Graph theory network function in Parkinson's disease assessed with electroencephalography

Abstract

Objectives

To determine what differences exist in graph theory network measures derived from electroencephalography (EEG), between Parkinson's disease (PD) patients who are cognitively normal (PD-CN) and matched healthy controls; and between PD-CN and PD dementia (PD-D).

Methods

EEG recordings were analyzed via graph theory network analysis to quantify changes in global efficiency and local integration. This included minimal spanning tree analysis. T-tests and correlations were used to assess differences between groups and assess the relationship with cognitive performance.

Results

Network measures showed increased local integration across all frequency bands between Control and PD-CN; in contrast, decreased local integration occurred in PD-D when compared to PD-CN in the alpha1 frequency band. Differences found in PD-MCI mirrored PD-D. Correlations were found between network measures and assessments of global cognitive performance in PD.

Conclusions

Our results reveal distinct patterns of band and network measure type alteration and breakdown for PD, as well as with cognitive decline in PD.

Significance

These patterns suggest specific ways that interaction between cortical areas becomes abnormal and contributes to PD symptoms at various stages. Graph theory analysis by EEG suggests that network alteration and breakdown are robust attributes of PD cortical dysfunction pathophysiology.

1.0 Introduction

Lewy-type synucleinopathy (LTS) is established to be prominent in the cerebral cortex at mid and later stages of Parkinson's disease (PD). Alzheimer's disease (AD) pathology and vascular lesions are also present in a portion of cases. Functional imaging and quantitative electroencephalography (QEEG) studies have revealed physiologic dysfunction of widespread cortical areas. However, the manner in which cortical or subcortical pathology disrupts interaction within and between cortical areas has received little study and remains to be established. An example of the importance for understanding PD cortical dysfunction is the concomitant cognitive decline that results in dementia for about 80% of patients (Svenningson et al., 2012; Caviness et al, 2011). Currently, the major pharmacologic therapy for PD-dementia (PD-D) is acetylcholinesterase inhibition (Caviness et al., 2011). This therapy is based on the theory that decreased cholinergic input to the cortex may also result in cognitive deficit, perhaps due to pathology in the nucleus basalis of Meynert. Such treatment is modest, transient, and better treatments arising from other pathophysiological mechanisms are desperately needed. One such mechanism is potentially how disruption of the degree and type of multiple interactions between cortical areas contributes to cortical dysfunction at different Parkinson's disease stages and cognitive deficits among these individuals.

Previous research has found electroencephalography (EEG) spectral and event-related changes (Caviness et al., 2007) as well as standard functional magnetic resonance imaging (fMRI) differences (Olde Dubbelink et al., 2014a) related to PD and its cognitive deterioration (Lewis et al., 2003; Gottlick et al., 2013). Such classic physiological methods use resting, local area activation or seed-based pairwise interactions. However, these approaches do not assess multiple brain area interactions simultaneously, i.e. network functional connectivity. Although “networks” have been discussed for decades in neurology, new advances in applied mathematics have recently allowed networks to be analyzed in a quantitative way. It has been demonstrated that normal brain function relies on the interaction of different brain regions that are integrated within a large-scale complex network (Stam and van Straaten, 2012). More specifically, the human cortex manifests in what is referred to as a “small world organization,” demonstrating a combination of local integration (i.e. local clustering of connections) and global efficiency (i.e. long-distance connections) (Reijneveld, et al., 2007). There are mutually and densely interconnected regions (hubs) that form a connectivity backbone crucial for efficient brain communication (Stam and van Straaten, 2012). Synaptic connections in the cortex are known to influence computational network models and PD pathology is known to affect synapses (Bellucci et al., 2012; Klassen et al., 2011; see Bullmore and Sporns, 2009 for review). Quantifying network physiology, for PD pathology as well as the evolution of cognitive decline, has the potential to provide understanding about how the abnormal multiple cortical area interaction occurs in ways not afforded by previous approaches.

Circuits that involve a specific phenomenon or task in PD patients, i.e. motor, behavioral, cognitive, have been examined in multiple investigations, but these studies are generally not whole network analyses using graph theory (Hacker et al., 2012; Segura et al., 2013; Wu et al., 2009; Ibarretxe-Bilbao et al., 2011; Wu et al., 2011). Moreover, whole network analysis of PD patients using graph theory has only been explored in a few studies, using fMRI, structural MRI, and magnetoencephalography (MEG) (Baggio et al., 2014; Olde Dubbelink et al., 2014b). EEG network analysis in PD and its cognitive decline can assess frequencies 100-1000× faster than fMRI and is widely available, making it optimal for sensitive and specific clinical and research use. In the current study, we hypothesized alterations and breakdowns in EEG derived whole network graph theory measures of local integration and global efficiency across PD clinical stages.

2.0 Methods

2.1 Standard Protocol Approvals, Registrations, and Patient Consents

The Banner Sun Health Research Institute (BSHRI) and Mayo Clinic institutional review boards approved all procedures and written informed consent was obtained from study participants.

2.2 Subjects

The Control and PD cohorts were studied as part of the Arizona Study of Aging and Neurodegenerative Disorders (AZSAND), a brain and body donation program under Banner-Sun Health Research Institute (website: www.brainandbodydonationprogram.org). The Control and PD cohorts in AZSAND undergo ongoing longitudinal assessments until death as described previously (Beach et al., 2009; Caviness et al., 2015). Briefly, controls and subjects with neurodegenerative disease annually have medical history and physical, clinical behavioral and movement assessments, and neuropsychological testing (Beach et al., 2009; Caviness et al., 2015). PD motor severity is assessed with the motor United Parkinson's Disease Rating Scale (UPDRS). The cognitive testing includes both the mini-mental status exam (MMSE) and more recently Montreal Cognitive Assessment (MOCA). The designation of a PD diagnosis (and cognitive state) or Control was made in consensus conference with movement disorder and cognitive neurologists, and neuropsychologists. Clinical PD diagnostic criteria were as per previous reports (Klassen et al., 2011; Caviness et al., 2007). For PD subjects, motor scales, PD age onset and duration, and medications were recorded, and PD medication was converted to levodopa equivalents as per previous reports (Klassen et al., 2011; Caviness et al., 2007). Subjects also undergo biennial digital EEG recording. The cohort data was selected from our database by searching diagnosis and eligibility criteria, along with the EEG recording and other data and consensus determinations associated in that assessment epoch. We excluded EEG examinations with excessive muscle artifact, deep brain stimulation, and barbiturate, benzodiazepine or anti-seizure medication on day of EEG to avoid influencing the EEG. We further excluded recordings with slightly fewer or different than the 21 electrodes used by the network analysis software (Brainwave) to eliminate possible effects of even slightly different recording electrode node sites.

2.3 Design of group comparisons

2.3.1 Healthy Controls versus PD-Cognitively Normal

Those PD subjects defined as PD-Cognitively normal (PD-CN) were not found to be cognitively disabled and did not fulfill criteria for mild cognitive impairment or dementia. Healthy Control subjects were age and sex-matched to PD-CN subjects. This was deemed important as to avoid potentially confounding network changes secondary to age and sex.

2.3.2 PD-Cognitively normal versus PD-Dementia (PD-D)

The same PD-CN group designated above was compared to PD-D subjects with adjustment for age and sex. Since PD subjects with mild cognitive impairment (PD-MCI) comprised a small group, comparison with PD-CN was considered a preliminary analysis only but was adjusted for age and sex. Movement Disorder Society guidelines were used for PD-MCI and PD-D (Emre et al., 2007; Litvan et al., 2012).

2.4 EEG recording

Recordings during monitored relaxed wakefulness (eyes closed) were obtained from control and for PD subjects in the “on” state using standard twenty-one 10-20 EEG positions, as previously described (Klassen et al., 2011; Caviness et al., 2007). Data were acquired using the Synamps2 system (Charlotte, NC, USA) at a sampling rate of 1000 Hz and a bandpass of 1 - 200 Hz. When muscle, eye movement/blink, or other artifacts were identified, the subject was coached until an optimum quality signal was obtained of approximately 120-150 seconds.

2.5 Graph Theory: Network Analysis

Graph theory is a field within mathematics that models networks, also referred to as graphs (van Steen, 2010). In graph theory, “nodes” represent a series of network locations that can be connected to one another. The number of nodes, and the existence and strength of connections between nodes form the basic network building blocks. Moreover, graph theory utilizes this “graph” to represent these individual pieces of the network and quantify their connections to each other and various pieces of the network. The graph theory model is applied to EEG by defining each EEG electrode as a node and synchronization between two electrodes as the connection between them. There are several key network measures in graph theory network analysis (see Table 1), and details about utilizing these important concepts from graph theory in the analysis is given below.

Table 1.

Functional connectivity and network outcome measures.

Data type	Measure	Definition	Physiological significance
Functional connectivity	PLI	Phase lag index. Measure of functional coupling between nodes. PLI values are used as edge weights in the network analyses.	A decreased value corresponds to decreased connectivity strength.
Weighted Network	Gamma	Normalized weighted clustering coefficient. Measure of the degree of connectivity between node neighbors.	A decreased value corresponds to decreased local integration.
	Lambda	Normalized characteristic path length. Measure of the average weight of all shortest paths between any two nodes of the network.	An increased value corresponds to decreased global efficiency.
	KappaW	Weighted degree divergence. Measure of the broadness of the weighted degree distribution, where weighted degree is the summed weights of all edges connected to a node. Calculated as the ratio of the average squared weighted degree and the average weighted degree.	A decreased value corresponds to a decrease of highly connected nodes or “hubs.”
	Modularity	Ratio of the intra- and intermodular connectivity strength where modules are subgraphs containing nodes that are more strongly connected to themselves than to other nodes*. Modularity is a measure of the strength of the modules.	A decreased value corresponds to decreased local integration.
Minimum spanning tree (MST)	MST BCmax	Maximum MST betweenness centrality. Maximum number of paths between any two nodes of the MST running through a single node.	A decreased value corresponds to decreased global efficiency and a decrease of “hub” strength.
	MST Diameter	Maximum distance (number of connections between two nodes) of the MST.	An increased value corresponds to decreased global efficiency.
	MST Eccentricity	Average maximum distance between any two nodes of the MST.	An increased value corresponds to decreased global efficiency.
	MST Leaf	MST leaf fraction. Ratio of number of nodes with only one link (i.e. degree 1 nodes or “leafs”) and the maximum possible number of leafs (number of nodes – 1) in the MST.	A decreased value corresponds to decreased global efficiency.

^*See Newman and Girvan, 2004; Guimerà, R. et al., 2004; and de Haan et al., 2012 for further detail and mathematical computations.

2.6 Analysis

The continuous EEG data was divided into non-overlapping 4096 point (4095 ms) epochs. Each epoch was visually inspected for artifacts, though rejection of artifacts was uncommon due to the monitoring of the online acquisition. For detecting blink and other eye movement artifacts, comparison was made to the vertical and horizontal eye movement channels. Epochs with muscle artifact were rejected if such artifact signal were present grossly. From these epochs, 8 artifact-free epochs were chosen for analysis. Graph theory network analysis was performed with Brainwave software (Stam, 2015). Briefly, the data was first analyzed via a synchronization detection method. For this, we used phase lag index (PLI), as it is less affected by volume conduction than more traditional measures like coherence, and is sensitive to true changes in synchronization (Stam et al., 2007). PLI is utilized as a measure of functional connectivity. On the basis of previous network analysis from the literature in a variety of neurological disorders, we chose both traditional graph theory weighted connectivity measures and those from “Minimal Spanning Tree” (MST) graph analysis (Stam et al., 2014). All chosen measures, their definition, and physiological significance are shown in Table 1. All measures were calculated for the following frequency bands: delta (2.5-4 Hz), theta (4-8 Hz), alpha1 (8-10 Hz), alpha2 (10-13 Hz), and beta (13-30 Hz). Utilizing Brainwave, the weighted network graph of connections, and relative distances, can be visualized in a node map for a given frequency band. Here, the network “map” for each group was calculated as an average of all individual epochs, for all participants, in that group. The measures derived from the Minimum Spanning Tree analysis have only been recently applied to brain networks, and it is thought to be possibly more reliable for comparison of networks across different groups since it does not require arbitrary parameter settings (Tewarie et al., 2015).

2.7 Statistics

Using ANOVA or the Pearson chi-square tests, differences in demographics among subject groups were assessed. To assess our primary objective, pairwise comparisons were assessed between Control versus PD-CN and PD-CN versus PD-D. Differences were assessed using the two-sample t test. Normality of the sampling distributions was assessed by plots of residuals from the fitted models for these compared measures. Adjusted differences were assessed using a general linear model with terms for group, age, and sex. Further, the relationship between network measures of all patients with PD and both the MOCA and MMSE were assessed via the Pearson correlation coefficient; this allows for the assessment of the relationship of these network measures with a range of cognitive performance. The distributions of the residuals from the models of network measures versus MMSE or MOCA are normally distributed, satisfying assumptions for the use of parametric analysis. Analysis of the confounding effect of motor severity (UPDRS) on another was performed by using multivariable Cox regression modeling. As an exploratory objective, an additional pairwise comparison was done between PD-CN versus PD-MCI. The results of this assessment are considered preliminary data, given the small sample size of the PD-MCI group, and the subtle differences in cognitive status. These differences were assessed in the same way as for the primary analysis above. For all analyses, significance was defined as p < .05.

3.0 Results

Table 2 shows demographics and basic clinical characteristics of the four subject groups. For the PD groups, UPDRS III motor score, Hoehn &Yahr stage, and disease duration were significantly different between PD-CN, PD-MCI, and PD-D, as seen in previous studies of our cohort but consistent with groups of advancing stage (Caviness et al., 2015). Levodopa equivalents and PD age onset were not different between PD groups.

Table 2.

Mean (SD) of demographic and basic clinical characteristics of groups.

	Control N= 57	PD-CN N= 57	PD-MCI N= 13	PD-D N= 18	p- value
Age (years)	77	74.9 (8.2)	77.0 (5.2)	80.6 (6.5)	<.001
UPDRS III score	N/ A	22 (14)	29 (12)	43 (17)	<.001
H&Y stage	N/ A	2.18 (1.08)	2.95 (0.91)	3.47 (0.86)	<.001
Levodopa equivalents (mgs)	N/ A	730 (370)	860 (460)	800 (400)	.50
PD age onset (years)	N/ A	65.1 (10.1)	62.9 (9.7)	63.6 (11.8)	.74
PD duration (years)	N/ A	9.7 (5.6)	13.8 (7.9)	17.0 (9.0)	<.001

Note: N/A = Not applicable.

Table 3 shows the group mean differences of network measures for Control compared to PD-CN. Plots of residuals from the fitted models indicated that the sampling distributions were normal. Mean PLI synchronization was modestly higher in the theta band but no differences in other frequency bands. There were significant differences for all weighted network measures as well as for all frequency bands pertaining to each weighted network measure. For PD-CN, Gamma (cluster index), Lambda (path length), and Modularity were all higher compared to Control. Kappa_w was lower in the delta and alpha bands but higher in the theta and beta bands. MST differences were only present in the slow band frequencies and less common than for weighted measures. For PD-CN, MST diameter was lower in delta and theta bands; MST eccentricity was only lower in the delta band; and MST leaf fraction was higher in both the delta and theta bands.

Table 3.

Values of network measures for age and sex-matched, healthy control participants and patients with Parkinson's disease who are cognitively normal (PD-CN).

	Delta		Theta		Alpha1		Alpha2		Beta
	Control	PD-CN	Control	PD-CN	Control	PD-CN	Control	PD-CN	Control	PD-CN
Functional connectivity
PLI	0.26	0.26	0.18	0.20*	0.29	0.29	0.22	0.22	0.10	0.10
Weighted network
Gamma	1.03	1.27*	1.04	1.34*	1.05	1.37*	1.05	1.34*	1.04	1.28*
Lambda	0.94	1.03*	0.94	1.12*	0.96	1.34*	0.95	1.16*	0.94	1.09*
KappaW	5.53	4.15*	3.87	4.32*	6.15	4.40*	4.64	4.32*	2.14	4.19*
Modularity	0.09	0.21*	0.08	0.19*	0.07	0.18*	0.08	0.19*	0.08	0.20*
Minimum Spanning Tree
MST BC	0.71	0.71	0.72	0.73	0.72	0.73	0.72	0.73	0.72	0.72
MST Dia.	0.44	0.42*	0.41	0.40*	0.41	0.41	0.41	0.40	0.41	0.41
MST Ecc.	0.35	0.33*	0.33	0.32	0.33	0.33	0.33	0.32	0.33	0.32
MST Leaf	0.54	0.56*	0.568	0.58*	0.58	0.59	0.57	0.58	0.56	0.56

^*Significant differences are noted by (p < .05).

Table 4 shows mean differences for PD-CN compared to PD-D. Comparisons of means were adjusted for age and sex. Plots of residuals from the fitted models indicated that the sampling distributions were normal. Here, all significant differences are in the alpha1 and alpha2 bands, except for a higher MST betweenness centrality for PD-D in the delta band. In the alpha1 band, functional connectivity was notably lower in the PD-D group. Also in the apha1 band, PD-D had lower gamma, lambda, and Kappa_w. Modularity was higher for PD-D in both alpha bands. Differences between PD-CN and PD-D group MST measures are also primarily found in the alpha1 frequency band, with MST betweenness centrality, diameter, and eccentricity all being higher for PD-D in the alpha1band. MST diameter was also higher in the alpha2 band in PD-D, while leaf fraction was lower.

Table 4.

Values of network measures for patients with Parkinson's disease who are cognitively normal (PD-CN) compared to those with dementia (PD-D), for which comparisons are adjusted for age and sex.

	Delta		Theta		Alpha1		Alpha2		Beta
	PD-CN	PD-D	PD-CN	PD-D	PD-CN	PD-D	PD-CN	PD-D	PD-CN	PD-D
Functional connectivity
PLI	0.26	0.26	0.20	0.21	0.29	0.24*	0.22	0.20	0.10	0.10
Weighted network
Gamma	1.30	1.30	1.30	1.40	1.40	1.30*	1.30	1.30	1.30	1.30
Lambda	1.00	1.10	1.10	1.20	1.30	1.20*	1.20	1.00	1.10	1.10
KappaW	4.10	4.20	4.30	4.30	4.40	4.20*	4.30	4.20	4.20	4.20
Modularity	0.21	0.21	0.19	0.19	0.18	0.21*	0.19	0.21*	0.20	0.20
Minimum Spanning Tree
MST BC	0.71	0.74*	0.73	0.73	0.73	0.71*	0.72	0.72	0.72	0.72
MST Dia.	0.42	0.41	0.40	0.40	0.41	0.43*	0.40	0.43*	0.41	0.41
MST Ecc.	0.33	0.32	0.32	0.32	0.32	0.34*	0.32	0.34	0.32	0.32
MST Leaf	0.56	0.57	0.58	0.57	0.59	0.55*	0.58	0.57	0.56	0.57

^*Significant differences are noted by (p < .05).

Table 5 shows the pairwise comparison of the preliminary analysis for PD-CN v. PD-MCI groups; here, significant differences exist between gamma and Kappa_w only, in the alpha1 frequency band, with functional connectivity differing in the delta and theta bands. While significant differences are not seen in as many measures, those seen do mirror those seen with comparison of PD-CN to PD-D.

Table 5.

Values of network measures for patients with Parkinson's disease who are cognitively normal (PD- CN) compared to those with mild cognitive impairment (PD-MCI).

	Delta		Theta		Alpha1		Alpha2		Beta
	PD-CN	PD-MCI	PD-CN	PD-MCI	PD-CN	PD-MCI	PD-CN	PD-MCI	PD-CN	PD-MCI
Functional connectivity
PLI	0.26	0.27*	0.20	0.22*	0.29	0.26	0.22	0.21	0.10	0.10
Weighted network
Gamma	1.30	1.30	1.30	1.30	1.40	1.30*	1.30	1.30	1.30	1.30
Lambda	1.00	1.10	1.10	1.20	1.30	1.20	1.20	1.10	1.10	1.00
KappaW	4.10	4.20	4.30	4.30	4.40	4.20*	4.30	4.30	4.20	4.20
Modularity	0.21	0.20	0.19	0.19	0.18	0.20	0.19	0.19	0.20	0.20
Minimum Spanning Tree
MST BC	0.71	0.71	0.73	0.73	0.73	0.72	0.72	0.72	0.72	0.72
MST Dia.	0.42	0.42	0.40	0.41	0.41	0.42	0.40	0.41	0.41	0.42
MST Ecc.	0.33	0.34	0.32	0.33	0.32	0.33	0.32	0.33	0.32	0.33
MST Leaf	0.56	0.56	0.58	0.57	0.58	0.56*	0.58	0.57	0.56	0.56

^*Comparisons are adjusted for age and sex; significant comparisons are indicated by (p < .05).

Figure 1 shows the unweighted average network connectivity maps for alpha1 and alpha2 for PD-CN and PD-D groups. A range of thresholds for viewing network connectivity were employed so as to compare the network connectivity between PD-CN and PD-D in these bands. The lines drawn between nodes (electrodes) represent connections that have PLI synchronization above the threshold (connectivity). The absence of the line does not suggest the absence of connectivity between those cortical areas, but rather no connectivity with strength above the set threshold. For the both the alpha 1 and alpha2 frequency bands, the maps show the dramatic decrease in connectivity for PD-D.

Figure 1.

Average network maps for the Alpha1 (above) and Alpha2 (below) frequency bands for each PD-CN (left) and PD-D (right). The Alpha1 maps are shown at thresholds set at .2, .25, and .3 PLI values. The Alpha2 maps demonstrated lower PLI values and are shown at thresholds set at .1, .15, and .2 PLI values. For both bands, the relative disconnection of PD-D network maps compared to PD-CN is evident.

Figure 2 shows Minimum Spanning Trees for the PD-CN and PD-D groups for the alpha1 and alpha2 frequency bands. The PD-D MSTs show a less “star-like” quality, reflecting less global efficiency of the network.

Figure 2.

Minimum Spanning Tree (MST) for PD-CN and PD-D groups for the alpha1 (left) and alpha2 (right) frequency bands. Note the less “star-like” quality of the PD-D MSTs, reflecting overall less global efficiency of the network in PD-D.

To further understand the relationship between network measures and cognitive status, correlations with network measures of patients with PD across cognitive status are calculated with both the MOCA and the MMSE (see Table 6). MMSE had significant correlation coefficients with network measures derived from the lower and upper alpha bands and delta frequency band. MOCA had significant correlation coefficients with network measures derived from the delta, theta, and lower alpha bands. However, when both corresponding relationships showed significance, the MOCA mostly showed higher correlation coefficients than the MMSE. This is particularly true in the alpha1 band, which is the EEG frequency band that also had the most dramatic mean differences between PD-CN and PD-D for network measures in Table 4.

Table 6.

Pearson correlations (R) of network measures and MMSE score, and MOCA score, for patients with PD across the spectrum of cognitive performance (includes patients with PD, PD-MCI, and PD-D).

	MMSE (N = 86)
	Delta	Theta	Alpha1	Alpha2	Beta
Functional connectivity
PLI	.10	−.20	.23*	.18	.11
Weighted network
Gamma	−.06	−.25*	.23*	.12	−.05
Lambda	−.05	−.20	.11	.20	.05
KappaW	−.15	−.17	.35*	.23*	−.03
Modularity	.04	.01	−.30*	−.25*	.00
Minimum Spanning Tree
MST BC	−.3*	−.10	.19	.14	−.08
MST Diameter	.23*	.03	−.17	−.22*	.01
MST Eccentricity	.21*	.02	−.19	−.21*	.06
MST Leaf	−.13	−.03	.15	.11	−.17

	MOCA (N = 51)
	Delta	Theta	Alpha1	Alpha2	Beta
Functional connectivity
PLI	.26	−.09	.19	.22	.13
Weighted network
Gamma	.21	.14	.36*	.32*	.22
Lambda	.1	−.09	.25	.27	.19
KappaW	.29*	.17	.44*	.27	.28
Modularity	−.37*	−.15	−.30*	−.14	−.15
Minimum Spanning Tree
MST BC	.01	.06	.26	.06	.22
MST Diameter	−.2	−.27	−.24	−.15	−.09
MST Eccentricity	−.2	−.28*	−.28*	−.13	−.09
MST Leaf	.24	.31*	.30*	.17	.12

^*Significant correlations are noted by (p < .05).

The multivariable modeling analysis showed that the significant correlations in Table 6 were not confounded by motor severity. In modeling, motor severity did not affect the value of these correlations by more than 10% and, in some instances, modestly strengthened correlations. For example, looking at the strongest correlations among the analysis, while adjusting for UPDRS score, the correlation of gamma in the alpha1 frequency band to the MMSE increased from .23 to .34 and increased the correlation with the MOCA from .36 to .42. Gamma in the alpha2 frequency band also increased from .12 to .18 when correlated with the MMSE and from .32 to .43 when correlated with the MOCA. Kappa_w similarly improved from .35 to .44 when correlated with the MMSE and .44 to .45 when correlated with the MOCA.

4.0 Discussion

Our results reveal EEG evidence of network alteration and breakdown in PD and PD cognitive decline. The sensitivity of network measures obtained via EEG to PD pathology, as well as subsequent decline in cognitive function, is an important contribution regarding pathophysiology and future clinical biomarker use. Interestingly, there is frequency band and network measure-specific sensitivity revealed in the significant differences and correlations. Overall, from the vantage point of physiology, there were significant network alterations in certain measures across all frequency bands in PD-CN compared to Control showing increased local integration. However, qualitatively different network changes suggesting network breakdown in the alpha band existed for PD-D versus PD-CN. The ability of EEG graph theory network analysis to show not only quantity of network change but also different types of network change highlight its ability to offer insight into cortical physiology changes across clinical states. We found that certain network measures acquired via EEG, such as Kappa_w, appear to be stronger surrogate biomarker candidates than other network measures for cognitive decline among individuals with PD. Data on the PD-MCI group suggests these measures may even be sensitive to subclinical or subtle changes in cognitive status.

Examining the network changes between Control and PD-CN, results show higher path length and increased local clustering and modularity for PD-CN. These robust and uniform findings across all frequency bands point strongly towards increased local integration with decreased global efficiency for the network. In graph theory, this is consistent with increased regularization and some loss of “small worldness” within the network. Kappa_w, a measure of highly connected nodes (or hubs), shows band specificity regarding its changes, being increased for PD-CN in the theta and beta bands but decreased in delta and alpha bands. This result shows a shift from hub activity in the band where the peak background frequency exists (i.e. alpha) to both higher and lower frequency bands. The integration of slow and fast band activity has been termed “cross-frequency coupling” and has been proposed to facilitate information transfer across large-scale brain networks (Canolty et al., 2010). Thus, the change from alpha/delta coupling to beta/theta coupling suggests a change in cortical network operation in PD. Furthermore, the combination of changes seen in the MST graph measures of lower diameter and lower eccentricity with higher leaf number indicates that the PD networks are more integrated (star-like topology) than the networks of older healthy subjects. This is consistent with the results from the weighted graph measures mentioned above.

The major physiological correlate of these changes in network measures between Control and PD-CN groups suggests particularly local increased connectivity is emphasized in PD without cognitive impairment. Since these cognitively normal PD subjects have relative early PD, it is possible these differences are predominantly secondary to abnormal sub-cortical influences on cortical network organization. It is important to emphasize that these network changes do not seem to correlate with cognitive performance and as such, may not adversely affect cognition. Indeed, while normal cognitive function is dependent on normal functional connectivity, some network changes seem independent of those important for cognition. The clinical meaning of this increased local connectivity is not clear, but Brown and others have observed excessive synchronization between basal ganglia, cortex, and muscles (Brown, 2003). This activity is partially reversed by levodopa and deep brain stimulation and its presence has been correlated with bradykinesia in PD (Hammond et al., 2007). It is also possible that this increased local connectivity is compensatory for dysfunctional subcortical input coming via basal ganglia dysfunction. It has been suggested that this increased integration occurs in the network with higher energy costs, thus perhaps putting the network under stress (de Haan et al. 2012).

The comparison of network measures between PD-CN and PD-D groups shows differences that are qualitatively separable from Control and PD-CN groups (Table 3). First, even though there are numerous differences, every difference found between PD-CN and PD-D, except one, is within the alpha frequency range (alpha1 or alpha2). Second, many measures are in the opposite direction, when comparing Control versus PD-CN and PD-CN versus PD-D. The differences between PD-CN and PD-D reveal network breakdown characterized by the decreased functional connectivity, i.e. decreased clustering and less PLI synchronization in the background frequency band (alpha1). In graph theory, the shorter path length and decreased clustering in PDD correspond to evolution towards a “random network”. The smaller Kappa_w mean in PD-D supports the decreased connectivity strength of hubs. The changes MST graph measures showing decreases MST Leaf and MST BC, along with increases in MST Diameter and Eccentricity all support diminished global efficiency.

The major physiological correlate derived from the network measure changes between PD-CN and PD-D suggests network breakdown correlated with decreased cognitive performance. More specifically, in the frequency band that contains the background frequency for EEG activity, there are many network changes showing less connectivity and decreased global network efficiency with a loss of highly connected cortical hubs. Decreased hub connectivity has been found in many disorders (Crossley et al., 2014). Importantly, these changes are all in the hypothesized direction that is consistent with network breakdown in cognitive dysfunction. This is similar to network changes seen in Alzheimer's disease (Stam et al., 2009; Sanz-Arigita et al., 2010); perhaps suggesting that this pattern of network measure changes are important for tracking cognition, irrespective of the underlying pathology. In Figure 1, the network connectivity map of the PD-D group shows a visually depicted concept of cortical area disconnection, which may reflect cortical dysfunction that accompanies global cognitive deterioration. This disconnection has also been suggested to correlate to increases in delta frequency bandpower (Caviness et al., 2015). Braak pathology stages 4-6 are associated with cortical Lewy pathology and cognitive dysfunction (Braak et al., 2005). Thus, cortical LTS may play a causative role in cortical area disconnection that results in network breakdown in PD. It is remarkable that the comparison between the PD-CN and PD-MCI subject groups show these differences are already occurring with mild cognitive impairment (see Table 5). Indeed, comparing PD-CN to PD- MCI, there are decreases in Kappa_w and MST Leaf (alpha1 band), consistent with hub loss and diminished global efficiency, as well as decreases in gamma (alpha1), consistent with decreased local integration. Differences between PD-CN and MCI mirror those seen when comparing the PD-CN to the PD-D group in the current study and, again, to previous literature examining Alzheimer's disease via network analysis of MEG data (Stam et al., 2009). Interestingly, the PD-MCI group demonstrates no change in PLI synchronization. Thus, while the functional connectivity strength is relatively spared, there are early, detectable signs of network decompensation. Across the spectrum of PD, taking into account the differences that we have found, it is tempting to speculate that the evidence of network stress in our PD-CN group over time could translate to the network breakdown in our PD-D group. This general idea has been validated for AD using a simulation model (de Haan et al., 2012). Longitudinal followup of our cases is ongoing.

Previous studies using graph theory in PD have employed different methods and examined different PD populations (Baggio et al., 2014; Olde Dubbelink et al., 2014b; Skidmore et al., 2011; Gottlick et al., 2013; Pereira et al., 2015). Our results provide support of network changes in PD, but our findings revealed important and interesting differences. Also, the current study design provided different advantages. Olde Dubbelink and colleagues used MEG to study graph theory measures in PD (Olde Dubbelink et al., 2014b). This MEG study is the only investigation that the frequency bands in our EEG study can be compared with, since the few other graph theory studies in PD have been done with either structural MRI (no time series analysis) or fMRI which detects brain activity frequencies at about <0.15 Hz. They found differences between de novo PD patients and controls in delta band network measures. In contrast, we found changes between controls across all frequency bands, and our changes reflected increased local connectivity. None of the patients in the present study were de novo, and they had PD for a longer period of time with more motor severity. It would be interesting to confirm that the more advanced disease without cognitive impairment is responsible for this difference in findings. Longitudinally, Olde Dubbelink et al. observed decreased clustering and leaf number in theta, alpha1, and alpha2 bands. They further found that path length decreased in alpha2 band (Olde Dubbelink et al., 2014b). Although we did not follow our PD subjects longitudinally in this study, these differences are overall similar when comparing our PD-CN and PD-D groups. However, it is difficult to directly compare results, since their longitudinal study did not provide a comparison of PD-CN and PD-D, but rather used disease duration and time as the variables of interest. Indeed, their patients are de novo patients, and have mean disease duration of 0.92 years. In the present study, the mean disease duration is 9.7 years. The difference in subject groups and methods between their study and ours could have contributed to qualitative and quantitative result differences.

Studies by Skidmore et al., Gottlich et al., and Baggio et al. all used fMRI for graph theory network analysis (Baggio et al., 2014; Skidmore et al., 2011; Gottlick et al., 2013). Skidmore et al. and Gottlich et al. both found network alteration in PD, but the cognitive state of the patients was not given and thus hinders comparison with other studies (Skidmore et al., 2011; Gottlick et al., 2013). The study by Baggio and colleagues studied differences in cognitive states of PD patients, comparing cognitively normal and PD-MCI patients to healthy controls (Baggio et al., 2014). Interestingly, when looking at the whole cohort of PD patients, regardless of cognitive status, there were no differences in network measures between patients and healthy controls. However, when examining the cohort of MCI patients, they found increased local interconnectedness with reduced long-range connections. This finding suggests that there are also network changes at the low frequencies assessed by fMRI which seem complimentary to the changes observed at the faster frequencies of MEG and EEG. Graph theory was applied to structural measures of cortical thickness and subcortical volumes in an MRI study examining PD-MCI (Pereira et al., 2015). This study found disruptions in large- scale brain networks for PD-MCI, but it is difficult to directly compare structural findings with physiologic method measures derived from time series (e.g. MEG, EEG, fMRI). To date, there is no other EEG study assessing PD cognition stages utilizing the PLI graph theory approach with which the present study could be compared.

Another important contribution of the current work is the direct comparison of cognitive tests and their relationship to network parameters of patients with PD, along the continuum of cognitive decline. There are several insights gained in assessing correlations between network measures and cognitive exams- the MOCA and MMSE. There are strong correlations between the measures and both cognitive exams. Together with group differences, these positive correlations reinforce the concept that these measures are promising as biomarkers for cognitive decline. Importantly, the strong, significant correlations that exist are in the expected direction, suggesting declining global and local integration with increasing severity. It is interesting that the strongest correlation for both MOCA and MMSE is Kappa_w, suggesting that hub connectivity is a key requirement for cognitive performance. When comparing the correlations of the MOCA and MMSE to graph theory measures, there are somewhat stronger relationships with the MOCA (see Table 6). This is consistent with current thinking in the field regarding improved sensitivity of the MOCA, when compared to MMSE, for screening cognitive deficits in patients with PD (Hoops et al., 2009; Zadikoff et al., 2008; Nazem et al., 2009).

There are limitations to our study. We used only 21 EEG electrodes in this analysis to find evidence that such an EEG recording could show important differences. However, using more electrodes (nodes) could yield different results. There is a potential concern about the influence of levodopa, or other anti-Parkinsonian medications when comparing controls to the PD-CN group, and we are unable to determine if such medication affected outcomes. However, there were no significant differences between PD groups for levodopa equivalents. While the current interpretations are made based on a between-subjects design, a longitudinal design will be important further interpretation of these findings. A limitation of our exploratory analyses of the PD-MCI group is the small sample size. Larger cohorts of each cognitive subgroup of PD subjects will allow for stronger statistical power and estimation of effect sizes. Nevertheless, the preliminary findings in the PD-MCI group are interesting and so will warrant further study. Finally, it is possible that average clustering coefficient and path length values, and other measures, are dependent on the average connectivity in a graph. It is of note that we used normalized measures for the average clustering coefficient and average shortest path length to correct for the influence of differences in connectivity strength between groups. This normalization was done by comparing the clustering coefficient and path length to 50 surrogate random networks that were constructed from the original networks by randomly reshuffling the connectivity values with preservation of the number of connections of each node. Because of this preservation, kappaW and modularity, which by definition use the number of connections for each node, could not be normalized and were potentially influenced by the PLI values. In all bands, except for the theta band in the comparison of PD-CN and control and the alpha2 band in the comparison of PD-CN and PD-D, we did not find significant PLI differences between the groups; therefore, we regard the differences in Kappa_w and modularity between the groups as relevant and not driven by PLI differences only. In the theta band (PD-CN v. Control) and alpha1 band (PD-CN v. PD-D) we found a PLI difference, but in addition to differences in the Kappa_w and modularity there were differences in the normalized measures as well, indicating that the differences again were not driven only by the PLI difference.

5.0 Conclusions

In summary, our results reveal EEG evidence of network alteration and breakdown in PD and PD cognitive decline. The changes and differences between groups can be interpreted to be related to physiological changes. It is also the case that there are qualitative differences between the group comparisons in addition to quantitative differences. This includes differences in both measurements of weighted graph and MST graph network measures, emphasizing the important complimentary role of MST graphs which may be less biased for assessing between groups (Stam et al., 2014; Tewarie et al., 2015). Network over-activity could be speculated as network overload and distress in the PD-CN group. The network breakdown seen in the PD-D group may or may not be due to similar neuronal pathology to AD. Hence, it would be valuable to correlate network patterns with pathology and biochemistry, as PD-D cases exhibit heterogeneous findings including variable presence of AD pathology criteria (Caviness et al., 2011). Different modalities used in graph theory analysis of PD should be viewed as complimentary. Additionally, EEG is more widely available than MEG and examines brain frequencies 100-1000 times faster than fMRI. Thus, incorporation after full validation of such biomarkers into clinical practice could occur without cumbersome modifications to current clinical EEG practice. Importantly, this information also helps to generate hypotheses of pathophysiology of cognitive decline. Indeed, network connectivity, possibly through repairing synaptic problems in synucleinopathy, could form a basis for a therapeutic target (Caviness et al., 2011). This may help discovery pharmacological intervention that can act on the mechanisms that result in cognitive decline.

Highlights.

• Network measures showed correlation with cognitive performance in alpha band.

• Non-demented Parkinson's disease subjects showed increased local integration.

• Demented Parkinson's disease subjects showed network disconnection.