Identifying Patients with Epilepsy Having Depression/Anxiety Disorder Using Common Spatial Patterns of Functional EEG Networks

The presentation of epilepsy with depression/anxiety disorder (E-AD) is a complex comorbidity, and a systematic psychiatric screen poses a heavy economic burden on the patients. Typically, distinguishing between E-AD patients and epilepsy without depression/anxiety disorder (E-no-AD) needs multiple evaluations, which is incredibly difficult owing to the high costs. Therefore, there is an urgent need for a reliable assessment protocol to preemptively distinguish between the two types of patients. We collected the resting-state electroencephalography (EEG) signals from E-AD patients and E-no-AD patients and constructed the functional networks for the two groups of. The connectivity strength and the corresponding network properties of different frequency bands were statistically compared to explore the inherent differences. Based on the differences in functional networks, we evaluated the classification performance using different network features, including the network properties, principal component analysis (PCA) and spatial patterns of networks (SPN). Using the support vector machine, the differentiation of the 4416 EEG segments from E-no-AD patients and 6346 EEG segments from E-AD patients was realized using three different network features. The results revealed that the SPN feature of functional network connectivity was remarkably superior to the traditional network feature for distinguishing between the two patient groups, with an appreciable classification accuracy of 89.90% and sensitivity of 87.37% and specificity of 91.67%. These findings demonstrate the superiority of functional SPN features as a reliable way to characterize the differences between E-AD and E-no-AD patients and potentially provide insights regarding the complex mechanisms causing E-AD in patients.


Introduction
Epilepsy affects more than 70 million people globally, making it among the most common serious neurological disorders, and notably, active epilepsy in more than 75% of cases goes untreated, which represents a significant treatment gap, particularly in low-income and middle-income countries [1,2]. Depression and anxiety are social dysfunction diseases primarily manifested by low mood. A survey showed that patients with epilepsy had a predisposition to develop some mood disorders, with depression and anxiety being the most common comorbidities of epilepsy [3]. The morbidity and mortality were higher for epilepsy patients with depression/ anxiety disorders (E-AD) than for those without depression/anxiety disorders (E-no-AD) [4,5]. Owing to the close 1 3 relationship between epilepsy and depression/anxiety, both individuals and the society stand to greatly benefit from an effective method for diagnosing epilepsy comorbid depression [6,7,8]. The existing diagnoses are mainly based on the assessment of patient scales, and the long evaluation process leads to a huge economic burden on the patients. Conversely, the scale-based methods are greatly affected by the patients' state of mind and the physicians' experience, making the assessment very unstable.
Electroencephalography (EEG) is easily accessible and relatively inexpensive [9], which has led to its widespread use for the diagnosis and classification of epilepsy and mental disorders. The characteristics of EEG signals allow us to distinguish specific patterns, which aid the identification of patients with epilepsy [10,11], the detection of epileptogenic focus localization [12,13], and the prediction of depression and anxiety [14][15][16]. In theory, internal interactions among brain regions can be characterized using a functional network constructed from EEG signals; this would allow us to distinguish resting-state signals by differentiating specific network patterns [17][18][19]. In a clinical setting, EEG records can help estimate the degree and decide the optimal treatment of different diseases, which can help reduce the economic pressure [11,20].
In previous studies, the complex networks or potential pathological abnormalities were described by graph theory [21,22]. It is a reflection of a network and is represented by its connections and nodes. Several network properties can be used to describe the graphs, such as clustering coefficient, characteristic path length, local efficiency, and global efficiency. Overall connectedness is indicated by the characteristic path length, and the local connectedness of the graph is measured by the clustering coefficient [23][24][25]. However, the complete body of information within a network cannot be reflected in the properties represented by statistical measurements. This approach aims to use changes in connectedness to our advantage for classifying and identifying the anxiety state using EEG signals. Accordingly, not always can ideal performance be achieved by differentiation of specific brain networks on the basis of graph theory analysis findings. To this end, some classic EEG feature extraction methods have been proposed by researchers, such as the adaptive auto regressive model [26], wavelet transform [27,28], empirical mode decomposition [29,30], and the analysis of common spatial [31,32]. Currently, the analysis of spatial pattern of network (SPN) is acknowledged as effective methods to reflect spatial features in an EEG network. The spatial information remarkably affects two-class EEG signal classification; in a training data set, this algorithm computes spatial filters that minimize the variance of one class and maximize the variance of the other [33,34]. The superiority of such spatial information has been confirmed in previous related studies for several purposes, such as predicting epilepsy seizures [35,36], differentiating between patients with schizophrenia vs. healthy controls, and differentiating between psychogenic non-epileptic seizures vs. epileptic seizures [10,37].
As we mentioned above, a stable, clinically usable method for distinguishing between E-AD and E-no-AD patients can help doctors with early detection, timely treatment, and non-invasive screening. However, directly differentiating between the two groups when the patients are in a resting brain state remains challenging. As both E-AD and E-no-AD are associated with brain network abnormalities [38][39][40], identifying inherent properties of the resting state network may help distinguish between the two groups. Herein, the resting-state EEG signals were collected from 11 E-no-AD patients and 15 E-AD patients and were used to construct functional networks. We attempted to realize a stable differentiation of E-AD patients and E-no-AD patients based on the spatial features of functional networks. To compare the classification performance, we also estimated the accuracy of the classification based on the traditional network properties and the classic principal component analysis (PCA) from the functional networks. According to our final results, the SPN feature of functional network connectivity showed great superiority to the traditional network features for distinguishing between E-AD patients and E-no-AD patients.

EEG Dataset
We collected the resting-state EEG signals of patients with epilepsy from the General Hospital of Western Theater Command PLA. The EEG was performed with locations according to the international 10-20 system using 21 Ag-AgCl electrodes. Next, we sampled all signals at 500 Hz with a 50-Hz notch filter and a 0.01-100-Hz bandpass filter. The Ethics Committee of The General Hospital of Western Theater Command PLA approved all experimental protocols. All participants provided written informed consent, and the study was performed in compliance with the tenets of the Declaration of Helsinki. Two senior epileptologists having extensive experience in diagnosis and treatment of epilepsy established the diagnosis.
After careful review, patients to be included for analyses were selected by two senior epileptologists with extensive experience in diagnosis and treatment of epilepsy. They selected 15 E-AD patients (age, 18-57 years) and 11 E-no-AD patients (age, 19-46 years). The following were the inclusion criteria for the screened E-AD patients: (1) Each patient underwent an outpatient or inpatient video-EEG recording for ≤ 24 h. The EEG segments under the resting state were screened out by referring to the video recording done over a longer period, and these segments were confirmed by two senior experts. (2) A confirmed diagnosis of epilepsy was established by two senior experts; (3) Both groups of patients were assessed using the Hamilton Mood Scale or Self-Rating Anxiety/Depression Scale (SAS/SDS) by professional physicians [41,42] and were identified as having anxiety/depression disorder; and 4) the treatment did not involve the use of any mood-altering drugs [43]. The E-no-AD patients included herein were screened using all the same standards except for not being detected as having anxiety/depression disorder. First, using the reference electrode standardization technique, the data were transformed to the approximate zero reference for reducing the impact of the reference effect [44]. Continuous EEG data were randomly divided into five-second segments for each patient, and high-amplitude segments (> 100 μV) were excluded. Finally, 6346 and 4416 EEG segments for E-AD and E-no-AD patients were selected for the next step, respectively.
Coherence ( Coh ) was used to measure the connection strength between each pair of electrodes for establishing the functional network. In theory, for analyzing the cooperative, synchrony-defined cortical neuronal assemblies, Coh is the most commonly used metric. This metric represents a linear relationship at a specific frequency between two signals [ x(t) and y(t) ] on the basis of their cross-spectrum. Notably, to indicate the linkage strength between two network nodes, we herein adopted frequency-specific coherence. Coh was expressed using the following formula [46]: where P xy (f ) is the cross-spectrum between x(t) and y(t) , and P xx (f ) and P yy (f ) are the respective auto-spectra at frequency f estimated from the Welch-based spectrum at 0.1-Hz resolution. For each frequency band, the coherence matrices of all frequency points in this band were averaged to compute the coherence matrix.

Network Properties
For measuring the network topology property, we herein used several network measurements, such as global and local efficiency, characteristic path length, and clustering coefficient. We calculated the clustering coefficient ( CC ) as follows [47]: wherein k i is the degree of node i i , and ij is the weight between nodes i and j in the network. A network's characteristic path length ( L ) when L ij is the shortest path length between two nodes was calculated as follows: The global efficiency ( E g ) was computed using the following formula [48]: The local efficiency ( E i ) of node i was defined as follows: We used the Brain Connectivity Toolbox (http:// www. brain-conne ctivi ty-toolb ox. net/, Rubinov et. al) to calculate the above network properties. The authors have reported on more detailed descriptions of the network topology properties [48].

Principal Component Analysis
Notably, PCA was aimed to transform a number of correlated variables into a significant smaller number of uncorrelated variables, called principal components. It has a wide range of applications, such as de-noising signals, cluster analysis, feature reduction, and pattern recognition.
Let the centered data input vectors be , and s t linearly transforms each vector x t as: The eigenvalue of the principal components could be calculated as follows: where i is the eigenvalue of C . Then, the principal components of s i could be computed as follows:

Spatial Patterns Networks
For distinguishing between normal and abnormal EEG components, the use of common spatial pattern (CSP) analysis was proposed in the early 1990s [33,34]. SPN is primarily aimed at identifying the CSPs among various weighted brain network topologies. Therefore, as is the case with canonical CSP, the SPN-extracted spatial pattern is not in the physical data space but rather in the network space [49].
Let 1 and 2 be the N × N centered matrices for each subject; the spatial filters are the projections that maximize the following function [10,34]: Here Φ 1 and Φ 2 are the covariance matrices of the adjacency matrix for the two groups. The objective function can be written as follows upon the introduction of the Lagrange multiplier: Under the condition L = 0 , the generalized eigenvalue equation can be used to estimate the objective projection .

Pattern Recognition
Herein, to distinguish between E-AD patients and E-no-AD patients, we performed the network analysis and the feature extraction in the collected EEG datasets. As shown in Fig. 1, we constructed the functional networks from resting-state EEG segments. Then, for evaluating the classification performance, four traditional network properties (characteristic path length, clustering coefficient, local efficiency, and global efficiency), the PCA features of the network, and SPN features of the network were extracted. Then, we compared the classification performances to identify effective features for distinguishing between E-no-AD and E-AD patients. Notably, to explore the mechanism, the above comparisons were repeated in different frequency bands. For each feature, we introduced a support vector machine (SVM) classifier to learn feature distribution [50]. Next, to determine the Fig. 1 The procedures of differentiating between E-AD patients and E-no-AD patients based on resting-state EEG segments. a The process of constructing the functional networks from the training dataset.
b Building the functional networks from the testing dataset. c Extracting the network features from the functional networks for the differentiation of E-AD patients and E-no-AD patients optimized set of parameters, we implemented a grid search approach. It was ensured that the process for each step in the testing set was the same as that in the training set [51]. To ensure the reproducibility of the results, five-fold cross-validation was used to evaluate prediction results for validation.

Statistical Testing
Herein, we employed two-sample Student's t-test to compare the connectivity strength represented by the trial-averaged networks, wherein p < 0.01 was considered as the threshold for significance. We also used it to assess significant differences in network properties between E-AD patients and E-no-AD patients.
To evaluate the classification performance based on the different features, the five-fold cross-validation strategy was used for the testing process. In each evaluation, four-fifth of the segments in this dataset were used for training, and the other segments were used for testing. This process was repeated to ensure that all segments served as the testing dataset. We herein used three evaluation metrics, namely specificity ( SPE ), sensitivity ( SEN ), accuracy ( ACC ), to assess the specific classification performance. Sensitivity and specificity values indicate missed diagnosis and misdiagnosis rates of E-AD patients, respectively; the preference is to keep these rates low, which indicates an overall better performance. Accuracy represents the probability of correct identification in all cases.
where N AD and N NAD represent the total number of E-AD and E-no-AD patients, and n AD and n NAD represent the number of correctly identified EEG signals of these patients, respectively.

Results
In this study, we built functional networks in different frequency bands, and the classification performance was evaluated on the basis of network features, which included network properties, SPN features, and principal components of networks. First, we compared the network properties of the functional networks in different frequency bands for E-no-AD and E-AD patient groups. As shown in Fig. 2, we found the four network properties, namely characteristic path length, clustering coefficient, local efficiency, and global efficiency, did not exhibit significant differences between the two groups (p < 0.01), thus making it challenging to differentiate between the two patient groups only on the basis of network properties (as shown in Table 1). For comparison purposes, we also performed the classification on the basis of principal components and spatial features of networks. For most frequency bands, the principal components of functional networks were superior to the network properties in differentiating between E-no-AD patients and E-AD patients. However, principal component analysis (PCA) did not achieve ideal classification performance, and the best accuracy rate obtained was 66.31%. Therefore, the two abovementioned methods may not be suitable for the actual clinical diagnosis. Instead, the SPN features realized stable classification for E-no-AD and E-AD patients in different frequency bands. Considering the potential redundancy as well as the classification performance, we herein chose four pairs of spatial filters for SPN feature extraction. Further, we found that the SPN features in beta bands had the highest recognition rate with an accuracy of 89.90%.
To further explore the potential differences in functional networks between the patient groups, we checked for significant differences in the connectivity strength (p < 0.01) of functional networks in different frequency bands. The connectivity strength did not statistically significantly differ in most frequency bands (Fig. 3). However, in some particular frequency bands, abnormal connectivity strength was occasionally observed. To be specific, the connectivity strength between the frontal lobe and the parietal lobe was significantly decreased in E-AD patients when compared with E-no-AD patients in the beta-band functional network, which may have translated into the better classification performance of the beta-band SPN features. In addition, the brain regions recorded by the electrodes P4 and Fp1 were of great significance in distinguishing between the two patient groups. Conversely, similar statistical differences were noted along with the full frequency band as well; however, the trend overall was weakened by the influence of other frequency bands.
In the SPN filters, the nodes with stronger/weaker linkages for E-AD patients would be emphasized with larger absolute value (positive/negative). To identify the brain regions of great significance for differentiating between E-AD and E-no-AD patients, we extracted the first pair of SPN filters from the beta-band functional networks and mapped the weights of the SPN filters to the corresponding brain regions. As shown in Fig. 4, the first pair of SPN filters corresponded to the largest and smallest eigenvalues of the covariance matrix, and the corresponding SPN features are most discriminative for the two patient groups. Notably, the comparisons of both connectivity strengths and SPN filters (Figs. 3 and 4) potentially reflect differences in functional networks. Figure 3 shows statistical differences among all brain regions, while Fig. 4 emphasizes the discriminative properties of some crucial nodes extracted by the first pair of SPN filters. Furthermore, it should be noted that several non-statistical features underrepresented by statistical comparison may also have influenced the SPN filters. These factors indicated that even though both SPN filters and statistical comparisons revealed differences in the connectivity strengths between E-AD and E-no-AD patients, the two results were not completely identical.

Discussion
Previous studies have described the human brain as a largescale network having multiple brain regions that interact with one another to process complex and changing information. Therefore, when the brain region becomes the epileptic focus, other functions, such as epilepsy-related anxiety/ depression, are bound to be affected. In this study, we used resting scalp EEG signals and achieved stable classification of two types of epilepsy patients (E-no-AD patients and E-AD patients). By distinguishing among patients on the basis of the spatial features extracted by SPN filters, we differentiated between E-AD and E-no-AD patients with 89.90% accuracy. We further confirmed our conclusion in two ways: SPN features of functional networks could help distinguish between the two patient types with high accuracy, whereas the traditional network properties and principal component were not as effective in making this distinction.
Considering the possibility that the cognitive deficits incurred by the depression/anxiety disorder may be reflected in the brain regions related to these disorders, we attempted the use of network properties to assess and compare brain efficiency between E-AD and E-no-AD patients [52,53]. However, this classification achieved unsatisfactory accuracy in distinguishing the two patient types. Theoretically, the network properties are directly statistical measurements and cannot contain the complete information of network differences. The results shown in Fig. 2 also prove that the resting-state functional network properties were not significantly different. Conversely, to present the spatial topography of the abnormal components and extract abnormal components from EEG signals, PCA could be an effective method [54]. As shown in Table 1, the classification based on PCA components showed evidently better performance than that based on network properties. However, such differentiation did not meet the requirements for practical use.
In previous studies, the superiority of the spatial topology information in recognizing the abnormal networks has been established [55]. The analysis of spatial patterns is an effective way to extract the spatial features of networks. In theory, SPN filters can reflect critical nodes that exhibit significantly different connectivity strengths in the networks by endowing  Fig. 3 The significantly different (p < 0.01) connectivity strengths of the functional networks between E-AD and E-no-AD patients in different frequency bands. The red/blue lines representing the corresponding connection strength of E-AD patients is significantly stronger/weaker than for E-no-AD patients relatively larger weights. In the present study, appropriate SPN filters revealed a discriminative feature between E-AD and E-no-AD patients. As shown in Fig. 4, the brain regions exhibiting significant differences in the statistical comparison were emphasized with larger coefficients while the others were compressed with smaller coefficients. Accordingly, we achieved a reliable classification performance on the basis of such spatial features extracted by the spatial filters. We compared the classification performance of the SPN features in different frequency bands and found that the functional networks in the beta band could be crucial in distinguishing between E-AD and E-no-AD patients. It has been previously reported that the beta-band activity in resting-state is closely related to the development of neurological diseases [56,57]. Further, on the basis of the statistical comparison of the beta-band functional network shown in Fig. 3, we found the critical brain regions that may be closely related to the pathological differences between the two patient types. For instance, we found that the connectivity between frontal and parietal lobes was significantly weaker in E-AD patients than in E-no-AD patients, which could help distinguish between the two types of patients. In addition, a close relation between these brain regions and the depression/anxiety disorder has also been suggested [58].

Conclusion
Within the scope of this article, we propose that spatial features of functional networks can be applied to differentiate between E-no-AD and E-AD patients. Notably, our approach was based on the resting-state scalp EEG and did not require long-term clinical observation; this could greatly reduce the economic pressure of patients. However, this study has some limitations. First, the dataset used here is relatively small and needs further expansion for subsequent studies. Second, although the classification based on the SPN features of functional networks performed well in this study, its suitability to other neurological diseases needs to be determined. Besides, we herein did not implement several alternative methods to estimate the brain networks, such as Granger causality or transfer entropy for effective networks and multiple autoregressive models for functional network [59][60][61]. In future studies, a classification based on fused features of both effective and functional networks deserves an attempt for a higher accuracy rate as it may offer better discriminating abilities owing to the advantages of a high-dimensional feature space. Fig. 4 The scalp topologies for the first pairs of SPN filters extracted from beta-band functional networks. The brain regions with stronger/ weaker connectivity strengths for E-AD patients would be emphasized with a larger absolute value (positive/negative), which correspond to the deep red/blue areas