Estimating multimodal brain variability in schizophrenia spectrum disorders: A worldwide ENIGMA study

Abstract

Schizophrenia is a multifaceted disorder and associated with structural brain heterogeneity. Recent research s that underscored that profound understanding of structural brain heterogeneity is relevant to identify illness subtypes as well as informative biomarkers. However, our understanding of structural heterogeneity in schizophrenia is still limited. This comprehensive meta-analysis therefore investigated and compared the variability of multimodal structural brain measures for white and gray matter in individuals with schizophrenia and healthy controls. Using the ENIGMA dataset of MRI-based brain measures from 22 sites, we examined variability in cortical thickness, surface area, folding index, subcortical volume and fractional anisotropy, both at regional and global level. At the regional level, we found that schizophrenia patients are distinguished by higher heterogeneity in the frontotemporal network with regard to multimodal structural measures. Multimodal heterogeneity in these regions potentially implies different subtypes that share impaired frontotemporal interaction as core feature of schizophrenia. At the global level, the Person-Based Similarity Index (PBSI) analysis surprisingly revealed that schizophrenia patients are distinguished by a significantly higher homogeneity of the folding index, implying that certain gyrification attributes represent a uniform aspect of schizophrenia across subtypes. These findings underscore the importance of studying structural brain variability for a more holistic understanding of schizophrenia’s neurobiology, potentially facilitating the identification of illness subtypes and informative biomarkers. These results could guide future investigations and tailor precision medicine approaches for schizophrenia.

Introduction

Schizophrenia is characterized by an array of biological and symptomatic heterogeneity that remains incompletely understood^1–3. Notably, the biological heterogeneity of schizophrenia is reflected in structural irregularities of the brain^4–8 beside functional abnormalities^{9, 10}. Historically, neuroimaging studies including meta-analyses have predominantly focused on evaluating mean differences in structural brain measures between psychosis patients and healthy controls^11–13. However, recent studies also emphasized that psychosis patients or individuals at clinical high risk for psychosis exhibit different variability of structural brain measures in comparison to healthy controls^7,14–18: While higher variability of selective brain structures is potentially associated with schizophrenia subtypes, lower variability argues for a more stable feature of the disorder that is shared across illness subtypes⁷. Therefore, a shift in perspective towards the analysis of variability, above and beyond mean differences, promotes a more nuanced neurobiological understanding of schizophrenia, with clinical relevance for the identification of illness subtypes and biomarkers as well as for the analysis of medication effects^19–21. In addition to regional structural measures, global measures such as the Person-Based-Similarity-Index (PBSI)^{15, 22} have been applied for example to reveal higher divergence of general cortical thickness, surface area and subcortical volume in individuals at high risk for psychosis¹⁷. Against this backdrop, our meta-analysis study investigated a variety of MRI-based brain measures, both for gray and white matter structures and both at global and regional levels, to enable a comprehensive, multimodal assessment of neuroanatomical heterogeneity in schizophrenia. To attenuate limitations of individual studies with small sample size and to support the replicability of our findings, multimodal variability was explored using Meta- and Mega-analysis of a large dataset of the Enhancing Neuroimaging Genetics through Meta-Analysis (ENIGMA) consortium^{13, 23}.

Methods

Participants

The ENIGMA dataset comprised neuroimaging data from 22 sites, including 15 sites with individual-level data and 7 sites with group-level data. Sites with individual-level data provided cortical, subcortical and white matter measures for individual schizophrenia patients (SZ) and healthy controls (HC), while sites with group-level data provided mean and standard deviation of respective measures across SZ and HC. The dataset included 6138, 6138, 5701, 6139 and 2636 individuals for cortical thickness (CT), cortical surface area (SA), cortical folding index (FI), subcortical volume (SV) and the white matter measure fractional anisotropy (FA), respectively. Prior to data collection, each site obtained ethics committee approval, and participants provided informed consent or assent prior to participation.

MRI processing

Neuroimaging data were processed according to the standardised ENIGMA protocol(http://enigma.ini.usc.edu/protocols/imaging-protocols/). Motion correction, automated Talairach transformation, skull stripping, segmentation of the subcortical white and gray matter volumetric structures as well as intensity normalisation were incorporated by automated Freesurfer pipelines^24–31. Preprocessing procedures generated values of fractional anisotropy (FA) for white matter labels, values of subcortical volumes (SV) as well as values of cortical thickness (CT), surface area (SA) and folding index (FI) for cortical regions according to the Desikan-Killiany atlas³². As part of the ENIGMA quality control procedure, outliers (±2 SD from the mean) were identified and all images were visually inspected to remove poorly segmented regions. These quality control procedures led to minor fluctuation in sample size for each region of interest (ROI). The application of this protocol finally yielded 68 cortical ROIS, 18 subcortical ROIS and 63 DTI ROIS. Subjects younger than 18 years or older than 65 years were discarded from further analysis.

Statistical analysis

Variability ratio:

Statistical analyses were performed in R version 4.2.0 and Python version 3.9.16. Variability ratio (VR) was computed and visualized as forest plots in R using the packages metafor and tidyverse. In order to compare variability in regional neuroanatomical measures between schizophrenia patients and healthy controls, the natural logarithm of VR was calculated for each region of interest (roi) according to the following formula:

$$ln\left(\frac{{\sigma }_{SZ}}{{\sigma }_{HC}}\right)=ln\left(\frac{S{D}_{SZ}}{S{D}_{HC}}\right)+\frac{1}{2\left(n_{SZ}-1\right)}-\frac{1}{2\left(n_{HC}-1\right)}$$

where σ is the unbiased estimate of the standard deviation, SD the reported sample standard deviation and n the sample size for the respective group. The results of the variability ratio analysis were summarized in forest plots which exhibit the VR with 95% confidence intervals across all rois for a given measure (CT, SA, FI, FA and SV). For the forest plots, In(VR) values were back-transformed into a linear scale (VR) to facilitate interpretation: VR of 1 indicates equal variability in neuroanatomical measures between schizophrenia patients and healthy controls, while VR > 1 indicates greater and VR < 1 lower variability in schizophrenia patients, respectively. Brain surface plots in which the variability ratio for CT, SA, FI, FA and SV is illustrated with color-coded z-values were generated in Python, using the libraries mayavi, nibabel, surfer and matplotlib.

Coefficient of variation ratio:

In biological systems, variance usually scales with mean such that larger mean values are coupled with larger variance³³. For a given structure and measure, a between-group difference in variability could be therefore driven by a between-group difference in mean. For each measure, we therefore also investigated how mean and standard deviation (SD) differences between patients and controls vary across sites. While the majority of cortical, subcortical and white matter regions showed correlations between mean and SD differences close to zero, some regions showed moderate positive correlations (Supplementary Fig. 1). Even though there were also very few regions with negative correlations between mean and SD differences, the magnitude of these negative correlations were comparably low (Supplementary Fig. 2). To assess a potential impact of correlations between mean and SD on our variability ratio results, we additionally computed the coefficient of variation ratio (CVR) which incorporates between-group differences in mean to quantify between-group differences in variability⁷:

$$ln\left(\frac{\frac{{\sigma }_{Sz}}{{\bar{x}}_{Sz}}}{\frac{{\sigma }_{HC}}{{\bar{x}}_{HC}}}\right)=ln\left(\frac{\frac{S_{Sz}}{{\bar{x}}_{Sz}}}{\frac{S{D}_{HC}}{{\bar{x}}_{HC}}}\right)+\frac{1}{2\left(n_{Sz}-1\right)}-\frac{1}{2\left(n_{HC}-1\right)}$$

where $\bar{x}$ is the mean for the respective group. As for VR, the CVR analysis was illustrated in forest plots, ln (CVR) values in the forest plots were back-transformed into a linear scale and brain surface plots with color-coded CVR-z-values were generated.

Person-based similarity index:

In addition to the region-specific VR and CVR analyses, we calculated the person-based similarity index (PBSI) to explore between-group differences of variability for a given measure across brain regions^{16, 22}. Separately for the brain measures CT, SA, FI, FA and SV, the PBSI was computed for the patient and control group according to the following formula:

$$\mathit{\text{PBS}}{I}_i=\frac{1}{N-1}\sum _{i\ne j}\mathit{\text{cor}}\left(y_{i’}{y}_j\right)$$

For j≠i, the PBSI of the ith participant is the mean Spearman correlation between his brain measure y_i and the brain measure y_j of any other participant of the same group with N individuals. For a given participant i, y_i represents a vector in which all of his/her values for a given measure (e.g. CT) are concatenated. With regard to a given brain measure, the PBSI score of each participant therefore indicates his/her similarity to all other members of the same group (PBSI scores closer to 1 indicate higher similarity). The PBSI scores for CT, SA, FI, SV and FA were then compared between schizophrenia patients and controls using a Mann-Whitney U test with Bonferroni correction.

Results

Cortical thickness

Compared with controls, the schizophrenia group exhibited greater variability in bilateral inferior, middle and superior temporal regions as well as in bilateral superiorfrontal regions. From these frontotemporal regions, the largest effect sizes were found in the right middle (VR = 1.19, 95% CI: 1.08 – 1.30) and left superior temporal areas (VR = 1.16, 95% CI: 1.10 – 1.22). Unilaterally, the schizophrenia group demonstrated greater variability in the left supramarginal region (VR = 1.11, 95% CI: 1.03 – 1.19), in the right pars orbitalis region (VR = 1.07, 95% CI: 1.01 – 1.13), in the left fusiform region (VR = 1.09, 95% CI: 1.01 – 1.17) as well as in the right precentral region (VR = 1.10, 95% CI: 1.01 – 1.21). In comparison to schizophrenia patients, healthy controls did not show significantly higher variability in cortical thickness for any cortical region (Fig. 1). When the coefficient of variation ratio was computed instead of the variability ratio, all regions with higher heterogeneity in schizophrenia were preserved, and there was still no region with higher heterogeneity in controls (Supplementary Fig. 3).

Figure 1: Variability ratio of cortical thickness.

Lower panel: Variability ratio (VR) effect sizes for schizophrenia patients (SZ) vs. healthy controls (HC) are shown for different cortical regions and on a linear scale, statistically controlling for age and gender. Upper panel: VR effect sizes are projected onto the brain surface and color-coded as z-values. In these cortical maps, areas with higher VRs in patients appear in gradients of red, and areas with higher VRs in controls appear in gradients of blue. CI: Confidence Interval. Independent of the applied variability measure (variability ratio vs. coefficient of variation ratio), higher heterogeneity in schizophrenia was observed for the following regions: Bilateral superior, middle and inferior temporal region, bilateral superior frontal region, left supramarginal and fusiform region as well as right pars orbitalis and precentral region.

Cortical surface area

Compared with controls, we found greater variability in the schizophrenia group particularly in the following regions: Right superiortemporal region (VR = 1.12, 95% CI: 1.02 – 1.23), cortical areas around the right superior temporal sulcus (VR = 1.11, 95% CI: 1.03 – 1.20) and right supramarginal region (VR = 1.11, 95% CI: 1.04 – 1.18). Moreover, greater variability in the schizophrenia group was found in the left postcentral region (VR = 1.09, 95% CI: 1.01 – 1.17), in the right medial orbitofrontal region (VR = 1.09, 95% CI: 1.02 – 1.15), in the left superiorparietal region (VR = 1.07, 95% CI: 1.01 – 1.14), in the right middle temporal region (VR = 1.09, 95% CI: 1.01 – 1.17) as well as in the left precuneus (VR = 1.06, 95% CI: 1.01 – 1.11) and right lingual region (VR = 1.10, 95% CI: 1.01 – 1.19). Bilaterally, the superior frontal and transverse temporal areas showed higher variability in comparison to healthy controls (left superior frontal region: VR = 1.07, 95% CI: 1.01 – 1.12; right superior frontal region: VR = 1.06, 95% CI: 1.01 – 1.12; left transverse temporal region: VR = 1.07, 95% CI: 1.02 – 1.13; right transverse temporal region: VR = 1.08, 95% CI: 1.01 – 1.16 (Fig. 2). When compared to the schizophrenia group, the control group did not exhibit higher variability of the cortical surface area in any cortical region (Fig. 2). When we computed the coefficient of variation instead of the variability ratio, all regions with higher variability in schizophrenia were preserved, and there was still no region with higher variability in controls (Supplementary Fig. 4).

Figure 2: Variability ratio for cortical surface area.

Same conventions as for Fig. 1. Higher heterogeneity in schizophrenia was observed for the following regions, irrespective of the deployed variability measure: Bilateral superiorfrontal and transverse temporal regions, cortical areas around the right superior temporal sulcus, right superiortemporal region, right supramarginal region, left postcentral region, right medialorbitofrontal region, left superiorparietal region, left precuneus, right middle temporal region and right lingual region.

Cortical folding index

The schizophrenia group exhibited higher heterogeneity than controls in the right inferiortemporal region (VR = 1.33, 95% CI: 1.10 – 1.60) and lower heterogeneity in the left parahippocampal region (VR = 0.72, 95% CI: 0.56 – 0.91, Fig. 3). Higher heterogeneity of the right inferiortemporal region as well as higher homogeneity in the left parahippocampal region in schizophrenia was preserved when we applied the coefficient of variation ratio instead of the variability ratio (Supplementary Fig. 5). Of all gray matter measures, only the cortical folding index showed higher regional homogeneity in addition to higher regional heterogeneity when schizophrenia patients are compared to healthy controls.

Figure 3: Variability ratio for cortical folding index.

Same conventions as for Fig. 1. Irrespective of the applied variability measure, schizophrenia patients exhibited higher heterogeneity in the right inferiortemporal region, but lower heterogeneity in the left parahippocampal area.

Subcortical volume

Bilaterally, the variability of the lateral ventricle and of the inferior lateral ventricle was distinctly higher in the schizophrenia group than in the control group. From these ventricles, the highest effect size was found in the left inferior lateral ventricle (VR = 1.43, 95% CI: 1.26 – 1.61, Fig. 4). Bilaterally, we additionally found higher variability in pallidum and putamen, and unilaterally, the left nucleus accumbens, the left caudate and the right hippocampus showed greater heterogeneity in the schizophrenia group. From these brain regions, the highest effect sizes were found in the left nucleus accumbens (VR = 1.13, 95% CI: 1.04 – 1.22) and in the left pallidum (VR = 1.10, 95% CI: 1.04 – 1.17, Fig. 4). No subcortical region showed significantly higher heterogeneity in the control group (Fig. 4). When the coefficient of variation ratio was computed instead of the variability ratio, higher heterogeneity in schizophrenia was confirmed for all regions except for bilateral pallidum and putamen (Supplementary Fig. 6). As for the variability ratio, no subcortical region exhibited higher heterogeneity in controls when the coefficient of variation ratio was applied (Supplementary Fig. 6).

Figure 4: Variability ratio for subcortical volume.

Same conventions as for Fig. 1. Lower panel: Variability ratio (VR) effect sizes are shown for subcortical volumes. Upper panel: VR effect sizes are projected onto respective subcortical structures and color-coded as z-values. In these subcortical maps, areas with higher VRs in patients appear in gradients of red, and areas with higher VRs in controls appear in gradients of blue. Independent of the applied variability measure, schizophrenia patients showed higher heterogeneity for lateral ventricles, left nucleus accumbens, left caudate and right hippocampus.

Fractional anisotropy

Variability of the superior fronto-occipital fasciculus was greater in the control group in comparison to the schizophrenia group, both on the left (VR = 0.86, 95% CI: 0.78 – 0.94) and right side (VR = 0.85, 95% CI: 0.75 – 0.95, Fig. 5). In contrast, no white matter tract showed greater heterogeneity in the schizophrenia group when the variability ratio was used as variability measure. When the coefficient of variation was computed instead of the variability ratio greater bilateral heterogeneity of the superior fronto-occipital fasciculus in the control group was preserved (Supplementary Fig. 7), and several regions including the left uncinate fasciculus exhibited greater heterogeneity in the schizophrenia group (CVR = 1.16, 95% CI: 1.01 – 1.33, Supplementary Fig. 7).

Figure 5: Variability ratio for fractional anisotropy.

Same conventions as for Fig. 1. Lower panel: Variability ratio (VR) effect sizes are shown for fractional anisotropy (FA) of white matter structures. Upper panel: VR effect sizes are projected onto respective white matter structures and color-coded as z-values. In these maps, areas with higher VRs in patients appear in gradients of red, and areas with higher VRs in controls appear in gradients of blue. Irrespective of the deployed variability measure, higher homogeneity in schizophrenia was observed for the superior fronto-occipital fasciculus in both hemispheres.

Person-Based Similarity Index

At a global level, the Person-Based Similarity Index (PBSI) analysis revealed that heterogeneity of cortical thickness is similar for schizophrenia patients and healthy controls (p = 4.89e-01, Mann-Whitney U test, Bonferroni-corrected). However, schizophrenia patients showed less heterogeneity in cortical surface area, subcortical volume and especially in the Folding Index when PSBI scores are compared to healthy controls (Folding Index: p = 2.12e-16; Surface Area: p = 4.97e-06; Subcortical Volume: p = 8.33e-06, Mann-Whitney U test, Bonferroni-corrected). On the other hand, schizophrenia patients exhibited higher heterogeneity in fractional anisotropy (p = 2.65e-09, Mann-Whitney U test, Bonferroni-corrected).

Discussion

The results of this comprehensive meta-analysis afford an advance in the understanding of structural brain heterogeneity in schizophrenia, an aspect that has been insufficiently studied to date, yet holds significant implications for the identification of illness subtypes and informative biomarkers⁷. A strength of this study was the comprehensive heterogeneity assessment of multiple structural brain measures in the framework of the ENIGMA dataset, a substantial repository of MRI-based brain measures collected from various sites globally^11,12,23. Different structural brain measures are thought to be influenced by separate sets of genes during different stages of development^34–39, and the multimodal assessment and comparison of structural brain measures is instrumental to disentangle their potentially different alteration in schizophrenia¹³. This study therefore explored and compared variability in cortical thickness, surface area, folding index, subcortical volume, and fractional anisotropy, allowing for a thorough assessment of regional and global variability. On the other hand, the framework of the large ENIGMA dataset increases the robustness of results and thereby supports the replicability of findings in psychiatry and neuroscience^23,40–44.

One of our key findings at the regional level is greater multimodal variability in schizophrenia for frontotemporal structures: While cortical surface area and folding index were more heterogeneous in few temporal and frontal regions, higher variability of cortical thickness comprised widespread bilateral temporal areas in addition to a few frontal regions. Thus, these results suggest that schizophrenia is associated with multimodal structural variability in different components of a frontotemporal network, thereby extending previous findings according to which schizophrenia patients show on average thinning of frontal and temporal regions^12,13,45,46. A possibility would be that different subgroups exist that can be distinguished by different neuroanatomical correlates: While one subgroup may have structural irregularities in temporal rather than frontal regions, be it cortical thinning, folding abnormalities or a combination thereof, another subgroup may be characterized by the inverse structural pattern. In all cases, the result would likely be impaired frontotemporal interaction, a hallmark of psychotic disorders. At least with regard to mean-scaled variability, schizophrenia patients also showed more heterogeneity of fractional anisotropy in the left uncinate fasciculus which connects temporal with frontal regions. Structural deficits of the left uncinate fasciculus may therefore also contribute to aberrant frontotemporal interaction in a subset of patients. On the other hand, our study also showed that gyrification of the left parahippocampal region is more homogeneous in schizophrenia compared to controls. Gyrification characteristics of the left parahippocampal region may therefore occur uniformly across potential subtypes and represent a central attribute of the disorder. Interestingly, a part of the left parahippocampal region has been recently shown to be thinner and associated with hallucinations in schizophrenia spectrum individuals⁴⁷. In our study, increased variability was also observed in left nucleus accumbens and caudate, regions implicated in reward processing and goal-directed motor control, respectively^48,49. It seems conceivable that variability in these subcortical structures is linked to the heterogeneity of negative and catatonic symptoms that differ substantially across schizophrenia individuals^50,51. Overall, the largest effect sizes with regard to increased variability ratio were observed for the lateral ventricles on both sides. Since ventricular enlargement is among the best established neurobiological findings in schizophrenia patients^7,11,52–55, it is notable that higher variability in schizophrenia remained robust when the coefficient of variation with accounting for group differences in mean was used instead of the variability ratio. In contrast, a previous study comparing first-episode schizophrenia with healthy controls showed higher variability of lateral ventricles only for the variability ratio, but not for the coefficient of variation⁷. We presume that the difference to our study could be related to the longer illness duration of the patients in our study, which included patients with chronic schizophrenia.

Notably, at a global level, our findings challenges the prevailing view of increased brain heterogeneity in schizophrenia in comparison to healthy individuals^7,14,15,18. The Person-Based Similarity Index (PBSI) analysis^16,17 surprisingly revealed that schizophrenia patients are characterized by a significantly higher homogeneity of especially the folding index, a measure related to the gyrification of the brain. This observation supports the view that certain gyrification characteristics could represent a potential core feature of schizophrenia³⁹. Previous studies demonstrated that adolescent and young adult schizophrenia patients typically feature increased gyrification in widespread regions compared with healthy controls^56–66, possibly related to overexpression of fibroblast growth factor-2 gene in the dorsolateral prefrontal cortex and further molecular-level aberrations in schizophrenia^39,67. While gyrification patterns are determined early in development⁶⁸, it is plausible that certain folding characteristics might render the brain more susceptible to the later development of schizophrenia, thus presenting a potential predictive biomarker for the disorder. In line with this view, prior studies with limited sample size have shown that schizophrenia is associated with aberrant gyrification characteristics in specific cortical regions which are in part related to particular symptoms and deficits of the disorder^{58,59,62,63,65,69–73}. For instance, the severity of positive symptoms in schizophrenia has been associated with gyrification in the right frontal and temporolimbic region^63,64. Similarly, fronto-temporo-limbic disconnectivity in schizophrenia, which is related to delusions, hallucinations and disorganization symptoms^74,75, has been proposed to be linked with local gyrification patterns³⁹. Previous studies also compared PBSI scores for cortical thickness, subcortical volume and cortical surface area between schizophrenia or first episode psychosis patients and healthy controls^15,16. While some results are in line with ours, e.g., higher cortical surface area PBSI scores in psychosis patients¹⁶, others are different, e.g. higher cortical thickness PBSI scores in controls^15,16 instead of no significant difference in our study. Potential reasons for deviations from our results are the far more limited sample size in both studies as well as the lower illness duration of first-episode psychosis patients in Antoniades et al¹⁶. However, the PBSI analysis also revealed that schizophrenia patients exhibit greater global heterogeneity than controls with regard to fractional anisotropy, suggesting that different white matter abnormalities contribute to schizophrenia subtypes. White matter pathology has been associated with functional dysconnectivity in schizophrenia⁷⁶, and resulting impairment of neural network synchrony may underlie schizophrenia symptoms^77,78.

However, despite the robust findings, this study has some limitations. Not all data were individual-level data, which could potentially introduce bias. Moreover, not all measures were collected at all sites - in particular, the amount of fractional anisotropy data was lower in comparison to other measures. In addition, the parcellation of gray and white matter networks was coarse, and a more fine-grained parcellation in future studies may lead to more differentiated insights of structural variability.

In conclusion, this study provides novel insights into the heterogeneity of structural brain measures in schizophrenia. In light of increasing interest in schizophrenia subtypes and corresponding neurobiological characteristics^79–91, high variability in certain neuroanatomical structures may point to potential avenues for the development of more targeted treatment strategies. On the other hand, low variability of structural features such as gyrification may substantiate core features of schizophrenia that are shared across illness subtypes.

Our results hold potential to guide future research efforts and pave the way for precision medicine approaches in schizophrenia. Ultimately, this study underscores the critical role of variability in our understanding of complex disorders such as schizophrenia, a perspective that is essential for a more holistic approach to its neurobiology.

Figure 6: PBSI analysis.

PBSI for cortical thickness (CT), cortical surface area (SA), cortical folding Index (FI), subcortical volume (SV) and fractional anisotropy (FA).