrated that a higher antipsychotic medication dose is related to more severe corpus callosum abnormalities. A structural equation model demonstrated that patients with more WM abnormalities display higher symptom severity. Last, the results exhibited sex-specificity. Males showed a stronger association between duration-of-illness and WM abnormalities. Females presented a stronger association between WM abnormalities and symptom severity, with IQ impacting this relationship.

Our findings provide clear evidence for the interaction of demographic, clinical, and behavioral variables with WM pathology in SCZ. Our results also point to the need for longitudinal studies, directly investigating casualty and sex specificity of these relationships, as well as the impact of cognitive resiliency on structure-function relationships.

Introduction

Schizophrenia (SCZ) is a severe mental disorder with lifelong consequences for patients, their families, and society,^e.g.,1. With the advent of magnetic resonance imaging (MRI), it became possible to gain insight into the structural pathology of SCZ in vivo². Specifically, diffusion MRI (dMRI) allows the characterization of white matter (WM) microstructure and brain connectivity³. Most dMRI studies have utilized fractional anisotropy (FA) measure, and FA reductions are frequently interpreted as indicators of compromised WM (e.g., fiber density, myelination, or tract coherence)⁴. Earlier studies have reported widespread FA reductions across multiple WM areas^5-7 at all stages of SCZ^8-13.

While previous studies have established the importance of WM abnormalities for the pathology of SCZ, it is still unclear how the clinical course of SCZ interacts, influences, or is dependent upon observed abnormalities. Several demographic and clinical variables, including age, sex^{14, 15}, IQ^{16, 17}, duration-of-illness^18-22, medication^23-26, and symptom severity^{12, 27-29}, can be predictive of the clinical outcome in SCZ. However, we are far from being able to disentangle the primary pathophysiology of schizophrenia from the secondary effects of these variables.

Small sample sizes hamper most imaging studies³⁰, resulting in low statistical power and an inability to model the dynamic and complex interplay between WM structure, demographic, and clinical variables. Therefore, multicenter efforts have led to the creation of large scale dMRI datasets. Two principal approaches have been employed to combine multisite imaging data. Meta-analyses process each dataset separately and apply statistical approaches to pool summary statistics from all sites^{5, 31}. Harmonization approaches³², on the other hand, are applied at the signal level early in the preprocessing pipeline. They thereby remove scanner, software, and site-specific effects from the dMRI signal. Thus, after harmonization, the multisite dMRI data can be jointly analyzed as if acquired on the same scanner, allowing for the proceeding application of any data processing and model fitting.

Our group recently developed a novel dMRI data harmonization method and validated it utilizing several datasets³³. We previously utilized this method to harmonize multisite dMRI data to explore WM differences between 490 healthy individuals (HC) and 597 patients with SCZ³⁴. We modeled FA trajectories for both patients and HC, demonstrating age and region-specific WM abnormalities in SCZ³⁴.

The present study aims to extend our previous work by examining demographic and clinical variables associated with abnormal WM trajectories. To do so, we developed strategies to standardize retrospectively collected clinical and cognitive data across different studies and international sites. We predict each patient’s FA values based on healthy FA trajectories and utilize the difference between predicted and real FA values for all analyses. We hypothesize that a longer duration-of-illness and greater antipsychotic medication will be associated with the largest deviations from healthy FA. We apply a structural equation model (SEM) approach to test whether deviations from healthy FA trajectories predict symptom severity and whether IQ interacts with this structure-function association. Last, we examine if any postulated relationships exhibit sexual dimorphism to characterize sex-specific vulnerability to SCZ.

Methods

Participants, image acquisition and processing

A detailed description of participants, image acquisition, and preprocessing can be found in our previous work³⁴. Briefly, we utilized data from participants recruited and scanned at 13 independent international sites (Table 1, Supplementary Table 1). Data from all sites included patients with SCZ and matched HC, except for the Philadelphia Neurodevelopmental Cohort (PNC) data set. The PNC data, which only collected data from HC, was added to more precisely model normative FA trajectories for our analyses. The original studies’ principal investigators provided all data (except the PNC data set) following institutional IRB approvals to share and analyze de-identified data. The PNC data was downloaded following NIH approval.

Table 1:

Demographics

	Sample Size	Sex	Age (years)	Years of education (years)	Race	IQ	Handedness	Medication dosage	Duration-of-illness (years)	Age of onset (years)	Positive Symptoms	Negative Symptoms
		M: Male F: Female	Mean +/− std	Mean +/− std	1: African American 2: Caucasian 3: Other	Mean +/− std	(%) 0: Right 1: Either 2: Left	Current CPZ 1: 0 mg 2: 300 mg/day 3: 300-1000 mg/day 4: > 1000 mg/day	Mean +/− std	Mean +/− std	Mean +/− std	Mean +/− std
Everyone	597	M:380 F:217	31.31 +/− 12.05	13.18 +/− 2.35	1: 16% 2: 41% 3: 14%	100.73 +/− 16.85 Current IQ only: 102.11+/− 16.53	0: 67% 1: 2% 2: 11%	1: 14% 2: 34% 3: 31% 4: 4%	9.13 +/− 10.47 M: 9.38 +/− 10.71 F: 8.73 +/− 10.06 (T=.69, df=519, p<.49)	21.64 +/− 6.90 M: 20.96 +/− 6.09 F: 22.77 +/− 7.89 (T=−2.93, df=519, p<.004)	1.79 +/− .52	1.48 +/− .43
CMH	76	M: 48 F: 28	36.11 +/− 12.18	13.66 +/− 2.40	1: 15% 2:42% 3:7%	110.41 +/− 13.11	0:80% 1:1% 2:11%	1:9% 2:40% 3:40% 4:1%	13.60 +/−11.33	22.1 +/−5.67	1.43 +/−.40	1.27 +/− .29
MRC	42	M:32 1:10	37.67 +/− 10.83	13.05 +/− 2.08	1:48% 2:45% 3:7%	99.02 +/− 15.06	0:83% 1:2% 3:14%	1:5% 2:38% 3:41% 4:7%	16.76 +/−12.16	20.90 +/− 6.45	1.59 +/−.41	1.61 +/−.38
ZHH	24	M: 11 F:13	37.08 +/− 8.55	13.52 +/− 2.80	1: 25% 2: 29% 3: 25%	95.14 +/− 13.84	0:71% 1:8% 2:8%	2: 38% 3: 29% 4: 4%	17.67 +/− 11.20	19.86 +/− 7.57	1.51 +/− .44	1.63 +/−.36
Biceps	46	M: 29 F: 17	25.80 +/− 6.38	13.74 +/− 1.99	1:14% 2:28% 3:3%	109.30 +/− 14.10	0:40% 2:6%	2:52% 3:44% 4:2%	3.93 +/− 2.27	21.96 +/− 5.74	1.53 +/− .32	1.27 +/− .35
Bsnip Baltimore	51	M: 35 F: 16	33.86 +/− 10.69	12.71 +/− 2.16	1:47% 2:47% 3:6%	92.39 +/−17.80	0:90% 2:10%	2:22% 3:57% 4:6%	13.14 +/− 9.42	20.73 +/− 6.82	1.82 +/−.50	1.44 +/− .29
Bsnip Hartford	74	M:43 F:31	33.26 +/− 11.43	13.13 +/− 2.23	1:14% 2:72% 3:11%	97.01 +/−15.75	0:82% 1:3% 2:11%	2:37% 3:40% 5:23%	10.31 +/− 9.90	21.51 +/− 8.19	1.84 +/− .47	1.57 +/− .48
BWH-1	37	M:25 F:12	37.70 +/− 14.18	13.35 +/− 2.29	1:22% 2:35% 3:5%	102.68 +/− 15.21	0:54% 2:5%	1:11% 2:22% 3:16%	10.39 +/− 12.78	21.41 +/− 4.45	1.69 +/− .57	1.35 +/− .21
BWH-2	43	M:37 F:6	33.47 +/− 13.92	13.38 +/− 2.57	1:12% 2:58% 3:9%	108.65 +/− 16.24	0:67% 2:7%	1:12% 2:54% 3:9%	6.43 +/− 9.66	22. 58 +/− 6.47	1.60 +/− .57	1.46 +/− .38
AMC	48	M: 18 F:30	28.85 +/− 6.28		3:100%	Current IQ: 97.58+/−15.84		2:69% 3:23% 4:4%
Oxford	47	M:28 F:19	16.49 +/− 1.20			Current IQ: 89.41 +/− 17.76	0:43% 1:13% 2:6%	2:36% 3:47%	2.07 +/−1.59	14.20 +/− 1.64	2.34+/− .23
UB	46	M:28 F: 18	38.37 +/− 11.43	13.36 +/− 3.10	2:100%	IQ measured by nonverbal intelligence test, not included in analyses	0:100%	1:9% 2:35% 3:46% 4:11%	12.92 +/− 12.29	25.90 +/− 7.70	1.95 +/− .50	1.91 +/− .49
NSU	63	M: 46 F: 17	21.31 +/− 4.94	12.32 +/− 1.91			0:64% 2:33%	1:100%	2.17 +/− 3.58	19.13 +/− 4.62	2.29 +/−.39	1.37 +/− .43

Abbreviations: IQ= intelligent quotient, Std= standard deviation, CPZ= chlorpromazine equivalent dosage, Dataset: MRC - Maryland Psychiatric Research Center, CMH - Centre for Addiction and Mental Health, ZHH - Zucker Hillside Hospital, Biceps - Brain Imaging, Cognitive Enhancement, and Early Schizophrenia, BWH - Brigham and Women’s Hospital, Oxford - Warneford Hospital, Bsnip - Bipolar and Schizophrenia Network for Intermediate Phenotypes, AMC - Asan Medical Center, UB - University of Bern, NSU - North Shore University Medical Center

Multisite dMRI data underwent thorough quality control, and consistent preprocessing was employed across sites using our open-source software (https://github.com/pnlbwh/pnlutil). Next, our dMRI harmonization procedure was applied (https://github.com/pnlbwh/dMRIharmonization³³). The performance of our harmonization method has been well-validated in multiple studies^33-37. Please refer to the supplementary materials for the details of the harmonization approach (Supplementary Figure 1). After the harmonization procedure removed scanner-, site-, and sequence-specific effects from the dMRI signal while preserving individual anatomical variability (Supplementary Figures 2 and 3), we registered FA maps of all harmonized dMRI data to the Illinois Institute of Technology (IIT) Human Brain FA Atlas version 4.1. We decided on the IIT Human Brain Atlas, given that it combines several advantages over other templates, including high image sharpness, no visible artifacts, low noise level, and high spatial resolution³⁸. For each subject, average FA was estimated for the 14 largest WM tracts. Given that our previous studies on the same dataset did not show any effect of hemisphere^{34, 39}, we averaged FA values for left and right WM tracts. Additionally, we averaged FA of the cingulum bundle hippocampal and cingulate gyrus parts to get one cingulum bundle (CB) measure. Thus, the analyses for the present study were conducted on seven tracts: forceps minor, forceps major, CB, inferiorfronto-occipital fasciculus (IFOF), inferior-longitudinal fasciculus (ILF), superior-longitudinal fasciculus (SLF), and uncinate-fasciculus (UF) (Supplementary Figure 4).

Demographic and clinical data standardization

Given that the demographic and clinical data were collected retrospectively across multiple international studies, we determined the following criteria to standardize demographic and clinical information in the best possible manner. Standardization was carried out by the board-certified clinical psychologist J.W. and psychiatrist Ma.Ke., both with extensive clinical experience and numerous publications in psychometrics^{40, 41}.

The SCZ group included subjects diagnosed with schizophrenia, schizoaffective disorder, and schizophreniform disorder. For detailed information about diagnostic, inclusion-and exclusion criteria of the single studies, please see Supplementary Table 2. Handedness was determined as right, left, or both based on the Annett, Andreasen, or Edinburgh assessment methods^{42, 43}. We grouped race into either African American, Caucasian, or other/unknown based on the available information. IQ estimates were obtained using either a Wechsler Adult Intelligence Scale (WAIS) full-Scale IQ test, or a premorbid IQ reading test (Wide Range Achievement Test [WRAT], or the Wechsler Test of Adult Reading [WTAR]) (Table 1)^44-47. First, we investigated the combined premorbid and postmorbid IQ scores. For our secondary analyses, we restricted all IQ-related analyses to patients with premorbid IQ measurements only.

We derived Positive and Negative symptoms scores from either the Positive and Negative Symptom Scale (PANSS) or the Brief Psychiatric Rating Scale (BPRS)^48-50. After evaluating both the item and label definitions, we found seven positive symptoms (Delusions, Conceptual Disorganization, Hallucinations, Excitement, Grandiosity, Suspiciousness, and Hostility) and two negative symptoms (Blunted Affect and Emotional Withdrawal) to be equivalent in the PANSS and BPRS. Given different definitions of the individual scale scores of PANSS and BPRS, we did not use the original 7-point rating scales but transformed them into 4-point scales (0-no symptoms / 1-mild / 2-moderate / 3-severe). Finally, we averaged all seven positive symptoms to compose a positive symptom index for each patient. The same was done for the two negative items to compose a negative symptom index.

To determine the influence of medication on WM, we calculated the chlorpromazine equivalent dosage (CPZ) on the day of scanning, utilizing previously established conversions (Supplementary Table 3). Given that CPZ only approximates medication dosage, we opted to use a categorical CPZ approach introduced by Sohler et al.⁵¹ and separated patients into four groups (CPZ=0, CPZ<300mg/day, CPZ 300-1000mg/day, CPZ >1000mg/day).

Statistical analyses

Statistical analyses were performed using SPSS and Amos Version 24.

Based on the observations from our previous publication^{34, 39}, and multiple other lifespan studies^52-54, we generated FA trajectories from the HC data by conducting separate quadratic regressions for each tract, with age and age squared as the independent variables and FA of each tract as the dependent variable. We utilized the parameters from the model that was derived from the HC population only to predict FA for each tract of each patient with SCZ. Next, we calculated the difference between the predicted and real FA value for each tract and each patient with SCZ. We termed this difference FA deviation and used it for all further analyses. We utilized FA deviations instead of absolute FA values for several reasons. This approach allowed us to first regress out the effect of healthy aging on FA. Second, we were able to investigate the influence of demographic and clinical variables on WM pathologies in patients independently of the effects of healthy aging. Please note that larger FA deviations present lower absolute FA values in patients compared to age-matched HC and, consequently, more severe WM impairments.

All analyses were Bonferroni-corrected for multiple comparisons across seven tracts (p< 0.007), and two-tailed tests were utilized for all group comparisons.

Age and duration-of-illness

We used regression and mediator analyses to explore if age and duration-of-illness influenced the FA deviations in patients. First, we conducted regression analyses separately with either age or duration-of-illness as the independent variables, sex as a covariate, and the FA deviations as the dependent variable. Next, age and duration-of-illness were added together as independent variables to determine if one mediates the other’s influence on WM. Once more, sex was added as a covariate. Finally, mediator models (implemented via the mediator module PROCESS, Model 4⁵⁵ in SPSS ) were utilized to test for A) a direct effect of age on FA deviations and B) an indirect effect of age (via the mediator duration-of-illness) on FA deviations of each tract. Sex was included as a covariate for all intermediate outcome variables (age to duration-of-illness, duration-of-illness to FA deviation, and age to FA deviation). PROCESS automatically provides a bootstrap confidence interval for every indirect effect. The bootstrap confidence interval is constructed by randomly resampling n cases from the data with replacement and estimating the bootstrap sample’s indirect effects. After repeating this resampling procedure thousands of times, an empirical representation of the sampling distribution is built. A 95% confidence interval for the indirect effect above or below zero supports the claim for mediation ⁵⁶.

Please note that our decision to utilize FA deviations instead of raw FA values was motivated by our desire to investigate the influence of duration-of-illness and pathological aging in patients that would be separately of the well-known effect age has on healthy individuals’ WM structure. However, to compare our analyses with previous studies, we repeated the regression analyses with absolute FA values and present these results in the supplementary materials.

Medication

To further understand the effect of medication on WM abnormalities, we included the CPZ group as an independent variable into the regression model with the FA deviation of each tract as the dependent variable and sex and duration-of-illness as covariates.

Symptom severity and cognitive functioning

Several smaller studies explored the relationship between symptoms and WM structure^{10, 57-59}, focusing on specific symptoms, tracts, and populations. However, results have been relatively inconsistent, and the existing evidence has not been convincing enough to demonstrate that symptoms are associated with WM tracts. The heterogeneity in populations (e.g., different duration-of-illness, symptom profiles) is one reason for the limited cross-study replication. Therefore, we opted to follow a different approach to investigate the association between WM and symptom severity in psychosis. Leveraging our well-powered sample, we aimed to test if WM structure, when estimated across the whole SCZ spectrum, is associated with symptom severity. We postulated an SEM for the subset of patients with complete clinical data (n=362). SEM combines factor analysis and multiple regression analysis to study relationships between measured and latent variables. Given that we did not measure a variable of overall structural impairment or symptom severity, we postulated both as latent variables. The FA deviations of each of the seven tracts were reflective indicators for the latent variable structural impairment. Positive and negative symptom indices were reflective indicators for symptom severity. We applied these latent variables in regression analyses to test the structural impairment’s effect on symptom severity. Next, we added IQ to the SEM to test whether it influences structural impairments and symptom severity. To account for the non-normal distribution of the data, we calculated all data paths with an asymptotically distribution-free model.

Sex

We repeated all regression analyses for males and females separately. Furthermore, sex was included as a variable in the SEM to test if 1) the overall model fit improves, 2) if the proposed association between structural impairments, symptom severity, and IQ differs between sexes.

Results

For demographic information, see Table 1.

Age and duration-of-illness

Regression analyses with FA deviation as the dependent variable, sex as a covariate, and duration-of-illness as the independent variable revealed a significant influence of duration-of-illness on the forceps major (standardized regression coefficient [B]=.14, T=3.24, p<.001), forceps minor (B=.15, T=3.40, p<.001), IFOF (B=.14, T=3.17, p<.002), and SLF (B=.16, T=3.83, p<.001). For all tracts, a longer duration-of-illness was associated with higher FA deviations. The same regression using age as the independent variable showed significant effects for the forceps major (B=.12, T=2.87, p<.004) and SLF (B=.14, T=3.41, p<.001). Here, older ages were associated with higher FA deviations. After including age and duration-of-illness in the model, only the effect of duration-of-illness remained significant (forceps minor B=.21, T=2.83, p<.005; ILF B=.29, T=3.90, p<.001), again indicating that a longer-duration-of-illness was associated with more severe WM abnormalities.

Mediator analyses demonstrated an indirect effect (mediated through duration-of-illness) of age on FA deviations of the forceps minor, IFOF, and ILF (Figure 1). These findings suggest that older ages are related to more pronounced FA abnormalities, and a longer duration-of-illness mediates this effect of age on WM abnormalities.

Figure 1: Association between white matter, age, and duration-of-illness.

We conducted regression analyses to study the influence of age and duration-of-illness on white matter (WM) abnormalities. Fractional anisotropy (FA) deviations of each tract were the dependent variables; sex was the covariate. Age (A) or duration-of-illness (B) or both (C) were entered as independent variables. We observed an effect of age and duration-of-illness on several WM tracts. Significant associations are highlighted in green. Mediator analyses (D) demonstrated that the duration-of-illness is mediating the effect of age on FA. D I) shows the direct effect of age on FA deviations (p<.007 indicates statistical significance). D II) shows the indirect effect of age on FA deviations (mediated through duration-of-illness. Bootstrap intervals are shown in [], statistical significance is given if 0 is not included in the bootstrap interval.

* indicates statistically significant associations, e 1–11 represent the residual/error variances of the endogenous variables.

Abbrevations: inferior-fronto-occipital fasciculus (IFOF), inferior-longitudinal fasciculus (ILF), superiorlongitudinal fasciculus (SLF), uncinate fasciculus (UF), intelligence quotient (IQ).

Given the association between duration-of-illness and age-of-onset, we repeated our analyses, including age-of-onset, as an additional variable of interest. Last, we calculated all regression analyses utilizing absolute FA values instead of FA deviations to compare our results to previous studies. The additional analyses are presented in Supplementary Table 4. The supplementary results highlight that when using absolute FA values instead of FA deviations, duration-of-illness is still associated with FA of the forceps, IFOF, ILF, and SLF (even when correcting for age or age-of-onset).

Medication

The inclusion of the CPZ group to the regression model with FA deviation of each tract as the dependent variable, sex and duration-of-illness as covariates, and CPZ group as the independent variable revealed a significant effect of CPZ on the forceps major (B=.17, T=3.67, p<.001) (Table 2). Patients taking more medication showed a higher FA deviation, hence more prominent WM pathology within the largest inter-hemispheric WM connection.

Table 2:

Influence of CPZ on FA deviations

	CPZ	Duration-of-illness	Sex
Forceps major	B=.17, T= 3.67, p<.001	B=.10, T=2.07, p<.039	B=.14, T=3.23, p<.001
Forceps minor	B=−.07, T=−1.56, p<.119	B=.18, T=3.82, p<.001	B=.14, T=3.12, p<.002
Cingulate	B=−.03, T=−.54, p<.592	B=.00, T=.014, p<.989	B=.21, T=4.54, p<.001
IFOF	B=.05, T=.94, p<.348	B=.13, T=2.76, p<.006	B=.06, T=1.33, p<.185
ILF	B=−.08, T=−.1.58, p<.114	B=.12, T=2.40, p<.017	B=.02, T=.53, p<.595
SLF	B=−.04, T=−.83, p<.404	B=.17, T=3.54, p<.001	B=.15, T=3.29, p<.001
UF	B=−.00, T=−.12, p<.906	B=−.00, T=−.03, p<.979	B=.12, T=2.53, p<.012

Abbreviations: B= standardized regression coefficient, fractional anisotropy (FA), inferior-fronto-occipital fasciculus (IFOF), inferior-longitudinal fasciculus (ILF), superior-longitudinal fasciculus (SLF) and uncinate fasciculus (UF), chlorpromazine equivalent dosage in mg (CPZ), grouped into four: CPZ=0, CPZ<300mg/day, CPZ 300-1000mg/day, CPZ >1000mg/day; tracts remained significant after Bonferroni-correction (p<0.007) are marked in blue.

Symptom severity and cognitive functioning

The SEM model represented the influence of the latent variable of overall structural impairment on the latent variable of overall symptom severity. The model fit showed good overall fit as indicated by the goodness of fit index (GFI = 0.92) and adjusted goodness of fit index (AGFI = 0.86). While booth indices determine the fit between the data and the hypothesized model, the AGFI considers the number of indicators of each latent variable. Both indices range between 0 and 1, where .90 indicates a good model fit.

In line with our hypothesis of the latent variable of structural impairment, we observed a strong association between the FA deviations of all tracts and the postulated latent variable (B between 0.57 and 0.90, Figure 2). The positive symptom index was more impactful for the latent variable of symptom severity (B= 0.54) than the negative symptom index (B = 0.30). As postulated, structural impairment had a moderate influence on symptom severity (B = 0.42), suggesting that more severe WM pathology in patients leads to more severe symptoms. Last, IQ was significantly associated with structural impairments (B=−.26), showing that patients with lower IQ displayed more severe WM pathologies. The negative association between IQ and symptom severity did not reach significance (B=−.12) for patients (Figure 2). Restricting analyses to patients with premorbid IQ measurements did not have any effect on the results (Supplementary Table 5).

Figure 2: Association between white matter, symptom severity, and IQ (structural equation model).

We calculated a structural equation model (SEM) for the subset of patients with complete clinical data (n=362) to test if white matter abnormalities (WM) are related to symptom severity and if IQ influences WM and symptom severity. The fractional anisotropy (FA) deviation of each of the seven tracts were reflective indicators for the latent variable structural impairment. Positive and negative symptom indices were reflective indicators for symptom severity Figure 2 shows the standardized regression weights [B] between all reflective indicators and the latent variables and between the two latent variables and IQ. The results further demonstrate that: a) all FA deviations significantly contribute to the latent variable structural impairment (B between .57 and .90), b) positive symptoms contribute more (B=.58) to the latent variable symptom severity than negative symptoms (B=.30), c) structural impairments are positively related to symptom severity (B =.42), and d) IQ has a significant negative effect on structural impairments (B=−.26).

Abbreviations: B= standardized regression coefficient, fractional anisotropy (FA), inferiorfronto-occipital fasciculus (IFOF), inferior-longitudinal fasciculus (ILF), superior-longitudinal fasciculus (SLF) and uncinate fasciculus (UF).

Sex

When adding sex as a covariate in the regression model (independent variable: duration-of-illness, dependent variable: FA deviations), a significant effect of sex was observed on forceps major (B =.15, T=3.41, p<.001), forceps minor (B =.15, T=3.46, p<.001), CB (B =.19, T=4.32, p<.001), and SLF (B =.16, T=3.67, p<.001). Male patients displayed higher absolute FA values, and female patients higher FA deviations.

When analyzing males and females separately, the association of duration-of-illness and FA deviations remained significant for the males only (forceps major B =.19, T=3.55, p<.001; forceps minor B = .15, T=2.75, p<.006; IFOF B = .15, T=2.76, p<.006; SLF B =.20, T=3.61, p<.001) (Table 3). While a longer duration-of-illness was significantly associated with more severe WM abnormalities in males, the same was not true for females.

Table 3:

Influence of duration-of-illness for males and females separately on FA deviations

	Duration-of-illness males	Duration-of-illness females
Forceps major	B=.19, T=3.55, p<.001	B=.06, T=.85, p<.396
Forceps minor	B=.15, T=2.75, p<.006	B=.14, T=1.99, p<.048
Cingulate	B=−.02, T=−.41, p<.685	B=.03, T=.38, p<.702
IFOF	B=.15, T=2.76, p<.006	B=.11, T=1.59, p<.113
ILF	B=.09, T=1.70, p<.090	B=.09, T=1.37, p<.172
SLF	B=.20, T=3.61, p<.001	B=.11, T=1.59, p<.114
UF	B=.03, T=.53, p<.599	B=−.08, T=−1.05, p<.295

Abbreviations: B= standardized regression coefficient, fractional anisotropy (FA), inferior-fronto-occipital fasciculus (IFOF), inferior-longitudinal fasciculus (ILF), superior-longitudinal fasciculus (SLF) and uncinate fasciculus (UF); tracts remained significant after Bonferroni-correction (p<0.007) are marked in blue.

The results of the regression analyses with the CPZ group as the independent variable were not affected when investigating males and females separately. Both sexes displayed an association between higher medication dosage and higher FA deviations in the forceps major.

Finally, when separating males from females, the SEM model fit improved (GFI=0.92, AGFI=0.87). Interestingly, the influence of structural impairment on symptom severity was more significant for females than for males (females = 0.48, males = 0.33). These findings suggest that the association between WM pathology and symptom severity is stronger in females than in males. Additionally, only the females showed a significant effect of IQ on structural impairment (B = −0.44, males = −0.17) and symptom severity (B females =−0.52, males =−0.10) (Supplementary Table 6). Lower IQ in females was associated with both a greater degree of WM pathology and more severe symptoms.

Discussion

In the present study, the large sample size of harmonized dMRI data and standardized clinical data enabled us to investigate each patient’s deviation from healthy WM trajectories, instead of looking at absolute FA values. Our results revealed several relationships between WM abnormalities and demographic and clinical variables in SCZ. Specifically, we found that a more severe deviation from healthy WM is related to a longer duration-of-illness. This finding was independent of the physiological age and more prominent in males than in females. Next, we reported that more structural impairments are associated with more severe symptoms, with IQ modulating this association. Structural impairments, symptom severity, and IQ were thereby more robustly associated with one another in females than in males.

Age and duration-of-illness

We showed that the deviation from the healthy WM trajectories of several tracts became more pronounced with a longer duration-of-illness. While some smaller studies had reported an association between WM structure and duration-of-illness^{18-22, 60, 61}, others had not^{5, 18}. Conflicting results may be due to several factors, including a lack of power in small sample sizes, the confounding influence of healthy aging, and region-specific effects of duration-of-illness. We observed an effect of duration-of-illness only for interhemispheric and association, but not limbic tracts. This finding coheres with previous gray matter SCZ studies reporting no effect of duration-of-illness on limbic circuitries⁶². Our results also align with healthy aging studies, which report relative GM preservation in limbic areas^{63, 64}.

Medication

We observed that higher medication dose is associated with more severe corpus callosum abnormalities. Our finding is in line with several studies that report a medication effect only limited to a specific region⁶⁵. However, it is worth mentioning that the results regarding the impact of antipsychotic medication on brain structure are heterogeneous^{23-26, 66-71}. One explanation is that most studies, including ours, use normalized CPZ medication dose at the scan time. CPZ dosage at the scan time may not be ideal for evaluating the influence of lifetime medication exposure or distinguishing the effects of different antipsychotics⁷². Future studies that control for dose-years⁷³, lifetime antipsychotic dosage, and consider different types of medication are needed.

Symptom severity and cognitive functioning

While many previous studies reported an association between particular WM tracts and specific symptoms^{12, 27-29, 74-77}, we demonstrated a relationship between general structural impairment and symptom severity. Our results show that more WM abnormalities are associated with more severe symptoms (irrespective of age and region). Specifically, we observed that all WM tracts were associated with symptom severity. However, in line with some other studies, we found a more consistent association between brain structure and positive symptoms than between brain structure and negative symptoms^{12, 27-29, 74-77}.

We were also interested in exploring the potential role of IQ for structural impairment and symptom severity. IQ impairments are commonly seen in patients with SCZ before illness onset. Additionally, IQ is a good predictor of conversion and long-term clinical outcome^{7, 8, 64, 16, 78-80}. More specifically, studies have found that individuals with higher premorbid IQ may be more resilient to the illness than those with lower premorbid IQ⁸¹. Furthermore, a recent study reported that premorbid intelligence represents an important modifier between age-related loss of gray matter and cognitive changes in SCZ⁸². The authors suggested that premorbid IQ might be an indicator of cognitive reserve and that including the concept of cognitive reserve in future studies could explain inconsistencies in brain structure - cognition findings in the extant literature. Our approach of studying the impact of IQ on the structure-symptom relationship allowed us to extend this previous work. Indeed, we observed that IQ modulates both WM impairments and symptom severity in patients with SCZ. Future studies that include other indicators of cognitive reserve (e.g., education-occupational level and leisure activities) are needed, to better understand the potential role of cognitive reserve for SCZ^83-85.

Taken together, our finding of an association between generalized WM abnormalities, symptom severity, and IQ is supportive of the disconnection hypothesis⁸⁶. This prominent biological hypothesis states that SCZ is characterized by altered communication within, and abnormal integration between, brain systems. This abnormal communication, in turn, causes false inferences of predicted and sensory inputs^87-89. Supporting evidence for this hypothesis came from functional and PET imaging studies that reported widespread abnormal functional connectivity between different brain systems while resting and performing various tasks^90-93. Our findings add to the functional connectivity literature in further characterizing dysconnectivity as the core pathology in SCZ.

Sex

We observed a significant effect of sex on several tracts, including the corpus callosum, CB, and SLF. Of note, FA of limbic tracts, such as the CB, are also the most influenced by sex in healthy populations^{14, 15}. This regionally specific effect of sex is not surprising given that receptors for the sex-specific hormones, such as estrogens or testosterone, are predominantly expressed in limbic areas^{94, 95}, and that changes in sex hormones in these areas correlate with FA changes⁹⁶. Specifically, estrogen influences WM development and degeneration, and testosterone promotes myelination^{97, 98}.

Additionally, we demonstrated the influence of sex on the association of WM abnormalities, demographic, and clinical variables. We reported that only males showed a relationship between a longer duration-of-illness and more severe WM abnormalities. Clinical studies observed a similar pattern, reporting a more potent effect of duration-of-illness on males’ clinical outcomes than females’⁹⁹. We also found that for females, IQ had a more significant influence on WM abnormalities and symptom severity than for males. Previous work had shown that IQ is more strongly correlated with WM in healthy females than in males¹⁰⁰. Future studies need to establish why IQ is especially meaningful for female patients, and if that knowledge can be utilized to develop sex-specific psychotherapeutic treatment strategies.

Limitations and future directions

We acknowledge several limitations in the current study. Because of our study’s retrospective nature, not all relevant demographic and clinical information was available for all subjects. Additionally, clinical instruments across different international sites were quite diverse. We standardized the demographic and clinical data, following a few existing examples found in the literature. Until, however, such standardization methods are validated, our study presents only the first necessary step towards implementing demographic and clinical data harmonization standards. Next, several additional measures that likely influence the reported associations were not available. Those include the duration of untreated psychosis, substance abuse, lack of social communication, or comorbid medical disorders. Additionally, while we calculated CPZ dose at scan time, information about cumulative medication exposure was not available. As highlighted above, we think future studies must collect more detailed medication data to understand WM and medication’s complex association better. Finally, the present study is cross-sectional. While harmonization of data enabled us to model lifespan trajectories and SEM allowed for the assumption of causal relationships, one must be cautious when claiming causality using cross-sectional data. Using SEM, causal relationships between latent variables can be postulated based on previous knowledge. Hence, we hypothesized that WM impairments in SCZ could predict symptom severity. However, SEM can only indicate the plausibility of an assumed model, and to further test the causal assumption that WM impairments predict symptom severity, longitudinal studies are needed. Last, to expand the biological understanding of WM changes, these future studies should further investigate the other diffusion measures of WM (such as axial and radial diffusivity, or more advanced measures, such as NODDI or Free Water Imaging^{101, 102}).

Conclusion

Our work presents a well-powered investigation of the relationship between WM abnormalities, demographic, and clinical variables in patients with SCZ.

We observe sex and region-specific effects of duration-of-illness on WM abnormalities that indicate a progressive pathology in the WM of patients with SCZ. We also report a link between structural impairments and symptom severity, which is modulated by IQ. Interestingly, the associations between structural impairments, symptoms, and IQ are more pronounced in females. Our study provides a significant step towards understanding the interaction between demographic and clinical variables and WM pathologies in SCZ. It is crucial to understand, which pathologies are inherent to SCZ and independent of clinical trajectories, and which are dependent on demographic and clinical variables. Future studies can then be guided regarding which demographic and clinical information should be collected to characterize SCZ best. Second, and more importantly, this will increase our understanding of the vulnerability of patients’ subgroups to the pathophysiology of SCZ. Specifically, our analysis of sex-specific vulnerability to WM pathology will provide the first step towards sex-specific monitoring and treatment strategies. Finally, future longitudinal studies are needed to further explore which variables influence and interact with individual patients’ SCZ pathophysiology.