SCHIZOPHRENIA RESEARCH ELSEVIER
Schizophrenia Research 30 (1998) 127-135
Does dysplasia cause anatomical dysconnectivity in schizophrenia? Edward T. Bullmore *, Peter W.R. Woodruff, Ian C. Wright, Sophia Rabe-Hesketh, Robert J. Howard, Nasser Shuriquie, Robin M. Murray Departments of Psychological Medicine and Biostatistics and Computing, Institute of Psychiatry, De Crespigny Park, London, SE5 8AF, UK Received 7 June 1997; accepted 1 July 1997
Abstract
Evidence is reviewed that dysplastic brain development in the second half of pregnancy predisposes to schizophrenia. We suggest that an important corollary of aberrant development at this stage of ontogenesis is abnormal afferentation of the cortical plate, and that this may be macroscopically measurable in terms of abnormal correlational structure in adult brain imaging data. This prediction is tested by analysis of multiple cortical volume measures on magnetic resonance imaging data acquired from 35 male right-handed schizophrenics and 35 matched controls. There are no significant differences between groups in global, intra-hemispheric or inter-hemispheric correlational structure; but schizophrenics are shown to have significantly reduced dependencies between frontal and temporal lobe volumes, and frontal and hippocampal volumes, in the left hemisphere. We conclude that anatomical dysconnectivity (between frontal and temporal cortex) in schizophrenia may be caused by dysplasia. © 1998 Elsevier Science B.V.
Keywords: Afferentation; Connectivity; Correlation; MRI; Neurodevelopmental; Schizophrenia
1. Introduction
The last 15 years or so have witnessed the powerful renaissance of an idea, originally proposed more than a century ago (Clouston, 1891), that schizophrenia should be regarded as a longterm consequence o f abnormal or dysplastic brain development. A considerable body of evidence has recently accumulated from two main areas of schizophrenia research in support of this view. Epidemiological studies have shown: (i) that * Corresponding author. Tel: +44 (0) 171 740 5290; Fax: +44 (0)171 277 1390; e-mail:
[email protected] 0920-9964/98/$19.00 © 1998 ElsevierScience B.V. All rights reserved. PII S0920-9964 (97) 00141-2
psychomotor development m a y be deviant or delayed in children destined to develop schizophrenia in later life (Jones et al., 1994); (ii) that age of onset and severity of illness m a y be modulated by gender, with males tending to present earlier and experience a more severe form of the disease, as is typical of several other neurodevelopmental disorders (Castle and Murray, 1991); and (iii) that the later risk of schizophrenia m a y be increased by intra-uterine or early postnatal exposure to viral infection (Sham et al., 1992), or cerebral hypoxia as a complication of pregnancy or birth (McGrath and Murray, 1995). Neuropathological studies have demonstrated
128
E. 72 Bullmore et al. / Schizophrenia Research 30 (1998) 127 135
abnormalities of neuronal configuration in neocortical and limbic structures that are consistent with dysplastic cortical formation (Akbarian et al., 1993); and an absence of conspicuous gliotic reaction, such as would be expected if neuronal damage had been incurred in adulthood. Neuroimaging studies of brain structure, however, have arguably contributed relatively little to the rebirth of the neurodevelopmental model of schizophrenia. Neuroimaging has been used to demonstrate ventricular enlargement in schizophrenics at the time of their first presentation to clinical services; and that this ventriculomegaly fails to progress over the course of several years (DeLisi et al., 1992; Weinberger, 1995). A number of imaging studies have demonstrated abnormal regional or hemispheric asymmetries in schizophrenia (Bilder et al., 1994; Bullmore et al., 1995), which may reflect perturbed brain development (Crow et al., 1989). Finally, there are reported associations between schizophrenia and qualitatively recognisable abnormalities of brain structure, such as cavum septum pellucidum or callosal agenesis, that are undoubtedly of developmental origin (Lewis and Mezey, 1985). However, the majority of imaging studies have been concerned with measurement of regional volume deficits in cross-sectional samples of schizophrenics. While this line of research has established that schizophrenia is often associated with widespread but subtle decrements in cortical volume (Harvey et al., 1993), it does not address the pathogenetic question(s) of how (or when) these abnormalities arise. There are two main strategies by which neuroimaging could be used more decisively to test neurodevelopmental models of pathogenesis in schizophrenia. The first strategy is based on prospective study of individuals at high risk of schizophrenia. For example, one could scan children born to schizophrenic parents in the hope of demonstrating neuroanatomical abnormalities only in those children destined later to develop schizophrenia (Woodruff and Murray, 1994). The other strategy is less direct, but logistically less demanding. It is to predict theoretically the longterm neuroanatomical sequelae of developmental aberration predisposing to schizophrenia; and specifically to test those predictions by analysis of
structural imaging data acquired from adult schizophrenics. In this paper we are concerned exclusively with the second strategy.
1.1. A tripartite scheme for brain development The complex and individually variable process of brain development can be crudely subdivided into three main stages (Goldman-Rakic and Rakic, 1989; Jones, 1991); see Table 1. We have to address two related questions about the pathogenesis of schizophrenia in relation to this scheme. First, at which stage of the process is perturbed brain development most likely to predispose to schizophrenia? And second, what are the likely sequelae in terms of brain structure? On the basis of the existing neuropathological, epidemiological and imaging data, we suggest that it is during the second stage, i.e., during the second half of gestation, that aberrant brain development is most likely to predispose to schizophrenia. If that is so, then we can predict abnormalities of cortical lamination, symmetry and gyrification in adult schizophrenic brain structure. These 'predictions' simply echo the existing data on which they were based (and are therefore not very interesting); but we can further predict that second stage dysplasia will cause abnormal afferentation of the cortical plate.
1.2. Afferentation, connectivity and correlation Axonal projections begin to invade the cortical plate and establish synaptic territories from approximately the 20th week of gestation. Cortical afferentation by subcortical (e.g. thalamic) structures may precede afferentation by ipsilateral cortical regions, and later contralateral cortical regions. Connections between cortical regions are predominantly reciprocal and glutamatergic (Kerwin, 1993). Afferentation and the establishment of a primary repertoire of connections between brain regions has been described in the context of neural Darwinism as a competitive process (Edelman, 1987). Neurones that succeed in this 'struggle to connect' will mutually benefit from the trophic effects of glutamatergic synapsis (Kerwin and Murray, 1992); neurones that fail to establish
E.T. Bullmore et al. / Schizophrenia Research 30 (1998) 127-135
129
Table 1 A tripartite scheme for brain development Approximate timing
Main neurodevelopmental events
0 20 weeks
* formation of the neural tube and proliferative periventricular epithelium; * mitotic division of germinal cells and radial migration of neurones to form cortical plate and subcortical zone * afferentation of cortical plate by axonal projections; * establishment of primary synaptic connections between brain regions; * cytoarchitectonic differentiation of cortical plate; * appearance of macroscopic asymmetries and gyrification of cortical surface.
20 40 weeks
0-20 + years
* regressive phenomena (cell death, synaptic pruning, axonal retraction); * experience-dependent synaptic modification; * myelination.
20-40 weeks marks the second stage of cerebrogenesis; developmental aberration at this stage is proposed to determine structural brain abnormalities in adult schizophrenia.
primary connections will be rendered more vulnerable to developmentally later regressive processes, such as cell death. In short, neurones that securely establish primary synaptic connections with other neurones are more likely to be selected as components of adult neural systems than neurones that do not. Of course, we cannot directly test adult brain imaging data for evidence of abnormal corticocortical connectivity at a microscopic level; no in vivo imaging modality has sufficient spatial resolution. However, if we accept that reciprocal afterentation in utero of two cortical regions, say A and B, confers a mutually trophic and protective effect, we might expect the subsequent growth Of A and B, and the adult regional volumes of A and B, to be positively correlated (Fig. 1). This expectation is testable at the macroscopic level of brain imaging data; as is the related prediction that correlation between volumes of A and B may be abnormal in adults, such as schizophrenics, who (we suppose) have suffered dysplastic brain development in the second half of pregnancy.
2. Methods 2.1. Subjects
Thirty-five male patients with schizophrenia, diagnosed by DSM-III-R criteria (APA, 1987),
and 35 normal male volunteers, were recruited as part of a previously reported study (Woodruff et al., 1997). All subjects were right-handed. The groups were matched in terms of age, ethnicity, paternal social class, and mean weekly alcohol consumption. Four individuals (three controls and one schizophrenic) with outlying volume measurements on one or more regions of interest were not included in this analysis (which is why the sample size is slightly less than for the correlational analysis of 38 controls and 36 schizophrenics previously reported; Woodruff et al., 1997). 2.2. Regional volume measurement
Tl-weighted magnetic resonance imaging (MRI) data were acquired from all subjects on a 1.5 Tesla system with acquisition parameters as previously described (Woodruff et al., 1997). Volumes of the following nine regions of interest (ROIs) were bilaterally measured by interactive parcellation of each image: frontal lobe (F), dorsolateral prefrontal cortex (DLPFC), ventrolateral prefrontal cortex (VLPFC), temporal lobe (T), superior temporal gyrus (STG), parahippocampus (Pa), hippocampus (Hi), anterior cingulate gyrus (AC), and posterior cingulate gyrus (PC). All ROI volumes included both gray and white matter, except hippocampus. Inter-rater reliability of measurement was good; intra-class correlation coefficients >0.8 for all regions (Woodruff et al., 1997).
130
E. 72 Bullmore et al. / Schizophrenia Research 30 (1998) 127-135
In Utero
Adulthood
+ +
+
InterNeuronal (Micro) Links
InterRegional (Macro) Correlations
Fig. 1. Putative linkage between (microscopic) intra-uterine afferentation of the cortical plate and (macroscopic) adult brain correlational structure. Top row: Microscopically, cortical regions A and B are linked in utero by reciprocal, glutamatergic axonal projections, which confer mutually trophic and protective effects on growth of both regions. Macroscopically, the regional volumes of A and B are positively correlated in adulthood. Bottom row: Microscopically, A and B are both linked to C but not to each other. Macroscopically, the volumes of all three regions are positively correlated. This illustrates the potential lack of specificity in correlational analysis and motivates multiple regression modelling of the specific inter-dependencies of cortical regions.
2.3. Notation We denote the N x P data matrix for the controls as X c and the equivalent matrix for the schizophrenics as Xs. For both groups, the number of cases N = 3 5 , and the number of variables P = 18. The corresponding (P x P) correlation matrices are denoted C¢ and Cs, respectively. The ( N x P / 2 ) matrix obtained for each group by averaging right and left hemispheric measurements of the same regions is denoted, e.g. Xo.
3. Results
1979). PCA of the correlation matrices Cc and Cs yields two (P x 1) vectors of eigenvalues, or variances of the P principal components of each dataset. The eigenvalues 2i, i = 1..... P, can be graphically compared between groups by inspecting scree diagrams or eigenvalue spectra (Fig. 2). These plots suggest that there may be subtle differences in the global correlational structure between groups: the controls have a somewhat 'peakier' spectrum than the schizophrenics, implying greater inter-regional dependency or integration. The global integration of a multivariate dataset X can be more formally defined (Tononi et al., 1994) as: P
3.1. Global and hemispheric integration To assess dependencies between all regional volume measurements in each group's data matrix, we used methods based on information theory and principal component analysis (PCA) (Jones,
Z ln(,~,)
I(X) = - - - , 2
I(X)>_O,
(1)
where 2 i denotes the ith eigenvalue of the correlation matrix of X. It can be seen that if some of these eigenvalues are close to zero, indicating that
E. T. Bullmore et al. / Schizophrenia Research 30 (1998) 12 7-135
O
LU
".,.
~..~..
Fig. 2. Eigenvalue spectra. Solid line, controls; dotted line, schizophrenics.
the variables are substantially correlated with each other, the integration will be large; conversely, if none of the eigenvalues is close to zero, indicating that the variables are relatively independent of each other, then integration will be small. Furthermore, if X (N × P) is partitioned into two mutually exclusive subsets of variables, X'(N × P') and X"(N × ( P - P ' ) ) with P ' < P , and integration is independently estimated for each subset (as above), the mutual information (MI) between subsets is MI(X'X") = I ( X ) - ( I ( X ' ) + I(X")); MI(X',X") >_O.
(2) If the two subsets are statistically independent of each other then the mutual information between them will be zero; and, to the extent that they are not independent of each other, the mutual information between them will be greater than zero. We used these expressions to measure global integration, left hemispheric integration, right hemispheric integration, and the mutual information between hemispheres, in both groups (Table 2). It is clear that, regardless of diagnostic status, the mutual information between hemispheres is large in proportion to global integration (greater than 50%), suggesting that regional vol-
131
umes are strongly correlated between hemispheres. It is also evident that controls have slightly larger values for all these measures than schizophrenics; for example, global integration is approximately 7% greater in controls. Normal theory does not provide good distributional approximations for the variances of principal components, or eigenvalues of a PCA (Krzanowski, 1988). Therefore, we used a 'non-parametric' or 'distribution-free' approach to test the significance of these small differences between groups (Edgington, 1980; Manly, 1991). The basic principles exemplified below can be used (and were used in this study) to test any statistic of interest. We wished to test the null hypothesis that the observed difference in a statistic between groups was not determined by the diagnostic status of group members. We therefore randomly reassigned each individual to one of two groups (of the same size as the observed groups) and estimated the difference in the test statistic between randomised groups. This procedure was repeated many (> 100) times, resulting in many estimates of betweengroup difference in the test statistic under the null hypothesis, which could be sorted in ascending order to form a null distribution. If the difference in the test statistic between observed groups exceeded the (1-~)*100 percentile value of the null distribution, then the null hypothesis was refuted by a one-tailed test with probability of Type I (false positive) error = ~. For example, the absolute difference between observed groups in global integration was 0.5; this was tested against a null distribution sampled by 100 random reassignments of 70 subjects to two groups of size 35; the observed difference was less than the 95 percentile value of the null distribution; therefore the observed difference was not significant by a one-tailed test of size ~=0.05. By the same testing procedure, observed differences in right and left intra-hemispheric integration, and mutual information between hemispheres, were also compatible with the null hypothesis. In short, there were no significant differences between groups in cortico-cortical integration at global, hemispheric, or inter-hemispheric levels of anatomical organisation.
E.T. Bullmore et al. / Schizophrenia Research 30 (1998) 127-135
132
Table 2 Global and intra-hemispheric integration, and mutual information between hemispheres Integration
Controls Schizophrenics
Mutual information
Global
Left
Right
Inter-hemispheric
7.0 6.5
1.59 1.44
1.79 1.52
3.62 3.54
3.2. Inter-regional correlational analysis To assess between group differences in correlational structure with greater anatomical resolution, we compared the two matrices of inter-regional correlation coefficients derived from bilaterally averaged data, Cc and Cs (Table 3). These matrices reveal that most inter-regional correlation coefficients are positive in the control group (only the correlation between anterior and posterior cingulate is substantially negative); whereas there are more near-zero or negative correlation coefficients in the schizophrenic group. To identify correlations that were significantly greater in controls than schizophrenics, we applied the procedure for randomisation testing outlined above. In this case, the test statistic was the difference between groups in the correlation coefficient for a given pair of regions; the number of random reassignments was 200; and the size of a one-tailed test of the null hypothesis that interregional correlation was not greater in controls
than schizophrenics was ~ < 0.05. At this level, we expect less than 2 of the 36 tests to be 'significant' by chance. Six inter-regional correlations (mainly involving frontal and temporal structures) were found to be significantly more positive in the controls than schizophrenics (Fig. 3). A comparable analysis of correlation matrices separately derived for left and right hemispheric data is reported by Woodruff et al. (1997).
3.3. Inter-regional regression analysis A problem with correlational analysis is its relative lack of specificity. For example, a positive correlation between the volumes of regions A and B might suggest a direct inter-dependency, but could equally arise because both A and B are positively correlated with a third region, C (see Fig. 1). To investigate inter-regional relationships with greater specificity than correlational analysis allows, we used multiple regression. Each ROI was
Table 3 Inter-regional correlation matrix a F F DLPFC VLPFC T STG Pa Hi PC AC
l 0.63 0.13 0.61 0.62 0.43 0.19 0.00 0.12
DLPFC 0.57 1 0.27 0.67 0.43 0.56 0.37 0.08 0.26
VLPFC 0.53 0.67 1 0.23 0.06 0.40 0.01 0.19 0.09
T 0.24 0.23 0.14 1 0.72 0.63 0.24 0.16 0.30
STG - 0.03 0.08 0,16 0,64 1 0.35 0.15 0.13 0.09
Pa 0.38 0.27 0.28 0.33 0.18 1 0.43 0.24 0.05
Hi
PC
AC
0.01 0.21 - 0.04 0.29 0.24 0.35 1 0.33 - 0.05
- 0.08 0.12 0.23 - 0.31 -0.15 0.04 0.08 1 - 0.23
0.00 - 0.07 0.08 0.07 -0.12 -0.4 - 0.04 -0.07 1
Coefficients for controls are below the main diagonal; coefficients for schizophrenics are above the main diagonal.aBilaterally averaged ROI measurements: F, frontal lobe; DLPFC, dorsolateral prefrontal cortex; VLPFC, ventrolateral prefrontal cortex; T, temporal lobe; STG, superior temporal gyrus; Pa, parahippocampal gyrus; Hi, hippocampus; PC, posterior cingulate gyrus; AC, anterior cingulate gyrus.
E.T. Bullmore et al. / Schizophrenia Research 30 (1998) 127-135
133
AC
PC HiPa
ST( T. VLPI C DLPI!C
DLPFC VLPFC
+
STG
Pa
Hi
PC
AC
Fig. 3. Correlation matrix; regions abbreviated as in Table 3. Black elements indicate inter-regional correlation coefficientsthat were significantly greater in controls than schizophrenics. For a one-tailed test of each coefficient ~
