Evolution of brain network dynamics in early Parkinson’s disease with mild cognitive impairment

Abstract

How mild cognitive impairment (MCI) is instantiated in dynamically interacting and spatially distributed functional brain networks remains an unexplored mystery in early Parkinson’s disease (PD). We applied a machine-learning technology based on personalized sliding-window algorithm to track continuously time-varying and overlapping subnetworks under the functional brain networks calculated form resting state electroencephalogram data within a sample of 33 early PD patients (13 early PD patients with MCI and 20 early PD patients without MCI). We decoded a set of subnetworks that captured surprisingly dynamically varying and integrated interactions among certain brain lobes. We observed that the master expressed subnetworks were particularly transient, and flexibly switching between high and low expression during integration into a dynamic brain network. This transience was particularly salient in a subnetwork predominantly linking temporal-parietal-occipital lobes, which decreases in both expression and flexibility in early PD patients with MCI and expresses their degree of cognitive impairment. Moreover, MCI induced a regularly interrupted, slow evolution of subnetworks in functional brain network dynamics in early PD at the individual level, and the dynamic expression characteristics of subnetworks also reflected the degree of cognitive impairment in patients with early PD. Collectively, these results provide novel and deeper insights regarding MCI-induced abnormal dynamical interaction and large-scale changes in functional brain network of early PD.

Introduction

Mild cognitive impairment (MCI) is a frequent non-motor symptom of Parkinson’s disease (PD), affecting between 20 and 40% of patients with PD in the early phase of the disease (early PD) (Litvan et al. 2011; Aarsland et al. 2010). PD with MCI (PD-MCI) is defined as cognitive decline, commonly quantified as impairments in attention, memory, executive ability, and visuospatial skill individually or in combination (Litvan et al. 2012). Patients with PD-MCI have a significantly increased risk for development of PD dementia (PDD) compared with PD patients without MCI (PD-nMCI) (Williams-Gray et al. 2007; Kehagia et al. 2010; Goldman and Litvan 2011; Pedersen et al. 2013; Kawada 2015; Hobson and Meara 2015). High priority has been placed on early identification of PD patients at risk of developing PDD, which may help to provide more timely interventions. However, the neurobiological functional and structural damage of the brain in patients with early PD-MCI remains unknown.

The human brain can be described as a modular dynamic system that continuously reconfigures its functional connectivity patterns over time (Meunier et al. 2010; Protopapa et al. 2016; Liu et al. 2020). By analyzing the dynamic evolution of the brain, previous research reported that the brain network can be hierarchically superimposed by discrete modular organizations, and these modules are interconnected by the distributed functions of different brain regions (Bullmore and Sporns 2012; Zhang et al. 2016). Numerous studies have extracted the modular components of brain networks and reported that patterns of interaction between them must be flexibly reconfigured in response to driving changes in cognitive demands (Mattar et al. 2015; Kopell et al. 2014; Hutchison et al. 2013; Yuan et al. 2021). This suggests that the existence of modular interactions in dynamic brain networks and the strength of interactions between modules can be configured according to cognitive instructions. Therefore, we speculate that MCI leads to such abnormality of the functional modular interactions of dynamic brain functional connectivity networks in PD patients, and verifying this hypothesis is of great significance for explaining the brain neurobiological function and structural damage of PD-MCI.

Electroencephalography (EEG) signals, which contain spatial information of scalp distribution and high temporal resolution, have been widely used to analyze the characteristics of dynamic brain networks in recent years (Zhang et al. 2022; Hassan and Wendling 2018; Kucyi et al. 2020). We believe that the apriori EEG frequency band reflecting cognitive deficit can provide an effective guarantee for extracting modular features of the dynamic brain network in early PD-MCI. We have examined multitudes of related studies and found that the evidence for the spontaneous alpha rhythm observed on being related to cognitive deficit was more consistent compared to other frequency bands (Cozac et al. 2016; Sen Bhattacharya et al. 2011; Hinault et al. 2021; Yi et al. 2022). For example, studies showing alpha rhythm slowing as a correlation of cognitive deficiency and neuro-degeneration (Sen Bhattacharya et al. 2011; Hinault et al. 2021; Babiloni et al. 2021; Michels et al. 2021; Sargent et al. 2021). Moreover, decrease of alpha rhythm activity based on EEG study is identified as one of the biomarkers to distinguish PD patients with normal cognition from PD with MCI (Cozac et al. 2016). Furthermore, our previous work reveals that MCI resulted in abnormal characteristics of dynamic functional network in alpha band of patients with early PD (Yi et al. 2022). In this case, the construction of brain networks based on the alpha frequency band of EEG signals can provide effective support for exploring the modular distributed interaction characteristics of the dynamic brain functional network in early PD-MCI. However, there are two key challenges remain to be resolved.

The first key issue needs to be solved is, how to build a personalized dynamic brain network? Conventional EEG dynamic functional network construction is mainly based on the segmentation of EEG series by using the sliding window of equal length, and the epoch of EEG data in each time window is used to calculate each brain functional network (Kabbara et al. 2020; Li et al. 2019; Gu et al. 2020; Naim-Feil et al. 2018; Fang et al. 2020). However, the unaddressed concerns are that (1) there is no consensus on the optimum length of the sliding window; (2) it is not objective to extract each brain network with a uniform fixed window length. The fundamental property of the choice for these fixed window lengths is to characterize the functional connectivity networks limited to equidistant durations within the window period, while ignoring the temporal-stochastic dynamic switching properties of real quasi-stable-state brain network (Preti et al. 2017). Ideally, adaptive time-varying window sizes should be used to accurately capture the quasi-stable-state changes of time-varying functional connectivity networks in the brain, which cannot be satisfied by the conventional sliding window method. It has been proved that the time-dependent window algorithm captures the periodical information of brain quasi-homeostasis based on the characteristic of data at each sampling point (Zhuang et al. 2020). Therefore, the estimation of quasi-steady states of dynamic brain networks can be realized by using time-dependent sliding window algorithm. Empirical mode decomposition (EMD) is a data-adaptive analysis algorithm for non-stationary and nonlinear data, which allows the decomposition of time series into near-orthogonal modes (Huang 2005). EMD can divide window-sizes according to the data characteristics of each patient, facilitating the finding of information with significant differences in time-varying signals (Chen et al. 2010). Study has shown that this adaptive sliding-window method can obtain better dynamic information about cognitive function and behavior (Zhuang et al. 2020). In this study, we established such time-dependent sliding windows based on EMD algorithm to construct adaptive dynamic brain functional connectivity networks for different individuals. This satisfies the requirement of personalized dynamic brain network construction.

The notion that each large-scale brain network can be described as a repertoire consisting of multiple modules whose integration is manipulated to meet task demands (Chai et al. 2017) suggests another key challenge. It requires tools that can describe to what degree different modules are engaged in a certain quasi-stable-state brain network, and how that engagement changes over time. To address this challenge, we applied a data-driven unsupervised machine learning algorithm that utilizes modular-based matrix decomposition techniques to examine the modularized components of early PD-MCI dynamic brain functional connectivity networks and the dynamic distribution of their degree of participation. In this study, we define such modular components as functional subnetworks. Conceptually, functional subnetworks can be considered as topological patterns of interacting brain regions, which are reconstructed into primitive brain functional networks by weighted summation of different expression intensities during the dynamic evolution of brain functional connectivity networks.

In this work, we established personalized adaptive time-dependent sliding windows based on EEG’s alpha-frequency band and calculated the brain functional connectivity network within each window of patients with early PD-MCI/nMCI. In order to reveal the essential properties of dynamic functional brain networks, the unsupervised machine learning algorithm was used to extract the independent functional subnetworks that constitute dynamic functional brain networks and the weighted expression coefficients of each subnetwork reconstructed into the corresponding quasi-steady-state networks during the dynamic evolution of the brain networks. Finally, the abnormal dynamic characteristics of functional subnetworks in early PD-MCI were analyzed to reveal the essential features of the dynamic functional connectivity networks of early PD affected by cognitive impairment.

Methods and materials

Participants

In this study, a total of thirty-three early PD patients with the Hoehn and Yahr (H-Y) staging scale I–II (Goetz et al. 2004) were recruited from the Department of Neurology of Tianjin Medical University General Hospital. This study was approved by the Medical Ethics Committee of Tianjin Medical University General Hospital. All subjects understood the purpose of collecting the data and the significance of the study, and signed the written informed consent. The inclusion criteria for participants were as follows: (1) All patients were diagnosed with primary PD; (2) No history of psychiatric disorders and no audiovisual impairments; (3) All participants were stopped from medication for more than 12 h before EEG collection; (4) No head tremor symptom; (5) No history of head trauma with loss of consciousness. Detailed participant demographic and clinical characteristics are shown in Table 1. According to Montreal Cognitive Assessment (MoCA) scale score (Gill et al. 2008), the subject with a score of 18–25 is classified as MCI and 26–30 as nMCI (Chen et al. 2010). Based on this, in this work, there were 13 patients with early PD-MCI and 20 patients with early PD-nMCI.

Table 1.

Participant demographic and clinical characteristics

	Early PD-MCI (n = 13)	Early PD-nMCI (n = 20)	Group comparisons
Sex (Male/Female)	6/7	4/16	${\mathrm{\chi }}^2$. = 2.552; p = 0.110
Age (year)	62.000 ± 8.347	63.500 ± 6.295	t = − 0.588; p = 0.561
H&Y stage	1.500 ± 0.577	1.175 ± 0.438	t = 1.838; p = 0.098
Disease duration (year)	3.385 ± 2.725	4.575 ± 4.499	t = − 0.855; p = 0.399
MoCA scale scores	23.000 ± 1.826	27.700 ± 1.302	t = − 8.644; *p < 0.001

EEG recording and preprocessing

All patients with early-PD were seated comfortably in a quiet semi-dark room with their eyes closed but remained awake in a relaxed state and EEG signals were recorded between 9 a.m. and 11 a.m. Nineteen Ag/Agcl scalp electrodes are placed on the scalp in accordance with international standard 10–20 (active electrodes, SYMTOP, Beijing, China). Two reference electrodes are placed behind the ears. Additional channels are used to monitor electrocardiogram (ECG), electromyogram (EMG) with the electrodes placed on the skin between thumb and index finger on the back of both hands, as well as four electrooculograms (EOGs) for recording the horizontal, vertical, and blinking movements of the eyes, for further preprocessing. The impedances of all electrodes used for collection are kept below 5 kΩ. The sampling frequency is 500 Hz. Fast independent components analysis (fast-ICA) algorithm (Hyvarinen 1999) was applied to remove the noise artifacts, including, cardiac artifacts (from ECG), movement artifacts (from EMG), and ocular artifacts (from EOG) in the EEG signals for each subject respectively. More than 5 min preprocessed EEG data were kept for further analysis for each participant. Then the EEG signals was processed by 8–13 Hz (alpha frequency band) band-pass finite impulse response (FIR) filter.

Time-dependent functional connectivity networks construction

As for EEG signals of all channels $X_p={\left[x_1,x_2,\cdots ,x_m,\cdots ,x_{19}\right]}^T$, EMD decomposes the original EEG signals of each channel into a series of nearly orthogonal modes (Patrick and Goncalves 2011), which are named intrinsic mode functions (IMFs):

$$x_m(t)=\sum _{i=1}^N{c}_{\mathit{mi}}(t)+r(t)$$ 1

here, N is the number of IMFs, $x_m(t)$ denotes the EEG signals of m-th channel at time t, c_mi(t) indicates the i-th IMF of $x_m(t)$ and $r(t)$ is the corresponding residual function, generally representing a monotone function of the average trend or constant (Huang et al. 1998). In this case, $c_i(t)$ is the i-th IMF of each channel’s EEG signals. For each IMF, the instantaneous frequency $f_i(t)$ is calculated via the Hilbert transform by extending $c_i(t)$ into the complex domain (Huang 2005; Huang and Wu 2008), and the instantaneous period $p_i(t)$ is then denoted as $1/f_i(t)$.The average energy density $E_i$ per unit time of the i-th IMF is defined as:

$$E_i=\frac{1}{T}\sum _{i=1}^T{c}_i{(t)}^2$$ 2

In summary, the instantaneous period $p_i(t)$ captures the local non-stationarity of the EEG signals, and the average energy density $E_i$ quantifies the energy contributions of each IMF to the EEG signals. Following the above steps, the time-dependent instantaneous period $P(t)$ of each time process is defined as the average of $p_i(t)$ weighted by $E_i$ (Cordes et al. 2018):

$$P(t)=\frac{1}{{\sum }_{i=1}^N{E}_i}\sum _{i=1}^N(p_i(t)\times E_i)$$ 3

In order to ensure that the window-sizes of the 19 channels are the same, the largest time-dependent instantaneous period in the 19 channels at this moment is extracted, so the time-dependent sliding window-size T of the subject at each time point is finally defined as:

$$T=max(P_1(t),P_2(t),\cdots ,P_{19}(t))$$ 4

After dividing the time-dependent sliding windows, we adopted the phase lag index (PLI) algorithm (Stam et al. 2007) to obtain the brain functional connectivity networks (i.e., adjacency matrix) between 19 channels of each patient. The functional connectivity of a pair-wise EEG channels within the r-th time-dependent sliding window can be obtained by quantifying the asymmetry of the phase difference distribution between a pair of signal series, and it can be defined as:

$$F_r\left(i,j\right)=\left|〈sign\left[sin\left(\mathrm{\Delta }\varphi \left(t_k\right)\right)\right]〉\right|,k=1,2,\dots ,m_r,i\ne j$$ 5

where $F_r\left(i,j\right)$ defines the functional connectivity of electrical channel i and j within the r-th time-dependent sliding window, m_r is the number of time points within the r-th time-dependent sliding window, $\mathrm{\Delta }\varphi \left(t_k\right)$ is the phase differences calculated at k time points, and $〈.〉$ denotes the mean value inside. The values of PLI range from 0 to 1, where 0 indicates no coupling or zero lag in coupling between two signals and 1 stands for perfect phase locking (non-zero delay). Finally, the 19 × 19 adjacency matrix composed of PLI values of all pair-wise EEG channels was taken as the functional connectivity network F_r within the r-th time-dependent sliding window. Thus, a set of functional connectivity networks were obtained with v time-dependent sliding windows (named time-dependent functional connectivity networks), i.e., $F=\left\{F_1,···,F_r,···,F_v\right\}$.

Partitioning the dynamic functional networks

Orthogonal projective non-negative matrix factorization algorithm

Non-negative matrix factorization (NMF) algorithm (Lee and Seung 1999) is a data-driven, multivariate, and unsupervised machine learning technique, has been employed to extract functional subnetworks of dynamic functional brain networks (Khambhati et al. 2016, 2017, 2018). On this basis, we use the orthogonal projective NMF (OPNMF) algorithm (Sotiras et al. 2015; Patel et al. 2020) extended by NMF in order to ensure that the identified spatial components do not overlap with each other and each of them represents a distinct module (i.e., functional subnetwork).

First, we vectorized all the adjacency matrixes of subjects in each group into an entire matrix, named original given matrix A (size m × n). m is the number of edges in each functional connectivity network and n is the number of functional connectivity networks in all participants. NMF decomposes the original given matrix A into the product of two non-negative component matrixes W (component matrix, size m × k) and H (weight matrix, size k × n):

$$A\approx WH(W\ge 0,H\ge 0)$$ 6

where k denotes the number of functional subnetworks. OPNMF algorithm replaces H with $W^{T}A$. Indeed, OPNMF estimates W and H by solving the minimization problem:

$$∥A-W^{T}WA∥(W^{T}W=I;W\ge 0)$$ 7

where $∥.∥$ indicates the squared Frobenius norm. W is initialized by non-negative double singular value decomposition (Boutsidis and Gallopoulos 2008). I is the identity matrix which ensures orthogonality among the columns of W. After that, W is obtained by using a multiplicative update rule (Patel et al. 2020):

$$W_{\mathit{ij}}^{{}^{\prime }}=W_{\mathit{ij}}\frac{{(A{A}^{T}W)}_{\mathit{ij}}}{{(W{W}^{T}A{A}^{T}W)}_{\mathit{ij}}}$$ 8

Finally, the weighted matrix H is calculated by projecting the original given matrix A onto component matrix W:

$$H=W^{T}A$$ 9

The extraction process of functional subnetworks behind the functional brain networks is shown in Fig. 1. Each 19 × 19 adjacency matrix is a weighted symmetric matrix. The process of vectorization is to expand the upper triangular matrix of each adjacency matrix into a one-dimensional column matrix whose length is denoted by m (i.e., the number of edges in each functional connectivity network). The original given matrix A is obtained by concatenating the one-dimensional column matrices obtained by the time-varying brain network expansion of all subjects in the two groups. The matrix A is decomposed into a component matrix W and a weight matrix H via OPNMF algorithm, and the functional subnetwork matrix is visualized in the form of a functional network (the inverse process of vectorization). The values in the weight matrix H are named temporal expression coefficients. The OPNMF algorithm operates with non-negative double singular value decomposition initialization conditions (Patel et al. 2020), with max iterations was 10⁵ and tolerance was 10^–5.

Fig. 1.

Schematic overview of the functional subnetwork extraction process via OPNMF algorithm

Selection of the number of functional subnetworks by extracting reliable components

To determine the optimal dimension of the extracted component matrix, that is, the optimal number k of the extracted functional subnetworks, we evaluated the stability of the degree of variation of the extracted functional subnetworks and the accuracy of the reconstruction of the functional subnetworks into the original brain networks with each execution of the OPNMF algorithm. Stability is used to judge whether there will be great differences between functional subnetworks decomposed after each execution of the algorithm. High stability is to ensure the uniqueness of functional subnetworks decomposed and prevent major changes in the subnetworks generated by each execution. On the other hand, accuracy is quantified by computing the gradient of the reconstruction error. The error generated during the reconstruction of the functional sub-network into the original network is named reconstruction error, and the degree to which the reconstruction error changes when increasing the number of subnetworks is the gradient of the reconstruction error.

To assess stability, matrix A is randomly divided into two groups. The component matrices $W_a$ and $W_b$ are obtained by OPNMF algorithm, respectively. For the two component matrices obtained, the cosine similarity between every two rows of each component matrix is calculated (Patel et al. 2020), and then the similarity matrices c_W_a and c_W_b (c_W: size m × m) of the two component matrices are obtained respectively. Then, we compute the correlation coefficient between corresponding rows of c_W_a and c_W_b to quantify if the OPNMF decomposition groups a given functional connectivity into the same subset in each data split to form the corresponding subnetwork. Finally, the stability coefficient of the given subnetwork is evaluated by calculating the mean correlation across all functional connectivity. A stability coefficient of 1 indicates complete stability, while a stability coefficient of -1 indicates instability. The number of subnetworks is determined to be between 2 and 10. In order to obtain the stability index under the number of each subnetwork, the operation of randomly dividing matrix A is repeated 10 times. Moreover, the reconstruction error is expressed as:

$$error=∥A-WH∥$$ 10

Next, the gradient of reconstruction error is calculated, indicating the variation of reconstruction error as the number of subnetworks increases. The negative reconstruction error gradient indicates that the reconstruction error decreases with the increase of the number of subnetworks, while the absolute value of the reconstruction error gradient decreases, indicating that the change degree of reconstruction error decreases with the increase of the subnetworks. Finally, the optimal number of subnetworks is determined by combining the stability and reconstruction error gradient corresponding to each subnetwork number.

Measures of subnetwork fluctuation, strength, and dynamics

The subnetworks and the corresponding time expression coefficient matrices of subnetworks were obtained when the splicing matrix of dynamic functional brain networks of all subjects was decomposed by OPNMF algorithm. Here, the time expression coefficient quantifies the dynamic expression of each subnetwork and the variation of its connection strength during the process of reconstructing the original network. The higher time expression coefficient in any time-dependent window indicates that the corresponding subnetwork plays a greater role in the reconstruction of the original functional network. By calculating the fuzzy entropy of the time expression coefficients corresponding to each subnetwork of each subject, the fluctuation of the expression degree of subnetwork in the time-varying process was expressed (Chen et al. 2007).

The energy of the temporal expression coefficient quantifies the overall expression strength of the subnetwork in a period and represents the coupling strength of its mode in the whole dynamic process of reconstructing into the original network, which defined as: $Energy={\sum }_{n=1}^t{S}_n^2$. Here, t is the number of time-dependent windows, and $S_n$ is the temporal expression coefficients under each time-dependent window. Since the calculated value of Energy is large, it is expressed as log (Energy) for the convenience of observation and comparison.

Under one time-dependent window, the functional subnetwork with the highest temporal expression coefficient plays a dominant role in the process of integration into the original network. In order to quantify the dynamic characteristics of the strongest expressed subnetwork in the integration process, we define two characteristic parameters: the occurrence of each subnetwork in the dynamic process (that is, the percentage of time that each subnetwork presents its strongest expression throughout the dynamic process presents its strongest expression throughout the dynamic process), and duration of the strongest expression (that is, the average duration of each strongest expression of a subnetwork).

Statistical analysis

Statistical analyses were carried out by MATLAB software (MathWorks Inc., Natick MA, United States) and SPSS 25.0 software (IBM Inc., Chicago, IL, United States) for participant demographic and characteristic parameters. The significance of characteristic parameters in individual between the early PD-MCI and early PD-nMCI groups was tested using independent sample t-test (p < 0.05) with false discovery rate (FDR) corrected. Furthermore, Spearman’s correlation analysis was employed to calculate the correlation coefficients between MoCA scale scores and characteristic parameters of subjects. Corrected FDR acted on p values for multiple comparisons of the Spearman’s correlation coefficients and the significant level was accepted at p < 0.05.

Results

Selecting parameters for OPNMF-based subnetwork decomposition

The results of the stability and accuracy analysis are shown in Fig. 2. The optimal number of subnetworks of OPNMF decomposition was determined by stability coefficient and gradient of reconstruction error. Both the stability coefficient (red, left) and the reconstruction error gradient (blue, right) for k = 2 to 10 (k indicates the number of subnetworks). Overall, the stability coefficient is shown to have a roughly inverse relationship with the gradient of reconstruction error. When k ranges from 2 to 5, the stability coefficients are generally stable and tends to 1, while k > 5, notable drop offs and irregular variation in stability are observed. The reconstruction error gradient shows an obviously positive relationship with k, as more subnetworks allow the decomposed brain network to recover an increasing number of patterns. For each k, the value of reconstruction error gradient is the variation in reconstruction error from k − 1 to k, and the larger gradient of reconstruction error indicates higher accuracy. A sharp increase in the magnitude of reconstruction error gradient when k moving from 3 to 4, and then with the increase of k, the gradient of reconstruction error presents an overall upward trend, but the increase range remains gentle. In the case of ensuring higher stability and accuracy at the same time, these results suggest that k = 5 is the optimal selection for further analysis, and any increase or decrease in the number of subnetworks has a diminishing return on reconstruction effect.

Fig. 2.

Parameter optimization for subnetwork detection. Stability coefficient and reconstruction error gradient for granularities from 2 to 10 for the functional subnetwork decomposition analysis. Stability coefficient is indicated by the red line (on the left side of the coordinate) and gradient of reconstruction error is indicated by the blue line (on the right side of the coordinate). The black dotted line marks the value corresponding to the optimal number of functional subnetworks. The ‘5’ circled by the red dotted line is the optimal number of functional subnetworks

Functional subnetworks differ in expression

After determining the optimal number of subnetworks, we extracted and visualized 5 functional subnetworks of the brain functional connectivity networks of all early PD patients (including PD-MCI and PD-nMCI), and the significantly expressed connectivity patterns are depicted in spatial scalp map, as shown in Fig. 3. By comparing the connectivity in each subnetwork, we observed that all subnetworks are intrinsically distinct in distribution, and they capture interactions among specific brain regions respectively. Specifically, functional subnetwork 1 mainly captures the interactions among the temporal, parietal, and occipital lobes on both hemispheres; Functional subnetwork 2 mainly captures the interactions of the left hemisphere of the brain, while functional subnetwork 3 is completely opposite with functional subnetwork 2 and it mainly captures the interactions of the right hemisphere of the brain, noting that the connections are more concentrated in the frontal lobe; Functional subnetwork 4 is nearly opposite with functional subnetwork 1 and it mainly captures the interactions among the frontal, parietal, and temporal lobes on both hemispheres; Functional subnetwork 5 is the distinct module and it mainly captures the interactions between the brain regions of central axis and other brain regions on the scalp, and the connections are more concentrated in the brain regions of central axis. These five functional subnetworks with mutually different and extremely complementary connectivity topological properties constitute the original dynamic brain network with specific time-varying expression coefficients (i.e., the values of weight matrix).

Fig. 3.

Mapping of the connectivity in functional subnetworks. The size of a node corresponds to the number of brain regions that have functional connectivity to the brain region where the node is located

Strength, fluctuation, and dynamics of functional subnetworks

Up to this point, we have merely reported the connectivity topological properties of five functional subnetworks expressed collectively among the cohort of patients with early PD-MCI and patients with early PD-nMCI. The dynamic expression of functional subnetworks across the individuals also should be explored via the time-varying expression coefficients to decode the personalized characteristic abnormality of the functional modular interactions in dynamic brain functional connectivity networks of patients with early PD-MCI. In this case, there are three issue concerns: (1) Does MCI lead to abnormal expression strength of subnetworks in early PD patients? (2) Do subnetworks with high expression strength fluctuate more rapidly in time? (3) Does MCI alter the dynamics of subnetworks in early PD patients?

Recalling that, the Energy of the time-varying expression coefficient quantifies the overall expression strength of the subnetworks over a period. For clear representation, we compared the logarithm of Energy of each subnetwork between early PD-MCI and early PD-nMCI groups as shown in Fig. 4a. We observed that from the overall view of the five functional subnetworks, the expression strengths of the early PD-MCI patients were weaker than that of the early PD-nMCI group. In addition, the expression strengths of the five subnetworks in the early PD-nMCI group decreased in order of 1 to 5, while the expression strengths of the first three subnetworks in the early PD-MCI group were relatively close. These findings suggest a transition toward a greater precedence of temporal-parietal-occipital interaction network, left hemisphere interaction network (dominated by the left frontal lobe), and right hemisphere interaction network (dominated by the right frontal lobe), and a greater switching of the brain regions and the interactions subserving these networks throughout the dynamic process in PD-MCI. Furthermore, the logarithms of Energy of subnetworks 1, 2, and 3 in patients with early PD-MCI were significantly lower than those in patients without MCI. There was a significant positive correlation between the logarithms of Energy of subnetworks 1, 2, 3 and the scores of MoCA scale (Table 3).

Fig. 4.

Energy and entropy of subnetwork expression. The logarithm of Energy (a) and the Entropy (b) of the five functional networks in patients with early PD-MCI and early PD-nMCI. c Spearman correlations between the logarithm of Energy and the Entropy of each functional subnetworks. Significant differences between groups are indicated by asterisks (* denotes p < 0.05; ** denotes p < 0.01)

Table 3.

Spearman’s correlation analysis between the characteristics of subnetworks and the MoCA scale scores

Sub-network	Entropy—MoCA		Log (Energy)—MoCA		Occurrence—MoCA		Duration—MoCA
	R	p	R	p	R	p	R	p
1	0.781	< 0.001	0.653	< 0.001	0.721	< 0.001	0.661	< 0.001
2	0.584	< 0.001	0.369	0.034	− 0.199	0.267	− 0.190	0.291
3	0.700	< 0.001	0.448	0.009	0.227	0.205	0.602	< 0.001
4	0.491	0.004	0.030	0.871	− 0.752	< 0.001	− 0.841	< 0.001
5	0.578	< 0.001	0.153	0.395	− 0.871	< 0.001	− 0.769	< 0.001

Significant differences (p< 0.05 with FDR corrected) are marked in bold

These results indicated that MCI induces a significant decrease in the overall expression strength of subnetworks 1, 2 and 3 in early PD patients during the whole dynamic process. The decrease of Energy also corresponded to the decrease of time expression coefficients, which reflected the change of functional connectivity strength of brain interregional interaction in the corresponding subnetwork. Therefore, this result also suggests that MCI leads to a significant reduction in the overall functional connectivity strength of subnetwork 1, 2, and 3 in early PD patients throughout the dynamic process.

We further explore whether the fluctuation of time expression coefficient of each subnetwork is affected by MCI. As shown in Fig. 4b, the entropy values of five subnetworks in patients with early PD-MCI were significantly lower than that in patients with early PD-nMCI. This suggests that the fluctuation of each sub-component of the dynamic brain network in patients with early PD-MCI is significantly reduced. Spearman’s correlation analysis was carried out between the entropy values of time expression coefficient of all subnetworks and the scores of MoCA scale. The results showed that the entropy values of time expression coefficient had a significant positive correlation with the scores of MoCA scale (Table 3). This indicates that with the deepening of cognitive impairment, the fluctuation of all subnetworks reconstructed into the original network significantly decreases, and the stability of the connectivity strength variation of the corresponding pattern of brain interregional interaction also significantly increases. Furthermore, we observed that the logarithm of energy and entropy were highly positively correlated for all the subnetworks (Spearman’s correlation coefficient: p < 0.001, Fig. 4c): strongly expressed subnetworks have a greater tendency to change in their expression levels with higher fluctuation, while weakly expressed subnetworks have a larger tend to remain stable at their expression levels with less fluctuation.

The higher the corresponding time expression coefficient of the functional subnetwork during a time-independent window, the stronger the expression strength of the subnetwork in the original dynamic brain network. Therefore, the subnetwork with the largest temporal correlation coefficient among the five subnetworks plays a leading role in integrating into the original functional brain network under a time-independent window. We next explored whether the dynamic characteristics of dominant subnetworks in early PD were altered during their integration into primitive functional brain networks affected by MCI. Figure 5a depicts the combination mode of the dominant subnetwork of early PD-MCI group and early PD-nMCI group in the dynamic process of integrating into the original brain network. We observed that when the subnetwork 1 (temporal-parietal-occipital interaction network) became the master expression subnetwork, its occurrence and duration were significantly decreased in patients with early PD-MCI (Fig. 5b). In this case, the occurrence and duration of subnetwork 1 were positively correlated with MoCA scale scores (p < 0.001, Table 3). These results suggested that the connectivity of the temporal-parietal-occipital interaction network was reduced by cognitive impairment in patients with early PD. Moreover, the duration of subnetwork 3 (right hemisphere interaction network dominated by the right frontal lobe) was also reduced in early PD-MCI and significantly positively correlated with MoCA scale scores (p < 0.001, Table 3), while its occurrence was at the similar level with early PD-nMCI (p = 0.982, Table 2). This is reflected in Fig. 5c. Compared with PD-nMCI, subnetwork 3 of PD-MCI has a distinct distribution, that is, it expressed with short duration but frequent occurrence accompanied by phased non-dominant expression. In addition, when subnetwork 4 (frontal-parietal-temporal interaction network) and subnetwork 5 (central axis interaction network) were most expressed, their occurrence and duration were significantly higher in patients with early PD-MCI than in patients with early PD-nMCI, and they were negatively correlated with MoCA scale scores (p < 0.001, Table 3). It indicates that the expressions of these two subnetworks were increased by cognitive function impairment in early PD.

Fig. 5.

The dynamic characteristics of dominant subnetworks in early PD during their integration into primitive functional brain networks. (A) The overall dynamic transition process of dominant subnetworks in early PD-MCI group and early PD-nMCI group. (B) Occurrence and duration of dominant subnetwork. Significant differences between groups are indicated by asterisks (** denotes p < 0.01). (C) Clearer overall dynamic transition process of dominant subnetworks in two groups. The dynamic process of the dotted box is magnified to obtain (D) the dynamic process of two subjects with early PD-MCI

Table 2.

Independent t-test analysis of the characteristics of subnetworks between early PD-MCI and early PD-nMCI

Sub-network	Entropy		Log (Energy)		Occurrence		Duration
	t	p	t	p	t	p	t	p
1	− 5.873	< 0.001	− 5.253	< 0.001	− 5.931	< 0.001	− 8.978	< 0.001
2	− 3.287	< 0.001	− 2.944	0.013	− 0.052	0.960	− 0.540	0.593
3	− 4.765	< 0.001	− 2.515	0.029	− 0.023	0.982	− 5.710	< 0.001
4	− 2.407	0.022	− 0.230	0.820	7.541	< 0.001	9.305	< 0.001
5	− 2.978	< 0.001	− 1.109	0.345	7.299	< 0.001	5.619	< 0.001

Significant differences (p< 0.05 with FDR corrected) are marked in bold

Interestingly, we observed that the dynamic switching process among functional subnetworks of early PD-MCI had obvious regularity and high repeatability in participants, while it was inexistent in the early PD-nMCI (Fig. 5c, d). The repeatable regularity was manifested by when subnetwork 1 had just entered the mode of non-dominant for a relatively long time, subnetwork 2 then appeared in a mode of mutual switching between non-dominant and dominant for a period of time, and after this following effect occurred, subnetwork 3 entered a relatively long-term non-dominant mode followed by a short-term dominance of subnetwork 4, followed by a relatively long-term non-dominant mode, with subnetwork 5 showing a longer-term dominance mode at about the same time (Fig. 5d). Importantly, such regularity appeared during the dynamic integration of the subnetwork into the primitive brain network in all early PD-MCI patients.

Discussion

Using dynamic machine-learning-based novel techniques that decompose personalized functional brain networks into time-evolving functional subnetworks, we studied how dynamic patterns of subnetworks mapped onto dynamic original functional connectivity networks and evolved with development in patients with early PD-MCI and early PD-nMCI. We demonstrated that the dynamic brain networks in early PD tend to act collaboratively by five sub-networks (in order of decreasing expression strength): temporal-parietal-occipital interaction network, left hemisphere interaction network (dominated by the left frontal lobe), and right hemisphere interaction network (dominated by the right frontal lobe), frontal-parietal-temporal interaction network, and the central axis interaction network. In this case, MCI results in a significant decrease in the expression strength of the temporal-parietal-occipital interaction network and the left/right hemisphere interaction networks (dominated by frontal lobe) during the dynamic integration process. Moreover, we observed that subnetworks with high-energy are more fluctuant and flexible in expression, while low-energy subnetworks are more stable. Finally, we explored changes in the personalized dynamic characteristics of the master expression subnetwork during the process. As compared to the patients with early PD-nMCI, the patients with early PD-MCI displayed a distinct abnormal regularity in expression process at the individual level, that is, when the first four functional subnetworks have regular long-time inexpression or intermittent expression, the fifth is expressed. These results reveal how MCI affects underlying essential changes in dynamic brain networks of patients with early PD.

The functional brain networks have been factorized into five kinds of subnetworks which dynamically change across time in a non-negative combination manner. This exploration has several significances. First, functional brain networks have been described as the co-expression of multiple non-negative dynamic interaction patterns between distinct brain regions, with an intuitive interpretation that stronger functional connectivity is determined by more positive interaction patterns. Moreover, this enforced non-negativity of temporal expression coefficients of each subnetwork makes the dynamic expression process of different subnetworks over time interpretable. Second, the five functional subnetworks were obviously complementary and formed a relatively symmetrical combination covering the whole scalp brain network. Finally, we observed that all functional subnetworks are interactional across brain regions, rather than interconnections within individual brain regions. The evidence has shown that intrinsically correlated activity among distinct brain regions has been consistently observed during both resting state and task execution (Fox et al. 2005). Our results suggest that this feature also exists in the dynamic rules of brain networks in early PD patients and is not altered by MCI. However, the decreased expression strength of temporal-parietal-occipital interaction network and the left/right hemisphere interaction networks (dominated by frontal lobe) was significantly correlated with cognitive impairment in early PD patients (Fig. 4). It indicates that in patients with early PD-MCI, the interaction among temporal, parietal, and occipital lobes and the interaction between frontal lobe and other brain regions within the respective hemispheres are reduced. Filippi et al. reported that early PD patients with MCI experience atrophy of the occipital cortex (Filippi et al. 2020), and other studies have also observed that the temporal-parietal-occipital cortex of PD becomes thinner during the transition from normal cognition to MCI based on the magnetic resonance imaging (MRI) technique (Garcia-Diaz et al. 2018; Mak et al. 2015). Therefore, we speculate that this structural anomaly leads to a reduction in direct interactions among these brain regions. This also provides us with a new idea to further collect imaging data from these patients to validate this work.

It is universally acknowledged that the connectivity of functional brain networks dynamically varies even at the resting state (Bassett et al. 2011, 2013, 2015). And in resting-state networks, it has been previously noted that a stable and highly correlated behavior rapidly changes to uncorrelated activity in switching patterns (Hansen et al. 2015). This perfectly matches the switching between main-expressed functional subnetworks we have found. Noting that MCI induce a regular switching pattern of functional subnetwork. In a time-dependent window, the subnetwork dominated by the central axis of early PD-MCI patients is dominant when its expression coefficient is higher than that of the other subnetworks. In the whole process of integrating into the original functional brain network, the subnetwork dominated by the central axis is the least expressive of all the subnetworks, which indicates that the subnetwork dominated by the central axis is the lowest proportion of the original functional brain network. MCI leads to its regular expression in a period, even stronger than other functional networks, which is the unique feature of early PD-MCI dynamic brain network. In addition, the frontal region is involved in complex cognitive functions, such as switching attention, language, and decision making (Chrysikou et al. 2014), and delayed development of the frontoparietal regions has effects on the proactive cognitive control (Blackwell and Munakata 2014). In this work, MCI induced that the subnetwork dominated by the frontoparietal regions of early PD-MCI patients have higher occurrence and longer duration of expression, but its overall fluctuation is much lower than that of the patients with normal cognition. This abnormal characteristic reveals the slow frontoparietal interaction activities in early PD-MCI, and this injury significantly disrupted cognitive function of the frontoparietal regions.

The limitations of this work deserved to be discussed. Firstly, the relatively smaller number of EEG channels leads to the fact that we can only find the interregional interaction subnetworks in the functional brain network. However, whether multi-electrode array technology can reveal intraregional interaction subnetworks in a certain brain lobe that should exist in the inherent patterns will greatly help to crack the dynamic brain network. Secondly, the relatively small sample size is also a limitation. Since our novel findings in this work provide valuable insights into the nature of MCI-induced functional brain network abnormalities, we plan to recruit more eligible patients to enrich our sample. Moreover, although cognitive impairment is also evident in the resting state functional brain network, supplementing experimental tasks for cognitive function, combined with larger sample sizes and multi-electrode arrays, could better investigate the relationship between MCI severity and the abnormal properties of functional subnetworks in the functional brain network in early PD. Finally, we did not explain the role of the subnetwork dominated by the central axis in the early PD-MCI, because the current study could not provide direct results to explain this problem. This will become our next research question of great interest, and we will design a rigorous task paradigm to explore and answer this question.

Conclusion

In conclusion, our results suggest that there are five functional interaction subnetworks acting as the intrinsic patterns under the functional brain network of early PD patients with/without MCI. MCI induces a regularly interrupted, slow evolution of these five subnetworks in functional brain network dynamics in early PD. What’s more, the dynamic expression characteristics of these five subnetworks reflect the degree of cognitive impairment in patients with early PD. The results in this work provide a deeper insight into MCI-induced abnormalities in functional brain network dynamics and reveal the essential origins for the abnormal macroscopic functional brain network dynamics.

Data available statement

The datasets generated for this study are available on request to the corresponding author.

Declarations

Conflict of interest

The authors declare no competing interests.

Ethical statement

The studies involving human participants were reviewed and approved by Medical Ethics Committee of Tianjin Medical University General Hospital. All subjects understood the purpose of collecting the data and the significance of the study, and signed the written informed consent.