Computers in Biology and Medicine 56 (2015) 158–166

Contents lists available at ScienceDirect

Computers in Biology and Medicine journal homepage: www.elsevier.com/locate/cbm

Scalp EEG brain functional connectivity networks in pediatric epilepsy Saman Sargolzaei a,n, Mercedes Cabrerizo a, Mohammed Goryawala a, Anas Salah Eddin b, Malek Adjouadi a a b

Department of Electrical and Computer Engineering, Florida International University, Miami, FL, USA Florida Polytechnic University, Lakeland, FL, USA

art ic l e i nf o

a b s t r a c t

Article history: Received 28 March 2014 Accepted 22 October 2014

This study establishes a new data-driven approach to brain functional connectivity networks using scalp EEG recordings for classifying pediatric subjects with epilepsy from pediatric controls. Graph theory is explored on the functional connectivity networks of individuals where three different sets of topological features were defined and extracted for a thorough assessment of the two groups. The rater’s opinion on the diagnosis could also be taken into consideration when deploying the general linear model (GLM) for feature selection in order to optimize classification. Results demonstrate the existence of statistically significant (p o0.05) changes in the functional connectivity of patients with epilepsy compared to those of control subjects. Furthermore, clustering results demonstrate the ability to discriminate pediatric epilepsy patients from control subjects with an initial accuracy of 87.5%, prior to initiating the feature selection process and without taking into consideration the clinical rater’s opinion. Otherwise, leaveone-out cross validation (LOOCV) showed a significant increase in the classification accuracy to 96.87% in epilepsy diagnosis. & 2014 Elsevier Ltd. All rights reserved.

Keywords: Epilepsy Functional connectivity Graph theory Pediatric Scalp EEG

1. Introduction The Center for Disease Control and Prevention (CDC) estimates that more than 2.3 million adults and half a million children in the United States are affected by Epilepsy [13], which is characterized by the occurrences of recurrent seizures with unknown causes. This number is projected to dramatically increase every year with 0.15 million newly diagnosed epilepsy cases [19]. Although the impact of seizures varies from person to person, physical and mental functions of the affected person could be severely altered. A systematic approach in diagnosing epilepsy could improve the planning of the treatment process. Scalp electroencephalography (EEG) has been widely perceived as an effective tool for non-invasive brain studies in patients with epilepsy. Scalp EEG has gained significant prominence for assessing brain function in patients with different neurological disorders, among them epilepsy. These include surveying the response of the brain under the influence of drug therapies [4,38,44], source localization of epileptic seizures which exploits techniques in the time/frequency domains for analysis of individual EEG electrode recordings [2,3], to name a few. Assessment of functional connectivity network (FCN) in patients suffering with various neurological disorders (e.g. Epilepsy in our study) through n Correspondence to: Center for Advanced Technology and Education (CATE), Department of Electrical and Computer Engineering, Florida International University, 10555 West Flagler Street, EC 2220, Miami, FL 33174, USA. Tel.: þ1 305 348 4106. E-mail address: ssarg004@fiu.edu (S. Sargolzaei).

http://dx.doi.org/10.1016/j.compbiomed.2014.10.018 0010-4825/& 2014 Elsevier Ltd. All rights reserved.

modalities such as EEG recording has elicited new findings in ways of underlying distinctions that delineate epileptic from control populations [4,10,21,32,36,43,44,46]. The high temporal resolution of EEG renders it an indispensable tool in the primary diagnosis of epilepsy and in the visualization of characteristic temporal events like interictal spikes which are closely associated with epileptic foci [10,31]. Additionally, EEG has been utilized to distinguish focal and generalized neurophysiologic correlates of epilepsy [7]. However, visual inspection and interpretation of continuous temporal EEG recordings is tedious, time consuming and prone to human error. Furthermore, epilepsy diagnosis based on visual inspection of EEG has been shown to be very subjective to the expert opinion [32]. This has led to the general cohort of adopting various automated techniques for diagnosing epilepsy. Artificial neural network (ANN) based classifiers have received the most attention towards epilepsy diagnosis using scalp EEG recordings [11,26,48–50,60]. The general routine of ANN based techniques is to process each isolated EEG signal with the aim of extracting a set of discriminating features as input to train an ANN in the design of an optimal classifier and predictor of the diagnosis. Although conventional techniques offer a promising success rate in characterizing and classifying population groups, they suffer from various drawbacks. Epilepsy being a complex disease with multifactorial causes makes the diagnostic process much more complicated than simply relying on a training model. The need of a training phase for extracting discriminating features forces the models to be constrained in their applications to specialized cases,

S. Sargolzaei et al. / Computers in Biology and Medicine 56 (2015) 158–166

159

Table 1 Demographic characteristics of study subjects.

PC (n¼ 7) PE (n¼9) pnn

Age

Female/Male

12.86 7 3.39n 9.22 74.26 NS

3/4 4/5 NS

Subject ID

Age

Gender

Diagnosis

Sampling rate (Hz)

PC01 PC02 PC03 PC04 PC05 PC06 PC07 PE01 PE02 PE03 PE04 PE05 PE06 PE07 PE08 PE09

12 15 12 15 10 18 8 10 7 4 14 8 7 15 4 14

M F M F M F M F F F M M M F M M

– – – – – – – Left temporal dysplasia Left frontal region Right fronto-centro-temporal Generalized Right parietal Left frontal pole, posterior frontal lobe Left and right frontal Right fronto-centro-temporal Generalized

200 512 200 512 512 512 512 200 512 512 512 200 512 500 512 512

Fisher’s exact test was used to test for gender. PC¼ pediatric control; PE ¼pediatric epilepsy. M: Male; F: Female. NS: not significant. n

Data presented as mean7 S.D. where applicable. Student t test (with statistical significance threshold of 0.05) was used to test for age and number of segments.

nn

limiting their potential for generalization as more cases are treated. Also, performance trade-off between classifier metrics like sensitivity and specificity has to be considered for choosing the optimal model parameter. Feature based techniques tend to oversimplify the fact that human brain includes a complex web of neuronal interconnectivity and discrete anatomical regions that function together to generate brain activity [36]. This underlying brain infrastructure suggests that solutions to the diagnosis of epilepsy need to consider the whole brain functional connectivity network (FCN). Thus, FCN seeks to define a pattern of cross-correlation between discrete functionally characterized brain regions to give statistical importance to anatomical connectivity and subsequently determining inter-regional neurophysiological inferences [21,22,45]. The connectivity model utilized in the study to construct FCN is based on the strength of correlation that exists among scalp EEG time series. Connectivity assessment based on Pearson correlation coefficient is well defined for evaluating the inter connectivity among cortical brain regions in the form of time series [14,28,53]. Introducing graph theory applications to the field provided a more systematic infrastructure to compare brain networks with different conditions; however, care is required to be given in selecting the right protocol to hypothesize the existence of a connection between brain regions and in applying graph theory based strategies to reduce bias [47]. Different applications of graph theory and small world networks [13], with causality analysis combined with network analysis [45,55] and frequency coupling detection among isolated scalp EEG recordings [22]and electrocorticographic (ECoG) [34] are a few of the widely used model-based solutions for studying functional connectivity networks of brain using scalp EEG recordings. Spectral analysis of brain connectivity in frequency domain supported the importance of association among brain cortical regions in predicting epilepsy [16,17]. However, the dynamic multi-causal characteristics of epilepsy and seizure activities requires the connectivity analysis to be performed in the time domain as well [35] to evaluate the transient changes in the pattern of connectivity among brain regions.

With the exponential increase in the number of connectivity studies in patients with epilepsy, different changes in the topology of the FCN constructed using different network measurements were reported, however few studies [23,58] examine the changes in FCN’s of pediatric populations. The other important reason to focus on this population was the finding in [12] which supports that regular treatment of epilepsy may or may not treat childhood epilepsy. Epilepsy diagnosis in pediatric population is critical as most focal epilepsy has their onset during childhood and adolescence. The study objective was then first placed on introducing a novel technique to construct functional connectivity networks in pediatric population. The technique was developed based on finding a geometrical relation among the time series of non-invasive scalp EEG. Graph theory based measurements of network topology were then applied in a systematic approach to develop data-driven solutions for pediatric epilepsy diagnosis. Moreover, the data-driven approach introduced here involves no training phase in its development to classify patients with epilepsy from controls.

2. Materials and methods 2.1. Data acquisition 16 Children (7 pediatric control (PC) patients and 9 pediatric epilepsy (PE) patients) were recruited for this study. Scalp EEG obtained with varying sampling frequencies of 200 Hz, 500 Hz and 512 Hz from control subjects (4 males and 3 females) and subjects with epilepsy (5 males and 4 females) were recorded using the 10–20 electrode placement system with a referential montage. The raw digital EEG data used one reference electrode located in the midline of the scalp based on the 10/20 system. The electrode recordings were then referenced to the average of all referential recordings. The resting state data was collected from routine EEG recordings from epileptic patients and controls, without the imposition of long recording sessions. The EEG recordings for the study were recorded

160

S. Sargolzaei et al. / Computers in Biology and Medicine 56 (2015) 158–166

Scalp EEG Recording of Channels 1- 9 for a given Subject

Artifact Attenuation Blind Source Separation artifact removal using Independent Component Analysis

Functional Connectivity Network (FCN) Construction Pairwise Distances (

)

Adjacency Matrix (

Network Graph (

Graph ( ld: adn:

deg:

)

) Representation of

) Analysis and Feature vector ( Link Density of the graph Average Degree per node

) Extraction

Degree of the graph

adnn: Average of ADN of neighboring nodes acc: Average of Closeness Centrality abc: Average of betweenness Centrality Score of all nodes awcc: Average of Node weighted Clustering Coefficient rcm: Rich Club Metric sm: S Metric ac: Algebraic Connectivity of a graph ge: Graph Energy AVE: Average of Vertex Eccentricity GR: Graph Radius

Is rater’s opinion known?

Yes GLM-based Sequential Feature Selection (GLM-SFS)

No K-means Clustering (K=2) Fig. 1. Flowchart of the suggested decision support system to differentiate pediatric with epilepsy from pediatric control.

from both patients with epilepsy and controls were performed for screening purposes and none of the patients with epilepsy were taking medication at the time of the recordings. For an unbiased comparison of files, the segments from the patients with epilepsy were extracted from sections that do not include the seizure event, which may or may not contain abnormal discharges. Recordings were performed at Miami Children’s Hospital, Miami, Florida, USA, using XLTEK Neuroworks Ver. 3.0.5 equipment. The file segments from patients with epilepsy were only interictal (i.e. without seizure activity). The lengths of EEG epochs free of artifacts were varied from 9 s to 90 s. The choice of 9 s as the minimum segment size for this study followed the recommendation of [56] in segmenting the EEG signals in the studies focused on epileptic seizures. Table 1 summarizes the demographic characteristics of study subjects. The study was approved by the Institutional Review Board and consent forms were provided to the patients or legal representatives. 2.2. Preprocessing Continuous recorded data were digitized using an internal 22 bit analog to digital (A/D) converter and pre-processed to attenuate the

effect of unwanted sources in the signal. Eye movement, muscle artifacts, eye blinks, electromyography (EMG) and electrocardiography (ECG) were labeled for artifact rejection purposes through implementation of blind source separation artifact removal [30] by Independent Component Analysis (ICA). Fig. 1 outlines the main steps involved in the design of the automated decision support system to cluster a group of subjects into two categories (controls vs. patients with epilepsy). 2.3. Functional connectivity network construction The primary aim of the “Functional Connectivity Network Construction” block depicted in Fig. 1 was to determine the pattern of functional communication among isolated channels to establish a brain state. Functional connectivity among brain regions is another layer of defining an association between cortical areas rather than the structural connectivity. Structural connectivity is based on inspecting the existence of direct medium among brain regions found by structural imaging. Functional connectivity as defined in this study is based on the association among the EEG recordings across the brain cortical regions.

S. Sargolzaei et al. / Computers in Biology and Medicine 56 (2015) 158–166

161

Table 2 List of topological features to extract from graph corresponding to FCN. Feature

Description

ld adn deg adnn acc abc awcc rcm sm ac ge

Link density of the graph Average degree per node Degree of the graph Average of ADN of neighboring nodes for every vertex Average of closeness centrality Average of betweenness centrality score of all nodes Average of node weighted clustering coefficient Rich club metric S metric Algebraic connectivity of a graph Graph energy

First and second moment statistics for feature vectors across group of subjects Feature Control group ld adn deg adnn acc abc gcc rcc smg acg eng

Epilepsy group

Student t-test

51.9 7 13.23 934.29 7 238.1 934.29 7 238.16 20.217 1.25 0.0017 0.0005 07 0 1.05 7 0.001 51.9 7 13.23

58.86 7 20.88 1059.5 7 375.8 1059 7 375.82 19.487 0.79 0.0027 0.005 07 0 1.05 70.002 59.337 19.48

3:53  108 7 1:5  108 741.647 234.29 1964 7 405.1

4:5  108 7 1:8  108 923.32 7 342.77 21427 751.4

p ¼ 0.45 p ¼ 0.46 p ¼ 0.45 p ¼ 0.18 p ¼ 0.5 p ¼ NaN p ¼ 0.52 p¼ p ¼ 0.18 p ¼ 0.25 p ¼ 0.5

Fig. 2. Correlation coefficient is calculated and color-coded from the matrix whose rows are study subjects and columns are the graph measures (Left) Matrix of pair-wise Correlation Coefficients among features of FCN Graph. (Right) Average correlation coefficient for each feature across other features. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

A variety of measures exist for defining the association among two time-series of neurophysiological signals [39]. Metrics such as time domain cross-correlation [35], Phase lag index [17], synchronization likelihood [16] have been widely applied. Measures based on cosine similarity which is a normalized inner product are also widely accepted and applied in natural language processing and in information retrieval of text documents [29] and multi-channel biological signals [57]. It is also shown that inner product based measurements e.g. cosine similarity is computationally efficient in its applications involving high dimensional sparse data [15,20,52]. The use of more sophisticated and non-linear measures which are not based on simple cross-correlation would be definitely an option; however it is shown that almost similar efficiency was achieved with both linear and non-linear measures of mutual association among multi-channel signals when analysis of ictal and interictal biological signals is performed in epilepsy research [5,37,42].

To construct the FCN, each electrode recording is assigned to a ! vector in discrete time domain denoted as xi ; i ¼ 1; 2; …; m where m is the number of electrodes. These vectors thus represent the EEG signals in subject space, and their quantities at each time step are the corresponding magnitude of the signal at that time dimension. The dimension varies across subjects depending on the length of the recording. A pairwise geometrical distance θij ; i; j ¼ 1; 2; …; m was defined as in Eq. (1) to serve as a metric in order to describe the extent by which the recordings of a pair of electrodes share mutual information in the current brain functional state.   xi :xj i; j ¼ 1; …; m ð1Þ θij ¼ π  cos  1 jjxi jjjjxj jj This equation is essentially a distance measure calculated as a cosine similarity measure. Distances were normalized into ½0; π =2 range to give a more sensible definition of vector co-variations in a

S. Sargolzaei et al. / Computers in Biology and Medicine 56 (2015) 158–166

Source Electrode

C3 C4 O1 O2 Cz F3 F4 F7 F8 Fz Fp1 Fp2 P3 P4 Pz T3 T4 T5 T6

C3 C4 O1 O2 Cz F3 F4 F7 F8 Fz Fp1 Fp2 P3 P4 Pz T3 T4 T5 T6

60

50

40

30

20

10

C3 C4 O1 O2 Cz F3 F4 F7 F8 FzFp1Fp2 P3 P4 Pz T3 T4 T5 T6

C3 C4 O1 O2 Cz F3 F4 F7 F8 FzFp1Fp2 P3 P4 Pz T3 T4 T5 T6

Target Electrode

Target Electrode

Source Electrode

Source Electrode

162

C3 C4 O1 O2 Cz F3 F4 F7 F8 Fz Fp1 Fp2 P3 P4 Pz T3 T4 T5 T6 C3 C4 O1 O2 Cz F3 F4 F7 F8 Fz Fp1 Fp2 P3 P4 Pz T3 T4 T5 T6 Target Electrode

Fig. 3. Visualization of constructed undirected functional connectivity networks (FCNs) and the corresponding plot of the connectivity distances for the average map across pediatric control group (left top) and pediatric epilepsy (right top). Bottom of the figure shows the results of student t test for the null hypothesis that assumes no statistically significant differences for the index pair of electrodes across the pediatric epilepsy and pediatric control groups. Rejection of the null hypothesis is highlighted with the black boxes when p o 0:00001, bonferroni adjusted for multiple comparison.

geometrical context. Distance of 0 rad corresponds to maximum correlation among two vectors while π =2 rad implies that the vectors are orthogonal and thus uncorrelated. Adjacency matrices are created by concatenating all the pairwise distances θij for each subject in a two-dimensional array, as shown in Eq. (2).   ASn ¼ θij ; i; j ¼ 1; …; m ð2Þ The adjacency matrix for subject n; n ¼ 1; 2; …; N is denoted by ASn . Adjacency matrices are thus one form for introducing the next level of automation in epilepsy classification based on graph analysis of FCN using scalp EEG measurements.

2.4. Graph analysis of FCN Pairwise relations between objects could be mathematically modeled and studied in terms of a graph. Nodes of the graph refer to the electrodes used to record EEG signals and their location in graph is determined from the coordinates in the 10–20 EEG electrode placement system. An edge is defined as the pairwise relation connecting a pair of nodes (electrodes in the context of the study). Pairwise angles which make up the adjacency matrix are termed as weights of graph edges for the study. The graph as defined here is an undirected form of FCN. A graph may not

include edges among pair of nodes when there is no communication between them. Graph theoretical analyses of multi-channel EEG have been evaluated for different schemes of complex brain networks affected by epilepsy [40,41,45,46,51], emphasizing that small world network model of the brain function could be altered by epilepsy. To automatically cluster the group of subjects into epileptic or controls using the network graph ðGSn ; n ¼ 1; 2; …; NÞ as defined from their adjacency matrices, a subset of topological features [8] were extracted from the network graph. Table 2 lists all the features extracted from the network graphs for each subject. Features given in Table 2 describe density-related, distance-based and spectral ƒ! features. A vector f Sn with its elements defined as the abbreviated features listed in Table 2 has been assigned to each subject in the following form: ƒ! f Sn ¼ ½ld adn deg adnn acc abc awcc rcm sm ac ge n ¼ 1; …; N

ð3Þ

The final step of the classification system based on graph features extracted from FCN as presented in Fig. 1 involves the use of the Kmeans clustering technique [27]. K-means clustering technique partitions the observation vectors ðf Sj ; j ¼ 1; 2; …; NÞ into K ¼ 2 clusters such that each observation belongs to the cluster with the closest mean. In mathematical terms, the algorithm tries to partition

S. Sargolzaei et al. / Computers in Biology and Medicine 56 (2015) 158–166

163

the N feature vectors (one vector per subject) into 2 classes C i ; i ¼ 1; 2 so as to iteratively minimize the within class sum of squared differences, over all classes defined as in Eq. (4) 2

argminC ∑

ƒ! ∑ jj f Sn  μi 2 jj

i ¼ 1 f Sn A C i

i ¼ 1; 2; n ¼ 1; …; N

PC

ð4Þ

ƒ! where f Sn refers to the feature vector assigned to subject Sn and μi is the centroid vector assigned to the class C i . The difference between ƒ! the feature vector f Sn and the centroid vector μi indicates the pointto-cluster centroid distance.

In this case, F SN is a 14  N matrix comprising the N vectors ƒ! f Sn ; n ¼ 1; …; N derived earlier in Eq. (3), and W are the weights to be estimated so as to minimize the error ε. The number of subjects involved in the study is represented by N.

3. Results and discussion Adjacency matrix ASn of a subject acts as the initial driver for the data-driven solution presented here for epilepsy diagnosis. Each element of the matrix is a signature of the extent by which two corresponding areas of the cortical brain, as represented by the two

PC

PE

PC

PE

Fig. 4. Graph representations of average of constructed FCNs over the pediatric control (PC) group and pediatric epilepsy (PE) group. Thickness of links (graph edges) shows the strength of connectivity among electrode pairs. 1

0.8

p value

The proposed algorithm as described earlier in Fig. 1 has the option to incorporate rater into the decision making process by adding a Sequential Feature Selection (SFS) block before the K-means clustering process. SFS fulfills two tasks: (1) it helps reduce the dimensionality of the observation vectors and therefore alleviating the computational burden of the algorithm; and when available (2) it integrate the rater’s opinion into the decision making process. The rule of thumb is that classification accuracy increases with increasing number of features involved, but eventually a larger number of features could decrease the accuracy due to small sample size relative to the number of extracted features [24,25]. Non-parametric classifiers may be more prone to presence of correlation among features, however correlated features requires more computation resources and add complexity to the training phase of pattern discovery. Fig. 2 displays the pairwise correlation coefficients between each pair of features extracted from the graph representing FCN of each subject. Correlation coefficient is calculated and color-coded from the matrix whose rows were study subjects and columns are graph measures. Analyzing correlation coefficients suggests the high dependencies among most of the features which could lead to the curse of dimensionality issue. To overcome this situation, SFS block based on General Linear Model (GLM) [1,9,33] subject to existence of clinical priori diagnosis was added to rank the features with their importance to the classification process. The GLM-SFS process consists of first assigning a binary numeral to each subject based on the rater’s opinion. The rater opinion is considered as the results of the screening process performed based on the clinical assessments. A binary numeral of ‘1’ signifies that the subject is a patient with epilepsy whereas a ‘0’ represents a subject from the control group. A N  1 logical ! vector R is thus generated for all N subjects to be used in a GLMbased Sequential Feature Selection (SFS). A solution is one that fits a GLM using all features as the model repressors to the response ! R as in Eq. (5). SFS then searches over all possible subsets of features and chooses the subset which gives the most accurate response in terms of mean square error. ! R ¼ F SN  W þ ε ð5Þ

Threshold of the connectivity strength

2.5. Sequential feature selection (SFS) based on general linear model (GLM)

PE

0.6

0.4

0.2

0

0

20

40

60

80

100

Threshold

Fig. 5. Statistical significance of the group difference for network graph measures as the function of threshold.

electrodes, are sharing common information in the current brain state. Adjacency matrix could be simply interpreted as a tool to investigate the neural interaction among different brain regions. The existence of operating functional sub-networks in the brain can be expressed by the inspection of a color-coded adjacency matrix. Furthermore a hierarchical clustering tree can be built on top of the matrix to assess the pattern of inter-connections among brain regions. A set of two samples t-test were performed to inspect the existence of statistical difference among each connection between source electrode and target electrode from the set of electrodes across the pediatric control group and pediatric epilepsy group. The results obtained are provided in Fig. 3. Existence of a statistically significant (p o 0:00001 Bonferroni adjusted for multiple

164

S. Sargolzaei et al. / Computers in Biology and Medicine 56 (2015) 158–166

Table 3 Clustering results with no prior knowledge on diagnosis. Condition Healthy

Epileptic

Clustered as epilepsy

1

8

Clustered as healthy

6

1

Sensitivity (%) 88.8

Specificity (%) 85.7

Positive predictive value (%) 88.8 Negative predictive value (%) 85.7 Accuracy (%) 87.5

Fig. 6. Constructed functional connectivity map (The threshold of 451 is applied as the connection strength) for (a) subject diagnosed with left frontal region epilepsy and (b) subject diagnosed with generalized epilepsy.

corrections) difference in the connection strength (degree) of a specific source-target pair of electrodes is shown as a black box unit, insignificant difference in the connection strength is shown as a white box unit. The regional homogeneity could be observed from the hierarchy since the electrodes with close proximity to each other in the EEG 10–20 montage are more likely to be grouped in one sub-network. Fig. 4 illustrates the FCN graph for average of patients with epilepsy subjects and the one corresponding to the average representative of the controls group. Different threshold values were applied and the results were shown for both groups. The thickness of the edges signifies the strength of commonalities in co-variation among the pair of electrodes. The main assumption for differentiating epileptic subjects from control subjects are the differences as can be visualized in their corresponding FCNs. A two-sided two-sample student t-test was conducted to compare FCN (pairwise distances) of control subjects and functional connectivity of epileptic subjects. There was a significant difference in the functional connectivity of the control subjects ðμC1 ¼ 51:91; SDC1 ¼ 9:41Þ from the functional connectivity of the epileptic subjects ðμC2 ¼ 58:861; SDC2 ¼ 4:931Þ; t ð340Þ ¼ 8:57; p o 0:01. Results show that a statistical difference is seen between the FCN (pairwise distances) between the patients and control groups. Table 2 provides the mean and standard deviation values of each extracted feature and their corresponding two-tailed p-values of differences between the means per feature over patients with epilepsy and controls, when no thresholding was applied on the FCN’s. However, to observe the network measure changes with

different threshold, a set of thresholds were applied on the FCN’s and the statistical significance of the groups’ difference of graph theoretical measures were computed and plotted in Fig. 5. The extracted feature vectors for the subjects were provided as input to the K-means clustering step to categorize subjects into one of the two groups. The classification technique developed using graph features from FCNs assumed no prior diagnostic knowledge rendering the technique completely data-driven. To count for the possible variation in the classification results due to the initialization of Kmeans, the algorithm was repeated over fifty times, however no variation was observed in the classification decision. Table 3 summarizes the results of the classification performed on the group of subjects with the accuracy, sensitivity and specificity to demonstrate the effectiveness of the technique. The confusion matrix provided in Table 3 demonstrates that the developed technique differentiates epileptic and healthy subjects with 87.5% accuracy considering no prior knowledge on the diagnosis. With the inclusion of the rater opinions on the diagnosis for each subject, SFS could find the subsets of features which outperform others in MSE sense. To evaluate the algorithm with the inclusion of the rater opinion, the Leave One Out Cross Validation (LOOCV) were done across all subjects. The accuracy in this case improved significantly to 96.87%. Introducing the methodology for constructing brain functional connectivity network in this study could be extended for its potential applicability in identifying different types of epilepsy. Examples of such unique connectivity patterns are associated to the clinical

S. Sargolzaei et al. / Computers in Biology and Medicine 56 (2015) 158–166

165

aspects of the different type of seizures as illustrated in Fig. 6. Initial assessment of the applicability of the algorithm in identifying different types of epilepsy showed the ability to cluster generalized epilepsy from focal epilepsy; however, a more thorough investigation is needed to confirm the association of key characteristics in the connectivity patterns in direct relation to the type of seizure, 3D source localization, and potentially the connectivity in the structural MRI for the different regions of the brain. The improvement gained by associating connectivity patterns to the conventional diagnostic schemes is the added power of clustering subjects into their corresponding patients with epilepsy or control groups in the absence of ictal events and with or without the presence of interictal spikes. However, it should be noted that statistically significant alterations in connectivity patterns constructed from resting state data of patients from control group do not necessarily inform on the type of neurological disorder but rather augment the prospects for a diagnosis of a specific disorder, in this case epilepsy. Our objective was rather to investigate how the pattern of functional connectivity could be associated with the epilepsy in the absence of ictal events. This would augment in the future our prospects for diagnosis by associating key characteristics of such connectivity patterns to algorithms that deal with 3D source localization or for delineating EEG recordings that could potentially lead to a seizure from control EEG that do not. It is of great interest for future work to study key changes that happen in connectivity patterns as a function of different neurological or psychiatric problems.

compare these connectivity maps in relation to specific frequency and temporal features [6,10] or associate them to fMRI data using the connectome, decision functions or PCA [18,54,59] for an optimal classification process. Through the application of developing scalp EEG based connectivity; these initial results did support the hypothesis that pediatric epilepsy disease impacts the connectivity patterns significantly relative to the control population. We do emphasize however that alterations in such connectivity patterns may or may not be associated with a given neurological or psychiatric disorder; however such alterations could be used to augment or validate the diagnosis as hypothesized through initial assessment by the physician or clinician as to the nature of the neurological disorder, in this case epilepsy.

4. Conclusion

References

A data-driven approach using Functional Connectivity Networks (FCNs) of scalp EEG recording in conjunction with graph theory feature analysis is introduced. FCNs are derived from the pairwise relations among scalp EEG electrodes using the 10–20 montage. The system uses topological features extracted from a graph corresponding to the FCN as predictive variables. The system ultimately categorizes the subjects into control and epileptic groups by defining the best model to predict the diagnosis. Results of hypothetical testing on the differences between the network graph of control and epilepsy subjects show that significant changes are observed in the brain FCNs for the patients with epilepsy. Implementation of the technique using the K-means clustering algorithm showed classification results with high accuracy reaching 87.5%. Sequential feature selection also improves the accuracy of the classifier with the inclusion of the rater’s opinion on the diagnosis to nearly 97%. Thus, through the use of this new FCN construction method as described here, the results demonstrate that through connectivity maps, the physiological manifestations of abnormal cortical excitability that underlie epilepsy could infer the occurrence of high-risk level (epileptic population) and low risk level (control population) leading to follow up procedures. The construct of these FCNs could be generalized to extend to the study of other common neurological disorders such as ADHD/ADD, Parkinson, and autism. It should be noted that the proposed method focuses on the extraction of parameters from EEG, which has been a common practice in epilepsy research with applications ranging from interictal spike detection, seizure detection, 3-D source localization, and prediction, among others. What this paper proposes is to use key parameters through connectivity maps in order to classify EEG files into epileptic and non-epileptic files. This constitutes the main contribution of the method proposed. Connectivity maps are studied here in context of their merit to delineate EEG data of patients with epilepsy form that of a control population. The contrast can also be drawn to see if such connectivity maps could enhance the prospects of a diagnosis with the ability to augment other methods seeking a similar goal. Such methods could

[1] M. Adjouadi, M. Cabrerizo, M. Ayala, K. Nunez, An approach to functional brain mapping using an inverse solution based on the principal component transform, Inverse Prob. Des. Optim. 2 (2) (2005) 124. [2] M. Adjouadi, M. Cabrerizo, M. Ayala, D. Sanchez, I. Yaylali, P. Jayakar, A. Barreto, Detection of interictal spikes and artifactual data through orthogonal transformations, J. Clin. Neurophysiol. 22 (1) (2005) 53–64. [3] M. Adjouadi, D. Sanchez, M. Cabrerizo, M. Ayala, P. Jayakar, I. Yaylali, A. Barreto, Interictal spike detection using the Walsh transform, IEEE Trans. Biomed. Eng. 51 (5) (2004) 868–872. [4] J.F. Alonso,M.A. Mañanas, S. Romero, J. Riba, M.J. Barbanoj, D. Hoyer, Connectivity analysis of EEG under drug therapy, in: Paper Presented at the Conference Proceedings:… Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Conference, 2006. [5] K. Ansari-Asl, L. Senhadji, J.-J. Bellanger, F. Wendling, Quantitative evaluation of linear and nonlinear methods characterizing interdependencies between brain signals, Phys. Rev. E: Stat. Nonlinear Soft Matter Phys. 74 (3) (2006) 031916. [6] M. Ayala, M. Cabrerizo, P. Jayakar, M. Adjouadi, Subdural EEG classification into seizure and nonseizure files using neural networks in the gamma frequency band, J. Clin. Neurophysiol. 28 (1) (2011) 20–29. [7] G.L. Barkley, C. Baumgartner, MEG and EEG in epilepsy, J. Clin. Neurophysiol. 20 (3) (2003) 163–178. [8] G. Bounova, O. de Weck, Overview of metrics and their correlation patterns for multiple-metric topology analysis on heterogeneous graph ensembles, Phys. Rev. E: Stat. Nonlinear Soft Matter Phys. 85 (1) (2012) 016117. [9] M. Cabrerizo, M. Adjouadi, M. Ayala, I. Yaylali, P. Jayakar, G. Rey, Integrated study of topographical functional maps based on an auditory comprehension paradigm using an eigensystem study and spectrum analysis, Brain Topogr. 17 (3) (2005) 151–163. [10] M. Cabrerizo, M. Ayala, M. Goryawala, P. Jayakar, M. Adjouadi, A new parametric feature descriptor for the classification of epileptic and control EEG records in pediatric population, Int. J. Neural Syst. 22 (2012) 02. [11] M. Cabrerizo, M. Ayala, P. Jayakar, M. Adjouadi, Classification and medical diagnosis of scalp EEG using artificial neural networks, Int. J. Innovative Comput. Inf. Control 7 (2011) 6905–6918. [12] J. Chen, V. Tsai, W.E. Parker, E. Aronica, M. Baybis, P.B. Crino, Detection of human papillomavirus in human focal cortical dysplasia type IIB, Ann. Neurol. 72 (6) (2012) 881–892. [13] C.f.D. Control, Prevention, Epilepsy in Adults and Access to Care—United States, 2010. MMWR. Morbidity and Mortality Weekly Report, 61(45), 909, 2012. [14] D. Cordes, V.M. Haughton, K. Arfanakis, G.J. Wendt, P.A. Turski, C.H. Moritz, M.E. Meyerand, Mapping functionally related regions of brain with functional connectivity MR imaging, Am J. Neuroradiol. 21 (9) (2000) 1636–1644. [15] I.S. Dhillon, P.K. Ravikumar, A. Tewari, Nearest Neighbor Based Greedy Coordinate Descent. Paper Presented at the Advances in Neural Information Processing Systems, 2011.

Conflict of interest statement The authors have no commercial, financial, or other relationship related to the subject of this paper that could constitute or suggest a conflict of interest.

Acknowledgments This work is supported by the National Science Foundation under grants CNS-0959985, CNS-1042341, HRD-0833093 and IIP-1230661.

166

S. Sargolzaei et al. / Computers in Biology and Medicine 56 (2015) 158–166

[16] L. Douw, M. De Groot, E. Van Dellen, J.J. Heimans, H.E. Ronner, C.J. Stam, J.C. Reijneveld, ‘Functional connectivity’is a sensitive predictor of epilepsy diagnosis after the first seizure, PLoS One 5 (5) (2010) e10839. [17] L. Douw, E. van Dellen, M. de Groot, J.J. Heimans, M. Klein, C.J. Stam, J.C. Reijneveld, Epilepsy is related to theta band brain connectivity and network topology in brain tumor patients, BMC Neurosci. 11 (1) (2010) 103. [18] A.S. Eddin, J. Wang, W. Wu, S. Sargolzaei, B. Bjornson, R.A. Jones, M. Adjouadi, The effects of pediatric epilepsy on a language connectome, Hum. Brain Mapp. (2014). [19] M.J. England, C.T. Liverman, A.M. Schultz, L.M. Strawbridge, Epilepsy across the spectrum: promoting health and understanding.: a summary of the Institute of Medicine report, Epilepsy Behav. 25 (2) (2012) 266–276. [20] J. Friedman, T. Hastie, R. Tibshirani, Regularization paths for generalized linear models via coordinate descent, J. Stat. Software 33 (1) (2010) 1. [21] K. Friston, C. Frith, P. Liddle, R. Frackowiak, Functional connectivity: the principal-component analysis of large (PET) data sets, J. Cereb. Blood Flow Metab. 13 (1993) (5–5). [22] K.J. Friston, Functional and effective connectivity in neuroimaging: a synthesis, Hum. Brain Mapp. 2 (1‐2) (1994) 56–78. [23] W. Gaillard, M. Berl, E. Moore, E. Ritzl, L. Rosenberger, S. Weinstein, S. Sato, Atypical language in lesional and nonlesional complex partial epilepsy, Neurology 69 (18) (2007) 1761–1771. [24] C. Gillies, N. Patel, J. Akervall, G. Wilson, Gene expression classification using binary rule majority voting genetic programming classifier, Int. J. Adv. Intell. Paradigms 4 (3) (2012) 241–255. [25] C.E. Gillies, X. Gao, N.V. Patel, M.-R. Siadat, G.D. Wilson, Improved feature selection by incorporating gene similarity into the LASSO, in: 2012 IEEE 12th International Conference on Paper Presented at the Data Mining Workshops (ICDMW), 2012. [26] L. Guo, D. Rivero, A. Pazos, Epileptic seizure detection using multiwavelet transform based approximate entropy and artificial neural networks, J. Neurosci. Methods 193 (1) (2010) 156–163. [27] J. Han, M. Kamber, Data Mining, Southeast Asia Edition: Concepts and Techniques, Morgan kaufmann, 2006. [28] C. Honey, O. Sporns, L. Cammoun, X. Gigandet, J.-P. Thiran, R. Meuli, P. Hagmann, Predicting human resting-state functional connectivity from structural connectivity, Proc. Natl. Acad. Sci. 106 (6) (2009) 2035–2040. [29] A. Huang, Similarity measures for text document clustering, in: Paper Presented at the Proceedings of the Sixth New Zealand Computer Science Research Student Conference (NZCSRSC2008), Christchurch, New Zealand, 2008. [30] T.-P. Jung, S. Makeig, C. Humphries, T.-W. Lee, M.J. Mckeown, V. Iragui, T.J. Sejnowski, Removing electroencephalographic artifacts by blind source separation, Psychophysiology 37 (02) (2000) 163–178. [31] S.-H. Kim, C. Faloutsos, H.-J. Yang, Coercively adjusted auto regression model for forecasting in epilepsy EEG, Comput. Math. Methods Med. (2013) 2013. [32] B.F. Kirmani, Importance of Video-EEG Monitoring in the Diagnosis of Epilepsy in a Psychiatric Patient. Case Reports in Neurological Medicine, 2013, 2013. [33] R. Kohavi, G.H. John, Wrappers for feature subset selection, Artif. Intell. 97 (1) (1997) 273–324. [34] M.A. Kramer, E.D. Kolaczyk, H.E. Kirsch, Emergent network topology at seizure onset in humans, Epilepsy Res. 79 (2) (2008) 173–186. [35] U. Lee, S. Kim, K.-Y. Jung, Classification of epilepsy types through global network analysis of scalp electroencephalograms, Phys. Rev. E: Stat. Nonlinear Soft Matter Phys. 73 (4) (2006) 041920. [36] M. Lowe, B. Mock, J. Sorenson, Functional connectivity in single and multislice echoplanar imaging using resting-state fluctuations, NeuroImage 7 (2) (1998) 119–132. [37] F. Mormann, T. Kreuz, C. Rieke, R.G. Andrzejak, A. Kraskov, P. David, K. Lehnertz, On the predictability of epileptic seizures, Clin. Neurophysiol. 116 (3) (2005) 569–587. [38] R. Oostenveld, P. Fries, E. Maris, J.-M. Schoffelen, FieldTrip: open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data, Comput. Intell. Neurosci. 2011 (2011) 1.

[39] E. Pereda, R.Q. Quiroga, J. Bhattacharya, Nonlinear multivariate analysis of neurophysiological signals, Prog. Neurobiol. 77 (1) (2005) 1–37. [40] S. Ponten, F. Bartolomei, C. Stam, Small-world networks and epilepsy: graph theoretical analysis of intracerebrally recorded mesial temporal lobe seizures, Clin. Neurophysiol. 118 (4) (2007) 918–927. [41] S. Ponten, L. Douw, F. Bartolomei, J. Reijneveld, C. Stam, Indications for network regularization during absence seizures: weighted and unweighted graph theoretical analyses, Exp. Neurol. 217 (1) (2009) 197–204. [42] M.A. Quraan, C. McCormick, M. Cohn, T.A. Valiante, M.P. McAndrews, Altered resting state brain dynamics in temporal lobe epilepsy can be observed in spectral power, functional connectivity and graph theory metrics, PLoS One 8 (7) (2013) e68609. [43] S. Sargolzaei, A.S. Eddin, M. Cabrerizo, M. Adjouadi, Resting state functional connectivity based on principal component transformation of cortical fMRI measurements, in: 2013 Sixth International IEEE/EMBS Conference on Paper presented at the Neural Engineering (NER), 2013. [44] J.M. Schoffelen, J. Gross, Source connectivity analysis with MEG and EEG, Hum. Brain Mapp. 30 (6) (2009) 1857–1865. [45] A.K. Seth, A MATLAB toolbox for Granger causal connectivity analysis, J. Neurosci. Methods 186 (2) (2010) 262–273. [46] D.J. Smit, C.J. Stam, D. Posthuma, D.I. Boomsma, E.J. De Geus, Heritability of “small‐world” networks in the brain: a graph theoretical analysis of resting‐state EEG functional connectivity, Hum. Brain Mapp. 29 (12) (2008) 1368–1378. [47] C. Stam, P. Tewarie, E. Van Dellen, E. van Straaten, A. Hillebrand, P. Van Mieghem, The trees and the forest: characterization of complex brain networks with minimum spanning trees, Int. J. Psychophysiol. 92 (3) (2014) 129–138. [48] A. Subasi, A. Alkan, E. Koklukaya, M.K. Kiymik, Wavelet neural network classification of EEG signals by using AR model with MLE preprocessing, Neural Net 18 (7) (2005) 985–997. [49] A. Subasi, E. Erçelebi, Classification of EEG signals using neural network and logistic regression, Comput. Methods Prog. Biomed. 78 (2) (2005) 87–99. [50] R. Thatcher, P. Krause, M. Hrybyk, Cortico-cortical associations and EEG coherence: a two-compartmental model, Electroencephalogr. Clin. Neurophysiol. 64 (2) (1986) 123–143. [51] E. Van Dellen, L. Douw, J.C. Baayen, J.J. Heimans, S.C. Ponten, W.P. Vandertop, J.C. Reijneveld, Long-term effects of temporal lobe epilepsy on local neural networks: a graph theoretical analysis of corticography recordings, PLoS One 4 (11) (2009) e8081. [52] B. Van Durme, A. Lall, Online generation of locality sensitive hash signatures, in: Paper Presented at the Proceedings of the ACL 2010 Conference Short Papers, 2010. [53] J. Vincent, G. Patel, M. Fox, A. Snyder, J. Baker, D. Van Essen, M. Raichle, Intrinsic functional architecture in the anaesthetized monkey brain, Nature 447 (7140) (2007) 83–86. [54] Z. Wang, Z. Zhang, Q. Jiao, W. Liao, G. Chen, K. Sun, Y. Liu, Impairments of thalamic nuclei in idiopathic generalized epilepsy revealed by a study combining morphological and functional connectivity MRI, PLoS One 7 (7) (2012) e39701. [55] C. Wilke, G. Worrell, B. He, Graph analysis of epileptogenic networks in human partial epilepsy, Epilepsia 52 (1) (2011) 84–93. [56] L. Wu, J. Gotman, Segmentation and classification of EEG during epileptic seizures, Electroencephalogr. Clin. Neurophysiol. 106 (4) (1998) 344–356. [57] J.-W. Xu, H. Bakardjian, A. Cichocki, J.C. Principe, A new nonlinear similarity measure for multichannel biological signals, in: International Joint Conference on Paper Presented at the Neural Networks, 2007. IJCNN 2007, 2007. [58] X. You, M. Adjouadi, M.R. Guillen, M. Ayala, A. Barreto, N. Rishe, D. Morris, Sub‐ patterns of language network reorganization in pediatric localization related epilepsy: a multisite study, Hum. Brain Mapp. 32 (5) (2011) 784–799. [59] X. You, M. Adjouadi, J. Wang, M.R. Guillen, B. Bernal, J. Sullivan, W.D. Gaillard, A decisional space for fMRI pattern separation using the principal component analysis—a comparative study of language networks in pediatric epilepsy, Hum. Brain Mapp. 34 (9) (2013) 2330–2342. [60] Z. Zainuddin, L.K. Huong, O. Pauline, Reliable epileptic seizure detection using an improved wavelet neural network, Australas. Med. J. 6 (5) (2013) 308.