Diffusion Tensor Imaging White Matter Endophenotypes in Patients With Schizophrenia or Psychotic Bipolar Disorder and Their Relatives

Data were collected from 513 subjects participating at two sites of the Bipolar-Schizophrenia Network on Intermediate Phenotypes study (B-SNIP): 104 healthy comparison subjects, 109 probands with schizophrenia, 35 probands with schizoaffective disorder (16 depressed, 19 manic), 63 probands with psychotic bipolar disorder, 95 relatives of the schizophrenia probands, 43 relatives of the probands with schizoaffective disorder (24 relatives of depressed probands, 19 relatives of manic probands), and 64 relatives of the probands with psychotic bipolar disorder. The groups compared in the study, after regrouping of the probands with schizoaffective disorder, are shown in Table 1. The participants were recruited through word of mouth, advertisements, and community support groups. All subjects provided written informed consent statements approved by review boards of Hartford Hospital/Yale University and the University of Maryland/Johns Hopkins University. The participants were independent of study groups analyzed in previous publications on DTI studies.

Consensus diagnoses were established by trained clinical raters and senior diagnosticians using clinical data and the Structured Clinical Interview for DSM-IV (SCID). The probands were clinically stable and had been taking stable doses of medications for at least 4 weeks. All of the bipolar subjects were psychotic, as determined by the presence of hallucinations or delusions during prior affective episodes (14) or current psychosis at the time of scanning. Current psychosis was judged according to the relatively permissive criterion of a score of 3 or higher on one or more of the following scales of the Positive and Negative Syndrome Scale (PANSS) (18): delusions, conceptual disorganization, hallucinations, and suspiciousness/persecutory delusions. Other positive symptoms (grandiosity, excitement) did not qualify as psychosis. The relatives were further classified by the presence or absence of symptoms of DSM-IV-TR cluster A and cluster B personality disorders with the Structured Interview for DSM-IV Personality (19). Relatives with cluster A or B symptoms were included even if they possessed one fewer symptom than is required to meet the full criteria for the axis II disorder. The comparison subjects and the probands’ relatives were free of current axis I disorders, as assessed by the SCID (those who had axis I disorders were excluded). All subjects were free of known neurological illness and were not substance abusers, as judged by negative urine toxicology screens. Though we strove to achieve balanced subject groups, matching was not exact. The subject population overall contained more African Americans at the Baltimore site and more Hispanics at the Hartford site, consistent with the 2010 U.S. Census. More relatives were female, likely because of family volunteer bias.

DTI data were obtained on 3-T scanners (Siemens, Erlangen, Germany). The scanners at Hartford (Allegra, head only) and Baltimore (Trio) used highly similar but not identical scanning sequences; both used single-shot spin-echo planar imaging (EPI) with a twice-refocused balance echo sequence to reduce eddy current distortions. Hartford used TR/TE=6,300/85 ms, field of view=220 m, b=1,000 s/mm² along 32 directions, 45 contiguous slices, three imaging series, and a voxel size of 1.7×1.7×3 mm. Baltimore used TR/TE=6,700/92 ms, field of view=230 m, b=1,000 s/mm² along 30 directions, 48 contiguous slices, two imaging series, and a voxel size of 1.8×1.8×3 mm.

All data processing was executed at Hartford by investigators who were blind to diagnosis and used the same methods and quality thresholds. Data were checked for noise, motion, and artifacts (seven subjects at Hartford and 34 at Baltimore were excluded). FSL software (www.fmrib.ox.ac.uk/fsl) was used for standard analysis including motion correction and eddy current correction, with coregistration to the B₀ image with gradient directions corrected. Fractional anisotropy was used to assess white matter integrity. To enhance reproducibility and overlap of white matter fibers, we calculated a common white matter skeleton and used tract-based spatial statistics (20).

Effects of scanner, age, sex, race, and ethnicity were included in the general linear model and thus removed from main effects. Analysis was performed both at the voxel level and by region of interest.

Regions of interest were delineated on the white matter skeleton obtained by using all 50 predefined regions from the International Consortium of Brain Mapping (21) white matter atlas. Those regions were then overlaid with the white matter skeleton. The atlas covers only deep white matter structures that are most reproducible on standard coregistration. Tract-based spatial statistics allow for reliable definition of more fiber tracts. To investigate those tracts, we manually defined 26 additional regions using anatomical information only.

Voxel-based whole-skeleton analysis was performed on the skeletonized data by using threshold-free cluster enhancement (22) to correct for multiple comparisons. Since the voxel-level data showed group differences in variance, permutation-based approaches were used instead of the standard general linear model. The region-of-interest data showed uniform variance at p<0.05 for all regions.

In a direct comparison of probands’ relatives and either probands or healthy comparison subjects, using regions of interest or threshold-free cluster enhancement whole-map analysis, we detected no significant differences at a p<0.01 corrected level. Statistical power in this initial analysis was hampered by the large number of voxels in the whole-brain analysis and/or by large anatomical regions, diluting focal regional differences. We therefore analyzed the relatives’ data within smaller regions, defined as those showing the most significant differences in the probands. We selected the 29 anatomically defined regions showing significant statistical differences when we compared both proband groups to the comparison subjects in both voxel-based and region-of-interest analyses. Within each such region, we defined smaller, focused regions of interest contained within the clusters identified with threshold-free cluster enhancement, around the maxima of contrast between all probands and the healthy comparison subjects. Of these 29 subregions; all showed differences in schizophrenia, and 14 also showed differences for psychotic bipolar disorder. Table 2 shows Montreal Neurological Institute (MNI) coordinates for the centers of the primary anatomical regions and for the focal subregions. Coordinates of the subregions in all of the defined regions are presented in Table S1 in the data supplement accompanying the online version of this article.

Our criteria for axis 2 cluster A or B characteristics were met by 17 relatives of the schizophrenia probands and 18 relatives of the bipolar probands. This allowed us to analyze clinical regression along the following scale: healthy comparison subjects, 1; healthy relatives without cluster A or B symptoms, 2; relatives with cluster A or B symptoms, 3; probands, 4. Additional general linear model analyses (at both region-of-interest and voxel levels) correlated this status with fractional anisotropy.

The Schizo-Bipolar Scale (2) was designed to investigate overlap between schizophrenia and psychotic bipolar disorder by using continuous measures based on phenomenology and illness course. The correlation between fractional anisotropy values and scores on the Schizo-Bipolar Scale were calculated in a separate general linear model analysis.

Additional general linear model analyses assessed white matter differences possibly attributable to medications. The subjects were assigned to eight medication groups (Table 1). A general linear model analysis was performed separately for each diagnostic group, contrasting individuals using medications from each of the eight categories with all other subjects.

Heritability analyses were conducted by using Sequential Oligogenic Linkage Analysis Routines (SOLAR) (31). This software determines the relative importance of familial and environmental influences on a phenotype by modeling the covariance among family members as a function of kinship. Scanner was used as a covariate.

We first analyzed global mean fractional anisotropy averaged over the whole white matter skeleton. Fractional anisotropy in both proband groups was significantly lower (p<0.001) than in the healthy comparison subjects (t=4.82, df=227, for schizophrenia; t=3.69, df=184, for psychotic bipolar disorder; effect size: d=0.64 and 0.54, respectively), with no differences between proband groups (effect size: d=0.04). Fractional anisotropy was lower in 29 regions (of 76 analyzed) in probands (Table 2); none differed significantly between schizophrenia and psychotic bipolar disorder. For all regions, effect sizes for the differences between proband groups were less than 0.2. The global fractional anisotropy average showed significantly higher (p<0.01) variance in the bipolar probands than in either the schizophrenia probands or the comparison subjects and nonsignificantly higher variance in the relatives of the bipolar probands (p<0.06). Figure 1 (top left) shows the distribution of differences between the healthy comparison subjects and the clinical groups (p<0.01, corrected). Many regions revealed fractional anisotropy differences for schizophrenia but not psychotic bipolar disorder. With the exception of one cluster, all voxels showing differences in psychotic bipolar disorder also showed differences in schizophrenia. None of the fractional anisotropy values was higher in a clinical group than in the comparison subjects.

Among the relatives, the whole-brain average for fractional anisotropy was lower in the schizophrenia relatives than in the healthy comparison subjects (t=2.62, df=221, p<0.01, d=0.35). Two anatomically defined regions (left anterior corona radiata and superior right corona radiata) differed significantly (p<0.05 corrected) between the schizophrenia relatives and comparison subjects (Table 2). Table S1 and Figures 1S and 2S in the online data supplement shows results of the focused post hoc analyses of the regions of interest. In the secondary analyses, fractional anisotropy in the schizophrenia relatives was significantly lower than in the healthy comparison subjects in 10 regions and higher than in the schizophrenia probands in 11 regions. Figure 1 (top right) shows the spatial distribution of those differences. In many regions, the relatives in the schizophrenia group closely resembled the probands. Only the corpus callosum genu in the schizophrenia relatives differed significantly from that region in both the schizophrenia probands and the healthy comparison subjects.

The relatives of the probands with psychotic bipolar disorder showed fewer significant effects, with only one region where their fractional anisotropy was lower than that of the healthy comparison subjects (superior aspect of the left posterior corona radiata) and five regions where their values were higher than those of the bipolar probands. No analyzed region showed significant differences between schizophrenia and psychotic bipolar disorder for either the probands or the relatives.

Most of the 29 regions differing between probands and healthy comparison subjects also manifested a significant clinical regression effect for schizophrenia; in psychotic bipolar disorder this effect was restricted to the corpus callosum genu (Table 2). Similarly, in voxel-based analysis (Figure 1, bottom left) most regions showing differences between probands and comparison subjects manifested similar clinical regression effects. As with these clinical contrasts, correlations between fractional anisotropy and clinical status were higher in schizophrenia than in psychotic bipolar disorder.

Fractional anisotropy values are displayed in Figure 2 for the anatomically defined genu of the corpus callosum, the region showing the strongest differences among groups.

To study effects of psychosis common to both diseases, we analyzed group differences (probands and relatives versus comparison subjects) and clinical regression correlations with the data for schizophrenia and psychotic bipolar disorder combined. The region-of-interest analysis of clinical regression and the analysis of probands versus comparison subjects showed that all regions showing significant differences for schizophrenia (Table 2) also showed lower values in the joint group analysis. In the comparisons of relatives and comparison subjects, only one region (left anterior corona radiata) showed significant difference in the combined analysis.

Few relatives met our criteria for cluster A or B personality disorders, so their comparisons with relatives lacking those personality characteristics had lower statistical power. Nevertheless, in the analysis comparing cluster A relatives to other relatives, two regions (Table 2) showed significant (p<0.05) differences for the relatives of probands with schizophrenia, and one region showed a difference for the relatives of the bipolar probands. Both the corpus callosum genu and right posterior thalamic radiation skeleton showed significant stepwise progressions of fractional anisotropy values from the healthy comparison subjects to relatives to cluster A relatives to schizophrenia probands, with the cluster A relatives resembling the probands. Similar trends were seen in other regions where probands differed from comparison subjects. No such differences were observed for relatives with cluster B traits.

Whole-brain mean fractional anisotropy was highly heritable (h²=0.45, p<0.001), as was fractional anisotropy for many anatomically defined regions (Table 2, Figure 3), not restricted to those showing between-group effects (see Table S1 in the online data supplement).

Analysis of anatomically defined regions of interest showed a weak association between scores on the Schizo-Bipolar Scale and fractional anisotropy values (t=2.11, df=181, p<0.05, uncorrected) in the left superior longitudinal fasciculus, peaking at MNI coordinates –33, –19, 41 (see Figure 3S in the online data supplement).

While the general linear model approach removes effects of confounding variables, interactions among sex, race or ethnicity, age, and diagnosis are relevant for understanding pathologic and clinical applications. Separate general linear model analyses were performed to examine those interactions. Age was the most important confounding factor; fractional anisotropy values decreased significantly with age in all groups and regions of interest. For the global whole-brain average, age interacted with disorder. The interaction was negative in the probands with psychotic bipolar disorder and their relatives (less decrease with age than in comparison subjects). In schizophrenia it was positive for the whole-brain average and for the superior posterior corona radiata, where fractional anisotropy values decreased more steeply with age than they did among the healthy comparison subjects.

In the analysis of whole-brain mean fractional anisotropy, sex and race showed significant differences, with men and European Americans showing higher values. The effects of sex and race on global fractional anisotropy did not interact with clinical diagnosis. At the regional level, sex showed an interaction (t>3.0, p<0.01) in five regions in the comparison of bipolar probands and healthy comparison subjects: the corpus callosum body (MNI coordinates 13, 3, 33), right and left anterior limbs of the internal capsule (MNI coordinates 23, 6, 20 and –12, 14, 17), right anterior corona radiata (MNI coordinates 23, 27, 5), and right inferior fronto-occipital fasciculus (MNI coordinates 29, 14, –7). The direction of the interaction was such that in women the decrease of fractional anisotropy in the bipolar probands was larger than in men.

We also analyzed radial diffusivity, mean diffusivity, and axial diffusivity. They showed no significant differences at the threshold used for fractional anisotropy. We additionally investigated data pooled in all voxels showing significant changes in fractional anisotropy. As expected, fractional anisotropy decreases were accompanied by increases in all three diffusivity measures. A significant (p<0.01) difference was observed; differences in radial diffusivity were larger than those for mean and axial diffusivity in both the schizophrenia and bipolar probands, suggesting that demyelination (not axonal injury) causes the observed white matter differences.

After we adjusted for age, mean whole-brain fractional anisotropy did not differ between scanners. Site showed no interaction with clinical diagnosis, and the spatial map of scanner effects did not correlate with clinical effect maps. Quantifying between-scanner reproducibility as a stand-alone measure is difficult. To estimate it, we ran general linear model analyses for each scanner separately and calculated general linear model contrast maps for each site. To quantify their agreement, we developed a randomization procedure in which data from both scanners were randomly split into groups the size of the scanned subjects at each site; the general linear model was run for each data set separately, and the degrees of reproducibility in the subsamples were compared. The between-scanner correlation (r) values were 0.60 for age, 0.12–0.18 for the main contrast of probands and healthy comparison subjects, 0.05–0.10 for the main contrast of relatives and comparison subjects, and 0.06–0.12 for the contrast of schizophrenia and bipolar probands. In randomization procedures, real comparisons fell between 25% (schizophrenia probands versus comparison subjects) and 6% (relatives of bipolar probands versus comparison subjects) of randomized splits. Thus, the between-scanner data agreement was not significantly worse (p<0.05) than would be expected from two data sets derived from the same scanner. General linear model analyses performed separately for each scanner showed the same, but less significant, findings than in the combined analysis.

We compared probands receiving a particular medication class with all other probands. Patients with psychotic bipolar disorder who took second-generation antipsychotic drugs showed significantly higher fractional anisotropy in the corpus callosum genu and body (t=2.93, df=80, p<0.01 uncorrected) than did the remaining bipolar probands. For schizophrenia, a significant effect was observed in the corpus callosum splenium for anticonvulsant mood stabilizers (t=–3.31, df=123, p<0.01), such that probands with schizophrenia who were taking these medications had lower fractional anisotropy values than other schizophrenia probands.