Introduction
Evidence for white matter disruption from diffusion tensor imaging (DTI) and for dysconnectivity from functional magnetic resonance imaging (fMRI) suggests impaired neuronal connectivity as a potential core feature of bipolar disorder (BD) (Strakowski et al. Reference Strakowski, Delbello and Adler2005, Reference Strakowski, Adler, Almeida, Altshuler, Blumberg, Chang, DelBello, Frangou, McIntosh, Phillips, Sussman and Townsend2012; Emsell & McDonald, Reference Emsell and McDonald2009; Vederine et al. Reference Vederine, Wessa, Leboyer and Houenou2011; Houenou et al. Reference Houenou, d'Albis, Vederine, Henry, Leboyer and Wessa2012; Nortje et al. Reference Nortje, Stein, Radua, Mataix-Cols and Horn2013; Skudlarski et al. Reference Skudlarski, Schretlen, Thaker, Stevens, Keshavan, Sweeney, Tamminga, Clementz, O'Neil and Pearlson2013; Vargas et al. Reference Vargas, López-Jaramillo and Vieta2013; Wessa et al. Reference Wessa, Kanske and Linke2014; Kumar et al. Reference Kumar, Iwabuchi, Oowise, Balain, Palaniyappan and Liddle2015). In particular, studies have reported white matter microstructural abnormalities in connections between prefrontal and anterior limbic structures (Versace et al. Reference Versace, Almeida, Hassel, Walsh, Novelli, Klein, Kupfer and Phillips2008, Reference Versace, Andreazza, Young, Fournier, Almeida, Stiffler, Lockovich, Aslam, Pollock, Park, Nimgaonkar, Kupfer and Phillips2014; Benedetti et al. Reference Benedetti, Absinta, Rocca, Radaelli, Poletti, Bernasconi, Dallaspezia, Pagani, Falini, Copetti, Colombo, Comi, Smeraldi and Filippi2011) supporting neural models of impaired emotion processing and regulation in patients with BD (Phillips & Swartz, Reference Phillips and Swartz2014).
Advanced diffusion MRI analysis techniques examine the brain in vivo to define parameters of neuroanatomical connectivity. Impaired structural connectivity can be assessed on a network scale using graph theory properties (Sporns et al. Reference Sporns, Tononi and Edelman2000). There have been few graph analysis investigations in BD, although reduced global efficiency in patients has been reported (Leow et al. Reference Leow, Ajilore, Zhan, Arienzo, GadElkarim, Zhang, Moody, Van Horn, Feusner, Kumar, Thompson and Altshuler2013; Collin et al. Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016). Studies that investigated regional segregation indicate inconsistencies (Leow et al. Reference Leow, Ajilore, Zhan, Arienzo, GadElkarim, Zhang, Moody, Van Horn, Feusner, Kumar, Thompson and Altshuler2013; Gadelkarim et al. Reference Gadelkarim, Ajilore, Schonfeld, Zhan, Thompson, Feusner, Kumar, Altshuler and Leow2014; Forde et al. Reference Forde, O'Donoghue, Scanlon, Emsell, Chaddock, Leemans, Jeurissen, Barker, Cannon, Murray and McDonald2015), which may be related to clinical heterogeneity or variation in the range of graph properties and in-house measures utilized. Therefore, further investigation into global integration and regional segregation of structural brain networks in BD is warranted.
Complex network analysis combines structural and diffusion MRI to model the brain as a neuroanatomical network and evaluate topological organization and properties of brain structure. A network at the macro-scale comprises cortical and subcortical structures represented by ‘nodes’, and white matter connections, represented by ‘edges’ (Bullmore & Sporns, Reference Bullmore and Sporns2009; Rubinov & Sporns, Reference Rubinov and Sporns2010; Sporns, Reference Sporns2012). Following network construction, graph theory properties characterize brain integration and segregation (Bullmore & Sporns, Reference Bullmore and Sporns2009, Reference Bullmore and Sporns2012; Rubinov & Sporns, Reference Rubinov and Sporns2010). These properties include global measures quantifying whole-brain integration and connectedness, and local measures to characterize segregation: the anatomical architecture of structural inter-connectivity with nearby regions (Bullmore & Sporns, Reference Bullmore and Sporns2009; Bassett et al. Reference Bassett, Brown, Deshpande, Carlson and Grafton2011). Therefore, graph analysis extends DTI analysis of microstructural white matter to characterize patterns of organization.
Assessment of neuroanatomical sub-networks, the connectivity between several brain regions, assumes that connections belonging to the same component are highly connected (Zalesky et al. Reference Zalesky, Fornito and Bullmore2010; Meskaldji et al. Reference Meskaldji, Ottet, Cammoun, Hagmann, Meuli, Eliez, Thiran and Morgenthaler2011). The Network Based Statistic (NBS) tests for connectivity effects in edge weights within sub-networks. Therefore, investigation of sub-networks may identify regionally specific structural dysconnectivity in BD.
Furthermore, key hubs with a high density of interconnections, termed the rich-club, appear to play a key role integrating brain functioning and may be impaired in brain disorders (van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011; Collin et al. Reference Collin, Sporns, Mandl and Van Den Heuvel2013; Crossley et al. Reference Crossley, Mechelli, Vértes, Winton-Brown, Patel, Ginestet, McGuire, Bullmore, Vertes, Winton-Brown, Patel, Ginestet, McGuire and Bullmore2013, Reference Crossley, Mechelli, Scott, Carletti, Fox, McGuire and Bullmore2014; Sporns & Van Den Heuvel, Reference Sporns and Van Den Heuvel2013). Hubs, or nodes that are highly connected, appear to be involved in executive function, the salience network and the default mode network when the mind is at rest (van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011; Crossley et al. Reference Crossley, Mechelli, Vértes, Winton-Brown, Patel, Ginestet, McGuire, Bullmore, Vertes, Winton-Brown, Patel, Ginestet, McGuire and Bullmore2013; Senden et al. Reference Senden, Deco, De Reus, Goebel and Van Den Heuvel2014). Rich-club structures identified in healthy human brain networks include regions previously implicated in mood regulation and BD, for example the hippocampus, striatum and thalamus (Hallahan et al. Reference Hallahan, Newell, Soares, Brambilla, Strakowski, Fleck, Kieseppä, Altshuler, Fornito, Malhi, McIntosh, Yurgelun-Todd, Labar, Sharma, MacQueen, Murray and McDonald2011; Houenou et al. Reference Houenou, d'Albis, Vederine, Henry, Leboyer and Wessa2012) and their connections (Emsell et al. Reference Emsell, Langan, Van Hecke, Barker, Leemans, Sunaert, McCarthy, Nolan, Cannon and McDonald2013a , Reference Emsell, Leemans, Langan, Van Hecke, Barker, McCarthy, Jeurissen, Sijbers, Sunaert, Cannon and McDonald b ; Nortje et al. Reference Nortje, Stein, Radua, Mataix-Cols and Horn 2013 ; Ellison-Wright et al. Reference Ellison-Wright, Nathan, Bullmore, Zaman, Dudas, Agius, Fernandez-Egea, Müller, Dodds, Forde, Scanlon, Leemans, McDonald and Cannon2014).
Previously, we demonstrated fractional anisotropy (FA) reductions in the corpus callosum and limbic pathways (Emsell et al. Reference Emsell, Langan, Van Hecke, Barker, Leemans, Sunaert, McCarthy, Nolan, Cannon and McDonald2013 Reference Emsell, Leemans, Langan, Van Hecke, Barker, McCarthy, Jeurissen, Sijbers, Sunaert, Cannon and McDonald b ) that were consistent with other DTI studies (Vederine et al. Reference Vederine, Wessa, Leboyer and Houenou2011; Nortje et al. Reference Nortje, Stein, Radua, Mataix-Cols and Horn2013). Therefore, we sought to further examine regional connections defined by nodes connecting the corpus callosum and cingulum (Emsell et al. Reference Emsell, Langan, Van Hecke, Barker, Leemans, Sunaert, McCarthy, Nolan, Cannon and McDonald2013a ). Sub-network and rich-club analysis techniques were employed to probe inter-connectivity within neural circuits potentially implicated in BD. The present study aimed to investigate dysconnectivity in BD through global, local, network component and rich-club connectivity measures in a large clinically homogeneous sample of patients with euthymic BD.
Method
Participants
A total of 42 participants with BD were included for this graph theory analysis. Of these participants with BD, 35 between 18 and 60 years of age were recruited from the local community as part of the Galway Bipolar Study (Emsell et al. Reference Emsell, Langan, Van Hecke, Barker, Leemans, Sunaert, McCarthy, Nolan, Cannon and McDonald2013a , Reference Emsell, Leemans, Langan, Van Hecke, Barker, McCarthy, Jeurissen, Sijbers, Sunaert, Cannon and McDonald b ). An additional seven individuals with remitted BD who participated in a follow-up imaging of a first-episode psychosis study and underwent an identical scanning procedure were included in connectivity analysis (Scanlon et al. Reference Scanlon, Anderson-Schmidt, Kilmartin, McInerney, Kenney, McFarland, Waldron, Ambati, Fullard, Logan, Hallahan, Barker, Elliott, McCarthy, Cannon and McDonald2014; Kenney et al. Reference Kenney, Anderson-Schmidt, Scanlon, Arndt, Scherz, McInerney, McFarland, Byrne, Ahmed, Donohoe, Hallahan, McDonald and Cannon2015). Next, we recruited 43 age- and gender-matched healthy volunteers from the local community. The 42 patients with BD type I were confirmed using the Diagnostic and Statistical Manual of Mental Disorders, 4th edition (DSM-IV) Structured Clinical Interview for DSM Disorders (American Psychiatric Association, 1994). Exclusion criteria for all participants included a history of medical or neurological illness, history of head injury resulting in loss of consciousness for over 5 min, history of substance abuse in the past year, learning disability, and oral steroid use in previous 3 months. Further exclusion criteria for controls included personal or family history of psychotic or affective disorder in first- or second-degree relatives. Additional patient exclusion criteria included a lifetime co-morbid DSM-IV Axis I disorder. All patients were euthymic at the time of scanning, defined as a score <7 on both the Hamilton Rating Scale for Depression and Young Mania Rating Scale (Hamilton, Reference Hamilton1960; Young et al. Reference Young, Biggs, Ziegler and Meyer1978). Ethical approval was obtained from the National University of Ireland Galway and University Hospital Galway research ethics committees. After a complete description of the study was presented to participants, written informed consent was obtained.
MRI acquisition
All participants were scanned with identical imaging acquisition parameters. Structural MRI data were acquired on a 1.5 Tesla Siemens Magneto Symphony Scanner using a four-channel head coil. A volumetric T1-weighted magnetization prepared acquisition of gradient echo (MPRAGE) sequence was acquired with the imaging parameters: repetition time 1140 ms; echo time 4.38 ms; inversion time 600 ms; flip angle 15; matrix size 256 × 256; an in-plane pixel size 0.9 × 0.9 mm2; slice thickness of 0.9 mm.
Diffusion MRI data were acquired using an eight-channel head coil with an echo planar image diffusion sequence acquired with parallel imaging, 64 optimized diffusion gradient directions with b = 1300 s/mm2, seven non-diffusion weighted images, repetition time = 8100 ms, echo time = 95 ms, field of view = 240 × 240 mm2, matrix = 96 × 96, in-plane voxel size of 2.5 × 2.5 mm2, slice thickness = 2.5 mm, 60 slices.
Pre-processing
All MR images were corrected for subject motion and eddy current distortions using the diffusion MRI analysis software toolbox ExploreDTI v.4.8.3 (Leemans et al. Reference Leemans, Jeurissen, Sijbers and Jones2009). The b-matrix was rotated to preserve diffusion orientation information within voxels during subject motion correction (Leemans & Jones, Reference Leemans and Jones2009). Quality assessment for all diffusion MR images examined scans for potential artifacts including hypointensities, shift in images, and signal dropout. We rated MR images on a quality scale from mild to severe. Participants with poor MR image quality were excluded from subsequent analyses.
Whole-brain tractography
Whole-brain white matter tractography reconstructed the series of streamlines used to define the ‘edges’ in complex network analysis. White matter pathways were reconstructed using ExploreDTI v.4.8.3 (Leemans et al. Reference Leemans, Jeurissen, Sijbers and Jones2009). Robust estimation of the diffusion tensor was implemented using the RESTORE approach (Chang et al. Reference Chang, Jones and Pierpaoli2005). A deterministic constrained spherical deconvolution algorithm accounted for crossing fibres present within voxels (Tournier et al. Reference Tournier, Calamante and Connelly2007; Jeurissen et al. Reference Jeurissen, Leemans, Jones, Tournier and Sijbers2011). Fibre tracking initiated in each voxel and continued with a step size of 1 mm until the following threshold was exceeded: fibre orientation distribution >0.15, angle threshold curvature >30°, minimum length <20 mm, and maximum length >300 mm. A spherical harmonic order of Lmax = 8 was applied.
Generating connectivity matrices
The series of tractography streamlines were mapped through cortical and subcortical structures to produce a weighted and undirected 90 × 90 connectivity matrix for each subject. Connectome maps did not correct for changes in region-of-interest structural volume. Visual inspection using MRIcron confirmed registration of the cortical parcellation atlas to T1 images (Rorden et al. Reference Rorden, Karnath and Bonilha2007).
Selection of nodes
The Automated Anatomical Labeling (AAL) Atlas parcellated cortical and subcortical volumes into 90 regions (Tzourio-Mazoyer et al. Reference Tzourio-Mazoyer, Landeau, Papathanassiou, Crivello, Etard, Delcroix, Mazoyer and Joliot2002). The AAL Atlas is a macro-anatomical parcellation atlas based on a single-subject brain template set in Montreal Neurological Institute space (Tzourio-Mazoyer et al. Reference Tzourio-Mazoyer, Landeau, Papathanassiou, Crivello, Etard, Delcroix, Mazoyer and Joliot2002). The AAL Atlas applies spatial coordinates and associated volume for 90–120 cortical and subcortical structures. Node definition excluded the cerebellum resulting in a 90-node parcellation scheme (45 nodes bilaterally).
Selection of edges
Analysis of undirected and weighted edges included streamline count between nodes and mean FA between nodes. Averaged FA values between two nodes defined the FA edge weight (Levitt et al. Reference Levitt, Alvarado, Nestor, Rosow, Pelavin, McCarley, Kubicki and Shenton2012). To extend the previous DTI study that reported widespread FA reductions, we implemented FA edge weights in graph theory and sub-network analysis. Additionally, analysis of rich-club connections employed streamline count edge weights to examine effects of nodes rich in connections. Streamline count represents the total number of reconstructed streamlines interconnecting two nodes. Additionally, graph thresholding was applied to remove spurious streamlines, which when unaccounted for lead to unintended false positives. Connection matrices were thresholded at a density value 0.2, which resulted in equivalent connection densities between groups but allowed connection weights to vary, minimizing false-positive streamline count (Fornito et al. Reference Fornito, Zalesky, Pantelis and Bullmore2012).
Network metrics
The Brain Connectivity Toolbox contains the set of functions used to produce graph theory measures (Rubinov & Sporns, Reference Rubinov and Sporns2010). Graph theory analysis implemented weighted undirected edge strengths across all analyses. Global and regional measures probe properties of integration and segregation (Bullmore & Sporns, Reference Bullmore and Sporns2009, Reference Bullmore and Sporns2012; Rubinov & Sporns, Reference Rubinov and Sporns2010). The metrics selected for analysis in the study are described in Table 1.
Table 1. Weighted graph properties investigated in this study
Nodal analyses were selected a priori from a previous DTI analysis in this cohort (Emsell et al. Reference Emsell, Langan, Van Hecke, Barker, Leemans, Sunaert, McCarthy, Nolan, Cannon and McDonald2013a , Reference Emsell, Leemans, Langan, Van Hecke, Barker, McCarthy, Jeurissen, Sijbers, Sunaert, Cannon and McDonald b ). Nine bilateral nodes were selected (listed in Table 3) as endpoints of prefrontal white matter, cingulum and callosal splenium connections (Emsell et al. Reference Emsell, Langan, Van Hecke, Barker, Leemans, Sunaert, McCarthy, Nolan, Cannon and McDonald2013a ).
NBS
Collectively impaired interconnections or sub-networks investigated with the NBS toolbox characterize network differences by identification of particular inter-regional connections or components affected in one group of individuals relative to another. Studies investigating sub-networks in a number of brain disorders applied NBS analysis; however this technique has yet to be investigated in BD (Zalesky et al. Reference Zalesky, Fornito and Bullmore2010, Reference Zalesky, Cocchi, Fornito, Murray and Bullmore2012). The NBS identifies an experimental effect at the cluster level by performing mass univariate testing controlling for family-wise error rate. First, statistical significance threshold was selected at p < 0.05. Next, permutation testing performed 5000 permutations. The NBS requires selection of supra-threshold connections: as this threshold setting is quite arbitrary, investigation across three supra-threshold values was employed, as has been most commonly implemented (Zalesky et al. Reference Zalesky, Fornito and Bullmore2010, Reference Zalesky, Cocchi, Fornito, Murray and Bullmore2012). Finally, all connected components’ supra-threshold was compared between groups (Zalesky et al. Reference Zalesky, Fornito and Bullmore2010).
Rich-club coefficient
Next, we carried out an exploratory investigation of the ‘rich-club’ coefficient among cortico-subcortical connections. The ‘rich-club’ refers to a set of nodes that are rich in connections and densely inter-connected among themselves forming a club (McAuley et al. Reference McAuley, da Fontoura Costa and Caetano2007; van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011). We investigated weighted rich-club connectivity differences between groups, as well as rich-club structural membership. The rich-club coefficient is defined by the following equation:
$$\phi (k){\rm} = \displaystyle{{2E \gt k\; \; \; \; \; \; \; \;} \over {N \gt k(N \gt k - 1)}}$$
where by, E > k indicates the weighted number of streamline connections greater than k present within a subgraph degree >k, as N > k indicates the number of nodes in the subgraph (McAuley et al. Reference McAuley, da Fontoura Costa and Caetano2007; Collin et al. Reference Collin, Sporns, Mandl and Van Den Heuvel2013). The measure φ reflects the level of interconnectivity between nodes. The rich-club identifies structural connections with a high value of k, removing all lower-degree connections. The rich-club nodes will have a high k and high φ (McAuley et al. Reference McAuley, da Fontoura Costa and Caetano2007; van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011; Collin et al. Reference Collin, Kahn, De Reus, Cahn and Van Den Heuvel2014).
Normalized rich-club coefficient
A normalized rich-club coefficient indicates that these densely interconnected structures were connected based on more than chance alone. A normalized coefficient is adjusted to a number of comparable random networks by preserving the degree distribution (van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011). Normalized rich-club analysis uses a number of rewiring iterations of the preserved degree distribution to ensure that effects are not due to chance. The weighted normalized rich-club coefficient is given by the following equation:
$$\phi w_{norm} (k) = \displaystyle{{\phi w(k)} \over {\phi w_{random} (k)}}$$
A weighted normalized rich-club coefficient Фw norm(k) was computed as the weighted rich-club parameter Фw(k) over a set of m = 500 random networks of equal degree. As the sufficient number of rewirings lacks standardization with re-arrangements (m) ranging from 100 to 1000 (van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011; Daianu et al. Reference Daianu, Jahanshad, Nir, Jack, Weiner, Bernstein and Thompson2015; Kocher et al. Reference Kocher, Gleichgerrcht, Nesland, Rorden, Fridriksson, Spampinato and Bonilha2015), selection of a number of random rewirings (m = 500) revealed a standard deviation that converges below 0.001. The number of appropriate rewiring iterations (m = 500) was set at 10 to ensure that normalization was met. By definition, Фw norm >1 over a range of k implies the existence of a rich-club set (McAuley et al. Reference McAuley, da Fontoura Costa and Caetano2007; van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011).
Rich-club membership
Validation methods confirmed rich-club membership. Rich-club members defined at the statistical significant network between patients and controls (k = 56) displayed approximately 10–12 highly connected nodes. We identified rich-club members at Фw norm >1 across range k at a group threshold of 60 and 70% of participants to determine rich-club structures most consistent across individual networks. Additionally, the top 10 highest weighted degree nodes confirmed that rich-club members were not dependent on one hub definition alone.
Statistics
Statistical analysis of global and regional metrics (degree, clustering coefficient, betweenness centrality, characteristic path length, global efficiency and local efficiency) applied multivariate analysis of covariance tests, co-varied for age and gender, using IBM SPSS statistics software version 22 (IBM SPSS Amos, 2012). Calculations of global values from nodal measures averaged values across all 90 nodes to generate a global average for each measure. We corrected for global connectivity to assess whether findings were indicative of reduced connectivity globally or potentially affected topological organization. Global analyses and regional comparisons underwent false discovery rate (FDR) correction for multiple comparisons (Benjamini & Hochberg, Reference Benjamini and Hochberg1995).
Permutation testing was used to assess between-group effects in rich-club connectivity. We performed 9999 Monte Carlo resamples using R software (RStudio Team, 2012). Multiple comparisons corrected for 28 possible values of k density were implemented using the FDR method (Benjamini & Hochberg, Reference Benjamini and Hochberg1995).
We applied partial correlations co-varying for age and gender to assess the relationship between clinical symptoms scales, illness duration and lithium use, and significant graph theory metrics. Global efficiency was correlated with rich-club density co-varying for age and gender to determine a possible relationship between global integration and hub inter-connectivity.
Results
The sociodemographic and clinical details of the participants are outlined in Table 2. Participants and healthy volunteers were age and gender matched. On average, patients had a lower number of years of education than healthy volunteers. The mean age of onset of illness in patients was 28 years of age. Of the patients, 33 took mood stabilizers at the time of scanning, with most using lithium (29); 22 patients took antipsychotics, with most using olanzapine (15); and eight patients used antidepressants. Four participants in the BD group were unmedicated at the time of the MRI scan.
Table 2. Clinical and demographic variablesa
Data are given as mean (standard deviation) unless otherwise indicated.
a Participants were age and gender matched. Years of education differed between groups. All participants in the bipolar group were confirmed prospectively euthymic with clinical rating scales of mania and depression less than a score of 7.
Global and regional graph theory metrics
Analysis of global properties revealed statistically significant group differences whereby the BD group displayed increased characteristic path length and reduced global efficiency and clustering coefficient compared with the healthy volunteer group when connections were weighted by FA (Fig. 1 and online Supplementary Table S1). Seven of the regions connected by fronto-limbic and parieto-occipital pathways revealed significantly reduced connectivity surviving multiple comparison correction. In BD, reduced clustering and local efficiency predominantly incorporated the superior and middle frontal nodes and superior and middle occipital nodes, when defined by FA (Table 3). Findings did not change when corrected for global connectivity, defined by the global density metric. Preserved measures of degree and density indicate that significant differences in topology may be the primary feature of BD.
Fig. 1. Global metrics defined by fractional anisotropy connection weight in bipolar disorder patients and controls show unadjusted values for (a) characteristic path length, (b) global efficiency and (c) clustering coefficient. The mean is represented by the middle line, with bars representing standard error.
Table 3. Regional connectivity differences between patients and controls a
a This analysis examined two different local graph measures in fractional anisotropy connections. Multivariate analyses of covariance were carried out between groups, co-varying for age and gender. Nine nodes bilaterally were selected a priori from a previous data-driven diffusion tensor imaging analysis (Emsell et al. Reference Emsell, Langan, Van Hecke, Barker, Leemans, Sunaert, McCarthy, Nolan, Cannon and McDonald2013a , Reference Emsell, Leemans, Langan, Van Hecke, Barker, McCarthy, Jeurissen, Sijbers, Sunaert, Cannon and McDonald b ) for analyses of local graph theory measures.
* Significant p value after false discovery rate correction.
NBS
The BD group displayed significantly weaker sub-network connectivity, with a single dysconnected sub-network identified for each threshold (2, 2.5, 3). The NBS provides two outputs: (a) the supra-threshold set of connections comprised in the graph component found to show a significant effect; as well as (b) a corresponding p value for each such network (Zalesky et al. Reference Zalesky, Fornito and Bullmore2010) (Fig. 2). We identified collective network dysconnectivity differences with supra-threshold connections (t = 2, p = 0.015), consisting of frontal, parietal and occipital connections in BD. Higher supra-threshold connections (t = 2.5, p = 0.017; and t = 3, p = 0.020) revealed structural dysconnectivity among parietal and occipital connections in patients compared with healthy controls.
Fig. 2. Density of nodal connections in participant groups determined by network-based statistic (NBS) analysis. Nodal alterations identified in NBS toolbox, with colour and size of sphere indicating increasing t-statistic thresholds. Dark blue nodes indicate Automated Anatomical Labeling (AAL) nodes which did not significantly differ at any threshold. Light blue nodes (T = 2) identify a reduction in density of connections in the bipolar disorder (BD) group compared with the control group. Similarly, yellow nodes (T = 2.5) and red nodes (T = 3) represent reduced density of connections in the BD group at higher thresholds. (a) Axial orientation; (b) coronal orientation; (c) left sagittal orientation; (d) right sagittal orientation.
Rich-club connectivity
In this analysis, rich-club organization was evident across a range of connection densities. The weighted rich-club connection density ranged from 27 to 64 possible densities, while the normalized weighted rich-club coefficient ranged from 28 to 56. The group comparison of normalized weighted rich-club connectivity effects demonstrated statistically significant differences before FDR multiple comparison correction across two possible connection densities (k = 55, Z = −2.236, p = 0.024; k = 56, Z = −2.654, p = 0.0067). After FDR correction for 28 possible densities, the highest density (k = 56) demonstrated a moderate to large effect size (Cohen's d = 0.59).
Rich-club membership
Following examination of rich-club connectivity effects, we investigated rich-club membership. Rich-club structures were selected by the group-averaged cortico-subcortical network for the statistically significant Фw norm(k) and identified nodes connected by this pathway. Rich-club connections shared by 60% of participants indicate differential hub participation between groups. Although we identified rich-club members connected by pathways at the 60% group threshold, we applied a more stringent threshold of 70% to identify what findings were consistent. Previously, pathways present in >50% of participants were taken into account (van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011). Additionally, the top 10 highest degree nodes were identified to validate the threshold selection (van den Heuvel & Sporns, Reference van den Heuvel and Sporns2011; Collin et al. Reference Collin, Kahn, De Reus, Cahn and Van Den Heuvel2014).
Rich-club members revealed the following hubs: superior frontal gyrus, middle cingulate gyrus, hippocampus, caudate, precuneus and thalamus (Fig. 3). In BD, rich-club membership at the group threshold of 60% indicated that the right frontal superior gyrus and right middle cingulate gyrus were not included in the rich-club in most BD patients; however, BD rich-club structures incorporate the left middle occipital gyrus. Of interest, when we assessed rich-club membership in 70% of participants, the most notable between-group differences in the BD group supported the absence of the right superior frontal gyrus and left thalamus as hubs, and additionally the left middle occipital gyrus was no longer integrated in patients (Fig. 3). Of note, dysconnections identified by the NBS analysis overlap with rich-club members, namely the left hippocampus, precuneus and thalamus.
Fig. 3. Rich-club membership in patients and controls. Rich-club membership in the participants with differences between groups colour coded. All nodes presented represent rich-club members common to all healthy volunteers, while nodes in red indicate structures absent in the rich-club in the bipolar group, i.e. right superior frontal gyrus (SFG) and left thalamus (Thal). Nodes in green represent rich-club structures recruited in the bipolar group, not present in the healthy control group. Size of sphere relates to rich-club structures common to 60% of participants, larger spheres common to 95–100% of participants. Cau, Caudate; Put, putamen; MCG, middle cingulate gyrus; Hip, hippocampus; P.Cun, precuneus; MOG, middle occipital gyrus; L, left; R, right.
Partial correlations show that rich-club connectivity is associated with global efficiency in all participants (r = 0.299, p < 0.006) (online Supplementary Fig. S1). Taken together, the normalized rich-club coefficient revealed trend connectivity effects as well as rich-club membership differences. Global efficiency appears to be related to rich-club connectivity, supporting the role of rich-club connections in global integration.
Clinical associations
Partial correlations did not reveal any significant association between graph theory properties and the clinical measures assessed: age of onset, illness duration, whether patients were taking lithium at the time of scan or had previously taken lithium, and years of medication use.
Discussion
This study provides novel evidence of distributed neuroanatomical dysconnectivity, using a range of graph theory metrics, as a trait-based feature of BD. In summary, we identified disrupted anatomical integration and regional segregation in fronto-limbic and posterior parietal regions and reduced inter-connection density of sub-networks that converged among posterior parietal circuits in the BD group compared with healthy volunteers. We also demonstrated absence of rich-club structures in BD, most consistently the right superior frontal gyrus. Therefore, this study confirms disrupted global anatomical integration and represents the first study to show impaired frontal hub integration using graph analysis in BD.
This investigation is consistent with three studies that identified reduced global efficiency in BD (Leow et al. Reference Leow, Ajilore, Zhan, Arienzo, GadElkarim, Zhang, Moody, Van Horn, Feusner, Kumar, Thompson and Altshuler2013; Gadelkarim et al. Reference Gadelkarim, Ajilore, Schonfeld, Zhan, Thompson, Feusner, Kumar, Altshuler and Leow2014; Collin et al. Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016). However, two other studies determined connectivity to be preserved globally (Forde et al. Reference Forde, O'Donoghue, Scanlon, Emsell, Chaddock, Leemans, Jeurissen, Barker, Cannon, Murray and McDonald2015; Wheeler et al. Reference Wheeler, Wessa, Szeszko, Foussias, Chakravarty, Lerch, DeRosse, Remington, Mulsant, Linke, Malhotra and Voineskos2015). Potential differences in results between the current investigation and other studies may be due to the parcellation scheme employed. In contrast to the present analysis which used volumetric parcellation, one study used cortical thickness to define nodes and did not demonstrate differences in connection density between healthy individuals and BD in a multi-centre study (Wheeler et al. Reference Wheeler, Wessa, Szeszko, Foussias, Chakravarty, Lerch, DeRosse, Remington, Mulsant, Linke, Malhotra and Voineskos2015). Furthermore, in contrast with the investigation by Forde et al. (Reference Forde, O'Donoghue, Scanlon, Emsell, Chaddock, Leemans, Jeurissen, Barker, Cannon, Murray and McDonald2015) that weighted connections by streamline count, we identified reductions in global efficiency and average clustering coefficient when connections were weighted by FA. Impaired global integration may be a general pathophysiological feature across brain disorders (Crossley et al. Reference Crossley, Mechelli, Scott, Carletti, Fox, McGuire and Bullmore2014; Fornito et al. Reference Fornito, Zalesky and Breakspear2015); therefore, regional connectivity may elucidate topological patterns involved in emotional dysregulation and its effect on whole-brain integration.
Uncorrected regional analyses of clustering and local efficiency identified widespread reductions among fronto-limbic, parietal and occipital connections, consistent with previous studies (Leow et al. Reference Leow, Ajilore, Zhan, Arienzo, GadElkarim, Zhang, Moody, Van Horn, Feusner, Kumar, Thompson and Altshuler2013; Gadelkarim et al. Reference Gadelkarim, Ajilore, Schonfeld, Zhan, Thompson, Feusner, Kumar, Altshuler and Leow2014; Forde et al. Reference Forde, O'Donoghue, Scanlon, Emsell, Chaddock, Leemans, Jeurissen, Barker, Cannon, Murray and McDonald2015). Furthermore, regional dysconnectivity of the left middle frontal gyrus and right superior medial frontal gyrus has been supported in a recent investigation in patients from families multiply affected with BD (Forde et al. Reference Forde, O'Donoghue, Scanlon, Emsell, Chaddock, Leemans, Jeurissen, Barker, Cannon, Murray and McDonald2015). Similarly, the present dataset demonstrates callosal dysconnectivity due to regional frontal patterns of disorganization. Additionally, parietal and default mode network dysconnectivity identified by regional and network-based analysis is corroborated by an investigation of path length-associated community estimation (Gadelkarim et al. Reference Gadelkarim, Ajilore, Schonfeld, Zhan, Thompson, Feusner, Kumar, Altshuler and Leow2014). This analysis indicates that connectivity abnormalities in BD appear to extend beyond fronto-limbic regions to other association areas of the brain (Vederine et al. Reference Vederine, Wessa, Leboyer and Houenou2011; Nortje et al. Reference Nortje, Stein, Radua, Mataix-Cols and Horn2013; Wise et al. Reference Wise, Radua, Nortje, Cleare, Young and Arnone2016). These reductions may be related to features of topological dysconnectivity as opposed to reduced connectivity globally, as findings were unchanged when corrected for global connectivity.
Furthermore, the NBS measure identified a collection of interconnections encompassing fronto-limbic and parietal/occipital connections (Gadelkarim et al. Reference Gadelkarim, Ajilore, Schonfeld, Zhan, Thompson, Feusner, Kumar, Altshuler and Leow2014). Moreover, weaker connectivity of network components in patients indicates that these impaired connections interact collectively, supporting BD as a dysconnection syndrome (O'Donoghue et al. Reference O'Donoghue, Cannon, Perlini, Brambilla and McDonald2015). Analysis of highest supra-threshold connections revealed dysconnectivity among the cuneus, precuneus and superior occipital connections. Interestingly, evidence from functional connectivity analyses supports a model of an affected posterior default mode network as well as parieto-occipital dysconnectivity (Strakowski et al. Reference Strakowski, Delbello and Adler2005, Reference Strakowski, Eliassen, Lamy, Cerullo, Allendorfer, Madore, Lee, Welge, Delbello, Fleck and Adler2011). Strakowski et al. (Reference Strakowski, DelBello, Adler, Cecil and Sax2000, Reference Strakowski, Adler and DelBello2002) proposed that self-referential thinking and interpretation of visual stimuli are affected in disturbances of this network, namely altered functional connectivity of the precuneus and cuneus. These structures are topologically central with high degree, which may be particularly affected in BD (Crossley et al. Reference Crossley, Mechelli, Scott, Carletti, Fox, McGuire and Bullmore2014; Gadelkarim et al. Reference Gadelkarim, Ajilore, Schonfeld, Zhan, Thompson, Feusner, Kumar, Altshuler and Leow2014).
Investigation of rich-club connectivity effects shows trend-level reductions in BD. The associated relationship between global efficiency and rich-club density suggests that widespread neuroanatomical dysconnectivity may be related to communication among these central structures (van den Heuvel et al. Reference van den Heuvel, Sporns, Collin, Scheewe, Mandl, Cahn, Goñi, Hulshoff Pol and Kahn2013; Crossley et al. Reference Crossley, Mechelli, Scott, Carletti, Fox, McGuire and Bullmore2014). Recent studies report reduced rich-club connectivity in schizophrenia and unaffected siblings with schizophrenia (van den Heuvel et al. Reference van den Heuvel, Sporns, Collin, Scheewe, Mandl, Cahn, Goñi, Hulshoff Pol and Kahn2013; Collin et al. Reference Collin, Kahn, De Reus, Cahn and Van Den Heuvel2014, Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016). In BD, one study to date has reported preserved rich-club connections (Collin et al. Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016). Results of the present cortico-subcortical rich-club connectivity analysis are in contrast to this study which was confined to cortico-cortical connections (Collin et al. Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016). The significance of the limbic system in BD argues for the inclusion of subcortical limbic structures in network mapping. Differences in subcortical volume in BD has a substantial body of literature to support its role in the aetiology and as a trait feature of the illness (Houenou et al. Reference Houenou, d'Albis, Vederine, Henry, Leboyer and Wessa2012; Strakowski et al. Reference Strakowski, Adler, Almeida, Altshuler, Blumberg, Chang, DelBello, Frangou, McIntosh, Phillips, Sussman and Townsend2012; Vargas et al. Reference Vargas, López-Jaramillo and Vieta2013; Quigley et al. Reference Quigley, Scanlon, Kilmartin, Emsell, Langan, Hallahan, Murray, Waters, Waldron, Hehir, Casey, McDermott, Ridge, Kenney, Donoghue, Nannery, Ambati, McCarthy, Barker and Cannon2015; Hibar et al. Reference Hibar, Westlye, van Erp, Rasmussen, Leonardo, Faskowitz, Haukvik, Hartberg, Doan, Agartz, Dale, Gruber, Krämer, Trost, Liberg, Abé, Ekman, Ingvar, Landén, Fears, Freimer, Bearden, Sprooten, Glahn, Pearlson, Emsell, Kenney, Scanlon, McDonald, Cannon, Almeida, Versace, Caseras, Lawrence, Phillips, Dima, Delvecchio, Frangou, Satterthwaite, Wolf, Houenou, Henry, Malt, Bøen, Elvsåshagen, Young, Lloyd, Goodwin, MacKay, Bourne, Bilderbeck, Abramovic, Boks, van Haren, Ophoff, Kahn, Bauer, Pfennig, Alda, Hajek, Mwangi, Soares, Nickson, Dimitrova, Sussmann, Hagenaars, Whalley, McIntosh, Thompson and Andreassen2016; van Erp et al. Reference van Erp, Hibar, Rasmussen, Glahn, Pearlson, Andreassen, Agartz, Westlye, Haukvik, Dale, Melle, Hartberg, Gruber, Kraemer, Zilles, Donohoe, Kelly, McDonald, Morris, Cannon, Corvin, Machielsen, Koenders, de Haan, Veltman, Satterthwaite, Wolf, Gur, Gur, Potkin, Mathalon, Mueller, Preda, Macciardi, Ehrlich, Walton, Hass, Calhoun, Bockholt, Sponheim, Shoemaker, van Haren, Pol, Ophoff, Kahn, Roiz-Santiañez, Crespo-Facorro, Wang, Alpert, Jönsson, Dimitrova, Bois, Whalley, McIntosh, Lawrie, Hashimoto, Thompson and Turner2016). Additionally, during tract reconstruction we implemented constrained spherical deconvolution to overcome the challenges of white matter reconstruction in subcortical areas.
Next, we address the consistency in rich-club structures and hubs affected in BD. Therefore, this investigation examined regions connected by the statistically significant pathway and detected hub involvement specific to BD. The investigation by Collin et al. (Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016) examined additional rich-club classifications using the top 20% betweenness-centrality nodes and top 10% highest degree nodes. These hubs consisted of portions of the bilateral cingulate, precuneus, superior frontal, parietal and temporal gyri, as well as pre- and post-central gyri and insular cortices (Collin et al. Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016). In the present study, we defined hubs connecting pathways present in most participants and validated by top 10% highest degree nodes, which confirmed these rich-club structures. Cingulate, precuneus and superior frontal structures appear to be consistent across investigations and hub definitions, potentially due to above-average connectivity and participation across both cortical and cortico-subcortical mappings (van den Heuvel et al. Reference van den Heuvel, Sporns, Collin, Scheewe, Mandl, Cahn, Goñi, Hulshoff Pol and Kahn2013; Collin et al. Reference Collin, Kahn, De Reus, Cahn and Van Den Heuvel2014, Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016). Additionally, rich-club members in this analysis were identified as hubs in a study of patients with schizophrenia, defined by betweenness centrality, when network analysis implemented the equivalent structural atlas (van den Heuvel et al. Reference van den Heuvel, Mandl, Stam, Kahn and Hulshoff Pol2010). Validation of this parcellation scheme merits reproducibility of these structures as critical hubs in cortico-subcortical networks.
In the current study, the superior frontal gyrus was absent from the BD group rich-club, suggesting differential involvement of this densely interconnected frontal structure. Hub deficits reported in a study of patients with schizophrenia support rich-club members affected in the current analysis, consistent with a less central role of frontal hubs in psychotic illnesses (van den Heuvel et al. Reference van den Heuvel, Mandl, Stam, Kahn and Hulshoff Pol2010). This absence was maintained when the rich-club pathway was defined among group thresholds of 60% of patients and 70% of patients. Models of neural circuitry in BD suggest that decreased grey matter volume and aberrant functional connectivity in the right ventrolateral prefrontal cortex are features of impaired emotional processing and regulation in BD (Phillips & Swartz, Reference Phillips and Swartz2014). Nodes anatomically connected with the superior frontal gyrus, also present in the rich-club network, include the caudate and thalamus (Haznedar et al. Reference Haznedar, Roversi, Pallanti, Baldini-Rossi, Schnur, Licalzi, Tang, Hof, Hollander and Buchsbaum2005; Li et al. Reference Li, Qin, Liu, Fan, Wang, Jiang and Yu2013). These deficits are consistent with a DTI study reporting reduced FA in the anterior thalamic radiation in both BD and schizophrenia patients (Sussmann et al. Reference Sussmann, Lymer, McKirdy, Moorhead, Muñoz Maniega, Job, Hall, Bastin, Johnstone, Lawrie and McIntosh2009). While thalamic function has been implicated previously in BD, volumetric analyses of the thalamus have been varied (Hallahan et al. Reference Hallahan, Newell, Soares, Brambilla, Strakowski, Fleck, Kieseppä, Altshuler, Fornito, Malhi, McIntosh, Yurgelun-Todd, Labar, Sharma, MacQueen, Murray and McDonald2011). Rich-club members presented in the 60% group threshold indicated above-average connectivity of the right middle cingulate gyrus in healthy controls, absent in the BD group. Moreover, the BD group rich-club additionally involves the left middle occipital gyrus, which may potentially represent a compensatory effect from disrupted frontal connectivity (Griffa et al. Reference Griffa, Baumann, Thiran and Hagmann2013). Differential involvement of rich-club structures and reduced connection density between these rich-club structures may well contribute to the global connectivity effects identified in this study.
Reduced local efficiency in prefrontal regions, anterior cingulate cortex, and absence of the superior frontal gyrus as a rich-club hub is consistent with impaired top-down connectivity of prefrontal and ventral–limbic emotional circuits as suggested by Strakowski et al. (Reference Strakowski, Adler, Almeida, Altshuler, Blumberg, Chang, DelBello, Frangou, McIntosh, Phillips, Sussman and Townsend2012) and Phillips & Swartz (Reference Phillips and Swartz2014). Aberrant connectivity of the amygdala was not identified in this study, although it has been considered a characteristic marker underlying emotional dysregulation (Strakowski et al. Reference Strakowski, Adler, Almeida, Altshuler, Blumberg, Chang, DelBello, Frangou, McIntosh, Phillips, Sussman and Townsend2012). Some reports have shown that amygdala activity and volume ‘normalize’ when patients are in remission (Hallahan et al. Reference Hallahan, Newell, Soares, Brambilla, Strakowski, Fleck, Kieseppä, Altshuler, Fornito, Malhi, McIntosh, Yurgelun-Todd, Labar, Sharma, MacQueen, Murray and McDonald2011; Phillips & Swartz, Reference Phillips and Swartz2014) and may account for the lack of significant dysconnectivity of this important limbic node. This investigation extends current models of BD with evidence of more widespread and posterior abnormalities, suggesting that dysconnectivity extends beyond these prefrontal networks. Notably, a meta-analysis of DTI studies in BD also suggests that posterior white matter dysconnectivity may be related to impaired cognitive functioning that persists in patients with BD (Nortje et al. Reference Nortje, Stein, Radua, Mataix-Cols and Horn2013).
Inter-hemispheric dysconnectivity was not specifically examined in this investigation, where it has been supported as a feature of BD in previous structural network studies (Leow et al. Reference Leow, Ajilore, Zhan, Arienzo, GadElkarim, Zhang, Moody, Van Horn, Feusner, Kumar, Thompson and Altshuler2013; Caeyenberghs & Leemans, Reference Caeyenberghs and Leemans2014; Gadelkarim et al. Reference Gadelkarim, Ajilore, Schonfeld, Zhan, Thompson, Feusner, Kumar, Altshuler and Leow2014; Collin et al. Reference Collin, van den Heuvel, Abramovic, Vreeker, de Reus, van Haren, Boks, Ophoff and Kahn2016).
Methodological selections must be considered when interpreting these network findings (de Reus & van den Heuvel, Reference de Reus and van den Heuvel2013; Fornito et al. Reference Fornito, Zalesky and Breakspear2013). As network analyses lack standardized recommendations and methodological criteria at this point; this network reconstruction carries challenges when interpreting and reconciling results across investigations (Fornito et al. Reference Fornito, Zalesky and Breakspear2013). Methodological considerations in this novel and evolving field include the specific choice of parameters for white matter tract reconstruction, crossing fibres, edge weights of FA measures and streamline count. Advancing from previous research by use of more biologically relevant connection weights as well as a template cortical parcellation may identify less variable effects (Fornito et al. Reference Fornito, Zalesky and Breakspear2013). The AAL template employs an identical cortico-subcortical parcellation atlas for each participant; therefore, subject-specific cortical and subcortical volume changes were not accounted for in this analysis. Subject-specific parcellation schemes have their own inconsistencies, e.g. FreeSurfer-generated nodes may be implemented with or without manual correction of the boundaries of subcortical segmentation (McCarthy et al. Reference McCarthy, Ramprashad, Thompson, Botti, Coman and Kates2015). Extension of this work would improve from a subject-specific parcellation technique integrated in network analysis reliably. The field would benefit from some degree of standardization in these approaches, which would assist in directly comparing results as they emerge from research groups.
A strength of the current analysis is the parcellation scheme employed in the brain mapping pipeline. A majority of complex network analyses limit their connectome maps to cortical connection maps, while this analysis extended to cortico-subcortical mapping. The scale at which the brain should be accurately mapped to be most biologically meaningful is not yet standardized (Fornito et al. Reference Fornito, Zalesky and Breakspear2013). Additionally, we attempt to explain hub participation differences in BD through examination of rich-club membership. The heterogeneity of network analysis techniques and graph properties available makes comparison between studies and meta-analyses challenging. The present investigation utilized certain commonly employed graph metrics, such as global efficiency and clustering coefficient, in order to facilitate comparison of our results with those of other studies. These graph properties were selected as they specifically quantify a node's influence in network integration and segregation. In addition, certain more novel metrics, involving sub-network and rich-club connectivity, were employed given their recent application to other disorders and the convergent evidence for regional neuroanatomical dysconnectivity as a core feature of BD.
Taken together, these analytical methods support previous DTI investigations, and extend further understanding of structural dysconnectivity in BD. The relationship of graph measures to pathophysiological mechanisms is an active area of examination (Fornito et al. Reference Fornito, Zalesky and Breakspear2013). The relevance of this study indicates dysconnectivity to be a pathophysiologically relevant trait-related feature of BD, with differentially affected rich-club structures being particularly informative to describe complex structural integration.
Conclusion
This multifaceted analysis employing graph theory metrics provides substantial additional evidence for anatomical dysconnectivity representing a trait feature of BD. This study supports reductions in global efficiency and local connectedness of limbic structures, and extends initial investigations of BD sub-networks, identifying weaker connected components incorporating anterior and posterior brain networks representing trait features of BD.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291717000058
Acknowledgements
This research was supported by the Hardiman Research Scholarship, National University of Ireland, Galway, Galway Ireland and the Health Research Board (HRA_POR/2011/100). We gratefully acknowledge the participants of this study and the radiographers of the University Hospital Galway Magnetic Resonance Imaging Department for their support during data collection.
The research of A.L. is supported by VIDI grant 639.072.411 from the Netherlands Organization for Scientific Research (NWO). B.J. is a postdoctoral fellow supported by the Research Foundation Flanders (FWO Vlaanderen). G.J.B. receives honoraria for teaching from General Electric Healthcare, and acts as a consultant for IXICO.
Declaration of Interest
None.
