Analysis of complexity and dynamic functional connectivity based on resting-state EEG in early Parkinson’s disease patients with mild cognitive impairment

Abstract

To explore the abnormal brain activity of early Parkinson’s disease with mild cognitive impairment (ePD-MCI) patients, the study analyzed the dynamic fluctuation of electroencephalogram (EEG) signals and the dynamic change of information communication between EEG signals of ePD-MCI patients. In this study, we recorded resting-state EEG signals of 30 ePD-MCI patients and 37 early Parkinson’s disease without mild cognitive impairment (ePD-nMCI) patients. First, we analyzed the difference of the complexity of EEG signals between the two groups. And we found that the complexity in the ePD-MCI group was significantly higher than that in the ePD-nMCI group. Then, by analyzing the dynamic functional network (DFN) topology based on the optimal sliding-window, we found that the temporal correlation coefficients of ePD-MCI patients were lower in the delta and theta bands than those in the ePD-nMCI patients. The temporal characteristic path length of ePD-MCI patients in the alpha band was higher than that of ePD-nMCI patients. In the theta and alpha bands, the temporal small world degrees of ePD-MCI patients were lower than that of patients with ePD-nMCI. In addition, the functional connectivity strength of ePD-MCI patients affected by cognitive impairment was weaker than that of ePD-nMCI patients, and the stability of dynamic functional connectivity network was decreased. This finding may serve as a biomarker to identify ePD-MCI and contribute to the early intervention treatment of ePD-MCI.

Introduction

Parkinson’s disease (PD) is the second most common age-dependent neurodegenerative disease (Zhang and Roman 2005; Ma et al. 2014). The incidence rate of PD increased with age, and more often occurred in middle and old age. The mean age of onset was 60. Currently, there are over 2 million PD patients in China and the prevalence of Parkinson’s disease in the world population is about 0.1–0.5% (Chen et al. 2016). In China, the prevalence of PD in the elderly over 65 years old is about 1%, and that over 70 years old is as high as 5–8%. The disease is related to genetic, environmental and nervous system aging and so on (Kalia and Lang 2015). The typical clinical manifestations of motor symptoms mainly include static tremor, bradykinesia, myotonia and postural and gait disorders (Postuma 2015). Non-motor symptoms such as depression, sleep disorders and cognitive impairment may also occur (Park and Stacy 2009; Chen et al. 2012). Parkinson’s disease is diagnosed based on typical clinical symptoms. In the prodromal stage, patients may have sleep disorders, olfactory disorders and other neuropathic diseases similar to the common diseases of the elderly. Therefore, after the diagnosis of PD, most ePD patients have developed mild cognitive impairment (Litvan et al. 2011). There are two main ways to diagnose cognitive impairment in Parkinson’s disease. Firstly, Parkinson’s disease should be identified. Secondly, the degree of cognitive impairment was assessed based on the Montreal Cognitive Assessment (MoCA) or the Addenbrooke’s Cognitive Examination (ACE) scales (Gill et al. 2008; Mioshi et al. 2006). Nearly 30% of PD patients had cognitive impairment, and cognitive impairment has become a risk factor for the development of Parkinson’s disease dementia (PDD) (Aarsland et al. 2017). The incidence of Parkinson’s cognitive impairment and Parkinson’s dementia increases with the progression of Parkinson’s disease (Bellomo et al. 2020). If Parkinson’s disease progresses to dementia, there will be no effective treatment for dementia and no chance of surgery to treat the motor symptoms of Parkinson’s disease. Therefore, there is an urgent need for reliable biomarkers of Parkinson’s cognitive impairment (PD-MCI) to assist physicians in diagnosing and monitoring patients with high risk and poor prognosis of PD-MCI.

Currently, neuroimaging techniques for the diagnosis of Parkinson’s disease include positron emission tomography (PET), functional magnetic resonance imaging (fMRI), electroencephalogram (EEG) and so on. However, in practical clinical application, PET research is radioactive and expensive, which hinders the application of this technology on a larger scale. fMRI has high spatial resolution, but the examination time is long, and the rich rhythm information of the brain cannot be obtained. Its low temporal resolution is also not conducive to the detection of brain activity on small time scales. Although EEG signal has low spatial resolution, it has high temporal resolution, convenient collection, low cost and contains rhythm information. It can well represent the spontaneous oscillating activity of the cerebral cortex. Depending on whether the brain responds to a particular stimulus, EEG signals can be divided into spontaneous and induced EEG. Resting state EEG is a typical spontaneous EEG. Our research work is based on the EEG signal of the resting state.

The brain is a highly complex nonlinear chaotic system. EEG signals also have nonlinear dynamic characteristics. The complexity analysis method can be used to explore the complex performance of EEG signals. The complexity of EEG signals contains rich dynamic information of cerebral cortex under different rhythms. Complexity analysis of EEG signals has shown that the brain was a highly variable dynamic activity. In addition, studies have found that the Lempel–Ziv complexity (LZC) of EEG signal in the mild cognitive impairment (MCI) group at the early stage of cognitive activity was significantly higher than that in the normal control group, and the LZC of MCI patients gradually decreased over time. Results in the normal group remained the same. The two sets of complex topographic maps had similar regularity in spatial distribution (Yan et al. 2013). Another study also found that the LZC value of MCI patients was between Alzheimer’s disease (AD) patients and normal group (Fernndez et al. 2010). One research found that PDD patients showed higher EEG signal complexity and lower functional connectivity strength in low frequencies (mainly in delta frequency) than patients who remained dementia-free and controls (Bertrand et al. 2015). An analysis of the complexity of EEG signals in MCI patients and those with mild, moderate and severe AD patients found that MCI patients showed higher complexity than other groups (Maturana-Candelas et al. 2019). Many studies have shown that cognitive decline was strongly associated with irregular loss of EEG signals (Abásolo et al. 2013; Gomez et al. 2016; Simons 2018; Cao et al. 2015). Whether the change of EEG signal complexity in PD-MCI population overlapped with MCI, AD or PDD remains to be investigated.

The human brain is a network structure composed of numerous neuron cells connected by a large number of synapses, which guides the efficient completion of brain functions. Synchronous EEG signals are output as neural electrical activity caused by physiological functions of the brain. It is also widely used to analyze brain structure and function (Cai et al. 2011). Brain networks can be constructed from the perspectives of morphology, function and mathematical model. Functional connectivity (FC) is measured by spatially separating the brain into different regions. Then the correlation coefficients of time series’ signals between brain regions were measured. Carmona et al. (2019) used the coherence of resting EEG to assess local, intra, and interhemispheric connectivity in PD patients with and without mild cognitive impairment (MCI). The results showed that PD subjects without MCI (PD-nMCI) had lower intra and interhemispheric coherence in the alpha2 band compared with controls. PD with MCI (PD-MCI) showed higher intra and posterior interhemispheric coherence in alpha2 and beta1, respectively, in comparison to PD-nMCI. PD-MCI presented lower frontal coherence in the beta band compared with PD-nMCI. In other research, through the construction of functional connectivity network based on Phase lag index (PLI), it was found that the functional connectivity of patients with mild cognitive impairment decreased significantly, and the degree of graph topology integration was low (Duan et al. 2020). Mehraram R et al found that the connectivity strength of PDD was lower than that of healthy subjects in the alpha band (Mehraram et al. 2020). Some studies have also found that the deterioration of cognitive function in PD patients maybe lead to the decline of functional connectivity (Hassan et al. 2017; Lopes et al. 2017; Peláez Suárez and Berrillo Batista 2021). PLI was first proposed by Stam et al. (2007). It was a measure of functional connectivity in the brain based on EEG signals. The study of static functional connectivity networks provided a basic understanding of basic neural function and mental illness (Bai et al. 2009; Rombouts et al. 2006; Fox et al. 2005; Greicius 2004; Smith et al. 2009). However, some researchers have found that FC can change periodically with time rather than static time during fMRI scanning (Allen et al. 2014; Chang and Glover 2010; Jones et al. 2012; Liu and Duyn 2013; Majeed et al. 2011; Sakoğlu et al. 2010; Woolrich et al. 2012). Alterations in dynamic functional connectivity have been found in a variety of neurological diseases such as schizophrenia, AD, PD, major depression disorders, and autism (Jones et al. 2012; Damaraju et al. 2014; Cordes et al. 2018; Holtzheimer and Mayberg 2011; Price et al. 2014). In the process of dynamic fluctuation of the brain, the functional structure of the brain changes with cognitive activities. A study found that PD-MCI patients had an increased number of state transitions in the low FC state and unstable network state (Díez-Cirarda et al. 2018). Therefore, abnormal network dysfunction may be a potential diagnostic marker in the ePD-MCI population.

It was known that patients with cognitive impairment in Parkinson’s disease had already shown motor symptoms, so if the group of PD patients with cognitive impairment is selected and compared with normal people, the differences between the two groups include the motor symptoms already present in addition to the cognitive impairment, which will interfere with the results of cognitive impairment. Therefore, ePD-MCI group and ePD-nMCI group were compared in order to exclude the influence of Parkinson’s underlying pathological abnormality.

The purpose of our study is to explore how cognitive impairments affect PD patients through the analysis of complexity and dynamic brain functional networks, and to combine the obtained dynamic characteristics as a basis for assisting in the diagnosis of ePD-MCI. The results may help to control and even predict cognitive impairment in ePD patients and timely intervention treatment.

Materials and methods

Subjects

The study was approved by the local ethics committee. The purpose and the research significance of the data were explained to all participants. After obtaining the informed consent, sixty-seven patients from the Department of Neurology of Tianjin Medical University General Hospital were enrolled. Table 1 reported demographic data and clinical scales from the ePD-MCI group and ePD-nMCI group. Results are expressed as mean (standard deviation). We diagnosed PD through the third part of Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) (Goetz et al. 2004). And the Hoehn and Yahr (H&Y) stages were determined by the third part of MDS-UPDRS. The subjects were all PD patients with H&Y staging scale I-II (Goetz et al. 2004). The age of onset of the patient was over 40 years old, and the possibility of early onset PD was excluded. Patients had been off medication for more than 12 h in order to collect EEG data in the absence of pharmacodynamic effects (including levodopa). The patients had no significant health abnormalities such as cerebrovascular conditions. The experiment subjects were divided into two groups with a total of 67 subjects. There were 30 patients with ePD-MCI ($\mathrm{n}=30$) and 37 patients with ePD-nMCI ($\mathrm{n}=37$). All patients participated in neuropsychological tests. There were no significant (n.s.) differences statistically in age, the REM sleep behavior disorder screening questionnaire (RBDSQ) (Wang et al. 2015). According to the MoCA scale (Gill et al. 2008), 21–25 were classified as mild cognitive impairment and 26–30 as cognitively normal (Nasreddine 2005).

Table 1.

Basic information of patients

Information	ePD-MCI ($\mathrm{n}=30$)	ePD-nMCI ($\mathrm{n}=37$)	p-value
Age	60.70 (8.39)	63.24 (5.62)	n.s.
H–Y stage	1.57 (0.55)	1.15 (0.39)	n.s.
Course of the disease	3.67 (2.66)	4.78 (5.37)	n.s.
MoCA scale	23.03 (1.59)	27.62 (1.26)	1.2145e−10
RBDSQ	0.33 (1.12)	0.30 (1.22)	n.s.

EEG recording and preprocessing

When collecting spontaneous EEG signals, subjects lay comfortably in a quiet semi-dark room, closed their eyes and kept awake. 19 Ag-AgCl electrodes (Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T7, T8, P7, P8, Fz, Cz, and Pz) were applied according to the international 10-20 system of electrode placement as shown in Fig. 1. The signal sampling frequency is 500Hz. The assessment scales used in the study were the third part of Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale sponsored by the Movement Disorders Association and the MoCA scale.

Fig. 1.

Diagram of EEG electrode distribution

Then the EEG signals of the collected subjects was preprocessed by MATLAB software (MathWorks Inc., Natick MA, United States). First, EEG data was filtered to 1 Hz–30 Hz by zero-phase-shift filtering. Fast Independent Component Analysis (FastICA) extracted different independent components of each channel. The components related to horizontal and vertical ocular artifacts were set to zero, and the EEG after removal of ocular artifacts was reconstructed. However, interference caused by head movement and body movement was difficult to be removed by filtering and FastICA, so it can be eliminated by manual screening. Finally, the processed EEG of each subject was obtained, and continuous EEG signals for the first 200s after artifact and noise removal were selected. After the completion of EEG signal screening, the frequency division operation was carried out through the band-pass filter, and the EEG signal was divided into the four sub-bands, namely delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz) and beta (13–30 Hz) bands. In this study, the complexity and dynamic functional connectivity of ePD-MCI patients were analyzed on the basis of the four sub-bands.

Complexity analysis

Fuzzy entropy is a method proposed for complexity analysis of time series by Chen et al. (2007). It was measured by calculating the probability of a new pattern arising from a change in dimension of the time series. As the value of fuzzy entropy increased, it meant that the complexity of EEG time series increased. The fuzzy entropy was independent of the data length and maintained relative consistency. By using the exponential function to fuzzy the similarity measure formula, the entropy fluctuation with the change of parameters was smoother. For an N sample time series $\{u(n):1\le n\le N\}$, given embedding dimension m, form M-dimensional vector sequences $\left\{U_t^m,,,t,=,1,,,\dots ,,,N,-,m,+,1\right\}$ as follows:

$$\begin{array}{c}\hfill U_t^m=\left\{u_t,,,u_{t+1},,,\dots ,,,u_{t+m-1}\right\}-\sum _{j=0}^{m-1}\frac{u_{t+j}}{m}\end{array}$$ 1

Define the distance between $U_{t1}^m$ and $U_{t2}^m$ as follow:

$$\begin{array}{c}\hfill \begin{array}{cc}\hfill d_{t1t2}^m& =d\left[U_{t1}^m,,,U_{t2}^m\right]\hfill \\ \hfill & =max\left\{\left|U_{t1+k}^m,-,U_{t2+k}^m\right|,:,0,\le ,k,\le ,m,-,1,,,t,1,\ne ,t,2\right\}\hfill \end{array}\end{array}$$ 2

where the fuzzy function is ${\text{e}}^{-{\left({\text{d}}_{\text{t}1\text{t}2}\right)}^{\text{n}/\text{r}}}$ (n and r are the gradient and width of the boundary of the fuzzy function respectively).

Define function ${\varphi }^m$ as:

$$\begin{array}{c}\hfill {\varphi }^m=\frac{1}{N-m}\sum _{t1=1}^{N-m}\frac{1}{N-m-1}\sum _{t1=1,t1\ne t2}^{N-m}{\mathrm{e}}^{-{\left({\mathrm{d}}_{\mathrm{t}1\mathrm{t}2}\right)}^{\mathrm{n}/\mathrm{r}}}\end{array}$$ 3

Finally, the value of the fuzzy entropy is

$$\begin{array}{c}\hfill FuzEn(m,n,r)=\underset{N\to \infty }{lim}\left[ln,\left({\varphi }^m\right),-,ln,\left({\varphi }^{m+1}\right)\right]\end{array}$$ 4

The temporal variability of functional connectivity strength between the whole brain and different brain regions was calculated by fuzzy entropy measurement.

Functional connectivity network

PLI is obtained by calculating the instantaneous phase of the EEG signal and the asymmetry of the phase difference distribution between two signals (Stam et al. 2007). The expression is shown in (5).

$$\begin{array}{c}\hfill \mathit{PLI}=\left|〈\text{sign},\left[\mathrm{\Delta },\varphi ,\left(t_k\right)\right]〉\right|,k=1,2,\dots ,N\end{array}$$ 5

where $\mathrm{\Delta }\varphi \left(t_k\right)$ is the time series of phase differences calculated at N time points. The range of PLI is from 0 to 1. The smaller the PLI value is, the smaller the coupling strength between the two signals is. Conversely, the greater the coupling strength. When the PLI is 0, it indicates that there is no coupling or coupling with a phase difference centered around 0 mod $\pi$ between the two signals, and when the PLI is 1, it indicates that the two signals are perfectly locked at a value of $\mathrm{\Delta }\varphi$ different from 0 mod $\pi$. The characteristic parameters include temporal correlation coefficient, temporal characteristic path length and temporal small world degree were used to describe the DFNs (Tang et al. 2010).

The weighted functional connectivity network of each subject was averaged to get the functional connectivity strength of the whole brain, namely the global mean functional connectivity strength(FCS).

Optimal sliding-window extraction

For the method of sliding-window building DFNs, the set of time window and steps length affected the difference between ePD-MCI group and ePD-nMCI group. Appropriate window size and step size are more helpful to find the differences between the two groups after constructing the DFNs. Therefore, the optimal time window is determined through the formulation of technical route.

Step 1: The sample frequency of EEG signal was 500 Hz. The time interval between the two data points was 0.002 s. Firstly, 500 data points to 2000 data points (1 s–4 s) were selected as window lengths, and every 10 data points (0.02 s) was an interval. The overlap between the windows is set to 0–0.5 times the window length. Finally, we got 906 different window length settings of the four sub-bands.

Step 2: In order to judge the dynamic characteristics of the DFNs based on slinding-window, the temporal variability of the edge of the dynamic networks built by EEG and the corresponding zero model were analyzed respectively. The dynamic performance of the DFNs is represented by its ratio. The larger the value of $k_{\mathit{PLI}}$, the stronger the dynamic characteristics of the functional network. The lower the value of $k_{\mathit{PLI}}$, the weaker the dynamic characteristics of the functional network. $k_{\mathit{PLI}}$ is expressed by the following formula:

$$\begin{array}{c}\hfill k_{\mathit{PLI}}=\frac{k_{\text{real}}}{1{/}_N{\sum }_{i=1}^N{k}_{\text{random}}(i)}\end{array}$$ 6

where N represented the N random time series ($N=50$), $k_{\mathit{real}}$ is the standard deviation of each edge obtained by the actual functional networks in the dynamic process, which is used to describe the degree of the coupling strength between signals fluctuates in the process of dynamic changes of brain functional networks. $k_{\mathit{random}}$ is the standard deviation of each edge obtained by the zero-models.

Step 3: The obtained $k_{\mathit{PLI}}$ values of different sliding-window types in each group were analyzed for the difference between ePD-MCI and ePD-nMCI patients in the two groups. The sliding window with the most obvious significance was determined separately in four sub-bands as the optimal time window for ePD-MCI and ePD-nMCI patients.

Zero-model

The amplitude adjusted Fourier transform (AAFT) method was used to construct the zero-model based on the surrogate data generated from the real EEG data (Kugiumtzis 2002). The construction of zero-model is as follows:

Step 1: Generating Gaussian Sequence y(t) with Pseudo Random Number Generator, and the length of random sequence was consistent with the original data x(t).

Step 2: The rank of the original sequence x(t) was used to rearrange y(t), so as to ensure that the rearranged sequence not only followed the order of x(t) but also had the Gaussian amplitude distribution.

Step 3: Obtaining another series $y^{\prime }(t)$ by Fourier transform and phase randomization.

Step 4: The original data were rearranged according to the rank of $y^{\prime }(t)$ to obtain the surrogate data $x^{\prime }(t)$ of the original data.

Step 5: 50 groups of random sequences were obtained based on Step 1–4, and the DFNs constructed through PLI of the obtained random time series were taken as the zero-model.

Dynamic graph theory metric

Connections in complex networks was inherently time-varying and can capture more dimensions than analysis based on standard static graph measurements. The concepts of temporal path and distance of time-varying graph were introduced to obtain dynamic network parameters, which were used to measure the stability and transferred efficiency of dynamic brain functional network (Rombouts et al. 2006). Temporal correlation coefficient of node i in DFN was defined as:

$$\begin{array}{c}\hfill C_i=\frac{1}{T-1}\sum _{t=1}^{T-1}\frac{{\sum }_j{A}_{i,j}(t)A_{i,j}(t+1)}{\sqrt{\left[A_{i,j},(t)\right]\left[A_{i,j},(t+1)\right]}}\end{array}$$ 7

it represented the information aggregation ability of nodes i at adjacent time points. The larger temporal correlation coefficient indicated that it had strong local connectivity stability.

Temporal characteristic path length was defined as:

$$\begin{array}{c}\hfill L=\frac{1}{N-1}\sum _{i,j}d(i,j)\end{array}$$ 8

Temporal small world degree was defined as:

$$\begin{array}{c}\hfill \mathit{SW}=\frac{C/c_{\text{rand}}}{L/L_{\text{rand}}}\end{array}$$ 9

Statistical analysis

The statistical analysis was performed by Analysis of Variances (ANOVA) to evaluate the differences between ePD-MCI group and ePD-nMCI group, and the results were considered as significant at the level of $p<0.05/N$. N was the Bonferroni correction factor, which referred to the number of results tested. Spearman’s correlation was used to show whether there was a certain correlation between the indicators and age or course of disease. The false discovery rate (FDR) was used to correct the statical results, and spearman’s correlation was accepted at $P_{\mathit{FDRcorrected}}<0.05$. We used the ANOVA, corr and mafdr function of MATLAB, other methods were independently written using MATLAB.

Diagnostic accuracy

In order to test the potential diagnostic utility of the feature markers obtained in this study, we used the SVM classifier with 5-fold cross validation to test through Python software. The classifier was trained by using the significant difference index obtained in the sub-bands, and the ROC curve and accuracy were obtained. We conducted diagnostic testing studies on diagnostic scenarios that resulted in significant differences in EEG and DFN analysis.

Results

First, we analyzed the complexity of EEG signals in 19 electrodes and whole brain between ePD-MCI group and ePD-nMCI group in the four sub-bands as shown in Figs. 2 and 3. We found that the EEG signal’s complexity in the whole brain of ePD-MCI patients was all significantly higher than that of ePD-nMCI patients in the four sub-bands as shown in Fig. 3a. The topographic maps of the left two columns in Fig. 2 showed the average complexity values of EEG data at each electrode of ePD-MCI group and ePD-nMCI group in the sub-bands. And the red region of the p-value topographic map in Fig. 2 (right) is the region of significant difference ($p<0.05$, ANOVA with Bonferroni correction) in the complexity of each electrode’s EEG data between the two groups.The results indicated that the brain regions with significant differences in the delta band were mainly distributed in the frontal lobe, temporal lobe and central region. And the complexity of ePD-MCI group was generally higher in these regions than ePD-nMCI group. In the theta band, the brain regions with significant differences were also mainly distributed in the frontal lobe, temporal lobe and central region, as well as the complexity of ePD-MCI group was generally higher in these regions than ePD-nMCI group.

Fig. 2.

Topographical distribution of the complexity of EEG signals for ePD-MCI patients (left) and ePD-nMCI patients (center) in the sub-bands. The red regions of the right column show the significant differences (Bonferroni adjusted $p<0.05$) of corresponding measure between two groups

Fig. 3.

Analysis of the global mean complexity of EEG signals in the sub-bands. a The difference analysis of global mean complexity between ePD-MCI patients and ePD-nMCI patients in the sub-bands. The results of difference are shown by the P value at the top of each subfigure, and the results with significant differences are highlighted in red font. b Complexity was correlated with age and course of disease respectively by Spearman analysis in the significantly different sub-bands. FDR correction was performed for P values. (Color figure online)

There was a significant difference in almost the whole brain regions of the alpha band, and the brain electrical signal complexity of ePD-MCI patients was generally higher than that of ePD-nMCI patients. The brain regions with significant differences in the beta band were mainly distributed in the frontal lobe region. And the complexity of ePD-MCI group was generally higher in these regions than ePD-nMCI group. Compared with ePD-nMCI patients, there were significant differences in the prefrontal lobes in all four frequency bands. And in Fig. 3b, we found that there was no correlation between the complexity and age or course of disease in the four sub-bands, which excluded the effect of age and course of disease on complexity of ePD-MCI patients.

Then we completed the determination of the optimal time window in the four sub-bands. The variance test based on $k_{\mathit{PLI}}$ indicator for ePD-MCI and ePD-nMCI groups was conducted in the sub-bands, and it was found that the length of sliding window corresponding to the largest difference ($p<0.01$, ANOVA with Bonferroni correction) between two groups was 1.4 s and the overlap degree was 30% in the delta band. The optimal sliding window with the largest difference ($p<0.01$, ANOVA with Bonferroni correction) in the theta band was the window length of 1.2 s and the overlap degree of 0. The optimal sliding window with the largest difference ($p<0.01$, ANOVA with Bonferroni correction) was 2.98 s window length and 50% overlap degree in the alpha band. There is no significant difference in the beta band under different sliding windows. Therefore, we set the window length within the reasonable range of traditional sliding window. The results of significant difference in $k_{\mathit{PLI}}$ indicator between ePD-MCI group and ePD-nMCI group in the four sub-bands were showed in Fig. 4.

Fig. 4.

The difference analysis between ePD-MCI patients and ePD-nMCI patients based on the optimal sliding-windows in the sub-bands. The vertical axis represents the $k_{\mathit{PLI}}$ indicator. The results of difference are shown by the P value at the top of the figure, and the results with significant differences are highlighted in red font. (Color figure online)

The DFNs of ePD-MCI group and ePD-nMCI group were constructed according to the obtained optimal sliding-window in the sub-bands, and the attribute characteristics based on dynamic graph theory were analyzed. As shown in Fig. 5, it could be seen that in the process of DFNs changing, the temporal correlation coefficients of ePD-MCI group in both delta and theta bands were significantly different ($p<0.05$, ANOVA with Bonferroni correction) from those of ePD-nMCI group, and they were lower than that of ePD-nMCI group. And the temporal correlation coefficient described the stability of the functional connectivity network in the process of dynamic. It indicated that the stability of functional connectivity network of ePD-MCI patients was lower than that of ePD-nMCI patients in the delta and theta bands.The temporal characteristic path length represented the average time connectivity of the DFNs. Lower temporal characteristic path length indicated efficient communication at the whole brain level. The temporal characteristic path length of ePD-MCI and ePD-nMCI groups was significantly different ($p<0.05$, ANOVA with Bonferroni correction) in the alpha band, and ePD-MCI group was generally higher than ePD-nMCI group. The results indicated that the average time connectivity of the DFNs increased due to the influence of cognitive in the alpha band, which led to lower communication efficiency than that in ePD-nMCI group. ePD-MCI group had lower temporal small world degree in the theta and alpha band than ePD-nMCI. It showed that brain information communication speed of ePD-MCI group was slower than that of ePD-nMCI group. Therefore, the brain of ePD-MCI patients may need to separate and integrate information at a higher connectivity capacity and energy cost than ePD-nMCI patients.

Fig. 5.

Dynamic graph measures based the DFNs between ePD-MCI patients and ePD-nMCI patients in the sub-bands. C represents temporal correlation coefficient, L represents temporal characteristic path length, and $\mathit{SW}$ represents temporal small world degree. The results of difference are shown by the P value at the top of each subfigure, and the results with significant differences are highlighted in red font. (Color figure online)

Figure 6a analyzed the difference of global mean functional connectivity strength (FCS) between the two groups in the four sub-bands. It was found that the strength of functional connectivity of ePD-MCI in the delta band was significantly lower ($p<0.01$, ANOVA with Bonferroni correction) than that of ePD-nMCI. And there was no correlation between FCS and age or course of disease in the delta band in Fig. 6b.

Fig. 6.

The functional connectivity strength analysis of whole brain. FCS represents functional connectivity strength. a The difference analysis of FCS between ePD-MCI patients and ePD-nMCI patients in the sub-bands. The results of difference analysis are shown by the P value, and the results with significant differences are highlighted in red font. b Spearman’s correlation analysis between FCS and age or course of disease in the delta band. FDR correction was performed for P values. (Color figure online)

As shown in Fig. 7a, the results indicated that temporal variability (FCTV) of functional connectivity based on whole brain in the delta, theta and alpha bands under dynamic fluctuations was significantly different ($p<0.01$, ANOVA with Bonferroni correction) between ePD-MCI patients and ePD-nMCI patients, and the temporal variability of global brain functional connectivity in ePD-MCI patients was higher than that in ePD-nMCI patients. We found no correlation between FCTV and age or course of disease in the bands with significant differences in Fig. 7b. We used the temporal correlation coefficient (C) to describe the stability of functional connectivity network in the dynamic process. The lower the temporal correlation coefficient, the worse the stability of DFNs. It was found that the temporal correlation coefficient of ePD-MCI patients in the delta and theta bands was significantly lower than that of ePD-nMCI patients. However, FCTV showed significant differences in delta, theta and alpha bands when describing the fluctuation of functional connectivity in the whole brain, and the fluctuation of functional connectivity in ePD-MCI patients was significantly higher than that in ePD-nMCI patients. Spearman’s correlation between C and FCTV in the delta and theta bands was performed as shown in Fig. 7c, and it was found that there was a significant negative correlation between the temporal correlation coefficient and FCTV. Because the differences between groups obtained by a single indicator are not necessarily convincing, while multiple indicators from different perspectives illustrate the same problem is more powerful. Together, these two indicators indicate that the functional network of ePD-MCI patients in the delta and theta bands is significantly non-stationary.

Fig. 7.

The temporal variability analysis of dynamic functional connectivity based on whole brain. FCTV represents the temporal variability of dynamic functional connectivity. a The difference analysis of FCTV between ePD-MCI patients and ePD-nMCI patients in the sub-bands. The results of difference analysis are shown by the P value, and the results with significant differences are highlighted in red font. b FCTV was correlated with age and course of disease respectively by Spearman analysis in the significantly different sub-bands. c Spearman’s correlation analysis between the temporal correlation coefficients and FCTV. FDR correction was performed for P values. (Color figure online)

The result in Fig. 8 indicated that the functional connectivity strength of ePD-MCI group in the P3, Pz, P4, P8 and O1 electrode regions was significantly lower ($p<0.05$, ANOVA with Bonferroni correction) than that of ePD-nMCI in the delta band. In the theta band, ePD-MCI groups functional connectivity strength was significantly lower ($p<0.05$, ANOVA with Bonferroni correction) than ePD-nMCI group only in the Cz and Pz electrode regions. The functional connectivity strength of ePD-MCI in C3 and P4 electrode regions was significantly higher ($p<0.05$, ANOVA with Bonferroni correction) than that of ePD-nMCI in the alpha band. The functional connectivity strength of ePD-MCI was significantly lower ($p<0.05$, ANOVA with Bonferroni correction) than that of ePD-nMCI in the electrode regions of Fp2 and F8 of the beta band.

Fig. 8.

Topographical distribution of the functional connectivity strength for ePD-MCI patients (left) and ePD-nMCI patients (center) in the sub-bands. The red regions of the right column show the significant differences (Bonferroni adjusted $p<0.05$) of corresponding measure between two groups

We analyzed the temporal variability and difference of the functional connectivity strength of each electrode between the two groups in the dynamic brain activity of the sub-bands as shown in Fig.9. In the delta band, the temporal variability of the functional connectivity strength of the ePD-MCI group in the electrode regions of Fp1, F3, Fz, F8, Cz, Pz, C4, P4, P7 and O1 in the brain activity process was significantly higher ($p<0.05$, ANOVA with Bonferroni correction) than that in the ePD-nMCI group. The temporal variability of functional connectivity strength in ePD-MCI group was significantly higher ($p<0.05$, ANOVA with Bonferroni correction) than that in ePD-nMCI group in the electrode regions of Fp1, F3, FZ, F4, C4, P4, P8, P3, P7 and O2 in the brain activity process of the theta band. In F7, FP1, FZ, C4, P4, P8, O1, O2, P7, P3 electrode regions of the alpha band, the temporal variability of functional connectivity strength in the ePD-MCI group during the brain activity process was significantly higher ($p<0.05$, ANOVA with Bonferroni correction) than that in the ePD-nMCI group. The temporal variability of ePD-MCI functional connectivity strength in the beta band F4, F8 and C4 electrode regions during brain activity was significantly lower ($p<0.05$, ANOVA with Bonferroni correction) than that in ePD-nMCI group.

Fig. 9.

Topographical distribution of the temporal variability of dynamic functional connectivity for ePD-MCI patients (left) and ePD-nMCI patients (center) in the sub-bands. The red regions of the right column show the significant differences (Bonferroni adjusted $p<0.05$) of corresponding measure between two groups

We analyzed whether the complexity of EEG signals affected the volatility of DFN as shown in Fig. 10. Since significant differences were found only in the delta, theta and alpha bands in whole brain, Spearman’s correlation analysis of complexity and FCTV was performed on the first three bands. It is found that there is a significant positive correlation, indicating that the fluctuation of EEG signals may affect the dynamic characteristics of the dynamic brain functional network.

Fig. 10.

Spearman’s correlation between complexity and FCTV in the delta, theta and alpha bands. FDR correction was performed for P values

Finally, the complexity, FCS and FCTV of the whole brain in the sub-bands with significant differences were classified by SVM, and 5-fold cross validation was performed. And we obtained the ROC curve and the accuracy of different indicators as shown in Table 2 and Fig. 11. The complexity of the four sub-bands was significantly different between groups, and the classification results of the average complexity of the whole brain of the four sub-bands used as the feature was that the AUC value was 0.78 and the accuracy was 0.73. The functional connection strength of the delta band was significantly different, so the FCS of the delta frequency band was used as the feature for classification, and the results was the accuracy of 0.64 and the AUC value of 0.67. FCTV in the delta, theta and alpha bands showed significant inter-group differences, which were used as features for classification, and the results was the accuracy of 0.63 and the AUC value of 0.68. Moreover, the classification effect of combination among increased complexity, decreased strength of functional connectivity and increased non-stationary of functional connectivity in the delta band was better. The accuracy was 0.80 and the AUC value was 0.85.

Table 2.

The accuracy of SVM classification with features

Features	Complexity	FCS	FCTV	Combination of features
Accuracy	$0.73\pm 0.13$	$0.64\pm 0.06$	$0.63\pm 0.04$	$0.80\pm 0.14$

Fig. 11.

The ROC curve and the AUC value for the features of complexity, FCS and FCTV through SVM classification of 5-fold cross validation

And ePD-MCI group was divided randomly into two groups for several times. It was found that the results of the significant differences obtained in this study did the different analysis between two groups. The difference analysis of indicators based on double-blind grouping did not find significant differences within groups.

Discussion

One hypothesis is that the prefrontal lobe is basically responsible for working memory. By analyzing the EEG signals’ complexity of different electrodes, we found that the EEG signals’ complexity of ePD-MCI patients in the four sub-bands was significantly increased, mainly in the prefrontal lobe. The increased complexity of EEG signals in this area may be a potential physiological indicator of mild cognitive impairment in Parkinson’s disease. There are few literatures on the analysis of PD-MCI by studying the complexity of EEG signals. We found the literature on EEG complexity analysis for PDD, and they found that the complexity of EEG signals increased abnormally in the low frequency band (mainly in the delta band) and decreased in the high frequency gamma band for PDD patients (Bertrand et al. 2015). When we searched for studies on non-PD patients with MCI, we found that the literature showed that MCI showed higher complexity of EEG signals in the resting state than normal people and mild, moderate and severe AD (Maturana-Candelas et al. 2019). In addition, Yan et al also found that the EEG signals of MCI patients were higher than normal people (Yan et al. 2013). In our study, it was also found that the EEG complexity of ePD-MCI patients increased significantly in the four sub-bands. However, there were research found that the complexity of EEG signals in AD patients was lower than that in healthy controls (Simons 2018; Cao et al. 2015). We suspected that the phenomenon of increased EEG complexity in MCI patients may be the brain compensating for a functional deficit. Evidence of the ability of brain plasticity to attempt to compensate for early neurodegenerative disease has also been reported in preclinical AD patients, even before symptoms begin to appear (Becker et al. 1996). Our study only involved the MCI of PD patients, and we hope to further cooperate with the hospital to obtain the EEG data of PDD patients to verify this speculation.

In the analysis of dynamic graph theory metric, the results of the temporal correlation coefficient showed that the stability of functional connectivity network of ePD-MCI patients was lower than that of ePD-nMCI patients in the delta and theta bands, and the temporal characteristic path length of the DFNs in ePD-MCI patients was higher than that in ePD-nMCI patients, which may lead to the decrease of communication efficiency of ePD-MCI patients. The stability of the dynamic brain functional network in ePD-MCI patients was reduced, which may lead to slower information exchange. The results indicate that cognitive impairment may cause functional network connectivity disruption. And Some scholars have studied the space-time characteristics of AD progression by constructing EEG dynamic brain network based on mutual information, and found that the connectivity of brain network changed during the transition from MCI to AD, resulting in the loss of connectivity regions (Morabito et al. 2015). In the analysis of the global functional connectivity strength of the DFNs in the four sub-bands, it was found that there were significant differences between patients with ePD-MCI and ePD-nMCI in the delta band, and the global brain functional connectivity of patients with ePD-MCI was lower than that of patients with ePD-nMCI. Moreover, in the delta band, the functional connectivity strength of ePD-MCI patients’ parietal lobe, posterior temporal region and left occipital lobe region was significantly lower than that of ePD-nMCI patients. Wang et al studied the temporal and spatial evolution of dynamic brain network connectivity change in patients with mild to moderate cognitive impairment and AD based on fMRI, and also found that the connectivity between networks showed a trend of weakening from early cognitive impairment to dementia (Wang et al. 2021), which confirmed our findings. And the temporal variability of global brain functional connectivity strength in the process of dynamic change was significant differences between ePD-MCI group and ePD-nMCI group in the brain regions of the delta, theta and alpha bands, and the temporal variability was higher than that in ePD-nMCI patients. The brain regions with abnormal functional connectivity fluctuations were located in the frontal lobe and occipital lobe. These results indicated that cognitive impairment in Parkinson’s patients may affect the strength of brain functional connectivity and stability of functional connectivity.

We discussed the reason why the EEG signal complexity of the ePD-MCI population was higher than that of the cognitively normal PD patients, because some studies found that the EEG signal complexity of the MCI patients in the non-PD cognitively impaired population was higher than that of the AD and HC groups, and we speculated that this might be caused by the compensation mechanism of the brain. In addition, we also found literature support for the decrease in the strength and stability of functional connectivity caused by cognitive impairment. Whether the results obtained in our study can be used as a diagnostic test basis in PD population, its external validity still needs to be further verified. All the data that met the criteria were included in the study, and we hope subsequent continue repeating work after the introduction of new data to verify the results are true and has generalizability. In the experiment, we didn’t have a healthy control group and only conducted a basic information survey on the subjects, so there was a lack of follow-up and re-evaluation to update the data. In addition, the cognitive degree was assessed only based on MOCA, lacking a more detailed neuropsychological assessment. ACE and other scales can be combined in the follow-up of the experiment. The experiment will be updated according to the limitations of this experiment.

Conclusion

In this paper, we mainly studied two aspects. We analyzed the complexity of EEG signal and the differences between ePD-MCI patients and ePD-nMCI patients and found that cognitive impairment led to the increase of brain region complexity in ePD patients. Then we built dynamic network based on sliding-window. In order to determine that the time series fluctuation of network connectivity was due to the dynamic change of real functional connectivity rather than the statistical uncertainty, we introduced the method of the AFFT zero model to determine the real dynamic fluctuations of functional connectivity. Then we distinguished the differences between the two groups through $k_{\mathit{PLI}}$ indicator to determine the optimal time window for the DFNs construction. We analyzed the complexity of EEG signals and the dynamic characteristics of the DFNs.

Together, we found that the complexity of EEG signals in ePD-MCI patients in the full frequency range is generally high, and ePD-MCI patients in the delta band showed increased EEG complexity, decreased functional connectivity strength and stability, which may be the biological marker of early PD mild cognitive impairment, and contribute to understand the neural mechanism of ePD-MCI.