Multimodal Study of Resting-State Functional Connectivity Networks using EEG electrodes position as seed

Electroencephalography (EEG) is the standard diagnosis method for a wide variety of diseases such as epilepsy, sleep disorders, encephalopathies, and coma, among others. Resting-state functional magnetic resonance (rs-fMRI) is currently a technique used in research in both healthy individuals as well as patients. EEG and fMRI are procedures used to obtain direct and indirect measurements of brain neural activity: EEG measures the electrical activity of the brain using electrodes placed on the scalp, and fMRI detects the changes in blood oxygenation that occur in response to neural activity. EEG has a high temporal resolution and low spatial resolution, while fMRI has high spatial resolution and low temporal resolution. Thus, the combination of EEG with rs-fMRI using different methods could be very useful for research and clinical applications. In this article, we describe and show the results of a new methodology for processing rs-fMRI using seeds positioned according to the 10-10 EEG standard. We analyze the functional connectivity and adjacency matrices obtained using 65 seeds based on 10-10 EEG scheme and 21 seeds based on 10-20 EEG. Connectivity networks are created using each 10-20 EEG seeds and are analyzed by comparisons to the seven networks that have been found in recent studies. The proposed method captures high correlation between contralateral seeds, ipsilateral and contralateral occipital seeds, and some in the frontal lobe.


Introduction
Electroencephalography (EEG) is a method that measures the electrical activity of the brain. EEG uses surface electrodes to measure the electrical brain signals (Niedermeyer et al., 1993;Tatum et al., 2008).
EEG is routinely used to diagnose or monitor the following medical conditions and diseases: differentiate epileptic seizures, pre-operative assessment for defining an eventually resectable epileptogenic zone for a successful surgery, tumors, stroke, differential diagnosis of paroxysmal events, analysis of encephalopathies, monitoring consciousness, sedated patients at risk of seizures, patients in whom the only way to monitor brain function is EEG (Extracorporeal membrane oxygenation, for instance), prognosis cardiac arrest and hypoxic ischemic encephalopathy (HIE), brain death diagnosis, psychomotor regression study, acoustic or language development, suspected specific genetic phenotypes, and sleep disorders (Bickford, 1987). EEG is used for research purposes in the following topics: neuromarketing, psychology (processes underlying attention, learning and memory).
The International Federation of Clinical Neurophysiology (http://www.icfn.info) adopted the standardization for EEG electrode placement called 10-20 electrode placement protocol (Jasper, 1958;Klem et al., 1999). This protocol standardized the physical placements and designations of 21 electrodes on the scalp. Using reference points on the skull in the nasion, preauricular points and inion (Figure 1), the head is divided into proportional positions to provide adequate coverage of all the brain regions. The 10-20 designation indicates the fact that the electrodes are placed at 10%, 20%, 20%, 20%, 20%, and 10% of the total nasion-to-inion distance along the midline. The other series of electrodes are also placed at similar fractional distances from the corresponding reference distances. The name of each electrode consists of a letter and a number. The letter refers to the region of the brain where the electrode is positioned (F: frontal, C: central, T: temporal, P: posterior, and O: occipital), and the number is related to the cerebral hemisphere (even numbers in the right hemisphere, and odd numbers in the left; Figure 1). In 1985, an extension to the original 10-20 system was proposed involving an increase in the number of electrodes from 21 to 74 ( Figure 2) (American Electroencephalographic Society, 1994;Chatrian et al., 1988;Klem et al., 1999;Nuwer et al., 1998). 10-20 EEG electrode placement system is considered for clinical use, and 10-10 is more used for research. Functional connectivity MRI (fcMRI) measures the intrinsic functional correlations between brain regions (Mueller et al., 2013;Van Dijk et al., 2010). The method is sensitive to the coupling of both distributed as well as adjacent areas (Biswal et al., 1995;Yeo et al., 2011). It is believed that lowfrequency fluctuations observed in the BOLD signals reflect the spontaneous neural activity and that the synchronized fluctuations in distinct brain regions, therefore, point to functional connections between them. The first functional connectivity network found was related to the human motor cortex (Biswal et al., 1995). Later different functional connectivity networks have been found, and these networks change in patients with multiple pathologies (neurological, psychiatric). This renders fcMRI an interesting technique to further our understanding of brain function in health and disease. There are several methods for processing the BOLD signal to obtain connectivity networks: ICA (Independent Component Analysis) (Beckmann et al., 2005;Calhoun et al., 2006), seed-based method using correlation between BOLD time series (Biswal et al., 1995;Margulies et al., 2007;Van Dijk et al., 2010), ALFF (Amplitude of Low-Frequency Fluctuations) (Zang et al., 2007). Yeo et al. (Yeo et al., 2011), using resting-state data of 1000 healthy individuals, with 1175 ROIs on the cortex, the correlation between the fMRI time series of each ROI, and a clustering algorithm demonstrated the existence of seven main functional networks and a finer solution for 17 functional networks. The seven functional networks are visual, somatomotor, dorsal attention, ventral attention, limbic, frontoparietal, and default mode networks ( Figure 3) (Yeo et al., 2011). The networks found by Yeo et al. (Yeo et al., 2011) have shown to be valid across multiple subjects and robust to changes in the data processing. The existence of these networks provides a key motivation for our research.
One difference between fMRI and EEG is that fMRI has an excellent spatial resolution and a low temporal resolution (seconds), while EEG has a high temporal resolution (milliseconds) with spatial limitations. Simultaneous acquisition of fMRI and EEG could be a solution for the temporal limitations of fMRI and spatial limitations of EEG by combining their features (Ullsperger et al., 2010;Zang et al., 2007). Many authors show different acquisition techniques and data analysis for simultaneous EEG and fMRI (Han et al., 2014;Duyn, 2012;Lazeyras et al., 2001;Menon et al., 2005;Ullsperger et al., 2010). Other difference between both techniques are: EEG measures directly the electrical activity of the brain and fMRI indirectly by measuring changes in blood flow. EEG can be acquired simultaneously with fMRI (high-temporal-resolution data with high-spatial-resolution respectively) (Han et al., 2014;Duyn, 2012;Lazeyras et al., 2001;Menon et al., 2005;Ullsperger et al., 2010).
The fMRI acquired -either simultaneously or not-with an EEG of the same patient, could be processed in many different ways, for instance, ICA (data-driven) (Beckmann et al., 2005;Calhoun et al., 2006), ALFF (Zang et al., 2007), and correlation analysis (seed-based method) (Biswal et al., 1995;Margulies et al., 2007;Van Dijk et al., 2010). In the latter case, it is better to process rs-fMRI using seeds because it is possible to process it in a similar way to EEG (temporal signals in both cases in same places), therefore making it possible to compare the results of both techniques.
In this technical article, we show the functional connectivity networks obtained using seeds relative to the position of 10-20 and 10-10 EEG electrodes, and the relationship of these networks to the 7 and 17 functional connectivity networks (Yeo et al., 2011).

fMRI data processing and time series analysis Image Preprocessing
Data processing was performed using the Analysis of Functional NeuroImages software (AFNI; http://afni.nimh.nih.gov/afni; Cox, 1996;Cox, 1997) and fMRIB Software Library (FSL; http://fsl.fmrib.ox.ac.uk/fsl/fslwiki; Jenkinson et al., 2012;Smith et al., 2004). Image preprocessing involved the following steps: discarding the first 4 EPI volumes from each resting state scan to allow for signal equilibration; slice-time correction for interleaved acquisitions; 3-D motion correction with Fourier interpolation (volumes in which head motions caused a displacement of more than 2 mm in the x, y, or z-direction, or in which 2° of any angular motion was observed during the course of the scan, were excluded); despiking (detection and removal of extreme time series outliers); spatial smoothing using a 6mm FWHM Gaussian kernel; mean-based intensity normalization of all volumes by the same factor; temporal bandpass filtering (0.009 -0.1 Hz); and linear and quadratic detrending.
FSL FLIRT was used for linear registration of the high-resolution structural images to the MNI152 template (Jenkinson et al., 2001;Jenkinson et al., 2002). This transformation was then refined using FNIRT non-linear registration (Andersson et al., 2007a;Andersson et al., 2007b). Linear registration of each participant's functional time series to the high-resolution structural image was performed using FLIRT. This functional-to-anatomical co-registration was improved by intermediate registration to a low-resolution image and b0 unwarping.

Nuisance signal regression
To control the effects of motion and physiological processes (related to cardiac and respiratory fluctuations), we regressed each participant's 4-D pre-processed volume on nine predictors that modeled nuisance signals from white matter, cerebrospinal fluid, the global signal, and six motion parameters. Each participant's resultant 4-D residuals volume was spatially normalized by applying the previously computed transformation to MNI152 standard space, with 2mm 3 resolution.
3. Calculate the magnitude and direction of the vector connecting the center of mass of the brain and the position of each electrode ( Figure 4C).
4. The magnitude of the vector is modified until the electrode sphere (4 mm radius) is completely within the brain by keeping α , β , γ angles unmodified ( Figure 4D). In Supplementary Table 1, the coordinates, brain lobes, hemisphere, brain region, Brodmann area and EEG electrode name of each 10-10 EEG ROIs are included. In Supplementary Table 2, the same data is included, but for the ROIs equivalent to the 10-20 EEG electrode system. In Figure 5, the 21 spherical ROIs are shown (10-20 EEG electrode system), and in Figure 6, the 65 spherical ROIs are shown over a brain surface (10-10 EEG electrode system). Figures 5 and 6 were created using BrainNetViewer software (http://www.nitrc.org/projects/bnv; Xia et al., 2013).

Subject-level Resting State Functional Connectivity (RSFC) analysis
For each participant, the representative time series for each 10-10 EEG (seed) ROI was extracted from their 4D residuals volume in standard space by averaging the time-series across all voxels within the ROI. We then calculated the correlation between each seed ROI time series. The resultant participantlevel correlation maps were transformed via Fisher-z to Z-value maps and transformed into MNI152 2mm standard space for group-level analyses.
A similar analysis such as the previous one has been carried out using the representative time series for each 10-20 EEG (seed) ROI.

Group-level RSFC analysis
For each seed, group-level analyses were carried out using a random-effects ordinary least squares model. Whole-brain correction for multiple comparisons was performed using Gaussian Random Field Theory (min Z>2.3; cluster significance: p<0.05, corrected). This group-level analysis produced threshold z-score maps of activity associated with each 10-10 EEG and 10-20 EEG seed.

Functional Connectivity networks comparison
To compare the similarity of functional connectivity mapping obtained using each 10-20 EEG electrode seeds and relative to the seven Yeo networks (Yeo et al., 2011), we compute the Sørensen-Dice similarity coefficient (Dice, 1945;Sørensen, 1948). Sørensen-Dice coefficient (also known as Dice index, Sørensen index, or F1 score) is a statistical index commonly used for comparing the similarity of two different samples and their outputs are in the interval between 0 and 1. The Sørensen coefficient was computed using the equation:

Results
We computed the correlation matrix of each 10-20 EEG electrode seed for each healthy volunteer. The mean matrix of the 45 correlation matrices of each healthy volunteer is depicted in Figure 7A.
The adjacency matrix is a square matrix which is used as a way of representing binary relations, and it shows more clearly the pairs of seeds that have correlation coefficient values greater than a given threshold. Figure 7B shows the adjacency matrix created by applying a threshold of 0.2 to the correlation matrix ( Figure 7A). There is strong connectivity between frontopolar and some frontal 10-20 EEG seeds. In Figures 7A and 7B appear contralateral strong connectivity between frontal, central, temporal, parietal, and occipital seeds. Also, P3 and P4 have strong connectivity with frontopolar and some frontal seeds. Occipital seeds have high connectivity between them. Figure 8A shows the correlation matrix of the 10-10 EEG electrode system seeds, and Figure 8B shows the adjacency matrix with a threshold of 0.2 (mean value of positive values of the correlation matrix). These results are similar to Figure 6, but with higher resolution (due to the fact that 10-10 EEG has more seeds than 10-20 EEG). Frontopolar, anterior-frontal, frontal, parieto-occipital, and occipital seeds have ipsilateral and contralateral strong connectivity. Some frontocentral, central, centroparietal and parietal seeds have contralateral high connectivity ( Figures 8A and 8B).
We computed Sørensen-Dice coefficient. In Supplementary    : Similarity of 10-20 EEG seeds relative to Yeo networks. The similarity of functional connectivity mapping obtained using each 10-20 EEG seeds relative to seven Yeo networks (Yeo et al., 2011). The figure shows a bar chart for each EEG seed with Sørensen-Dice similarity coefficient (Sørensen, 1948;Dice, 1945)

Discussion
Epilepsy and other neurological diseases are diagnosed with EEG. In this paper, we show rs-fMRI results and analysis processed in a similar way as EEG.
In most clinical cases, the 10-20 EEG electrodes scheme is used adding some 10-10 EEG electrodes, and, therefore, functional connectivity was processed in a similar way as EEG, i.e. using 10-20 and 10-10 EEG related seeds.
The 10-10 correlation matrix ( Figure 8A), suggests that: There is a high connectivity between frontal, frontopolar, anterior-frontal, and frontocentral EEG seeds, between ipsilateral and contralateral seeds (adjacency matrix with a threshold of 0.2 in Figure 8B). There is a high contralateral connectivity between centroparietal, and parietal EEG seeds. There is high connectivity between the occipital EEG seeds (between ipsilateral and contralateral seeds; Figure 8B). The occipital EEG seeds (O1, O2, Oz) only have connectivity with PO7 ( Figure 8B).

Analysis for each functional connectivity network.
In eleven EEG seeds (FP1, FP2, FPz, F3, F4, F7, Fz, T3, T4, P3, P4), the Sørensen-Dice coefficient (Figure 9) of the Default Mode Network is greater than 0.2, and the left hemisphere EEG seeds have a higher Sørensen-Dice coefficient than those of the right side (left lateralized functional connectivity network). Our results are consistents with (Agcaoglu et al., 2015) that specify the same lateralization of the Default mode network for young people, and similar left lateralization conclusions reports (Nielsen et al., 2013) and (Swanson et al., 2011) for healthy people.
The Frontoparietal Network is right-lateralized because it has a higher percentage of functional connectivity voxels in even numbered EEG seeds than in odd numbered seeds. The Frontoparietal Network coexists with the default mode network (more than 35% of maximum value Sørensen-Dice index; 0.143 and 0.186 respectively) in multiple frontal seeds (Fp1, Fp2; BA 10, F3, F4; BA 9, FZ; BA 8), anterior temporals (F7, BA 47) and parietal seeds (P3, P4, BA 39). Limbic Functional Network is also left-lateralized because in general, odd numbered EEG seeds have higher Dice index than the even numbered ones. And because, the percent difference between left and right mean Sørensen-Dice coefficient is more than 20%.

Ventral Attention Network
The Somatomotor functional network has the highest Sørensen-Dice index in Cz, (BA 6), C3 (BA 3) and C4 (BA 3) EEG seeds (0.544, 0.447, and 0.410 respectively). In general, we could conclude that it is an equally lateralized network (less than 8% difference between both hemispheres of Sørensen-Dice coefficient).

General Analysis
The fact that more than one network coexists in a single seed (see bar charts in Figure 9) can be explained because the brain is and works as a network of interconnections (or graph-connected) where the existing networks are not independent and must be connected to function as an integrated organ. If we assume that networks are independent, different functional areas such as memory, visual or, motor could not be inter-related, contradicting the fact that the brain works as a network of interconnections.
Seeds located according to 10-20 EEG system in frontopolar, frontal, anterior temporal and parietal regions show high regional connectivity in relation to the default mode and frontoparietal networks. The default mode network predominates in the left hemisphere, probably this being related to a higher representativity of this functional network to the left, whereas in the right hemisphere the frontoparietal networks predominate. The frontoparietal network, located between the default mode network and the dorsal attention network, has a role in goal-directed cognition (Spreng et al., 2013;Vincent et al., 2008), and in the integration of information from the dorsal attention network and the default mode network (Vincent et al., 2008).
Also, frontal contralateral seeds have high connectivity ( Figures. 7, 8 and 10). Similar results were published in an ICA-derived EEG functional connectivity study that shows the power spectra of each independent component (Brodmann area 10, Figure 1C in Chen et al., 2013), and other EEG connectivity based paper that specify that contralateral frontal connections are common (61%, Figure 3 in Lacruz et al., 2007).
Central level seeds (C3, C4, and Cz) have high connectivity between each other with significant somatomotor representation (primary motor-sensory area; Figures 10 and 11), and moderate connectivity with temporal posterior region seeds (Figures 10 and 11) coinciding with the dorsal attention network (functional area related to processing of external sensory stimuli), ventral attention network, and not connected with the default mode network region .
Seeds in the occipital region (visual area) have high ipsilateral and contralateral connectivity slightly connected to some dorsal attention network regions (precuneus, cingulate gyrus, and hippocampus Figure 10). The primary sensory areas (visual, somatomotor) and the limbic area would be evolutionarily older than other parts of the prosencephalon, which is consistent with a modular-type organization, probably determined phylogenetically. Primary sensory areas participate in simple networks with local networks preferably. The greater the local connectivity is, the less functional variability exists, i.e. the occipital region is a stable connectivity area between the individuals (robust networks) and, therefore, would have no variability with respect to their connections. Instead, areas represented by seeds in associative area regions, mainly prefrontal, temporal and parietal, coinciding with frontoparietal and attention networks, have connectivity above the global mean of the inter-subject variability . Fahoum et al. (Fahoum et al., 2012) shows that in posterior quadrant epilepsy there exist only deactivation clusters in bilateral PCC and precuneus, not activations.  Associative areas have high long-range connections that are related to greater functional variability. Apparently, distant connectivity is necessary for prominent functional variability, and this is not observed in species with smaller brains. On the other hand, association areas are more recent in the evolution of the human brain development, whose greater functional variability could also be explained by its later maturation during development, which would make these more vulnerable to external postnatal influences and less dependent on genetic factors (Mueller et al., 2013).
Being able to establish normal connectivity models through combination (or fusion) of methods -in this case, electric signals and hemodynamic data-allows us to address the temporal and spatial relationship of the observed connections, and, in the future, to correlate it with connectivity in epilepsy patients (or other brain disease), in order to establish abnormal association networks so as to not only try to refine the detection of seizure origin area using non-invasive methods, but also to predict surgical prognosis, response to drug therapy, or cognitive impairment, inter allia.
The functional connectivity analysis of the rs-fMRI using 10-20 EEG positioned seeds as proposed in this article makes it possible to obtain connectivity maps superimposed with networks described by Yeo  and Brodmann areas (Brodmann, 1909), determining comparable connectivity matrices.
Functional connectivity analysis of rs-fMRI using 10-10 or 10-20 EEG positioned seeds helps the physician to understand the connectivity alterations detected by fMRI using classical parameters of EEG localization. EEG is currently the standard method for monitoring or diagnosis of some diseases (Niedermeyer et al., 1993). Also, using the proposed scheme it is possible to perform a correlation analysis between the anatomical layout of EEG surface electrodes and Yeo functional connectivity networks (Rojas et al., 2016;Yeo et al., 2011).
Finally, the objective of our study was to provide a replicable model in which the arrangement of seeds using 10-20 EEG system allows us to combine methods for temporal and spatial location through noninvasive markers, in order to identify abnormal functional networks.