IEEE Signal Processing - May 2018 - 84

Data recording and preprocessing
As previously stated, EEG data can be recorded in a taskrelated or task-free (resting state) paradigm. Depending on the
context (clinical or cognitive research), these recordings can
be performed using dense electrode arrays (64-256 sensors)
either in patients or in healthy subjects [Figure 1(a)]. MEG
and EEG are quite similar techniques. From a biophysics
viewpoint, the phenomena at the origin of recorded electric
and magnetic fields are slightly different (EEG detects both
radial and tangential currents, while MEG detects tangential
currents only). Besides cost issues that stem directly from the
technical difficulty of measuring magnetic fields to the order
of 1 fT (10-15 T), differences are also related to the sensitivity of both methods to deep sources, to the impact of volumeconductor modeling on the reconstruction of sources, and to
ease of use. While this article will focus on the EEG source
connectivity method, the analysis steps remain the same for
both techniques.

Number of channels
The number of scalp electrodes is a crucial parameter for the
performance of EEG source connectivity methods. Different
studies have shown that the number of channels has a strong

impact on the quality of the localized sources [13] or the networks reconstructed from scalp EEG data [14]. The use of the
available systems (going from the former 19-32 channels to
the newer 64-256 channels) can have a dramatic impact on
the performance of the source reconstruction step (see the section "Reconstruction of EEG Sources"). There is growing evidence that increasing the number of EEG channels provides
greater accuracy in source estimations. The minimal number
of electrodes required is also related to the other parameters
used in the pipeline, mainly the algorithm used to reconstruct
the dynamics of brain sources. Many studies have shown that
at least 128 electrodes are needed to obtain satisfactory
results, typically when the minimum norm class of inverse
methods is used for localizing sources [13] or identifying
functional networks [14].

Preprocessing
EEGs are often contaminated by various physiological or nonphysiological sources of activity, e.g., cardiac signals, eye
movements and blinks, muscle activity, and head/cable movements. Removing these artifacts is a crucial step in producing
noise-free signals prior to applying EEG source connectivity
per se. The detection can be done visually or semi- or fully

Regional Time Series

Dense EEG (256 Channels)

Inverse Solution

Task-Free (Resting State)/
Task-Related Paradigm

Time
Dipole Orientation
and Location

Network Measures

(b)

68 ROIs

Network-Based
Index

(a)

Statistical Couplings

Patients
and/or
Healthy
Subjects

68 ROIs

Head Model

Clinical Tests
68 ROIs
(c)

Ed

Node

ge

Abnormal
(d)

FIGURE 2. The process from EEG recording and preprocessing to brain networks and applications. (a) EEG data can be recorded during task-related
(evoked response) or task-free (resting state) paradigms. (b) To reconstruct EEG sources, the lead field matrix (the contribution of each cortical source
to the scalp sensors) is required. It is computed from 1) a multiple-layer head model (volume conductor), which is obtained from MRI segmentation, and
2) the position of the scalp electrodes. The boundary element method is one of the available numerical methods. (c) Once the regional time series are
reconstructed, the functional connectivity can be estimated by computing the statistical couplings between these time series. (d) Once nodes and edges
have been defined, network topological properties (organization) can be studied by graph theory-based analysis.

84

IEEE SIgnal ProcESSIng MagazInE

|

May 2018

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