Cognitive phenotypes in parkinson's disease differ in terms of brain-network organization and connectivity

Participants and Assessment

The present study follows a cross-sectional study that aimed to identify cognitive phenotypes in Parkinson's disease. One hundred fifty-six Parkinson's disease patients participated in the study. All patients met the United Kingdom Brain Bank criteria for idiopathic Parkinson's disease [Gibb and Lees, 1988] and none was suffering from a neurological disease other than Parkinson's disease. Patients with moderate and severe dementia (defined as a score >1 on the Clinical Dementia Rating [Morris, 1993] and according to the Movement Disorders criteria [Emre et al., 2007]) and those older than 80 years were excluded. They were recruited among the outpatients of two independent European movement disorder centers, in Lille, France and Maastricht, the Netherlands. All participants gave their informed consent to participation in the study, which had been approved by the local institutional review boards (CPP Nord-Ouest IV, 2012-A 01317-36, ClinicalTrials.gov Identifier: NCT01792843).

All patients underwent a comprehensive neuropsychological assessment including tests for global cognition and standardized tests representing five cognitive domains: (1) attention and working memory (Digit span forward and backward [Wechsler, 1986], Symbol Digit Modalities Test [Smith, 1982]) (2) executive functions (Trail Making Test B/A ratio [Reitan, 1992], the interference index and the number of errors in the interference condition of a 50-item version of the Stroop word color test and a 1-minute phonemic word generation task performed in single and alternating conditions) (3) verbal episodic memory (Hopkins verbal learning test [Brandt and Benedict, 2001]), (4) language (the 15-item short form of the Boston naming test [Graves et al., 2004] and animal names generation task in 1 minute), (5) visuospatial functions (the short version of the judgment of line orientation test [Benton et al., 1978]). A cluster analysis (based on the k-means method) performed on the neuropsychological variables identified five phenotypes that were used for separating the participants according to their cognitive status: (1) cognitively intact patients with high level of performance in all cognitive domains, (2) cognitively intact patients with only slight mental slowing, (3) patients with mild to moderate deficits in executive functions, (4) patients with severe deficits in all cognitive domains, particularly executive functions, (5) patients with severe deficits in all cognitive domains, particularly working memory and recall in verbal episodic memory (for details, see [Dujardin et al., 2015]).

Detailed demographic and disease related variables were recorded likewise. All the patients' medications were checked and doses of antiparkinsonian medication were converted to levodopa equivalent daily dose according to the algorithm by Tomlinson et al. [2010]. Severity of motor symptoms was assessed by the score on the Movement Disorders Society—Unified Parkinson Disease Rating Scale (MDS-UPDRS)—part III [Goetz et al., 2008] and disease stage by the Hoehn & Yahr score [Hoehn and Yahr, 2001]. The severity of depression, apathy and anxiety symptoms was quantified with the 17-item Hamilton Depression Rating Scale [Hamilton, 1960], the Lille Apathy Rating Scale [Sockeel et al., 2006] and the Parkinson Anxiety Rating Scale [Leentjens et al., 2014], respectively. The presence and severity of hallucinations was assessed by the score on the item 1.2 of the MDS-UPDRS.

All patients were assessed after having received their usual anti-parkinsonian medication and were in their “best on” state.

MRI Acquisition

Patients were scanned at two sites (Maastricht and Lille) using the same 3T MRI scanner (Philips Achieva) with identical software versions and MR sequences. High-resolution 3D T1-weighted images were acquired with a magnetization-prepared gradient echo sequence (voxel size: 0.750 × 0.727 × 0.727 mm³; TR: 10.4 ms; TE: 4.76 ms; matrix size: 214 × 352 × 352 voxels). Resting-state functional imaging (rs-fMRI) was performed with a T2*-weighted EPI sequence lasting 10 min (voxel size: 3 × 3 × 3 mm³; TR: 2400 ms; TE: 30 ms; matrix size: 64 × 64 × 40 voxels; flip angle: 90°). Patients were required to remain quiet, stay awake and close their eyes.

One hundred thirty-four of the participants had an MRI scan. For the quality control of the data, all images were visually inspected by an investigator (CD) and rs-fMRI images with largely incomplete brain coverage, high movement peaks, ghosting or large motion artefacts were excluded. The displacements were computed using the six parameters (three rotations and three translations) from the rigid registration algorithm (for details, see section “Functional MRI pre-processing”). Rotational framewise displacements were adjusted to be expressed as a millimeter displacement for a typical distance from the center of 50 mm (Power et al., 2012). The sum of the absolute value of the six displacement measures were calculated for each frame and scrubbing frames were defined as those for which the value was greater than two standard deviations from the mean displacement. Subjects whose scrubbing frames represented more than 15% of all the frames were excluded. Five patients were excluded from the study because of brain lesions (thalamic lacunar infarctions, left prefrontal, bilateral orbito-frontal and two cases of temporal ischemic lesions), four patients with poor quality of T1 image and six patients with motion artefacts in rs-fMRI data. As a result, one hundred nineteen participants were included in the present study and divided into four groups according to the cognitive phenotypes (due to the low number of subjects, participants of the two groups with severe cognitive deficits were merged into one group): (1) 31 cognitively intact patients with high level of performance in all cognitive domains—“ G1,” (2) 31 cognitively intact patients with only slight mental slowing—“ G2,” (3) 43 patients with mild to moderate deficits, mainly in executive functions—“ G3,” (4) 14 patients with severe deficits in all cognitive domains—“ G4-5.”

Functional MRI Pre-Processing

First, pre-processing of fMRI data included the removal of the first three image volumes (to avoid T1 equilibration effects), rigid-body head motion correction to the first frame of data, and slice-timing correction using the first frame as the reference. Then, the distortion field, inherent to EPI images in the phase encoding direction and responsible for geometric and signal artefacts, was calculated using a pair of spin echo EPI scans with opposite phase encoding directions [Holland et al., 2010]. The FSL toolbox “topup” [Andersson et al., 2003] was used to estimate the distortion field, which was applied to correct all fMRI images.

Brain Parcellation and Registration to Functional MRI

Structural T1-weighted images were processed using the Freesurfer software (http://surfer.nmr.mgh.harvard.edu/). This included the pre-processing steps of non-uniform signal correction, signal and spatial normalizations, skull stripping and brain tissues segmentation [Dale et al., 1999]. Freesurfer software provided parcellation of anatomical regions of cortices (74 for each hemisphere) based on the Destrieux atlas and subcortical regions (eight for each hemisphere: nucleus accumbens, amygdala, caudate, hippocampus, pallidum, putamen, thalamus and ventral diencephalon). In total, 164 regions of interest (ROIs) were studied (Fig. 1). ROIs defined in T1 space had to be transformed into fMRI space where all calculations were made. T1 images were registered to the fMRI data using boundary-based registration [Greve and Fischl, 2009]. This method is more accurate than common registration algorithm, such as correlation ratio and normalized mutual information. The first frame of the pre-processed fMRI data was chosen as the template image. The cross-modality registration was achieved by minimizing the misalignment between the cortical grey-white boundaries in the anatomical and fMRI template image through six degrees of freedoms. Finally, ROIs and brain tissues segmentation masks (white matter and cerebrospinal fluid) were resampled to fMRI space. Parcellation and registration were manually checked for errors by visual inspection.

Network Construction

Two additional processing steps on fMRI data were added before the computation of the functional connectivity. Firstly, the physiological signal fluctuations were removed by a brain tissue-based correction approach called CompCor [Behzadi et al., 2007]. This method required white matter and cerebrospinal fluid masks extracted from the T1 image in the previous section. After a first regression to remove linear/quadratic trends (to account for scanner drift) and six motion parameters, the top five principal components from both masks were regressed out from each BOLD signal. Secondly, residual fMRI data were low-pass filtered, such that only frequencies below 0.1 Hz were selected.

A network or graph G = (V,E) is composed of nodes V and edges E between nodes. Herein, the functional network was obtained from the defined and registered ROIs and BOLD signal. The nodes represented the one hundred sixty-four ROIs and the edges represented the linear correlation between the mean BOLD time-courses in two given ROIs. For each participant, a 164x164 symmetric weighted network was constructed. Finally, as age, formal education duration and participating centre differed across groups, the weighted network was corrected for all three by regressing out these factors of each participant from weighted edges.

Statistical Analyses

The network analyses were performed at two scales: network scale and edges scale. At the network scale, functional organization indexes of brain networks were computed using GRETNA toolbox (https://www.nitrc.org/projects/gretna/): the functional segregation, the functional integration and the degree distribution. At the edge scale, the alterations in functional connectivity were investigated using network-based statistics.

Network Topology Analysis

Functional segregation and integration (graph theory measures)

The corrected weighted network was thresholded at different levels of sparsities ranging from 10% to 40% using increments of 1%, in keeping the highest weighted edges. At each threshold, global network metrics were calculated: the clustering coefficient C_p, the mean path length L_p, the global efficiency E_g and the local efficiency E_loc. The area under the curve (AUC) for each network metric was calculated to provide a summarized scalar for topological organization of brain networks independent of a single threshold selection [Achard et al., 2006] . The metrics' definitions are briefly described below.

The clustering coefficient C_p of G quantifies the extent of local clustering or cliquishness of a network [Watts and Strogatz, 1998]. It is expressed as follows [Onnela et al., 2005]:

(2)

where k_i is the degree of node i ( ), and is the weighted value of the edge, which is scaled by the mean of all weights to control each participant's cost at the same level.

The mean path length L_p of G quantifies the ability for information to be propagated in parallel. It is expressed as follows:

(3)

where L_ij is the shortest path length between any pair of nodes i and j and is defined as the sum of the edge weight w_ij along this path.

The E_g of G quantifies the global efficiency of the parallel information transfer in the network. It is expressed as follows:

(4)

The E_loc of G quantifies the fault tolerance of a network, showing how efficient the communication is among the first neighbours of node i when it is removed. It is expressed as follows:

(5)

Where G_i denotes the subgraph composed of the nearest neighbours of node i.

Between-group differences in AUC values of network metrics were investigated in a one-way ANOVA after the normality of the data distribution had been checked in a Kolmogorov-Smirnov test. Levene's test showed non-homogeneity of variances within groups, so Welch ANOVA using SPSS 20.0 software (SPSS Inc., Chicago, IL, USA) was performed on data. Metrics showing main effect of group difference in ANOVA model were further evaluated by post-hoc tests (Games-Howell test). A significance threshold of P < 0.01 was applied to each test and the FDR was used to correct for multiple comparisons.

Degree distribution

The degree distribution P(k) of a graph is the fraction of nodes with degree k, and can be used to assay the hierarchy of hubs in the network. The analysis was performed on the sparsest binary network (10%) to eliminate the weaker noisy connections [Achard et al., 2006]. Brain networks may be fitted for three candidate models based on the frequency distribution of their node degrees:

(6)

(7)

(8)

Where K denotes degree, α the power law exponent and the β exponential cutoff.

These three statistical models were evaluated using Akaike's information criterion. Between-group differences in α and/or β values were assessed by one-way ANOVA and post-hoc tests at a significance threshold of P < 0.01, FDR corrected.

Demographic, Clinical, and Neuropsychological Characteristics

Table I shows the demographical and clinical characteristics of the four patient groups and results of neuropsychological assessment are shown in Table II. As significant between-group differences were observed for age, duration of formal education and number of subjects by participating centre, further analyses were adjusted on these variables. Despite a significant group effect on severity of the apathy symptoms and frequency of hallucinations, these variables were not considered as confounding factors since they have been frequently associated with cognition in Parkinson's disease [Aarsland et al., 2014; Dujardin et al., 2009]. Table I shows that most of the patients with hallucinations only had illusions or not-formed hallucinations (scoring 1 on the item 1.2 of MDS-UPDRS) or formed hallucinations without loss of insight (scoring 2 on the item 1.2 of MDS-UPDRS).

Network Topology

There was a significant difference between groups as determined by one-way ANOVA for global efficiency E_g (F = 14.10, P = 0.000), clustering coefficient C_p (F = 8.47, P = 0.001) and local efficiciency E_loc (F = 4.75, P = 0.008) (Fig. 2). A Games-Howell test revealed that E_g was significantly lower for G1, G2 and G3 than G4-5. Also, G1 showed lower E_g values than G2 and G3. The C_p values were significantly higher for G1 and G2 than G4-5. Also, G1 showed higher Cp values than G3. The E_loc values were significantly higher for G1 and G2 than G4-5. No statistically significant difference was found between G2 and G3 for all network metrics.

The best fitting using Akaike's information criterion was obtained with an exponentially truncated power law. Post-hoc tests showed that the power law exponent α was significantly increased but the exponential cutoff β was significantly reduced in G4-5 (Fig. 3). These distribution parameters indicate that the probability of high degree hubs is reduced in G4-5.

The leave-one-out cross-validation classification procedure was applied using the four network measures and the performance is summarized in Table III. A classification accuracy of 75% was obtained for all subjects. However, the accuracy varied between groups. It was good for groups G1 and G4-5 with a sensibility of 0.90 and 0.86 and a specificity of 0.97 and 0.98, respectively. More difficulties were observed to classify the subjects of G3 and especially those of G2.

The scatterplots showing the correlation between each of the network metrics and the cognitive score are presented in Figure 4. The cognitive score was positively correlated with the C_p, E_loc, and L_p values and negatively correlated with the E_g values. Hence, high level of performance at the neuropsychological tests was associated with more extended local clustering, higher ability of information to be propagated in the network, more efficient communication among neighbouring areas but less global efficiency of the network.

Functional Connectivity

The ANOVA revealed significant functional differences between groups (NBS corrected, P < 0.01) in the middle and inferior frontal gyri, the anterior cingular gyrus, the post central gyrus, the paracentral gyrus, the posterior insular area, the fusiform gyrus, the cuneus and superior occipital gyri, the calcarine sulcus, the striatum and the thalamus in the right hemisphere and in the inferior (opercular part) frontal and medial orbital frontal areas, the temporo-occipital areas, the striatum and the thalamus in the left hemisphere (Fig. 5 and Table IV). Further group comparisons are illustrated by the connectograms (Figs. 6 and 7) and revealed between-group differences in the architectural features of hub nodes and their connections (Table IV). Patients in G1 had higher hub connections of the right associative frontal and occipital areas than in G2. Compared with G3, G1 had also higher hub connections of the associative frontal, parietal, occipital areas, of the limbic cingular gyrus and temporal areas in the right hemisphere, as well as higher hub connections of a part of the associative temporal area in the left hemisphere. Compared with G4-5, G1 had higher hub connections of the associative frontal, temporal, occipital areas, the sensorimotor areas, the insular area, and the limbic cingular gyrus in the right hemisphere and of the insular sulcus, the associative temporal and occipital areas, the accumbens nucleus and hypothalamic area in the left hemisphere. Compared with G3, G2 had higher hub connections of the limbic and associative temporal areas in the right hemisphere and higher hub connections of the associative frontal sulcus and temporo-occipital areas in the left hemisphere. Compared with G4-5, G2 had higher hub connections of the associative frontal, parietal and temporal areas, of the limbic cingular and temporal areas, the primary visual area and the thalamus in the right hemisphere and higher hub connections of the associative temporal and occipital areas, of sensorimotor areas in the left hemisphere. Compared with G4-5, G3 had higher hub connections of the associative frontal areas, of the limbic frontal and cingular areas, limbic anterior cingular, the primary sensorimotor areas, the associative temporal and occipital areas in the right hemisphere and higher hub connections of the limbic frontal and cingular areas, in the left hemisphere.

Effects of Cognitive Impairment on Brain Network Topology

As shown by the progressive decrease of the values of clustering coefficient and local efficiency as cognitive impairment increased, there was a reduction of the functional segregation of the brain with cognitive impairment. According to Rubinov and Sporns, functional segregation is “the ability for specialized processing to occur within densely interconnected groups of brain regions” [Rubinov and Sporns, 2010]. This suggests that cognitive deficits in Parkinson's disease could be related to a progressive decrease in the density of connections between brain areas involved in specialized networks engaged in cognition. This is in line with the results of previous studies investigating functional connectivity of specific networks in Parkinson's disease. Hence, Tessitore et al. reported a decreased functional connectivity within the DMN in Parkinson's disease patients compared with healthy controls, which was correlated specifically with performance at cognitive tests but not with disease duration, motor impairment, or levodopa therapy [Tessitore et al., 2012]. In a 3-year follow-up of 36 Parkinson's disease patients and 12 healthy controls, Olde Dubbelink et al. showed that functional connectivity decreased more in Parkinson's disease patients, especially in posterior parts of the brain and these changes were correlated with cognitive decline [Olde Dubbelink et al., 2014].

The increase of global efficiency and the trend toward a decrease in path length with higher cognitive impairment is more difficult to interpret. Indeed, these are measures of global network functioning, or functional network integration, and reflect “the ease with which brain regions communicate” [Rubinov and Sporns, 2010] and one may expect that these network metrics changed in the opposite direction when cognitive impairment increases. This apparently contradictory result could be explained as follows. Firstly, to check if it was an artefact or a result related to Parkinson's disease as a whole, we compared our results with that of a group of healthy subjects coming from another study but matched for sex, age and education with our participants (demographic details are shown in Supporting Information Table S1). Again, we found that global efficiency was lower and path length higher in that group than in the groups with cognitive impairment (Supporting Information Fig. S1). Moreover, there was no difference between this healthy control group and our cognitively intact group (G1) suggesting that the effects we observed on brain network measures were not related to Parkinson's disease, per se, but to cognitive impairment. The same pattern of results was observed for the degree distribution (Supporting Information Fig. S2). Secondly, this phenomenon could be interpreted in using all measures. Indeed, networks can be modelled within a framework of regular, random and complex networks. Regular networks have high local and low global connectivity, while random networks show the opposite pattern with low local and high global connectivity. The human brain network is inherently complex with an optimised balance between local and global connectivity, the so-called small-world network [Sporns, 2011]. Increased global efficiency together with decreased mean path length and less fat-tailed degree distribution indicate excessive network integration in cognitively impaired patient groups. These network features associated with decreased clustering coefficient and local efficiency are characteristic of a brain topology that shifts from a small-world towards a random network topology. Randomness increases with age and Smit et al. showed that the small-world pattern of adult age is gradually replaced by a more random topology at higher ages [Smit et al., 2010]. Moreover, it is also an indicator of network alterations in disease states [Stam, 2014]. Hence, the increase in global efficiency and the trend toward a decrease in path length we observed here with higher cognitive impairment may reflect an increase in randomness as cognitive impairment worsens in Parkinson's disease. Randomness of the brain network may thus be a marker of cognitive decline in Parkinson's disease. This is also confirmed by the pattern of correlations we observed between the network metrics and the composite cognitive score since more extended local clustering, higher ability of information to be propagated in the network, more efficient communication among neighboring areas but less global efficiency of the network were associated with higher level of performance at the neuropsychological tests.

When comparing PD-MCI and healthy controls, Baggio et al. reported a reduction of inter-network connections in PD-MCI, while local interconnectedness increased [Baggio et al., 2014; Baggio et al., 2015b]. This seems at odds with our own results. However, these studies used different approaches. Firstly, in our study, the groups of patients were not determined a priori but resulted from a cluster analysis identifying different levels of severity in cognitive impairment. Secondly, some differences in our methodology could explain our results with those of Baggio et al. (Baggio et al., 2014). Although the spatial normalization to MNI atlas is a common step in neuroimaging, their approach may cause inaccuracies of spatial transformation due to inter-subject anatomical variations that exist between the MNI template and the transformed data. To avoid this spatial normalization step, we worked in the individual subject space. Moreover, whereas the statistical analyses at the network level differed in both studies, we used AUC values calculated in a sparsity range of the connectivity matrix instead of testing each sparsity level.

Effects of Cognitive Impairment on Brain Functional Connectivity

The NBS analyses showed that alteration of brain network functional connectivity progressively increased, as cognitive impairment worsened. In Parkinson's disease patients with only slight mental slowing, weakened connections concerned the dorsolateral frontal cortex and the extrastriate cortex, two highly associative areas. Examination of the connectogram (Fig. 7 top-left) shows that these alterations mainly concerned interhemispheric connections to the frontal, temporal and parietal cortex. Changes in the fronto-parietal attentional network [Petersen and Posner, 2012] agree with the attention/executive functions deficits usually described in Parkinson's disease patients [Kehagia et al., 2013] but loss of connections from the extrastriate cortex to the temporal and parietal areas may be an early marker of cortical posterior dysfunction leading to deficits in memory and visuospatial functions [Kehagia et al., 2013] and increasing the risk of progression to dementia [Williams-Gray et al., 2007]. Such changes in brain functional connectivity may thus exist before performance at neuropsychological tests assessing these functions significantly declines. Markers of functional connectivity may thus be used in further cohort studies in order to determine whether they contribute to early identification of Parkinson's disease patients at risk of cognitive impairment or dementia.

When comparing cognitively intact patients with only slight mental slowing with patients with mild to moderate deficits, mainly in executive functions, weakened connections concerned interhemispheric connections from the posterior cortex to the parietal, temporal, cingular, insular and frontal regions (Fig. 7 bottom). Again, this suggests that an alteration of the functional connections from the posterior cortex may exist before deficits on tests usually assessing cortical posterior functions are observed. Moreover, this alteration seems to contribute to the dysexecutive syndrome through the long-range connections of the posterior cortex to the frontal, parietal and cingular cortex. Only a follow-up of these three groups of patients may determine the predictive value of these biomarkers.

When cognitive deficits are severe in all domains, the functional network was widely altered.

The changes in brain functional connectivity were analyzed and discussed here taking into account the spectrum of severity of cognitive decline we had previously observed. Further analyses are obviously needed to determine whether deficits in specific cognitive domains are associated with alterations within specific functional sub-networks.

Limitations

Several limitations of the current study should be considered. Firstly, due to our study design (data-driven approach in order to identify cognitive phenotypes in Parkinson's disease), we did not include a healthy control group and G1 (cognitively intact patients) was considered as our reference group. Hence, we were not able to identify changes in functional connectivity in this group compared with healthy controls. Nevertheless, our additional analysis on global network parameters (see Supporting Information) included a healthy control group and exhibited no difference between that group and G1, reinforcing the assumption that patients in G1 are cognitively intact, behave as healthy controls and can be considered as a reference group.

Secondly, due to the small number of subjects in the two groups with severe cognitive deficits, we decided to merge these two groups to perform the functional analyses although they were characterized by different cognitive profiles according to the cluster analysis. This has prevented any possibility of identifying connectivity differences between those groups. Moreover, despite the fact that the two groups with severe cognitive deficits were merged, the number of subjects of this category remains small. This is partly due to the exclusion of patients with moderate and severe dementia from the study in order to avoid missing data. By consequence, patients with severe cognitive deficits are underrepresented. Thirdly, patients were assessed after having received their usual anti-parkinsonian medication (“best on” state) and dopaminergic medication may have influenced connectivity patterns. However, there were no group differences in terms of levodopa equivalent daily dose and it is unlikely that this factor may explain the observed differences. Finally, apathy and hallucinations were not considered as confounding factors in our analyses although previous studies showed alterations of resting-state functional connectivity in PD patients with apathy [Baggio et al., 2015a] or repetitive, complex visual hallucinations [Yao et al., 2014]. However, we considered that lack of initiative, reduction of interests and loss of insight may be symptoms of cognitive impairment since both apathy and hallucinations are embedded with cognitive impairment in Parkinson's disease. Adjusting these variables, in addition to reducing statistical power, would have removed useful information from our analyses.

Further Methodological Considerations

Several issues remain to be addressed. First, in this study, the regions of interest were obtained according to the parcellation of Fischl. Given the dependence of topological properties of brain networks and ROI definition [Wang et al., 2009], it would be interesting to use several parcellation systems to confirm the results. Nevertheless, this parcellation is well validated and no spatial normalization to MNI atlas is needed. Second, the head motion of subjects might have confounding effects on the final results of network analysis. To evaluate whether the results of network and edge analyses were affected by head motions, an autoregressive model AR(p) of order P = 1 was learned on the time series of the motions. The correlation between the AR(1) model coefficient of each time-course and network or edge measures was calculated by using a multivariate model [Salvador et al., 2008]. No significant correlations (P > 0.05) were found. Third, the brain network topology measures were calculated on binarized graphs that were constructed by thresholding the functional connectivity matrices. Although the use of binarized graphs reduced the complexity of network analysis, it also removed some detailed information. Further analyses could be conducted by using weighted correlation values (“weighted” graphs) instead of binarized graphs.