Computers in Biology and Medicine 41 (2011) 1110–1117

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/cbm

Review of advanced techniques for the estimation of brain connectivity

measured with EEG/MEG
V. Sakkalis n
Institute of Computer Science, Foundation for Research and Technology, Science & Technology Park of Crete, Vassilika Vouton, P.O. Box 1385, Heraklion 71110, Crete, Greece

a r t i c l e i n f o a b s t r a c t

Brain connectivity can be modeled and quantified with a large number of techniques. The main objective of
Keywords: this paper is to present the most modern and widely established mathematical methods for calculating
Human brain connectivity connectivity that is commonly applied to functional high resolution multichannel neurophysiological signals,
Functional connectivity including electroencephalographic (EEG) and magnetoencephalographic (MEG) signals. A historical timeline
Effective connectivity of each technique is outlined along with some illustrative applications. The most crucial underlying
Multivariate times series assumptions of the presented methodologies are discussed in order to help the reader understand where
Coherence each technique fits into the bigger picture of measuring brain connectivity. In this endeavor, linear, nonlinear,
Wavelet coherence
causality-assessing and information-based techniques are summarized in the framework of measuring
Nonlinear synchronization
functional and effective connectivity. Model based vs. data-driven techniques and bivariate vs. multivariate
Phase synchronization
Generalized synchronization methods are also discussed. Finally, certain important caveats (i.e. stationarity assumption) pertaining to the
Information based techniques applicability of the methods are also illustrated along with some examples of clinical applications.
Phase level value & 2011 Elsevier Ltd. All rights reserved.
Partial directed coherence
Alzheimer’s
Autism
Alcoholism
Schizophrenia

1. Introduction pathways tracking over extended regions of the brain, which are in
accordance with general anatomical knowledge [3]. Magnetic
There has been a growing interest in studying both normal and Resonance Imaging (MRI) and especially Diffusion Tensor Imaging
pathological brain function with respect to identifying variations in (DTI) can be used to examine structural connectivity and convey
activation within and interactions between brain areas. Under- information concerning the white matter fiber tracts. Techniques
standing and modeling brain function is based not only on the for measuring neuroanatomical connectivity are discussed in other
correct identification of the active brain regions, but also on the articles within this special issue.
functional interactions among the neural assemblies distributed Functional connectivity is defined as the temporal correlation (in
across different brain regions. The aforementioned concepts are terms of statistically significant dependence between distant brain
addressed in theoretical neuroscience, as the functional segregation regions) among the activity of different neural assemblies [4]. Many
(activation of specialized brain regions/neural assemblies) and neurophysiologic signals can be assessed with functional connectivity
integration (coordinated activation of very large numbers of neural techniques, including signals derived from single unit and local field
assemblies distributed across different cortical areas that constitute potential (LFP) recordings, Electroenchaphalography (EEG), Magne-
large-scale distributed systems of the cerebral cortex) principles [1]. toencephalography (MEG), Positron Emission Tomography (PET) and
Integration of cerebral areas can be measured by assessing Functional Magnetic Resonance Imaging (fMRI).
brain connectivity. Brain connectivity can be subdivided into Effective connectivity is a relatively new concept defined as the
neuroanatomical (or structural), functional and effective connectivity. direct or indirect influence that one neural system exerts over
Neuroanatomical connectivity is inherently difficult to define given another [5]. It describes the dynamic directional interactions among
the fact that at the microscopic scale of neurons, new synaptic brain regions. Effective connectivity can be estimated from the signals
connections or elimination of existing ones are formed dynamically directly (i.e. data-driven) or can be based on a model specifying the
and are largely dependent on the function executed [2]. But for the causal links (i.e. model-based combination of both structural and
sake of simplicity structural connectivity may be considered as fiber functional connectivity).
Several different modalities can be used to assess brain
connectivity. fMRI is widely used mostly due to the large avail-
n
Tel.: þ302810391448; fax: þ 302810391428. ability of MRI scanners. fMRI provides a high spatial resolution
E-mail address: [email protected] (1–10 mm), while EEG/MEG has more limited spatial resolution

0010-4825/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compbiomed.2011.06.020
 V. Sakkalis / Computers in Biology and Medicine 41 (2011) 1110–1117 1111

(1–10 cm). On the other hand, fMRI has a limited temporal studies have developed novel approaches for determining the brain
precision (  1 s), primarily due to the limitations of the hemody- sources that underlie the spatial and temporal patterns of EEG and
namic response, while EEG/MEG has high temporal precision of MEG signals [8,9]. In 2003, dynamic causal models were introduced
the EEG and MEG techniques ( o1 ms). Because functional and for fMRI [10]. Later, this basic idea was extended to EEG and MEG
effective connectivity techniques are largely dependent on calcu- [11,12]. The key to Dynamic Causal Modeling (DCM) technique is
lating the correspondence of neural signals over time, techniques that the response of a dynamic system can be modeled by a
such as EEG and MEG, which have excellent temporal resolution, network of discrete but interacting neuronal sources described in
are optimal for calculating such connectivity. terms of neural-mass [13–15] or conductance-based models [16].
This review focuses on the most promising methodologies for
assessing functional and effective connectivity from EEG or MEG 2.1.2. Data-driven effective connectivity techniques
signals. The introductory section provides an overview of brain In contrast to model-based technique, data-driven techniques do
connectivity, whereas Section 2 provides a historical and meth- not assume any specific underlying model or prior knowledge
odological perspective of different families of functional and concerning spatial or temporal relationships. Granger-casuality (GC)
effective connectivity techniques. Section 3 discusses the merits is one of the prototypical data-driven effective connectivity techni-
and the limitations of these techniques. The underlying assump- ques. GC is based on the assumption that causes precede their effects
tions of each technique are also discussed along with some in time. If a signal can be predicted by the past information from a
illustrative clinical paradigms. Finally, the fourth section con- second signal better than the past information from its own signal
cludes this review and points out future research directions. then the second signal can be considered causal to the first signal. GC
is a time-domain approach, but in 1982 Geweke [20] applied this
concept in the frequency domain. Geweke’s work enabled the
2. Methods analysis of coupling between EEG frequency bands that have a
well-known biomedical significance. As GC developed, the concept
From the early 1960s [6], scientific research focusing on brain was generalized from bivariate to multivariate signals [21,22].
connectivity has been increasing. Throughout this time, develop- Recently the Directed Transfer Function (DTF) [23] and Partial Directed
ing methods to efficiently and accurately quantify brain connec- Coherence (PDC) [24] techniques were developed out the GC method.
tivity has been, and still remains, a challenging problem. In this DTF and PDC are equivalent when applied in bivariate cases, but in
section we provide an overview of the most widely used techni- the multivariate case PDC is able to detect not only direct but also
ques and portray some of the most representative measures in indirect pathways linking interacting brain regions. PDC is briefly
each of the following categories: described below.

 Effective connectivity (Section 2.1)

o Model-based (Section 2.1.1) & data-driven (Section 2.1.2) 2.1.2.1. Partial directed coherence (PDC). PDC is based on the
techniques concept of partial coherence [25], a technique that quantifies the
 Functional connectivity (Section 2.2) relationship between 2 out of n signals while avoiding volume
o Linear (Section 2.1.1), nonlinear methods (Section 2.2.2) conduction (the most critical issue of traditional coherence) by
and information-based techniques (Section 2.2.3) accounting for the influence of interactions from all other n  2
signals. PDC extends the concept for partial coherence by
In-depth technical details of each method are provided in measuring directional (i.e. causal) influences. PDC is formulated
relevant references. using MVAR models. Suppose that a set of n simultaneously
observed time series xðtÞ ¼ ½x1 ðtÞ,. . .,xn ðtÞT is adequately
represented by an autoregressive model of order p:
2.1. Effective connectivity
p
X
Neural assemblies synchronize and interact dynamically in local xðtÞ ¼ Ar xðtrÞ þ eðtÞ ð1Þ
r¼1
or distant regions in order to accomplish perceptual, motor or 2 3
cognitive functions [17]. Such functions reflect complex interactions a11 ðrÞ  a1n ðrÞ
that include anticipation of the stimulus, attention to the stimulus 6 ^ & ^ 7
where Ar ¼ 4 5 is the coefficient matrix at time
and preparation for its associated actions [18]. Such an interaction an1 ðrÞ  ann ðrÞ
process can be realized through bidirectional or unidirectional
lag r, and eðtÞ ¼ ½e1 ðtÞ,. . ., en ðtÞT is a multivariate Gaussian white
coupling. The former case resembles mutual synchronization, where
process having zero mean and covariance matrix R. The
both systems adjust their rhythms to each other, whereas the latter
autoregressive coefficients aij(r), i,j¼1,y,n represent the
case reflects causal interaction between the driver (initiating external
influence of xj(t–r) on xi(t). Non-zero coefficient values can be
force) and the response (the driven system). Dynamic Causal
considered as information flow from signal j to signal i. GC is a
Modeling (DCM) [10] and Granger-causality [19] belong to this family
time-domain approach. PDC provides a frequency-domain
of techniques. These techniques will be discussed within this section.
description of GC [25–27].
Let the matrix Aðf Þ ¼ IAðf Þ ¼ ½a1 ðf Þa2 ðf Þ    aL ðf Þ with ele-
2.1.1. Model-based effective connectivity techniques ments aij ðf Þ representing the difference between the n-dimen-
Neurobiologically evidence and plausible theories generated from sional identity matrix I and the matrix Aðf Þ . The elements aij ðf Þ of
this evidence can form theoretical models that describe how brain Aðf Þ form the Fourier transform of the elements aij ðrÞ of the
areas interact and influence each other. This idea is the basis for P
coefficient matrix Ar, i.e. aij ðf Þ ¼ pr ¼ 1 aij ðrÞeıð2p=pÞrf Furthermore,
model-based effective connectivity. Using this technique competing ai ðf Þ, i¼ 1,2,y,n denote the columns of Aðf Þ. Then the PDC from
neurobiological models and hypotheses can be evaluated. This channel j to channel i is given by
technique allows proposed causal interactions to be assessed. aij ðf Þ
Neurobiological data are considered mixtures of independent pij ðf Þ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi ð2Þ
brain sources that are spatially and temporally correlated within aH ðf Þaj ðf Þ
the context of the specific brain state being investigated. Although, where H denotes the transpose and complex conjugate operation.
this idea dates back to the mid-1980s [7], during the past decade, Thus, PDC ranks the relative strength of causal interaction with
1112 V. Sakkalis / Computers in Biology and Medicine 41 (2011) 1110–1117

respect to a given channel while fulfilling the following normal- within each coherence estimate affect the resolution of the
2 PL 2
ization properties: 0 r 9pij ðf Þ9 r1 and i ¼ 1 9pij ðf Þ9 ¼ 1, for all measure.
1rj rn. An alternative method for calculating coherence is the Wavelet
Coherence (WC) [32]. This approach requires a-priori information
about the coupling range in time and frequency, in order to
2.2. Functional connectivity
allocate the optimal time–frequency resolution.1 WC is a function
of both time and scale that can be mapped to specific frequency
2.2.1. Linear connectivity
bins, broadly referred as pseudo-frequencies. The mapping pro-
In the 1960s, linear brain connectivity began to be measured
cedure requires the calculation of the leading dominant frequency
using cross-correlation of pairs of EEG signals [6, 28]. Higher correla-
of the scaled wavelet basis function. WC is particularly suited to
tions indicate stronger functional relationships between the related
quantifying time varying coherence, since it uses a shorter
brain regions. In order to measure linear connectivity in the
window for higher frequencies and a longer one for lower
frequency domain, the use of Magnitude Squared Coherence (MSC)
frequencies, thus avoiding the constant size windows as in the
or coherence was introduced. Coherence allows the spatial correla-
STFT coherence case. Similarly to the coherogram, WC produces
tions between signals to be measured in different bands [29].
the so-called ‘‘scalogram’’, as depicted in Fig. 1.
Coherence is sensitive to both change in power and change in phase
An interesting enhancement to the calculation of WC is the
relationships. In other words, if either power or phase changes in one
definition of a probability distribution of the calculated coherence
of the signals, the coherence value is affected. If there is no variation
values that can be used to define the 95% confidence level. In
over time in the original relationship between the two signals, the
order to apply these ideas on real EEG signals one may set a
coherence value remains unity [30]. This means that coherence does
population specific background spectra (or control-task spectra)
not give direct information on the true relationship between the two
defined as the mean time-averaged wavelet power spectrum for
signals, but only on the stability of this relationship with respect to
each EEG channel and scale averaged over all subjects performing
power asymmetry and phase relationship. Correlation, on the other
a control task [33]. Having derived this threshold, it is possible to
hand, may be calculated over a single epoch or over several epochs
indicate significant regions of increased or decreased coherence
and it is sensitive to both phase and polarity, independent of
over the scalogram and form a single measure per scale that
amplitude. However, under normal physiological conditions, no
reflects the Significant Wavelet Coherence (SWC). Basically, we are
strong and abrupt power asymmetries would be expected to occur.
able to obtain the coherence values over those time- and
Thus, the influence of power on coherence should be negligible and
frequency band-localized regions where significant coherence is
results similar to those produced by correlation would be expected
indicated by taking the coherence averages over certain bands
for the coherence measures.
and significant time intervals (contours depicted as dashed lines
in Fig. 1). An interesting study that successfully utilizes this
2.2.1.1. Magnitude squared coherence (MSC). Cross-correlation and
approach in extracting the variability of neural interconnections
MSC are the most commonly used linear synchronization
in schizophrenia patients, as compared to healthy controls [34], is
methods and are defined as follows:
discussed later in the clinical application section.
Consider two simultaneously measured discrete time series xn
and yn, n ¼1,y,N. Then the cross-correlation function (Cxy) is 2.2.2. Nonlinear coupling techniques
defined as Nonlinear methods are not designed to outperform linear
methods but rather provide complementary information under
N t
1 X certain and rather strict assumptions. Nonlinear measures for
Cxy ðtÞ ¼ ððxn xÞ=sx Þððyn þ t yÞ=sy Þ ð3Þ
Nt n ¼ 1 measuring the dynamics of an EEG signal were developed based
on deterministic chaos [35]. Nonlinear neural time series analysis
where x and sx denote mean and variance, respectively, while t is
was motivated by the fact that many crucial neural processes have
the time lag. MSC or simply coherence is the cross spectral
nonlinear characteristics (e.g. the regulation of voltage-gated ion
density function Sxy, which is simply derived via the FFT of
channels corresponds to a steep nonlinear step-function relating
Eq. (3), normalized by their individual autospectral density
membrane potential to current flow). In the early 1980s, the
functions. However, due to finite size of neural data one is forced
concept of synchronization was introduced to measure neural
to estimate the true spectrum, known as periodogram, using
connectivity. Synchronization is based on interacting chaotic
smoothing techniques (e.g. Welch’s method [31]). Thus, MSC is
oscillators [36,37]. Synchronization may be understood as an
calculated as
adjustment of rhythms of oscillating objects due to their weak
2
9/Sxy ðf ÞS9 interaction [38]. In neuroscience studies, synchronization is mainly
gxy ðf Þ ¼ ð4Þ
represented by the concepts of the phase- and generalized-synchro-
9/Sxx ðf ÞS9/Syy ðf ÞS9
nization [39]. Phase-synchronization (PS) [40] is most commonly
where /US indicates window averaging. The estimated MSC for a seen in gamma frequency large-scale oscillations that enter into
given frequency f ranges between 0 (no coupling) and 1 (max- precise phase-locking over a limited period of time when the
imum linear interdependence). subject is engaged in cognitive tasks. PS is also considered an
One of the major assumptions when using coherence is important mechanism in certain diseases, such as the genesis of
stationarity of signal. But, if the Short Time Fourier Transform epileptic phenomena [41]. One representative method capable to
(STFT) is used instead of the classical Fast Fourier Transform obtain a statistical measure of the strength of PS in different areas
approach to calculate coherence, then the stationarity assumption of the brain is the Phase Locking Value (PLV) [39, 42].
can be relaxed and coherence may be calculated around a
number of time instants. This technique produces the so-called
‘‘coherogram’’, which forms a three dimensional matrix of 2.2.2.1. Phase synchronization—PLV. The PLV approach assumes
time and frequency vs. coherence. However, stationarity is that two dynamic systems may have their phases synchronized
still required within each time interval for which coherence
is calculated, meaning that in practice one should carefully decide 1
Time-frequency resolution is constrained by the uncertainty principle: the
on the optimal section length (window) over which each coher- wider the windows, the better the frequency resolution, at the expense of timing
ence estimate is measured. Window length and overlapping information, and vice versa.
 V. Sakkalis / Computers in Biology and Medicine 41 (2011) 1110–1117 1113

Fig. 1. The squared WC time–frequency transformed scalogram. The 5% significant regions over the time-scale transform are indicated by the contours (green dashed
outline). The outer elliptical region at the edges of the second graph indicates the cone of influence in which errors (edge effects) may be apparent due to the
transformation of a finite-length series EEG signal [48]. The relative phase relationship is also shown as arrows (with in-phase pointing right, and anti-phase pointing left).
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

even if their amplitudes are zero correlated [43]. PS is defined as delay vectors need to be constructed out of the time series using
the locking of the phases associated to each signal, such as the following procedure known as time-delay embedding [47]:
9fx ðtÞfy ðtÞ9 ¼ const ð5Þ xn ¼ ðxn ,. . .,xnðm1Þt Þ and yn ¼ ðyn ,. . .,ynðm1Þt Þ ð10Þ
where n ¼1,y,N, and m and t are the embedding dimension and
In order to estimate the instantaneous phase of a signal, time lag, respectively. Let rn,j and sn,j, j¼1,y,k, denote the time
Hilbert transform (HT) may be used to form the analytical signal indices of the k nearest neighbors of xn and yn, respectively.
H(t) as For each xn the mean squared Euclidean distance to its k
neighbors is defined as
~
HðtÞ ¼ xðtÞ þixðtÞ ð6Þ
~
where xðtÞ is the HT of x(t), defined as 1Xk
RðkÞ
n ðXÞ ¼ ðxn xrn,j Þ2 ð11Þ
Z 1 kj¼1
1 xðt 0 Þ 0
~ ¼ PV
xðtÞ dt ð7Þ
p 1 tt
0
and the Y-conditioned squared mean Euclidean distance RðkÞ n ðX9YÞ

where PV denotes the Cauchy principal value. The analytical is defined by replacing the nearest neighbors by the equal time
signal phase is defined as partners of the closest neighbors of yn. If the set of reconstructed
vectors (point cloud xn) has an average squared radiusRðXÞ ¼
~
xðtÞ P
fðtÞ ¼ arctan ð8Þ ð1=NÞ N ðN1Þ
n ¼ 1 Rn ðXÞ, then RðkÞ ðkÞ
n ðX9YÞ  Rn ðXÞ 5 RðXÞ when the
xðtÞ
systems are strongly correlated, while RðkÞ ðkÞ
n ðX9YÞ  RðXÞ bRn ðXÞ if
Therefore for two signals x(t), y(t) of equal time length with
instantaneous phases fx ðtÞ, fy ðtÞ the PLV bivariate metric is they are independent. Hence, an interdependence measure is
defined as defined as [46]:
 
 NX  1 X N
RðkÞ
n ðXÞ
 1 1 iðf ðjDtÞf ðjDtÞÞ  SðkÞ ðX9YÞ ¼ ð12Þ

PLV ¼  e X Y  ð9Þ N n ¼ 1 RðkÞ
 n ðX9YÞ
N j ¼ 0 

where Dt is the sampling period and N is the sample number of Since RðkÞ ðkÞ
n ðX9YÞ bRn ðXÞ by construction, S ranges between 0
each signal. PLV takes values within [0,1], where 1 indicates (indicating independence) and 1 (indicating maximum synchro-
perfect phase synchronization and 0 indicates lack of nization). Another normalized and more robust version of S is
synchronization. defined as [45]:

2.2.2.2. Generalized synchronization. After the successful 1 X

N
Rn ðXÞRðkÞ
n ðX9YÞ
NðkÞ ðX9YÞ ¼ ð13Þ
application of PS in EEG analysis, another type of Nn¼1 Rn ðXÞ
synchronization, namely the Generalized Synchronization (GS),
was developed. GS represents how well neighborhoods (i.e. For an in-depth mathematical reasoning and historical over-
recurrences) of one chaotic attractor maps onto the other. view of the aforementioned techniques the interested reader is
Attractor mapping is considered to be a robust way of assessing referred to [48], while a detailed comparison of the aforemen-
the extent of GS [39, 44–46], even if it is prone to stationarity tioned synchronization methods applied in epileptic data analysis
shortcomings. To form such attractors from the raw EEG data, is presented in [39].
1114 V. Sakkalis / Computers in Biology and Medicine 41 (2011) 1110–1117

2.2.3. Information-based techniques assumption is violated, a stationarity independent measure such as

Information-based techniques are sensitive to both linear and PLV can be used. In addition, a novel and promising technique
nonlinear statistical dependencies between two time series. capable of decomposing a multivariate time series into its stationary
The most representative method is the Cross Mutual Information and nonstationary part known as stationary subspace analysis can be
(CMI) that measures the mutual dependence between two signals utilized to overcome these implicit stationarity constraints [53].
by quantifying the amount of information gained about one signal
from measuring the other, as a function of delay between these two
3.3. Multivariate modeling considerations
signals. CMI has been used in method for diagnosing Alzheimer’s
disease and Schizophrenia [49,50], as discussed in Section 4.
There is a growing interest in extending interdependence
Another information-based method for assessing the depen-
analysis from bivariate to multivariate signals. This is important
dence between time series is based on the degree of predictability
since pairwise analysis is likely to find spurious correlations in cases
of each of the two time series as a function of the other [39]. More
where one driver drives two responses. In this case both responses
specifically, the Minimum Description Length (MDL) principle is
may have a common driver, even if the responses appear to be fully
based on the idea that the best model for representing a signal is
independent. Several of the techniques are multivariate, such as the
the one with the shortest possible code length. According to the
GC, DTF, PDC and the GS measures. However, all these methods
MDL principle, the savings in code length of one signal due to the
depend on the reliability of the fitted MVAR model and especially
knowledge on the other is a measure of dependence between
the model order and epoch length. If the order is too low, the model
the two processes [51].
misses the dynamic nature of the signal, whereas if it is too high
Information-based measures may also assess causality as
overfitting mainly emphasizes noise. A number of methods such as
already discussed in Section 2.1.2.
the Akaike Information Criterion (AIC) [54] can be used to deter-
mine the optimal model order.
3. Discussion
3.4. Linearity/nonlinearity assumptions
This section illustrates the different underlying assumptions
and limitations of each family of methods, in order to help the Chaotic systems appear to have noisy behavior, which is actually
reader decide upon the best candidate method for a particular ruled by deterministic laws. Although the nonlinear measures
research study. presented are capable of identifying nonlinear interdependences,
they are highly susceptible to noise and, in the case of GS, the
3.1. Model-based vs. data-driven techniques embedding parameters. Another crucial issue is the requirement of
rather long stationary epochs. However, even though neurons are
The different underlying assumptions of both model-based theoretically highly nonlinear devices, strong evidence of chaos has
and data-driven techniques need to be considered when selecting not been found in EEG data [55]. Hence, at the present time, there is
one of these methods for a specific problem. Model-based a wide consensus that the EEG signal is not chaotic, at least in low-
approaches (e.g. DCM) are based on well-defined biophysical dimensions. Another classical misconception is that nonlinear tools
models of neuronal dynamics. In this case one should choose may replace linear ones. The opposite seems to be more valid;
the best model (or set of interacting models) and predefine or linear measures are more robust and perform well even in non-
experiment with a large number of different parameters in order linear cases [39]. Nevertheless, nonlinear analysis should be used to
to test a preset hypothesis. The uncertainty in predefining these complement linear ones in order to capture and provide additional
parameters and the large number of possible combinations of information hidden in linear approaches [39].
parameters is the main drawback of model-based techniques.
Established methodologies may assist in determining the best
3.5. Source estimation imaging vs. surface electrode connectivity
possible model [52]. However, it is very possible that no single
model exists but rather multiple models may be equally appro-
Another great concern lies in the problem of acquiring mixed
priate for a given data set.
activity captured from more than one brain region, when using
Data-driven methods do not assume any specific underlying
surface sensors (electrodes) to capture the signals. This can cause
spatial or temporal relationship. Such methods can be used in
spurious connectivity patterns. There are a number of recently
assessing connectivity when no a-priori structural knowledge is
proposed source imaging techniques that can account for volume
available.
conduction effects and can be applied in some of the techniques
reviewed in this paper. Most prominently, linear decomposition
3.2. Stationarity considerations
techniques such as Principal Component Analysis (PCA) [56] and
Independent Component Analysis (ICA) [18,57] [58], which
Most of the methods presented assume stationarity. For a process
attempt to invert the mixing process, can be utilized prior to
to be stationary, the mean, variance and autocorrelation structure
further connectivity analysis.2 On the other hand, there is also the
cannot change over time. Generally, an EEG distribution is consid-
possibility to examine only the imaginary part of the cross-
ered as a multivariate Gaussian process even if the mean and
spectrum since only the real part is affected by instantaneous
covariance properties change from segment to segment. Therefore,
effects when using spectrum-based techniques [59]. In that way
strictly speaking, an EEG signal is quasi-stationary since it is
there is no need for signal decomposition.
stationary only within short intervals. During mental and physical
Finally, it should be noted that graph-theoretic concepts can
activities this assumption can easily be violated since the state of the
also be used to visualize and quantify brain network topologies
brain can change in alertness and wakefulness. In addition, care must
using the presented connectivity measures [60,61].
be taken when examining EEG signals from epilepsy patients as
Table 1 summarizes the characteristics of the most widely
transitions between pre-ictal and ictal states often occur in such
accepted methods discussed in this review.
cases. WC provides a balance between a data segment long enough
to provide good frequency resolution and short enough to satisfy the
condition of stationarity. In conditions in which the stationarity 2
ICA can be also used in neuroimaging to study connectivity directly.
 V. Sakkalis / Computers in Biology and Medicine 41 (2011) 1110–1117 1115

Table 1 monitor the effects of intravenous scopolamine injection [64] in

Comparison of representative methods for estimating brain connectivity. AD. In this latter study, interhemispheric and left intrahemi-
spheric coherences were found to significantly decrease in the
DCM MSC STFT WC PLV GS GC PDC
COH Geweke theta band frequency band. This suggests that MEG and func-
tional connectivity measures may provide a tool for monitoring
Linear X X X X neurological disorder progression associated with cholinergic
Nonlinear X X abnormalities. More recently GC, phase synchrony and nonlinear
Info-based X X
Model-based X
generalized synchronization based measures have also been
Data-driven X X X X X X X tested in AD [65]. GC was able to discriminate patients from
Causality assessing X X X age-matched control patients achieving 82.9% classification rate
Multivariate X X X (in a leave-one-out classification scheme). In addition, CMI when
Stationarity X X
applied in 15 AD patients and age-matched normal controls was
independent
Functional X X X X X X X successful in identifying EEG abnormalities in AD patients with
connectivity functional impairment of information transmission in long
Effective connectivity X X X cortico–cortical connections [49].
Schizophrenia has been another very promising application
DCM: Dynamic Causal Modeling;
MSC: Magnitude Squared Coherence;
domain because connectivity analysis is able to test the discon-
STFT COH: Short Time Fourier Coherence; nection hypothesis of schizophrenia [66]. CMI was used to
WC: Wavelet Coherence; evaluate the information transmission of different cortical areas
PLV: Phase Locking Value; in 10 schizophrenic patients and age-matched controls [50].
GS: Generalized Synchronization;
Interhemispheric and intrahemispheric CMI values in schizophre-
GC: Granger Causality;
PDC: Partial Directed Coherence. nics were significantly higher than normal controls suggesting left
temporal lobe deficit and inter- and/ or intrahemispheric over-
3.6. Clinical applications connectivity in schizophrenics. PLV was also able to reflect
perceptual binding deficits especially in the 40 Hz frequency
The continuous advancement of neuroscience methods range [67,68]. More recently, working memory experiments (from
applied in EEG/MEG has been successful in capturing the under- 20 stable patients with schizophrenia and controls) based on WC
lying processes of several neurological disorders and neurode- assessment were able to successfully study and support the
generative diseases [62]. Studies incorporating the various ‘‘disconnection syndrome’’ hypothesis when examining the
connectivity methods discussed in this paper are presented in gamma frequency band [34]. In a similar working memory
this section. As will become evident below, there is no single framework, using mutual information applied in MEGs from 28
optimum method for assessing brain connectivity. Efficiency people with schizophrenia, the importance of beta-band oscilla-
greatly depends on the application and the underlying assump- tions for long-distance functional connections in brain networks
tions of each connectivity method. was highlighted [69].
Epilepsy is a common neurological disorder that has been Similarly to schizophrenia, an underconnectivity hypothesis also
extensively studied using both EEG and MEG. One clinically useful applies in autism [70]. A very recent study [71] achieved an overall
application is localization of the epileptogenic brain activity to performance of 87.5% accuracy of discriminating a group of 8 autis-
better define a surgically lesion [63]. In childhood epilepsy, visual tic individuals vs. healthy controls, based on GC connectivity
inspection of EEG tracings or traditional spectral analysis may not measures. Another work suggests that the usage of nonlinear
show differences between children with a history of seizures and methods and specifically the coarse-grained entropy synchroniza-
normal controls. Yet, more sensitive connectivity methods, which tion applied in sleep EEG may enable differences in children and
are able to identify subtle abnormalities, may be useful in the infant brain connectivity to be detected. More specifically it was
evaluation of neurophysiological activity and guiding clinical man- found that synchronization was significantly lower in children with
agement [39]. Different measures of quantifying synchronous autism than in a group of typically developing children [72],
oscillatory activity (MSC, MDL, PLV and GS) were evaluated in supporting the theory that the autistic brain exhibits low functional
[39] using a three-stage assessment framework. Initially, the non- connectivity. More recently, the nonlinear complexity of resting
linear methods were validated on coupled nonlinear oscillators EEG computed with modified multiscale entropy was able to
(artificial signals) under increasing noise interference; second, distinguish typically developing children from a group of infants
surrogate data testing was performed to assess the possibility of at high risk for autism spectrum disorder [73]. The study involved
nonlinear channel interdependencies in the acquired EEG data; and 79 different infants (46 at a risk for ASD and 33 controls) and
finally, synchronization on the actual data was measured. This reached almost 100% classification accuracy for boys at age
approach concluded that in a real case scenario, one should use 9 months and 70–90% at ages 12 and 18 months.
both a PS measure (e.g. PLV) and a GS measure, as well as linear As a final example of an application, alcoholism is presented.
connectivity methods since their underlying assumptions are Impaired cognitive functioning and specifically intrahemispheric,
different. For example, the PLV method performed better when posterior coherences are found to be significantly increased in the
applied to phase-synchronized oscillators but underperformed alpha and beta frequency bands both in long-term abstinent and
when examining general synchronized oscillators. non-abstinent alcohol-dependent subjects [74]. Alcoholics experi-
In Alzheimer’s disease (AD) many studies find that reduced ence cognitive deficits while performing complex cognitive tasks as
brain signal synchrony can facilitate early diagnosis. CMI was expressed also in generalized synchronization studies [75]. Lastly,
studied in [49] in order to quantify information transmission both linear and nonlinear interdependence measures (MSC, PLV and
between different cortical areas in 15 AD patients. Information the GS presented method) were also investigated in alcoholics
transmission was found to be lower between distant electrodes in during mental rehearsal of pictures [76]. The results were in
the right hemisphere and between interhemispheric electrodes. accordance with previous psychophysiology studies suggesting that
This suggests a functional impairment of information transmis- an alcoholic has impaired synchronization of brain activity and loss
sion in long cortico–cortical connections in AD patients, consis- of lateralization during the rehearsal process, most prominently in
tent with previous research. MEG coherence can also be used to alpha and lower beta frequency bands.
1116 V. Sakkalis / Computers in Biology and Medicine 41 (2011) 1110–1117

4. Conclusion [16] C. Morris, H. Lecar, Voltage oscillations in the barnacle giant muscle fiber,
Biophys. J. 35 (1) (1981) 193–213.
[17] A. Delorme, et al., From single-trial EEG to brain area dynamics, Neurocom-
A variety of advanced brain connectivity methodologies are puting (2002) 1057–1064.
reviewed in this manuscript. Although the majority of these [18] M. Zervakis, et al., Intertrial coherence and causal interaction among
techniques are currently research-based many may be clinically independent EEG components, J. Neurosc. Meth. 197 (2) (2011) 302–314.
[19] C.W.J. Granger, Investigating causal relations by econometric models and
useful in the near future for evaluating cortical dysfunctions in cross-spectral methods, Econometrica (1969) 424–438.
cases where classical EEG evaluation is inadequate. The use of [20] J. Geweke, Measurement of linear dependence and feedback between multi-
model-based/data-driven, bivariate/multivariate, causality-asses- ple time series, J. Am. Stat. Assoc. (1982) 304–313.
[21] J.F. Geweke, Measures of conditional linear dependence and feedback
sing, linear/ nonlinear and information-based techniques allows between time series, J. Am. Stat. Assoc. (1984) 907–915.
the analysis of complex cortical interactions from different, novel [22] Y. Hosoya, Elimination of third-series effect and defining partial measures of
perspectives. However, the accuracy of the results highly depends causality, J. Time Ser. Anal. (2001) 537–554.
[23] M.J. Kaminski, K.J. Blinowska, A new method of the description of the
on the underlying assumptions of each approach, as well as the
information flow in the brain structures, Biol. Cybern. 65 (3) (1991) 203–210.
application under consideration. Although analysis of brain func- [24] K. Sameshima, L.A. Baccala, Using partial directed coherence to describe
tional has evolved significantly during the last decades and a neuronal ensemble interactions, J. Neurosci. Meth. 94 (1) (1999) 93–103.
variety of methods addressing both functional and effective [25] L.A. Baccala, K. Sameshima, Partial directed coherence: a new concept in
neural structure determination, Biol. Cybern. 84 (6) (2001) 463–474.
connectivity are currently available, there is no single optimum [26] S. Wehling, et al., Assessment of connectivity patterns from multivariate time
technique to universally assess brain connectivity. series by partial directed coherence, Chaos Complexity Lett. (2007) 413–433.
In the years to come, more sophisticated and powerful meth- [27] L. Baccala, K. Sameshima, D.Y. Takahashi, Generalized partial directed
coherence, in: Proceedings of the 15th International Conference on Digital
ods will be developed which empower our current understanding Signal Processing, Cardiff, 2007, pp. 16–166.
of functional brain connectivity. Future methods for assessing [28] M. Brazier, J. Casby, Cross-correlation and autocorrelation studies of electro-
cortical connectivity patterns with greater spatiotemporal accu- encephalographic potentials, Electroencephalogr. Clin. Neurophysiol. 4 (2)
(1952) 201–211.
racy include multimodal fusion approaches integrating modalities [29] G. Pfurtscheller, C. Andrew, Event-related changes of band power and
that provide excellent temporal resolution (e.g. EEG and MEG) coherence: methodology and interpretation, J. Clin. Neurophysiol. 16 (6)
with modalities that offer better spatial resolution (e.g. PET and (1999) 512–519.
[30] J.S. Bendat, A.G. Piersol, 2nd ed.,Engineering Applications of Correlation and
fMRI). Further development of neuroconnectivity methodologies Spectral Analysis, XIV, J. Wiley, New York, 1993 458p.
would include combining both neuroanatomical information [31] M. Ding, et al., Short-window spectral analysis of cortical event-related potentials
derived from diffusion tensor imaging and high temporal resolu- by adaptive multivariate autoregressive modeling: data preprocessing, model
validation, and variability assessment, Biol. Cybern. 83 (1) (2000) 35–45.
tion functional connectivity approaches. Such methodologies will
[32] J. Lachaux, et al., Estimating the time-course of coherence between single-
be suitable for capturing the dynamic evolution of the time- trial brain signals: an introduction to wavelet coherence, Neurophysiol. Clin.
varying connectivity patterns that reflect certain cognitive tasks 32 (3) (2002) 157–174.
or brain pathologies. [33] V. Sakkalis, M. Zervakis, S. Micheloyannis, Significant EEG features involved
in mathematical reasoning: evidence from wavelet analysis, Brain Topogr. 19
(1–2) (2006) 53–60.
[34] V. Sakkalis, et al., Time-significant wavelet coherence for the evaluation of
schizophrenic brain activity using a graph theory approach, Conf. Proc. IEEE
Conflict of interest statement Eng. Med. Biol. Soc. 1 (2006) 4265–4268.
[35] E.N. Lorenz, Deterministic nonperiodic flow, J. Atmos. Sci. (1963) 130–148.
None declared. [36] L. Pecora, T. Carroll, Synchronization in chaotic systems, Phys. Rev. Lett. 64
(8) (1990) 821–824.
[37] A. Pikovsky, On the interaction of strange attractors, Z. Phys. B: Condens.
Matter. (1984) 149–154.
References [38] A. Pikovsky, M. Rosenblum, J. Kurths, Synchronization: A Universal Concept
in Nonlinear Sciences (Cambridge Nonlinear Science Series), Cambridge
[1] K.J. Friston, Modalities, modes, and models in functional neuroimaging, University Press, 2003 432.
Science 326 (5951) (2009) 399–403. [39] V. Sakkalis, et al., Assessment of linear and nonlinear synchronization
[2] A. van Ooyen, Competition in the development of nerve connections: a measures for analyzing EEG in a mild epileptic paradigm, IEEE Trans. Inf.
review of models, Network 12 (1) (2001) R1–47. Technol. Biomed. 13 (4) (2009) 433–441.
[3] M.A. Koch, D.G. Norris, M. Hund-Georgiadis, An investigation of functional [40] J. Lachaux, et al., Measuring phase synchrony in brain signals, Hum. Brain
and anatomical connectivity using magnetic resonance imaging, NeuroImage Mapp. 8 (4) (1999) 194–208.
16 (1) (2002) 241–250. [41] E. Niedermeyer, F.H. Lopes da Silva, 4th ed.,Electroencephalography: Basic
[4] A.A. Fingelkurts, A.A. Fingelkurts, S. Kähkönen, Functional connectivity in the Principles, Clinical Applications, and Related Fields, XI, Williams & Wilkins,
brain—is it an elusive concept? Neurosci. Biobehav. Rev. 28 (8) (2005) 827–836. Baltimore, 1999 1258 p..
[5] B. Horwitz, The elusive concept of brain connectivity, NeuroImage 19 (2 Pt 1) [42] E. Pereda, R. Quiroga, J. Bhattacharya, Nonlinear multivariate analysis of
(2003) 466–470. neurophysiological signals, Prog. Neurobiol. 77 (1–2) (2005) 1–37.
[6] W.R. Adey, D.O. Walter, C.E. Hendrix, Computer techniques in correlation and [43] F. Mormann, et al., Mean phase coherence as a measure for phase synchro-
spectral analyses of cerebral slow waves during discriminative behavior, Exp. nization and its application to the EEG of epilepsy patients, Phys. D: Nonlin.
Neurol. 3 (1961) 501–524. Phenom. (2000) 358–369.
[7] M. Scherg, D. Von Cramon, Two bilateral sources of the late AEP as identified [44] J. Arnhold, et al., A robust method for detecting interdependences: applica-
by a spatio-temporal dipole model, Electroencephalogr. Clin. Neurophysiol. tion to intracranially recorded EEG. Phys. D: Nonlin. Phenom. 134 (1999).
62 (1) (1985) 32–44. [45] R. Quian Quiroga, et al., Performance of different synchronization measures
[8] R. Grech, et al., Review on solving the inverse problem in EEG source analysis, in real data: a case study on electroencephalographic signals, Phys. Rev. E
J. Neuroeng. Rehabil. 5 (2008) 25. Stat. Nonlin. Soft Matter Phys. 65 (4 Pt 1) (2002) 041903.
[9] K.J. Friston, R.J. Dolan, Computational and dynamic models in neuroimaging, [46] J. Arnhold, et al., A robust method for detecting interdependences: applica-
NeuroImage 52 (3) (2010) 752–765. tion to intracranially recorded EEG, Phys. D: Nonlin. Phenom. 134 (1999).
[10] K.J. Friston, L. Harrison, W. Penny, Dynamic causal modelling, NeuroImage 19 [47] F. Takens, D.A. Rand, L.S. Young, Detecting strange attractors in turbulence,
(4) (2003) 1273–1302. Dynamical systems and Turbulence, Lecture Notes in Mathematics (1981)
[11] J. Daunizeau, S. Kiebel, K. Friston, Dynamic causal modelling of distributed 366–381.
electromagnetic responses, NeuroImage 47 (2) (2009) 590–601. [48] V. Sakkalis, M. Zervakis, Linear and Nonlinear Synchronization Analysis and
[12] S. Kiebel, et al., Dynamic causal modeling for EEG and MEG, Hum. Brain Visualization during Altered States of Consciousness, in: G.R. Naik (Ed.),
Mapp. 30 (6) (2009) 1866–1876. Recent Advances in Biomedical Engineering, IN-TECH, Vienna, 2010.
[13] F. Lopes da Silva, et al., Model of brain rhythmic activity. The alpha-rhythm of [49] J. Jeong, J.C. Gore, B.S. Peterson, Mutual information analysis of the EEG in
the thalamus, Kybernetik 15 (1) (1974) 27–37. patients with Alzheimer’s disease, Clin. Neurophysiol. 112 (5) (2001) 827–835.
[14] R. Moran, et al., A neural mass model of spectral responses in electrophysiol- [50] S.H. Na, et al., EEG in schizophrenic patients: mutual information analysis,
ogy, NeuroImage 37 (3) (2007) 706–720. Clin. Neurophysiol. 113 (12) (2002) 1954–1960.
[15] B. Jansen, V. Rit, Electroencephalogram and visual evoked potential genera- [51] J. Rissanen, M. Wax, Measures of Mutual and Causal Dependence Between
tion in a mathematical model of coupled cortical columns, Biol. Cybern. 73 Two Time Series, IEEE Transactions on Information Theory, IEEE Press
(4) (1995) 357–366. Piscataway, NJ, USA, 1987, pp. 598–601.
 V. Sakkalis / Computers in Biology and Medicine 41 (2011) 1110–1117 1117

[52] R. Moran, et al., Dynamic causal models of steady-state responses, Neuro- [65] J. Dauwels, et al., A comparative study of synchrony measures for the early
Image 44 (3) (2009) 796–811. diagnosis of Alzheimer’s disease based on EEG, NeuroImage 49 (1) (2010)
[53] P. von Bünau, et al., Finding stationary subspaces in multivariate time series, 668–693.
Phys. Rev. Lett. 103 (21) (2009) 214101. [66] K.J. Friston, Schizophrenia and the disconnection hypothesis, Acta Psychiatr.
[54] H. Akaike, A new look at statistical model identification, IEEE Trans. Autom. Scand. Suppl. 395 (1999) 68–79.
Contr. (1974) 716–723. [67] K.M. Spencer, et al., Abnormal neural synchrony in schizophrenia, J. Neurosci.
[55] J. Theiler, P. Rapp, Re-examination of the evidence for low-dimensional, 23 (19) (2003) 7407–7411.
nonlinear structure in the human electroencephalogram, Electroencephalogr. [68] A.R. Haig, et al., Gamma activity in schizophrenia: evidence of impaired
Clin. Neurophysiol. 98 (3) (1996) 213–222. network binding? Clin. Neurophysiol. 111 (8) (2000) 1461–1468.
[56] L. Marzetti, C. Del Gratta, G. Nolte, Understanding brain connectivity from [69] D.S. Bassett, et al., Cognitive fitness of cost-efficient brain functional net-
EEG data by identifying systems composed of interacting sources, Neuro- works, Proc. Natl. Acad. Sci. USA 106 (28) (2009) 11747–11752.
Image 42 (1) (2008) 87–98. [70] M.A. Just, et al., Functional and anatomical cortical underconnectivity in
[57] G. Gómez-Herrero, et al., Measuring directional coupling between EEG autism: evidence from an FMRI study of an executive function task and
sources, NeuroImage 43 (3) (2008) 497–508. corpus callosum morphometry, Cereb. Cortex 17 (4) (2007) 951–961.
[58] V. Sakkalis, et al., Assessment of neural dynamic coupling and causal [71] L. Pollonini, et al., Functional connectivity networks in the autistic and
interactions between independent EEG components from cognitive tasks healthy brain assessed using Granger causality, in: Engineering in Medicine
using linear and nonlinear methods, Conf. Proc. IEEE Eng. Med. Biol. Soc. 2008 and Biology Society IEMBS’10 32nd Annual International Conference of the
(2008) 3767–3770. IEEE, 2010.
[59] G. Nolte, et al., Identifying true brain interaction from EEG data using the [72] R. Kulisek, et al., Nonlinear analysis of the sleep EEG in children with
imaginary part of coherency, Clin. Neurophysiol. 115 (10) (2004) 2292–2307. pervasive developmental disorder, Neuro Endocrinol. Lett. 29 (4) (2008)
[60] V. Sakkalis, V. Tsiaras, I. Tollis, Graph analysis and visualization for brain function 512–517.
characterization using EEG data, J. Healthcare Eng. 1 (3) (2010) 435–459. [73] W. Bosl, et al., EEG complexity as a biomarker for autism spectrum disorder
[61] E.G. Christodoulou, et al., BrainNetVis: an open-access tool to effectively risk, BMC Med. 9 (2011) 18.
quantify and visualize brain networks, Comput. Intell. Neurosci. 2011 (2011) [74] G. Winterer, et al., EEG phenotype in alcoholism: increased coherence in the
747290. depressive subtype, Acta Psychiatr. Scand. 108 (1) (2003) 51–60.
[62] V. Sakkalis, Applied strategies towards EEG/MEG biomarker identification in [75] E.A. de Bruin, et al., Moderate-to-heavy alcohol intake is associated with
clinical and cognitive research, Biomark. Med. 5 (1) (2011) 93–105. differences in synchronization of brain activity during rest and mental
[63] D. Cosandier-Rimélé, et al., Computational modeling of epileptic activity: from rehearsal, Int. J. Psychophysiol. 60 (3) (2006) 304–314.
cortical sources to EEG signals, J. Clin. Neurophysiol. 27 (6) (2010) 465–470. [76] V. Sakkalis, et al., Optimal brain network synchrony visualization: application
[64] D. Osipova, et al., Effects of scopolamine on MEG spectral power and coherence in an alcoholism paradigm, Conf. Proc. IEEE Eng. Med. Biol. Soc. 2007 (2007)
in elderly subjects, Clin. Neurophysiol. 114 (10) (2003) 1902–1907. 4285–4288.