Regional gray matter reductions associated with mild cognitive impairment in Parkinson’s disease: A meta-analysis of voxel-based morphometry studies

Mild cognitive impairment (MCI) is inconclusively associated with regional gray matter (GM) abnormalities in Parkinson's disease (PD). We aimed to quantitatively evaluate whole-brain voxel-based morphometry (VBM) studies that have investigated brain GM changes in PD patients with MCI (PD-MCI). Seed-based d Mapping, a well-validated coordinate-based meta-analytic approach, was utilized. We included 20 VBM studies that reported 22 datasets containing 504 patients with PD-MCI and 554 PD patients without MCI (PD-NCI). The most reliable finding identified in this meta-analysis was that patients with PD-MCI exhibited greater GM atrophy in the left anterior insula than those with PD-NCI. Our findings further suggest that several moderators (age, gender, educational level, disease stage, severity of motor disability, and the severity of cognitive impairments) in PD-MCI individuals, as well as scanner field-strength, may drive heterogeneous GM changes across studies. GM abnormalities in the anterior insula, an important cognitive hub involved in switching between neural networks, contribute to understanding the neural substrates of MCI in PD, which may serve as a biomarker of PD-MCI.


Introduction
Mild cognitive impairment (MCI) is a common non-motor symptom of Parkinson's disease (PD) which, even early in the course of the illness, affects approximately 27% of nondemented PD patients [1]. PD with MCI (PD-MCI) is diagnosed with a subjective complaint of cognitive decline that is not normal for age but accompanies essentially intact functional activities and which commonly manifests as impairments in executive abilities, attention, visuospatial skills, and memory, individually or in combination [2]. MCI in PD significantly affects patients' and their caregivers' quality of life [3] and causes significant societal and financial impacts [4]. PD-MCI increases substantially the risk for development of PD dementia (PDD), and is thus considered a transitional stage prior to PDD [5][6][7][8]. Interest in the early identification of individuals at risk of developing PDD, which may help to deliver more timely interventions, has grown over the past decade. However, much remains unknown about the neurobiological dysfunctions underlying MCI in PD [9].
In the last two decades, neuroimaging techniques have improved our understanding of the pathophysiology of cognitive impairments in Table 1 Characteristics of studies included in the meta-analysis.  patients with PD and may also help to establish potential biomarkers for the disease [10]. Among the available methods, voxel-based morphometry (VBM) is a well-established structural magnetic resonance imaging (MRI) technique used for the detection of between-group differences in brain anatomy at the whole-brain level. VBM analysis is an automated approach that does not require a priori definitions of anatomical borders, thus affording an opportunity for discovery of previously unidentified structural changes [11]. This non-biased nature has encouraged the wide-spread use of VBM in the quantification of regional gray matter (GM) changes in PD. Several meta-analyses of VBM studies have identified GM deficits in the frontal, parietal, occipital, and insular lobes in patients with PD relative to healthy controls [12][13][14].
Despite some consistent neural signals, results from VBM studies regarding MCI-related GM alterations in patients with PD remain heterogeneous, in part due to clinical and neuropathological heterogeneity. This renders interpretation of VBM-based results challenging. Multiple studies have reported MCI-specific GM reductions in patients with PD-MCI relative to patients without MCI (PD-NCI) in both cortical and subcortical regions [15][16][17][18]. In contrast, other VBM studies have failed to detect any significant MCI-linked GM differences [19][20][21].
To achieve consistent findings, Xu and colleagues conducted a stimulating coordinate-based meta-analysis (CBMA). This analysis included five VBM studies published prior to December 2015 [22]. The authors found that patients with PD-MCI relative to patients with PD-NC exhibited GM reductions in the left superior temporal lobe extending to the left insula, and the left superior frontal lobe. Of note, these results should be regarded as preliminary due to the small number of VBM studies included. Additional VBM studies in PD-MCI have been published subsequently. Inclusion of more studies in a CBMA increases its statistical power [23] and allows for more comprehensive, complementary analyses of multiple findings. Critically, the exploration of sources of heterogeneity [24], which is critical to form more accurate conclusions [25], is possible only with greater statistical power.
To identify the most consistent and robust GM alternations that are associated with cognitive impairment in patients with PD-MCI (versus PD-NCI participants), we aimed to conduct an updated CBMA of VBM studies by applying seed-based d mapping (SDM), a CBMA technique [24,26], that has been fully validated in previous neuroimaging studies of the alterations of brain function or structure across neuropsychiatric disorders [27][28][29][30]. In addition, we aimed to explore the potential confounding effects of age, gender, education level, disease duration, disease stage, severity of motor disability, and severity of cognitive impairment on GM alterations, which had not been examined previously.

Search strategies and eligibility criteria
The PubMed, Web of Science, and Embase electronic databases were systematically searched on March 3, 2018 and again on February 1, 2018. The following predefined keywords were used to identify pertinent studies: ("voxel-based morphometry" OR "vbm" OR "gray matter" OR "grey matter" OR "voxel*" OR "structural magnetic resonance imaging") AND ((("mild cognitive impairment" OR "mci" OR "cognitive disorder" OR "neurocognitive disorder" OR "cognitive decline" OR "cognition") AND ("Parkinson's disease" or "Parkinson")) or "PD-MCI"). Reference lists from relevant reviews and original articles were also manually checked for additional eligible articles not indexed by the database search. We did not place restrictions on language or date.
A study was included if it met the following criteria: 1) was an original research paper published in a peer-reviewed journal; 2) involved patients with PD-MCI and patients with PD-NCI that met the agreed diagnostic criteria [1,2,31-38]; 3) employed a whole-brain VBM analysis to investigate GM volume or density difference between patients with PD-MCI and those with PD-NCI; and 4) reported stereotactic coordinates and space used, per Montreal Neurological Institute (MNI) or Talairach space. Studies that did not meet the above criteria or met any of the following exclusion criteria were not included: 1) it primarily included subjects that did not satisfy the criteria for PD-MCI or PD-NCI; 2) it applied a region of interest (ROI) approach; 3) neuroanatomical coordinates were not available even after contacting the authors; and 4) less than seven patients were included in each group [39]. The study with the largest group size was included in case where two or more studies included identical PD-MCI samples.

Data extraction
The following information was extracted from each study included: first author's family name, study year, diagnosis (PD-MCI or PD-NCI), sample size, number of male participants, mean age, mean education level, mean Unified Parkinson's Disease Rating Scale, Part III (UPDRS-III) score, mean Hoehn and Yahr (H&Y) stage, medication state, mean illness duration, mean Mini-Mental State Examination (MMSE) score, and diagnostic criteria for PD-MCI ( Table 1). The imaging-specific data, such as MR field strength, MR sequence, voxel size/slice thickness, imaging processing software package, template, processing methods, modulation, smooth kernel, covariate, and statistical threshold were also recorded. Peak stereotactic coordinates for all significant clusters, their space, and their corresponding effect sizes (t-statistics or Z-scores) were extracted to perform a voxel-wise meta-analysis. Z-scores for the peak coordinates from each study were converted to t-statistics via an online statistics converter (https://www.sdmproject.com/utilities/? show=Statistics).
To minimize data entry errors, study screen and data extraction were independently performed and checked by two researchers (DZ and CC). Any disputes were resolved by a third researcher (WCS).

Meta-analysis
The STATA software package (version 12.0 for Windows, Stata Corporation, College Station, TX, USA) was used to meta-analyze and compare sociodemographic (age, gender distribution, and educational level) and clinical (UPDRS-III score, H&Y stage, illness duration, and MMSE score) data between PD-MCI and PD-NCI groups using randomeffects models.
The SDM software package (version 5.15 for Windows, https:// www.sdmproject.com/software/) was used to conduct a voxel-wise meta-analysis and identify regional GM differences between PD-MCI and PD-NCI groups. SDM Details of this SDM procedure are described in a tutorial (https:// www.sdmproject.com/software/tutorial.pdf) and in previous literature [24,26]. First, an effect size map and an effect-size variance map of GM differences between patients with PD-MCI and patients with PD-NCI for each VBM study were recreated based on peak coordinates and their corresponding t-values using an anisotropic non-normalized Gaussian kernel. Following this, a mean map was produced using a voxel-wise calculation of the random-effects mean of the study maps, accounting for sample size, intra-study variance, and between-study heterogeneity. Finally, significant results were thresholded using the default parameters: uncorrected p = 0.005, peak Z height = 1, and cluster extent = 10 voxels, which were found to optimally balance false positives and negatives and were approximately equivalent to a corrected pvalue of < 0.05 [24,26].
To examine the robustness of significant results, jackknife sensitivity and heterogeneity analyses were conducted. Statistical significance for the above analyses was determined using the following default thresholds, as has been done previously [24,26]: p = 0.005, peak height Z = 1, and cluster extent = 10 voxels. Egger's test was additionally used to detect potential publication bias towards significant results in the meta-analysis. To examine the potential effects of moderator variables, such as age, gender, education level, disease duration, UPDRS-III score, H&Y stage, MMSE score, and scanner field strength, on outcome heterogeneity, meta-regression analyses were also performed. Significant results for these meta-regressions were thresholded at uncorrected p = 0.0005, a peak height Z = 1, and a cluster extent = 10 voxels, which were suggested by the authors of the SDM method [24,26].
As illustrated in Fig. 2, patients with PD-MCI relative to patients with PD-NCI exhibited decreased GM in the following two clusters: one in the left posterior insula (extending to the superior temporal gyrus, Rolandic operculum, Heschl's gyrus, and the middle temporal gyrus), and the other in the left anterior insula (extending to the inferior frontal    gyrus, orbital part) ( Table 3). Jackknife sensitivity analyses revealed that two clusters were preserved in at least 19 combinations of the datasets ( Table 4).
Analyses of heterogeneity revealed that there was significant statistical heterogeneity among studies conducted in the left superior temporal gyrus (extending to the Rolandic operculum, Heschl's gyrus, and the posterior insula), which overlaps with one of the clusters in the left posterior insula (extending to the superior temporal gyrus, Rolandic operculum, Heschl's gyrus, and the middle temporal gyrus) identified in the meta-analysis (Fig. 3, Table 5).
Egger's tests to examine publication biases showed non-significant results in the two clusters from the meta-analysis (p > 0.05) ( Table 3).
Meta-regression analyses revealed that those of an older mean age in the PD-MCI sample exhibited more prominent GM decreases in the left superior temporal gyrus (extending to the Rolandic operculum, Heschl's gyrus, posterior insula, and middle temporal gyrus) as compared to those of a younger age. A higher percentage of males in the PD-MCI sample were also associated with larger decreases to GM in the left superior temporal gyrus (extending to the posterior insula). Higher mean education levels in the PD-MCI sample were associated with greater GM increases in the left middle and inferior temporal gyri. Studies including the PD-MCI sample with higher mean UPDRS-III score exhibited more GM decrease in the left superior temporal gyrus (extending to the Rolandic operculum, Heschl's gyrus, and posterior insula). Studies with a higher mean H&Y stage or a lower average MMSE score in the PD-MCI sample reported larger GM decreases in the right dorsolateral prefrontal cortex and left ventrolateral prefrontal cortex. Additionally, studies utilizing a higher field strength scanner detected more pronounced GM decreases in the left posterior insula (extending to the superior temporal gyrus, Rolandic operculum, and Heschl's gyrus). In contrast, no significant correlation was observed between mean illness duration and GM abnormalities in the PD-MCI sample. The results of these meta-regression analyses are summarized in Table 6.

Discussion
Using the SDM approach, we conducted an updated and comprehensive CBMA of 17 VBM studies, including 19 comparisons between 440 patients with PD-MCI and 488 patients with PD-NCI. This quantitative meta-analysis revealed that, relative to patients with PD-NCI, patients with PD-MCI demonstrated more GM atrophy in the following two clusters: the left posterior insula (extending to the superior temporal gyrus, Rolandic operculum, Heschl's gyrus, and the middle temporal gyrus), and the left anterior insula (extending to the inferior frontal gyrus, orbital part). Although jackknife sensitivity analyses revealed high replication and Egger's tests detected no publication biases in these regions, a heterogeneity analysis revealed significant between-

Table 3
Clusters of regional gray matter reduction in patients with PD-MCI relative to those with PD-NCI. study variability in GM differences in the left superior temporal gyrus (extending to the Rolandic operculum, Heschl's gyrus, and posterior insula), which significantly overlapped with the first cluster identified from the meta-analysis. To examine potential sources of heterogeneity observed in this meta-analysis, we performed several meta-regression analyses. These indicated that many confounding factors might account for this heterogeneity, including participant age, education level, ratio of males, and severity of motor disability (UPDRS-III score) in the PD-MCI samples, as well as the scanner field strength used. The most robust finding identified in the present current metaanalysis had to do with GM atrophy in the left anterior insula (extending to the inferior frontal gyrus, orbital part) in patients with PD-MCI relative to patients with PD-NCI. The anterior insula is a crucial node in the salience network that mediates dynamic interactions between other large-scale brain networks, such as the default mode network and the central executive network. The insula is thus considered to serve as an integral hub involved in attentional processing and cognition [53,54].
Despite its crucial functions, the role of the anterior insula in PD is often underestimated [55]. In recent years, a growing body of evidence has suggesting a critical role for the anterior insula in PD, particularly with regard to its contributions to non-motor symptoms, including cognitive impairment [55,56]. Criaud et al. investigated the functional role of the insular cortex in parkinsonian features by conducting a quantitative meta-analysis of functional neuroimaging studies that revealed the anterior insula's role in cognitive function [56]. A previous positron emission tomography imaging study suggested that striatal dopamine denervation combined with insular dopamine D2 receptors loss might underlie MCI in PD, and in particular the dysregulation of executive dysfunction [57]. Executive dysfunction is one of the most common and early cognitive deficits associated with PD [58]. Furthermore, a loss of dopaminergic function in the salience network and the medial temporal lobe contributed to memory impairments in PD [59].
The present study adds further evidence that brain GM changes in the left anterior insula are associated with MCI in PD. No significant heterogeneity in GM changes in the left anterior insula was observed here, and changes to this region were not confounded by patient age, gender, educational level, disease stage, severity of motor disability, illness duration, or severity of cognitive function. This suggests that left anterior insula changes are both robust and specific to MCI in PD Yes, the region(s) is reported; No, the region(s) is not reported; 1, Parkinson's disease with amnestic cognitive deficits; 2, Parkinson's disease with non-amnestic cognitive deficits; 3, PD-MCI at the 1·5-SD threshold; 4, PD-MCI at the 2-SD threshold.  patients. One potential mechanism for its disruptions in PD is the anterior insula's vulnerability to alpha-synuclein pathology [60]. According to the staging hypothesis proposed by Braak et al. [61], the insula is affected in later stages in PD. Before pathological changes, alterations of regional function or brain networks may have already existed. Atrophy in the anterior insula indicates neurodegeneration responsible for disruptions to the integral functions of this region as a cognitive hub, underlying the deficient switching between the neural networks that contributes to cognitive deficits in PD [55].
Interestingly, we found a left-lateralization of brain GM atrophy in PD-MCI, consistent with what was reported in a previous meta-analysis by Xu et al [22]. Interestingly, a recent study by Claassen et al used an analysis of cortical thickness to suggest the particular susceptibility of the left insula to atrophy in 109 patients with early stage PD [62]. None of the original studies included in the present meta-analysis elucidated structural lateralization in PD-MCI cases. The mechanism underlying this lateralization remains unclear at present and warrants further investigation. Lateralization may be an early pattern indicating broader brain structural alterations in PD-MCI because bilateral GM atrophy is involved in PDD and later in the disease [22,62,63].
Unlike the robustness of GM atrophy in the left anterior insula revealed in the present study, significant heterogeneity in GM atrophy in the left posterior insula (extending to the superior temporal gyrus, Rolandic operculum, Heschl's gyrus, and middle temporal gyrus) was identified. We observed significant differences among studies in their participant characteristics and clinical variables. Meta-regression analyses revealed that age, ratio of males, education level, and severity of motor disability among PD-MCI samples, as well as scanner field strength, were important confounding factors that affected GM changes in these regions, suggesting that they are not specific to MCI in PD patients. Of these, a greater mean age, more male participants, and higher average UPDRS-III scores in PD-MCI patients were associated with more GM atrophy in the left superior temporal gyrus/Rolandic operculum/Heschl's gyrus/middle temporal gyrus/posterior insula. A multicenter pooled analysis revealed that MCI in PD patients was associated with older age, male gender, and more severe motor symptoms [64]. Additionally, a higher field strength scanner is prone to detect more GM atrophy in the left posterior insula/superior temporal gyrus/ Rolandic operculum/Heschl's gyrus. Interestingly, a higher average education level in the PD-MCI sample was associated with GM increases in the left middle/inferior temporal gyri, indicating its potential neuroprotective effects. A systematic review and meta-analysis revealed that a higher education level, as a proxy for cognitive reserve, is associated with significantly better cognitive performance in individuals with PD [65]. Critically, two prior meta-analyses did not explore the potential effects of demographic and clinical aspects on GM changes in PD-MCI, indicating a lack of attention to these details in the broader literature [22,63]. Given the heterogeneity among individuals with PD-MCI revealed in the present study, future work should pay special attention to the effects of these potentially modulating variables.
The findings in our updated meta-analysis are not entirely consistent with those of two previous meta-analyses [22,63]. Several disparities may contribute to this inconsistency. The current meta-analysis included more VBM studies than two prior meta-analyses, thus increasing the statistical power. The results of the meta-analysis conducted by Mihaescu et al. may also be biased, as it synthesized multiple studies using both VBM and cortical thickness measures, taken together [63]. However, these two approaches account for different elements of GM (cortical gray matter volume is a product of cortical surface area and cortical thickness [66]), potentially contributing to greater variance in the results. In addition, VBM enables a whole-brain analysis approach, while cortical thickness measures are based on an analysis of the cortical surface and thus do not offer insights into subcortical region changes [67]. Thus, to minimize heterogeneity among analytical methodologies, we included only whole-brain VBM studies here. Furthermore, our meta-analysis offers a more comprehensive view of the moderating effects of multiple demographic and clinical variables on GM changes in PD-MCI than the two previous meta-analyses mentioned here. Collectively, these strengths render the findings of the current meta-analysis more conclusive than those of previous ones.
Despite its strengths, the present study also has several limitations which warrant some discussion. First, we utilized the coordinates and effect sizes in published studies rather than those derived from original structural images, which may have biased our findings. However, obtaining these original images can be difficult or impossible. Currently, CBMA, the approach used here, remains the most useful means for identifying brain structural or functional convergence across studies. Second, PD-MCI is a heterogeneous clinical syndrome [1,2]. We noted variations in diagnostic criteria for PD-MCI. The present meta-analysis identified a common brain structural pattern of GM alterations associated MCI in PD across multiple studies. We were, however, limited by the information available in these prior studies and were unable to analyze structural differences between the multiple subtypes of PD-MCI. Future studies would benefit from discerning between the various neuroimaging differences and clinical phenotype differences among PD-MCI patients, which argue for recruiting distinct, subtype-homogeneous samples using unified diagnostic criteria. Third, the results of the metaregression analyses of UPDRS-III score and H-Y stage on GM alterations in the PD-MCI samples should be interpreted with caution because we did not consider the medication state. Fourth, limited by the information available from original studies, the current meta-analysis could not analyze the effect of dopaminergic medication on GM alterations. Fifth, although we only included whole-brain VBM studies in the meta-analysis, the heterogeneity in imaging acquisition, preprocessing, postprocessing, and statistical methods used in VBM studies might be other potential reasons of inconsistence. Future studies should use standardized imaging acquisition and analytical protocols. Finally, the findings of the present meta-analysis are based on the results of cross-sectional studies. This is a limitation, as longitudinal studies allow for the exploration of brain structural predictors and the better determination of which patients with PD-MCI may or may not progress to PDD.

Conclusions
This is the latest and most comprehensive meta-analysis that first identified the most robust GM atrophy in the anterior insula in patients with PD-MCI relative to patients with PD-NCI. The finding suggest that structural abnormities in the anterior insula, a key cognitive hub involved in switching between neural networks, may serve as a neural substrate or biomarker for MCI regardless of clinical PD phenotype. The anterior insula may also serve as a specific region of interest in future investigations of PD. Furthermore, we highlight the critical need for future studies of GM changes in PD to account for potentially confounding factors such as patient age, gender, educational level, disease stage, severity of motor disability, and severity of cognitive impairment.

Competing interests statement
The authors have no conflict of interest to report.

Author contributions statement
PLP and HCS conceived the project. DZ, CC, WCS, ZPW, and ZQY searched and selected the studies, analyzed the data, and prepared figures. CC undertook the statistical analysis. ZD, WCS, PLP, HCS, ZYD, and JGZ participated in the interpretation of data. DZ and CC wrote the manuscript. PLP and HCS revised the manuscript. All authors read and approved the final manuscript.