Salience network structural integrity predicts executive impairment in alcohol use disorders

The neural bases of cognitive impairment(s) in alcohol use disorders (AUDs) might reflect either a global brain damage underlying different neuro-cognitive alterations, or the involvement of specific regions mostly affected by alcohol neuro-toxic effects. While voxel-based-morphometry (VBM) studies have shown a distributed atrophic pattern in fronto-limbic and cerebellar structures, the lack of comprehensive neuro-cognitive assessments prevents previous studies from drawing robust inferences on the specificity of the association between neuro-structural and cognitive impairments in AUDs. To fill this gap, we addressed the neuro-structural bases of cognitive impairment in AUDs, by coupling VBM with an in-depth neuropsychological assessment. VBM results highlighted a diffuse pattern of grey matter reduction in patients, involving the key-nodes of the meso-cortico-limbic (striatum, hippocampus, medial prefrontal cortex), salience (insular and dorsal anterior cingulate cortex) and executive (inferior frontal cortex) networks. Grey matter density in the insular and anterior cingulate sectors of the salience network, significantly decreased in patients, explained almost half of variability in their defective attentional and working-memory performance. The multiple cognitive and neurological impairments observed in AUDs might thus reflect a specific executive deficit associated with the selective damage of a salience-based neural mechanism enhancing access to cognitive resources required for controlled cognition and behaviour.

TMT-B)), memory (digit span, immediate and delayed prose memory), working-memory (10-and 30seconds interference-memory), executive functions (TMT-B, cognitive estimation, abstract reasoning, phonemic fluency, clock drawing, and overlapping pictures), as well as perceptive and praxis skills (praxis abilities, spontaneous drawing and copy drawing tests). The TMT-B test is listed twice because it involves two domains, i.e. attention and executive functioning (particularly working-memory and switching). The battery results in a score for every task, alongside an overall score of global cognitive status.

Analysis of neuro-cognitive data
For each test, we first checked the normality of the score distribution across the whole sample. Based on the results of this assessment, we then examined age and group effects by means of parametric (Pearson's correlation index and two-sample t-test, respectively) or non-parametric (Spearman's correlation index or Mann-Whitney t-test, respectively) statistical tests. For the tests showing a significant effect of age or education, we additionally run an Analysis of covariance (ANCOVA) to assess group differences on cognitive performance after removing their effect. We applied a primary statistical threshold of p<0.05, one-tailed due to a priori hypotheses of cognitive impairment in AUDs [2][3] (see Introduction), and then performed a correction for multiple comparisons based on FDR.
We investigated superordinate cognitive domains, transcending specific tasks, in which performance was impaired in patients. After assessing the suitability of the correlation matrix (Keiser-Meyer-Olkin Measure of Sampling Adequacy = 0.61; Bartlett's test of sphericity < 0.001; Supplementary  Table S7) we performed a principal component analysis on the 15 ENB2 raw scores. Due to the ambiguity of the scree plot (Supplementary Figure 1), we determined the number of components to be retained by applying the Kaiser-Guttman criterion (i.e., components with eigenvalue>1) (Supplementary Tables S8-S9). An orthogonal rotation (Varimax) was used to facilitate the interpretation of components [4][5]. To investigate group differences in cognitive performance, we used an ANOVA (with Bonferroni correction for multiple comparisons) on the resulting factor scores for each subject/component.

MRI data acquisition
We used a 3 Tesla General Electrics Discovery scanner (GE Healthcare, Milwaukee, WI), equipped with an 8channels head coil, to collect a high-resolution 3D T1-weighted IR-prepared FSPGR (BRAVO) brain scan (TR=8.2 ms, TE=3.2 ms, FA=12°, TI=450 ms, NEX=1, FOV=24 cm, reconstruction matrix=256 x 256, slice thickness=1 mm, 152 slices) acquired along a semi-axial orientation parallel to the AC-PC plane, covering the whole brain from the occipital foramen to the region above the vertex. A T2-weighted image was also collected, and interpreted by an experienced neuroradiologist for diagnostic purposes.
All the T1-weighted images were first inspected for motion artefacts and gross anatomical abnormalities, and the origin of the plane was set on the anterior commissure. By means of the CAT12 toolbox, images were then corrected for bias-field inhomogeneities, spatially normalized using the DARTEL algorithm [6] and segmented into GM, white matter (WM) and cerebrospinal fluid (CSF) [7]. Compared with the SPM12 pipeline, CAT12 improves the segmentation process via the estimation of partial volume effects [8], adaptive maximum a posteriori estimations [9] and a hidden Markov Random Field model [10]. We did not apply the jacobian modulation of segmented GM images, which corrects for volume change during spatial normalization, since this procedure has been shown to decrease the sensitivity to morphometric abnormalities [11]. Our results thus involve GM density, i.e. GM volume relative to WM and CSF volume. Finally, we applied a smoothing kernel of 8 mm (FWHM) to the normalized segmented GM images.
After preprocessing, all images underwent both a visual check for artefacts and an automated check of sample homogeneity based on the degree of correlation across images. Both procedures led to retain all images, since no clear outlier was identified. In addition, we found no group difference in terms of two quality parameters describing the image properties before CAT12 processing, i.e. "noise contrast ratio" (p=0.1919) and "weighted average image quality rating" (IQR; p=0.1960).
We performed whole-brain statistical analyses by means of the General Linear Model approach, as implemented in SPM12. Namely, we modeled: a) a two-sample t-test, to assess a significant decrease of GM density in alcoholic patients vs. controls; b) multiple regressions, to assess a significant relationship between GM density and performance in the domain/task displaying the greatest impairment in patients; c) full factorial models (two-sample t-test plus a behavioral covariate) to assess significant group differences in the relationship between GM density and performance (i.e. a significantly different regression slope in in patients vs. controls). For both b) and c), in separate analyses we modelled either the factor score of the basic-level executive component, or response-time of the TMT-A task to examine its contribution to the component. Multiple regressions and full factorial models highlighted, respectively, quantitative or qualitative group differences in the relationship between GM density and domain/task performance. We modeled age to remove its potentially confounding effect. We modelled age to remove its potentially confounding effect, and applied an internal GM threshold of 0.15 to prevent artefacts on the GM-WM border due to voxel misclassification. We used conjunction-null analyses [12] to assess the predicted anatomical overlap between the regions in which GM density was both reduced in patients vs. controls, and related to executive or TMT-A performance.
Since the above analyses involved two behavioural measures, we adjusted our primary statistical threshold to p<0.025 corrected for multiple comparisons with FDR (as implemented in SPM12) at the voxel or cluster level. We applied threshold-free cluster enhancement (TFCE; [13]) with 5000 permutations per contrast and correction for multiple comparisons. This approach has been shown to increase the sensitivity of VBM findings [14].

Region-of-Interest analyses
We then aimed to investigate whether, and to what extent, the pattern of cognitive impairment observed in alcoholic patients is explained by the degree of regional GM atrophy. To this purpose, we first used the SPM toolbox Marsbar (http://marsbar.sourceforge.net/) to create binary masks of the clusters displaying the different main effects reported above, i.e. a) GM atrophy in patients vs. controls; b) common effect of GM atrophy and correlation with basic-level executive performance. Using the SPM toolbox REX (http://web.mit.edu/swg/software.htm), the average GM density in these regions was extracted for each subject, and entered in offline analyses. Namely, we used average GM density in the observed clusters as simultaneous predictors of a multiple regression model, to assess their global and relative efficacy for predicting executive performance.

Overlap with the salience attentional network
Based on the results of the above analyses, we examined the spatial overlap between the regions displaying common effects of interest (i.e. atrophy in patients, correlation with executive and TMT-A performance) and those included in the salience network [15] in charge of switching between the default mode and executive control networks [16][17]. We grounded this investigation in the meta-analytic approach implemented by the Neurosynth toolbox (http://neurosynth.org), which allows to quantify the specificity and consistency of brain activity in association with a given cognitive process [18].
The current Neurosynth database includes coordinates from 11406 published studies, associated with over 413429 reported activations, automatically extracted from the available literature regardless of specific brain regions or processes of interest. This information is used to generate meta-analytic maps for several thousand psychological terms [19], i.e. 3107 as of January 2018, which can be accessed via a search interface displaying all studies containing a specified keyword. This procedure implements two types of brain-process inferences, i.e. forward (probability of observing activity in a region given knowledge of the psychological process) and reverse (probability of a psychological process being present given knowledge of activation in a specific brain region) [20]. These two approaches highlight, respectively, how consistently a task/state activates a region, and how specific its activation is to the task/state, with the latter being closer to the notion of decoding mental states from brain activity [21].
We used the Neurosynth interface to produce reverse inference maps associated with the "salience network". The resulting spatial map, based on 60 previous studies and 2327 activations, was first used to assess the spatial overlap between this network and the regions underlying the predicted effects of interest in AUDs. To this purpose, we used the Marsbar toolbox, as described above, to create spatial maps corresponding to the conjunction of all our four effects of interest, i.e. inclusion in the salience network alongside significant GM atrophy in AUDs, correlation with executive performance and with TMT-A response time. Then, to evaluate the extent to which the morphometric properties of the resulting brain regions accounts for cognitive performance, we replicated the procedure described above to extract average GM density from the commonly involved voxels, for subsequent offline multiple regression analyses.

SUPPLEMENTARY TABLES S1-S9
Supplementary The   The table shows the correlations between the ENB variables and the estimated components after the Varimax rotation. Correlation coefficients < 0.3 are not reported.