IEEE Signal Processing - May 2018 - 93

As EEG source connectivity is still a relatively new field
compared to fMRI connectivity analysis, more methodological efforts are still needed to completely overcome issues such
as mixing and spatial leakage. We also advise the use of multimodal recordings, such as EEG/fMRI, which can benefit from
the excellent spatial resolution of the fMRI and the excellent
time resolution of the EEG and can help to cross-validate the
results from both techniques.

Consistency of inverse/connectivity measures
Although all reported EEG studies include two main steps (an
EEG inverse problem followed by a source connectivity estimation), they differ from a methodological perspective. Indeed,
different algorithms were used to reconstruct cortical sources.
In these algorithms, various mathematical assumptions are used
for the regularization of an ill-posed inverse problem. The main
assumptions relate to sources with minimum energy, time/space
sparsity, and possible correlation between the reconstructed
sources. A plethora of functional and effective connectivity
measures were also proposed to measure statistical couplings
between regional time series. Therefore, the question naturally
raised is: What combination of inverse/connectivity methods
should be used to enhance global performance and to guarantee
the relevance of results in terms of identified brain networks?
Unfortunately, there is no answer to this question. As each of
the inverse and functional/effective methods has its own
assumptions and characteristics, there is no consensus, yet,
about the best combination.
This crucial issue has been addressed in various studies
showing that the selected methods (i.e., the inverse solution and
connectivity measure) directly impact the topological/statistical properties of networks identified from EEG surface signals.
Recently, Mahjoory et al. evaluated the effect of the anatomical
templates, head models, inverse solutions, and software implementations. The authors showed the variability between the
inverse solution algorithms (mainly LCMV and wMNE). Also,
the functional connectivity measures were much more consistent across the variables as compared with measures obtained
with effective connectivity methods.
We also conducted two comparative studies regarding the
choice of the optimal inverse/connectivity combination method. In both studies, our intent was to maximize the a priori
information (ground truth) regarding the brain networks that
were supposed to be identified from dense EEG. In the first
[14], we focused on a widely used cognitive task (picture naming) for which a strong literature background was available,
essentially coming from fMRI studies. In the second study
[38], we pursued a modeling approach in which epileptogenic
networks were used to simulate dense-EEG data that were subsequently used to evaluate EEG source connectivity. We then
compared the network obtained by each inverse/connectivity
combination with the reference network, using a network-similarity index recently proposed in our team.
Interestingly, both comparative studies led to the same conclusion: a strong variability was observed among the tested
combinations, but the results provided by the wMNE/PLV

combination displayed consistency and always exhibited the
best performance in terms of matching between the estimated
and the reference network. This result might be explained by
the fact that wMNE relies on reasonable physiological assumptions (the position and orientation of sources). The only mathematical assumption is that the solution has the lowest energy.
This assumption could also be interpreted physiologically in
terms of minimal energy cost in the brain during task performance or at rest [50]. Regarding the second step, the PLV
method estimates the PS between EEG oscillations. Therefore, this method is in line with the concept of communication through coherence in the brain in which synchronization
between locally generated signals is a crucial mechanism in
brain function. In the context of EEG source connectivity, the
PLV method, in particular, and, more generally, the PS methods, precisely reflect the underlying synchronization between
the brain signals generated by distant sources [25]. Altogether, these features may explain the good performance of this
combination of methods, particularly within the assessment of
brain networks involved in cognitive activity.

Clinical impact
A growing body of evidence suggests that brain disorders are
related to alterations in functional connections between brain
regions that disrupt the normal large-scale brain network
organization and function [2], [51]. A first conclusion from this
tutorial is that the extraction of valuable information about
pathological brain networks from EEG is challenging but
obtainable. A second remark is that clinical practice will certainly change in coming years. Furthermore, although the
combination of EEG source connectivity with network science
is still a young research field, results reported over the last few
years are, clinically, very promising. It is likely that the use of
novel tools allowing for the characterization and quantification of identified networks (which is the case in modern network science using graph theory-based analysis) will develop
and spread to clinics.
However, most of the studies reported and discussed in this
review were generally performed on relatively small groups of
patients. Because of the diversity of methodological approaches (e.g., candidate inverse-solution algorithms and functional/
effective connectivity measures, and the impact of the number
of electrodes) and the number of possible conditions (e.g., taskrelated versus task-free paradigms), a comparison of results is
still difficult. Further studies on larger cohorts of patients will
certainly contribute to standardizing the analysis conditions.
Within the context of epilepsy, one of the main clinical challenges is the delineation of the epileptogenic zone (EZ), electrophysiologically defined as the primary zone of organization
of ictal discharges. In drug-resistant partial epilepsies when
surgery can be indicated, resectioning the EZ showed it to be
sufficient to significantly reduce the occurrence of seizures
and even to lead to freedom from them. Yet, there exists no
available technique able to precisely define the EZ. In this context, the EEG source connectivity method showed encouraging
results to estimate epileptogenic networks from noninvasive

IEEE Signal Processing Magazine

May 2018

Table of Contents for the Digital Edition of IEEE Signal Processing - May 2018