Deep-learning-optimized microstate network analysis for early Parkinson’s disease with mild cognitive impairment

Abstract

Graph-theory-based topological impairment of the whole-brain network has been verified to be one of the characteristics of mild cognitive impairment (MCI). However, two major challenges impede the further understanding of topological features for the personalized functional connectivity network of early Parkinson’s disease (ePD) with MCI. The uncertain of characteristic frequency band reflecting the abnormality of ePD-MCI and the setting of fixed length of sliding window at a second level in the construction of conventional brain network both limit a deeper exploration of network characteristics for ePD-MCI. Thus, a convolutional neural network is constructed first and the gradient-weighted class activation mapping method is used to determine the characteristic frequency band of the ePD-MCI. It is found that 1–4 Hz is a characteristic frequency band for recognizing MCI in ePD. Then, we propose a microstate window construction method based on electroencephalography microstate sequences to build brain functional network. By exploring the graph-theory-based topological features and their clinical correlations with cognitive impairment, it is shown that the clustering coefficient, global efficiency, and local efficiency of the occipital lobe significantly decrease in ePD-MCI, which reflects the low degree of nodes interconnection, low efficiency of parallel information transmission and low communication efficiency among the nodes in the brain network of the occipital lobe may be the neural marker of ePD-MCI. The finding of personalized topological impairments of the brain network may be a potential characteristic of early PD-MCI.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11571-023-10016-6.

Introduction

Parkinson’s disease (PD) is a neurodegenerative disease mainly characterized by motor symptoms (Shulman et al. 2011; Yu et al. 2022). However, a substantial percentage of PD patients have non-motor symptoms as well. Cognitive dysfunction is a common non-motor symptom observed in PD patients and its severity varies from mild cognitive impairment (PD-MCI) to dementia (Emre et al. 2007; Litvan et al. 2011). Clinical findings reveal that nearly a quarter of early PD (ePD) patients experience mild cognitive impairment (MCI) (Schneider and Kortagere 2022) and ePD patients with MCI are at an increased risk of developing dementia (Wallace et al. 2022; Pagonabarraga and Kulisevsky 2012). Thus, pay more attention to the ePD patients at risk of MCI may help to provide more timely interventions. However, the neural basis underlying functional abnormalities of the brain in patients with ePD-MCI remains unknown, which brings a great challenge to explore the special feature of ePD-MCI.

The human brain can be regarded as a complex network (Faskowitz et al. 2018), which enables highly efficient information transmission. Previous research reported that the depletion of nigrostriatal dopamine in PD-MCI affected numerous broadly distributed neural circuits (Weingarten et al. 2015; Wang et al. 2023), therefore, brain network abnormality was a component of PD neuropathology (Strafella 2013; Ostaszewski et al. 2016). Numerous studies about pathological brain network have suggested that PD-MCI is related to abnormal functional connectivity in the brain network, and pointed that the topological characteristics of abnormal brain functional connectivity networks might provide a great potential in understanding brain’s cognitive impairment in PD (Wu et al. 2009; Sharman et al. 2013; Putcha et al. 2015; Zhang et al. 2015). Therefore, exploring the topological characteristics of abnormal brain functional connectivity network in PD-MCI is essential.

Neuroimaging techniques have an ability to characterize the pathological substrates of neurodegenerative diseases such as PD (Pievani et al. 2011). Electroencephalography (EEG), as one of these techniques, is a noninvasive and cost-effective approach to capture brain dynamics in high temporal resolution (Khanna et al. 2015) and has shown good potential in identification of cognitive deterioration in patients with PD (Koch et al. 2019). Although we have the effective technique to explore the characteristics of brain functional networks, we still encounter two major challenges. The first challenge is the determination of the characteristic EEG frequency band. Some 6-Hydroxydopamine hydrobromide (6-OHDA) rodent experiments found that lower frequency (0.5–4Hz) oscillations could be a marker of akinetic symptoms in PD (Whalen et al. 2020). And it was also believed that alterations in signal power in the theta (4–8Hz) and lower alpha (8–10Hz) bands were potential biomarkers of cognitive state in PD (Babiloni et al. 2011; Fonseca et al. 2013). Additionally, brain network activity was explored in specific conventional bands (delta (0.5–4Hz), theta (4–8Hz), alpha (8–12Hz), beta (12–30Hz) and gamma (30–80Hz)) to reveal characteristic patterns of frequency (Newson and Thiagarajan 2019). But the protocols of the classical frequency division have been demonstrated not to be optimal (Elgendi et al. 2011). In our previous study, a deep learning frame for characteristic frequency selection was proposed, where structured power spectral density with spatiotemporal frequency properties were used as the inputs to train the convolutional neural network (CNN) model and revealed 3.5–11Hz was a more accurate characteristic frequency band for ePD (Chu et al. 2021). Therefore, to explore and extract the ePD-MCI brain network features based on the optimal frequency band, deep learning is first used in this work.

The second challenge is the limitation of the data segmentation for constructing brain functional network. Traditional EEG brain functional network construction primarily uses sliding windows to split the original EEG time series, and then utilizes the data within each time window to construct a brain functional network (Kabbara et al. 2020; Gu et al. 2020; Fang et al. 2020). Although conventional sliding window is a good method, the sliding window length for constructing the brain networks is usually limited to a fixed length and second-level time scale (Wang et al. 2020). The fixed time window for functional network construction lacks of flexibility; while time window in a relatively large time scale may induce a loss of detailed dynamics (Vidaurre et al. 2018). Our previous study has found that microstate analysis could exhibit the sub-second dynamics of the brain activity (Chu et al. 2020). EEG microstate refers to a series of distinct quasi-stable states, which are defined by topographies of electric potentials recorded in a multichannel array over the scalp, which remain stable for 80–120 ms before rapidly transitioning to a different microstate (Khanna et al. 2015). The nature of the microstate persistence and transition suggests microstate might to be a basis for constructing the brain function connectivity networks. Therefore, we consider to use the microstate as a criterion for window delineation so as to gain more details of brain connectivity at the sub-second time scale.

Four characteristic EEG microstate topographies have been extracted by some classical procedures and clustering algorithms (Khanna et al. 2014; Pascual-Marqui et al. 1995). By convention, the four topographies were labeled EEG microstate classes A to D (Koenig et al. 2002). Previous studies have found that the occurrence and frequency of these EEG microstate classes profoundly varied in different populations and mental states. For example, microstate class A was more, class D less prominent in patients and risk groups of several mental disorders compared to healthy controls (Andreou et al. 2014; Koenig 2016; Koenig et al. 1999; Lehmann et al. 2005; Milz 2016; Nishida et al. 2013; Rieger et al. 2016; Strelets et al. 2003; Tomescu et al. 2014). The length of microstate classes A and B shortened and the occurrence of class C increased with age in the awake eyes closed resting state (Koenig et al. 2002). These studies suggested that differences between healthy controls and patients with mental disorders can be reflected in the differences of features among the microstate classes, so we speculate that there may also be differences in brain networks constructed with different microstate classes. In order to construct the brain networks that can accurately reflect the characteristics of ePD-MCI, we will build brain function networks under the four microstate classes.

In this work, we first build a CNN model and apply the gradient-weighted class activation mapping (Grad-CAM) method for model visualization to determine the optimal frequency band for the analysis of ePD-MCI, and then carry out the microstate analysis to obtain microstate sequences of the whole brain activity for each participant. Based on each subject’s microstate sequences, personalized functional connectivity networks are constructed, and the graph-theory-based topological characteristics and their clinical correlation with cognitive impairment are evaluated. Finally, the abnormal topological characteristics of brain functional networks in ePD-MCI were analyzed to reveal the essential features of the brain functional connectivity networks of ePD-MCI. The proposed individual microstate function networks will facilitate the characterization of ePD-MCI properties, and become a reliable biomarker for clinically auxiliary diagnosis.

Materials and methods

General information about participants

In this study, a total of 33 primary PD patients were recruited from the Department of Neurology, General Hospital of Tianjin Medical University. The inclusion criteria included (i) the Hoehn and Yahr (H&Y) staging scales of the patients were from 1 to 2.5, (ii) patients were stopped from any medication for more than 12 h before EEG acquisition. The exclusion criteria were as follows: (i) Participants with head tremor symptom, (ii) psychiatric disorders and (iii) head trauma with loss of consciousness. All participants provided written consent, and this study was approved by the Medical Ethics Committee of Tianjin Medical University General Hospital.

For each patient, we performed the Montreal Cognitive Assessment Scale (MoCA) test to assess cognitive function with the lower scores suggesting more severe cognitive impairment. Based on this, 33 PD patients (23 females, aged 64.0 ± 6.5 years; 10 males, aged 60.5 ± 8.1 years) were divided into two groups: the ePD-MCI group (13 subjects with MoCA scale scores of 21–25) and ePD-nMCI group (20 subjects with MoCA scores of 26–30) (Nasreddine et al. 2005; Fujiwara et al. 2010; Lifshitz et al. 2012). The basic information of patients was shown in the Table 1.

Table 1.

Description of the patients

	ePD-MCI	ePD-nMCI	statistical analysis
Age (mean ± SD)	62.000 ± 8.347	63.500 ± 6.295	T = 0.588; p = 0.561
Sex (male/female)	6/7	4/16	${\chi }^2$=2.552; p = 0.110
H&Y stage	1.500 ± 0.577	1.175 ± 0.438	T = −1.732; p = 0.098
Course of disease	3.385 ± 2.725	4.575 ± 4.499	T = 0.855; p = 0.399
MoCA scores	23.000 ± 1.826	27.700 ± 1.302	T = 8.644; p < 0.01

Note: ${\chi }^2$ and T values are the statistical value obtained by chi-square test and independent-samples test respectively. p values are the probability obtained by corresponding statistical test methods respectively (statistical significance was accepted at p < 0.05). Bold font represents significant differences

EEG recording and processing

Resting state EEG data were recorded for at least 10 min when all patients sat in a quiet semi-dark room with their eyes closed, and they were all awake and in a relaxed condition. The international 10–20 system electrode placement was used on the scalp of the patients. There were nineteen channels including FP1, FP2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2. The reference electrodes were placed at the bilateral earlobes (A1, A2) and ground electrode was placed on the forehead. Additional channels were used to record electrooculogram (EOG), electromyogram (EMG) and electrocardiogram (ECG) signals, these signals could monitor the corneo-retinal standing potential, electrical activity produced by skeletal muscles and the heart's electrical activity respectively for further processing. The contact impedance was kept below 5 kΩ during the recording. The sampling rate was 500 Hz.

The ocular, cardiac, and movement noise artifacts were eliminated using the fast independent components analysis (fast-ICA) algorithm (Hyvärinen 1999). EEG signals were decomposed by fast-ICA algorithm into several independent components. Pearson's correlation coefficients were computed independently for each independent component and the noise artifacts. We judged an independent component with an absolute value of correlation coefficient larger than 0.5 to be significantly correlated with a specific artifact signal, and we zeroed out these components to generate de-noised EEG signals with noise artifacts eliminated. EEG signals was processed by a 1–45 Hz band-pass Finite Impulse Response (FIR) filter to eliminate the interference of high frequency noises. In addition to contributions from neural activity, the signals contain interfering signals from other sources (Jiang et al. 2023). Some artifacts generated by body movement and technical artifacts were hard to eliminate using fast-ICA and filtering, therefore, they were rejected by manual screening. Finally, 5 min of de-noised data from each patient were saved for subsequent analysis. The preprocessing was performed by the EEGlab toolbox in MATLAB software (MathWorks Inc., Natick MA, United States).

Optimal frequency band acquisition based on deep learning

CNN structure and model visualization

We defined continuous 3s EEG data as an epoch (each epoch contained 1500 time points). The first 100 epochs of EEG data for each patient were selected to calculate the power spectral density (PSD) of 89 frequency points at 1–45Hz with a step size of 0.5Hz.

The structure of the CNN model was shown in Fig. 1. We constructed three convolution layers for the hidden layer of CNN model based on the classical application (Ullah et al. 2018; Hosseini et al. 2017; Lih et al. 2020), and during the continuous debugging of the architecture of CNN model, we did some changes, where some widely used layers, such as the pooling layer, dropout layer, or batch normalization layer, were removed. This is because that excessive model parameters they generated caused overfitting of the CNN model, which reduced model generalization. Furthermore, a normalization layer was added between two convolutional layers to boost convergence speed of the model, and full-connect layers were added before the SoftMax layer to provide an intuitive representation of the data distribution. Meanwhile, we set the kernel size of the convolutional layer to 1 × 3.

Fig. 1.

CNN structure and model visualization. C-L: convolution layer, N-L: normalization layer, NL-A: nonlinear layer activation, KS: kernel size (1 × 3), P-L: pooling layer, D-L: dropout layer, BN-L: batch normalization layer, O-L: output layer, FC: full-connect layer. 4@89 × 19 means we used four convolution kernels to extract the characteristics of the input with dimension 89 × 19. Obtain the feature maps and the weight of each feature in the last convolutional layer by back propagation and take a weighted sum of them. Finally, visualization results are obtained by a ReLU

The datasets contained 3300 samples (33 subjects × 100 epochs) divided into a training set and a test set at a ratio of 4:1. To evaluate the performance of the model by forming five-fold cross-validation, we randomly separated all samples from the training set into five equal sub-sample sets. During the training, three subsamples (1980 samples) were used as training data, and the rest one (660 samples) was used as validation data of the model test, the process was iterated four times to ensure all subsamples could be engaged in the training and testing phases. And the ultimate performance was evaluated by the test set. We set learning rate as 0.001. In order to evaluate the overall performance of the CNN model, we assessed accuracy (ACC) and area under the curve (AUC) in our study. ACC was the percentage of patients that were correctly classified, and AUC was the probability that a positive sample (ePD-nMCI) and a negative sample (ePD-MCI) were randomly selected to identify the probability that the positive sample score was higher than the negative sample score.

Based on the CNN model, we adopted Grad-CAM method (Selvaraju et al. 2017; Chu et al. 2021) for model visualization to obtain the significant feature which could discriminate ePD-MCI and ePD-nMCI. Grad-CAM could visualize CNN without modifying the network structure or retraining. Since the activation feature graph of the last convolutional layer of the CNN model contained high spatial information, the recognition results of ePD-nMCI and ePD-MCI were determined by the weight of each feature map in the last convolutional layer. Therefore, by using the Grad-CAM and normalizing the Grad-CAM results on a scale of 0–1 with a threshold of 0.45, the activation feature graph and the weight of each feature graph could be obtained to visualize the CNN model, then we could identify the optimal frequency band which could discriminate ePD-MCI and ePD-nMCI.

EEG microstate analysis

Based on the extracted optimal frequency band for ePD-MCI, we performed the pre-processing step of filtering the artifact-free EEG data. Then the pre-processed data were imported into CARTOOL software for EEG microstate analysis to obtain the microstate sequences of each subject. Global field power (GFP) is a measure of scalp potential intensity that is based on potential differences between the potential of all electrodes at each sampling point and their average potential, leading to a scalar value of the field intensity at each sampling point (Skrandies 1989). GFP is defined as:

$$\mathit{GFP}(\text{t})=\sqrt{\frac{\sum _{i=1}^n{[v_i(t)-\overline{v}(t)]}^2}{n}}$$ 1

where $i$ denotes each electrode, $n$ represents the number of electrodes of the EEG signals, $v_i(t)$ is the electric potential of the EEG signals $v$ at the electrode $i$ and $\overline{v}(t)$ is the average electric potential of all electrodes of the EEG signals $v$ (Jia et al. 2021). Since the GFP depicts the intensity of the electric field at each instant, it is frequently used to assess the global brain reaction to an event or to characterize the fast variations in brain activity (Khanna et al. 2015).

In addition, local maxima of GFP curve represents instants of strongest field strength and highest topographic signal-to-noise ratio, therefore, we select the instantaneous voltage amplitude values of all electrodes at the GFP peaks for clustering analysis (Khanna et al. 2015). For microstate segmentation, modified k-means algorithm (Murray et al. 2008), a clustering algorithm, is used to compute EEG microstates. We performed the microstate analysis in four steps (Fig. 2A–D).

Fig. 2.

Schematic diagram of the microstate analysis process. A Calculate global field power (GFP) of the EEG signal at each time point. B Obtain the corresponding electric potential distribution topographic at the local peak point of GFP. C Use the modified k-means clustering algorithm to get the EEG microstate classes. D Represent microstate sequences by four microstate maps. E Use the microstate sequences to determine the microstate windows, the length of each microstate window depends on the corresponding microstate duration. F The microstate windows of the same microstate class are clustered together to build microstate-class-based brain networks

Personalized brain network construction

Microstate windows

Based on the microstate analysis, we took the obtained microstate sequences as the basis for dividing the microstate windows (Fig. 2E), the microstate windows of the same microstate class were clustered together to build microstate-class-based brain networks (Fig. 2F). The length of microstate windows usually focused on the sub-seconds level, the length of the microstate window was variable and the windows next to each other did not overlap. The microstate sequences showed a typical transition property and short-time continuity, which characterized the dynamical characteristics of ePD-MCI. Therefore, we renewed the data segmentation strategy based on the properties of microstate sequences, and built the brain functional network under four microstate classes. We built the brain functional networks with the microstate windows of microstate class A, B, C and D respectively. In that case, we could obtain the corresponding function brain networks under these four microstate classes, and we defined them as microstate network A, B, C and D respectively.

Brain functional network construction

For constructing a brain functional network, many algorithms have been proposed by researchers in recent years (Schreiber 2000; Baccalá and Sameshima 2001). However, most of these methods are unable to contend with volume conduction. To effectively solve this problem, Stam et al. (2007) proposed the phase lag index (PLI). PLI, a method of using phase information to construct brain functional networks, omits EEG amplitude information, so volume conduction problem induced by transmission of EEG signals from intracranial to cortical electrodes can be avoided. PLI is defined as:

$$\mathit{PLI}=\left|〈\text{sign}\left[\mathrm{\Delta }\varphi \left(t_k\right)\right]〉\right|$$ 2

where $\mathrm{\Delta }\varphi$ is the phase difference at time point $t_k$ between two signals. The value of PLI ranges from 0 to 1, where PLI = 0 indicates that there is no coupling between the two-time series and PLI = 1 indicates completely synchronized of the two signals. The larger value of the PLI, the higher the degree of synchronization coupling. A functional connectivity matrix containing the PLI values between each pair of electrodes can be obtained. In order to avoid redundant connections, it is necessary to remove the connections with less correlation, so we perform the binarization processing. The following calculation of complex network attributes will be based on the binarized functional connectivity matrix. We adopt the absolute threshold approach to binarize functional connectivity matrix, it describes the selection of those network edges that exceed an absolute threshold 0.47, with (in the binary case) all surviving connections set to 1 and all other network connections set to 0 (Van den Heuvel et al. 2017).

Topological properties analysis based on graph theory

In recent years, the study of static and dynamic brain networks has developed rapidly. The graph theory approach establishes a mathematical framework to simulate paired communication between network nodes. In addition, it also can be applied to functional and structural connections in neuroscience. Graph-based network analysis reveals meaningful information about the topology of human brain networks (Liu et al. 2021). The brain functional networks constructed in this work were all unweight networks, and three indicators were used to describe the brain network topological properties, these indicators are defined as follows:

1. Clustering coefficient

The clustering coefficient, is a characterization of the dispersion degree of the nodes in the network, which can reflect the local connectivity in the network and represent a certain brain neural activity. Clustering coefficient is defined as:

$$C_i=\frac{2t_i}{k_i(k_i-1)}$$ 3

where $C_i$ represents the clustering coefficient of the $i_{\mathit{th}}$ node, $t_i$ is the number of neighboring nodes of the $i_{\mathit{th}}$ node, and $\frac{k_i(k_i-1)}{2}$ denotes the maximum number of connections a node can occupy.

• (2)Global efficiency

The global efficiency refers to the degree of distribution of the global network composed of all nodes in the entire brain function network, which is considered an important indicator to measure the efficiency of parallel information transmission between neurons or in the entire network. Global efficiency is defined as:

$$E_g=\frac{1}{N_G(N_G-1)}\sum _{\begin{array}{c}i,j\in G\hfill \\ i\ne j\hfill \end{array}}{d}_{\mathit{ij}}^{-1}$$ 4

where $G$ represents global networks, $N_G$ denotes the number of all nodes in the global network, $d_{\mathit{ij}}$ is the shortest distance between node $i$ and node $j$ in the global network.

• (3)Local efficiency

The local efficiency measures the communication efficiency among the nodes. The local efficiency is an important parameter related to the local communication capacity, which is defined as:

$$E_l(i)=\frac{1}{N_{G_i}(N_{G_i}-1)}\sum _{j,k\in G_i}{d}_{\mathit{jk}}^{-1}$$ 5

where $E_l(i)$ represents the local efficiency of $i_{\mathit{th}}$ node, $d_{\mathit{jk}}$ is the length of the shortest path between $j$ and $k$, $N_{G_i}$ means the number of the networks between the $i_{\mathit{th}}$ node and other nodes which connect with it.

Statistical analysis

Statistical analyses for demographic data, clinical data and characteristic parameters were performed by SPSS 25.0 software (IBM Inc., Chicago, IL, United States) and MATLAB software (MathWorks Inc., Natick MA, United States). We used the chi-square test to compare the gender differences between ePD-nMCI and ePD-MCI groups, and independent-sample t-test to compare the differences in subjects’ age, H-Y stage, course of the disease and MoCA scores at a 0.05 significant level. Independent-sample t-test was used to evaluate the differences between ePD-nMCI group and ePD-MCI group. In order to quantify the linear relationship, Spearman's correlations between clinical MoCA scales and the topological parameters in each microstate network were tested. The significant Spearman's correlation between the topological parameters and the MoCA scales was accepted at P < 0.05. Corrected false discovery rate (FDR) acted on P-values for multiple comparisons.

Results

Performance of CNN model

Five-fold cross-validation results of CNN model on the test sets indicated that the ePD-nMCI and ePD-MCI could be accurately classified with mean ACC of 99.475% (lowest ACC: 99%; highest ACC: 99.875%) and mean AUC of 99.495% (lowest AUC: 98.85%; highest AUC: 100%), and the test set with ACC of 99.51% and AUC of 99.485% illustrated the excellent performance of CNN model in distinguishing ePD-MCI and ePD-nMCI. Therefore, using spatial frequency features as input data to train CNN model could achieve great discrimination effect, which was attributed to that frequency features were endowed with spatial scale to deepen the dimension of this feature. In order to explore the characteristic basis for CNN model for achieving such excellent classification performance, we further adopted the Grad-CAM algorithm to realize the characteristic frequency band visualization. Performance of CNN model was shown in Fig. 3.

Fig. 3.

Performance of CNN model. A The ACC of CNN models. B The AUC of CNN models. First five different color column represents the result of five cross-validation of the model

Optimal frequency band acquisition based on CNN

As shown in Figs. 4 and 5, in the characteristic map, there were seven frequency points (1, 1.5, 2, 2.5, 3, 3.5, and 4Hz) with significant distinguishing contribution. It was obvious that there was the ability to reflect ePD-MCI abnormality in this frequency band. Therefore, this frequency band was defined as the optimal frequency band indicating the characteristics of abnormal brain network in ePD-MCI. Based on the above features, we studied the topological properties of the brain network in this defined optimal frequency band and explored brain regions involved in cognitive functions.

Fig. 4.

The 89 topographic maps with personalized characteristics with Grad-CAMs. The color represents the weight of the identification region in the input space for identifying the category ‘ePD-MCI’

Fig. 5.

Optimal characteristic frequency bands and their corresponding distribution of characteristic brain regions. The color represents the weight of the identification region

Brain network topological properties analysis

Group differences of topological parameters between ePD-MCI and ePD-nMCI were investigated in different brain regions among four microstate networks, which was shown in Fig. 6.

Fig. 6.

Compare the topological properties of ePD-MCI and ePD-nMCI

Among the four microstate networks, significant differences of topological parameters were distributed in these characteristic brain regions: occipital lobe, frontal lobe and central lobe, and we observed a significant decrease in the topological parameters (clustering coefficient, global efficiency, local efficiency) of ePD-MCI compared with ePD-nMCI, which meant that in these brain regions, the low degree of nodes interconnection, low efficiency of parallel information transmission and low communication efficiency among the nodes. In addition, it was obvious that the significant differences in topological parameters of brain regions were mainly concentrated in microstate network A, and especially in the occipital lobe of microstate network A, therefore, we speculated that abnormal function in occipital lobe of microstate network A might be involved in the cognitive process.

Correlation between topological parameters and cognitive level

The MoCA scale was a widely used screening assessment for detecting cognitive impairment in clinical practice (Costa et al. 2012). We took the MoCA scale to assess the patient’s cognitive impairment level and explored the Spearman’s correlation between the MoCA scale scores and the topological parameters. Figure 7 showed the result of Spearman’s correlations.

Fig. 7.

Clinical MoCA correlations. Spearman’s correlation analysis between values of topological parameters and MoCA scale scores under four microstate networks. p represents the significance level

In microstate network A, a significantly positive correlation between clustering coefficient and MoCA scale scores was observed (r = 0.4506, p = 0.00851), and a significantly positive correlation between local efficiency and MoCA scale scores was observed (r = 0.4132, p = 0.01685). In microstate network B, global efficiency significantly correlated with MoCA scale scores (r = 0.3932, p = 0.0236). Similarly, global efficiency showed a strong positive correlation with MoCA scale scores in microstate network C (r = 0.4385, p = 0.01125). From the results, we observed that significant correlations were all concentrated in the occipital lobe of microstate network A, and with the decline of the cognitive level, the corresponding topological parameters of occipital lobe significantly decreased. Therefore, we got a conclusion that topological properties of occipital lobe could reflect the degree of cognitive impairment.

The superiority of the personalized brain networks compared with conventional sliding-window-networks

In order to verify that the brain networks based on the strategy of microstate windows could better reveal the abnormal brain network topological properties of ePD-MCI, we further constructed brain networks using the conventional sliding windows in the same optimal frequency band and calculated the corresponding network topological parameters. In order to avoid the contingency of the sliding windows, we used several typical window lengths and step lengths: (1) window length of 500, step length of 500; (2) window length of 1000, step length of 500; (3) window length of 1000, step length of 1000.

We observed that the conventional strategy (1) could identify the significant difference of clustering coefficient between ePD-MCI and ePD-nMCI and the significant difference of local efficiency both in occipital lobe. The conventional strategy (2) could identify the significant difference of clustering coefficient in occipital lobe and the significant difference of local efficiency in occipital lobe and parietal lobe. The conventional strategy (3) could identify the significant difference of clustering coefficient in occipital lobe and the significant difference of local efficiency in occipital lobe and parietal lobe.

While using microstate windows, the significant differences of clustering coefficient of the brain network were mainly indicated in microstate network A and D. Specifically, the significant differences of clustering coefficient were distributed in frontal lobe, central lobe and occipital lobe for microstate network A and occipital lobe for microstate network D. The significant difference of local efficiency of the brain network mainly occurred in microstate network A and was distributed in occipital lobe. The significant differences of global efficiency of the brain network were mainly indicated in microstate network A, B and C. Specifically, the significant differences of global efficiency were distributed in occipital lobe and central lobe for microstate network A, occipital lobe for microstate network B and occipital lobe for microstate network C. By further calculating the correlation between network topological parameters with significant differences and the scores of MoCA scales, significant correlated brain region in different topological parameters was marked by the asterisk in the Table 2. The detailed results of the conventional sliding-window-network strategy were given in Supplement Material. By comparison, it could be found that microstate windows strategy had superiority in explore the pathological brain network topological properties of ePD-MCI.

Table 2.

The comparison between microstate windows and three common conventional sliding windows

	Conventional sliding window			Microstate window
	SL:500 WL:500	SL:500 WL:1000	SL:1000 WL:1000
Clustering coefficient	Occipital lobe	Occipital lobe*	Occipital lobe*	Frontal lobe_A Central lobe_A Occipital lobe_A* Occipital lobe_D
Local efficiency	Occipital lobe*	Occipital lobe* Parietal lobe*	Occipital lobe* Parietal lobe*	Occipital lobe_A*
Global efficiency	None	None	None	Occipital lobe_A Central lobe_A Occipital lobe_B* Occipital lobe_C*

Note: The asterisk means the brain region with significant correlation between MoCA scales and topological parameters. Brain region_A/B/C/D means the brain region in microstate network A/B/C/D. SL: step length, WL: window length

Discussion

In this work, to deal with two major challenges in the construction and analysis of brain networks, we proposed a novel deep learning paradigm to determine the characteristic EEG frequency band of ePD-MCI, and constructed the personalized functional brain networks. Firstly, the CNN was adopted to detect the optimal frequency band for identifying ePD-MCI, and we proved that CNN has convincing performance for distinguishing ePD-MCI and ePD-nMCI. The Grad-CAM visualization results revealed 1–4 Hz frequency band that was the most effective in identifying ePD-MCI. The band (1–4Hz) was thus defined as the optimal frequency band for reflecting the abnormality of ePD-MCI. After identifying the optimal frequency band based on CNN model, we preprocessed the raw EEG data in this optimal frequency band to produce clean and artifact-free data. Then, microstate analysis was carried out on the preprocessed data. Based on the microstate sequences, the personalized brain functional network was constructed. Finally, the graph-theory-based topological parameters of the microstate brain network were calculated and their clinical correlation with cognitive impairment were evaluated.

Previous studies suggested that the microstate windows had the following benefits. First, the microstate window provided a transcendental framework for describing the transiently stable states of brain dynamic activity (Chu et al. 2020). Second, based on microstate analysis of synchronous EEG and fMRI data, previous studies had shown that the emergence of a certain class of the microstate would maintain its state for a while, and then the microstate would transit to another class. The whole dynamic process had a specific correspondence with the alternating switching between brain functional networks (Xu et al. 2020). Third, microstate windows could reflect the characteristics of brain functional networks corresponding with the same class of the microstates. The division rules of brain functional networks based on the prior microstate class properties facilitated to describe the spatiotemporal variations of the topological characteristics of the brain functional networks for PD-MCI patients. Therefore, our study proposed an improved brain network analysis strategy of microstate windows based on microstate sequences. Additionally, in the construction of the brain network based on conventional sliding window method, there was still few uniform reference standards, and the chosen of window length remained a great challenge. Several studies had shown that large sliding windows might hamper the detection of sudden changes in brain networks due to the reduced temporal resolution (Carbo et al. 2017; Cai et al. 2020). In our study, we compared the conventional sliding windows and the microstate windows, finding that the brain networks based on the strategy of microstate windows could capture more topological properties which were attributed to the benefits of the strategy above.

From our results, we could observe that the microstate network A exhibited significant characteristics. Previous study used EEG microstate analysis to investigate the relationship between microstate networks and cognitive control network, their results suggested that microstate network C was negatively correlated with cognitive control, and that microstate networks A and B indicated more control, but higher cognitive control was not conducive to idea generation and normal function (Jia et al. 2021). This interpretation would agree with the results found in our study, ePD-MCI exhibited significant characteristics in microstate network A. This would suggest a higher cognitive control, which was not conducive to idea production and normal cognition function, leading to some degree of cognitive impairment. In addition, several studies using simultaneous functional magnetic resonance imaging (fMRI) and EEG recordings demonstrated that resting state networks assessed by fMRI correlated significantly with the temporal progression of all four EEG microstates (Koenig et al. 1999; Lehmann et al. 2005). EEG microstate classes A, B, C, and D were associated with auditory network (Gschwind et al. 2016; Britz et al. 2010) or sensorimotor (Gschwind et al. 2016), visual system network (Britz et al. 2010; Yuan et al. 2012), saliency network (Gschwind et al. 2016; Michel and Koenig 2018), and attention network (Gschwind et al. 2016; Michel and Koenig 2018), respectively. Besides, PD patients showed higher frequency of auditory alterations due to dopaminergic deficits (Iwaki et al. 2014; Lamas et al. 2017). All in all, microstate class A was the area implicated in auditory processing network and these clinical features confirmed our result of the significant topological properties we observed in microstate network A.

Besides, our analysis of the results showed that we found that the differential results as well as the significant Spearman’s correlations were all concentrated in the occipital lobe regions. Previous studies argued that occipital lobe atrophy was a key marker to distinguish dementia-PD (PDD) and non-dementia PD patients (Amboni et al. 2015; Hanganu et al. 2013). Similarly, we found that the ePD-MCI and ePD-nMCI patients had most differences in occipital lobe, and the topological parameters in occipital lobe had significant correlation with MoCA scores. In addition, Mammone et al. (2017) showed that for MCI subjects, a significant positive correlation could be observed between MoCA variation and the variation of clustering coefficient, which implied that the variation of clustering coefficient and cognitive levels were correlated. Another study proposed that global efficiency and the clustering coefficient of networks were reduced in MCI, which suggested cognitive control was associated with these network topology (Berlot et al. 2016), and decreased local efficiency was related to poorer overall cognitive performance across time, while decreased local efficiency was also associated with poorer verbal fluency (Amidi et al. 2017). These research findings provided the strong support for our results. Overall, with the decline of the cognitive level, the topological parameters of the corresponding brain region (occipital lobe) decreased, which suggested topological properties of occipital lobe could reflect the degree of cognitive impairment (relying on MoCA scale correlations). And this paper also validated the effectiveness of the new paradigm for brain network construction as well as an optimal frequency band selection method we proposed.

The limitation in this work was that we analyzed EEG signals of 13 ePD-MCI and 20 ePD-nMCI, which was still a small-scale clinical experiment. In order to verify the broad applicability of the conclusions, the number of experimental subjects still need to be increase and results need to be tested repeatedly on a large amount of data.

Conclusion

We concluded that the cognitive impairment for ePD patients could be expressed by microstate-correlated brain network temporal and spatial characteristics in low-frequency range (1–4Hz). And the brain function abnormal of occipital lobe could better reflect the differences in cognition impairment between ePD-MCI and ePD-nMCI. Our results revealed the topological properties of the brain network reflecting cognitive impairment in ePD-MCI. The results played a profound role in enhancing our understanding of cognition-related brain dysfunctions of PD. Meanwhile, our proposed characteristic frequency-band-selection tool and personalized brain network construction method provided a novel methodology for brain functional network research based on EEG data.

Supplementary Information

Below is the link to the electronic supplementary material.

Data availability

Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.

Declarations

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.