HAL Id: hal-00423717
https://hal.archives-ouvertes.fr/hal-00423717
Submitted on 12 Oct 2009
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
On the ”Dependence” of ”Independent” Group EEG
Sources; an EEG Study on Two Large Databases
Marco Congedo, Roy John, Dirk de Ridder, Leslie Prichep, Robert Isenhart
To cite this version:
Marco Congedo, Roy John, Dirk de Ridder, Leslie Prichep, Robert Isenhart. On the ”Dependence” of
”Independent” Group EEG Sources; an EEG Study on Two Large Databases. Brain Topography: a
Journal of Cerebral Function and Dynamics, Springer Verlag, 2010, 23 (2), pp.134. �10.1007/s10548-
009-0113-6�. �hal-00423717�
1
Brain Topography, 2009
Author copy
On the “dependence” of “independent” group EEG
sources; an EEG study on two large databases.
Marco Congedo
a
, Roy E. John
b
, Dirk De Ridder
c
, Leslie Prichep
b
, Robert Isenhart
b
a
Gipsa-lab, National Center for Scientific Reserach (cnrs), University Joseph Fourier, University
Stendhal, Grenoble Institute of Tecnology, Grenoble, France.
b
Brain Research Laboratory, New York University Medical School, Department of Psychiatry
c
Brain Research center Antwerp for Innovative and Interdisciplinary Neuromodulation (BRAI²N) &
Dept of Neurosurgery, University and Hospital of Antwerp, Belgium
Corresponding Author: M. Congedo, cnrs, Gipsa-lab, 961 rue de la Houille Blanche, Domaine Universitaire - BP 46 - 38402,
Grenoble, France. E-mail: marco.congedo@gmail.com, phone: +33(0)476826252, fax: +33(0)476574790
2
Abstract. The aim of this work is to study the coherence profile (dependence) of robust eyes-closed
resting EEG sources isolated by group blind source separation (gBSS). We employ a test-retest strategy
using two large sample normative databases (N=57 and N=84). Using a BSS method in the complex
Fourier domain, we show that we can rigourously study the out-of-phase dependence of the exctracted
components, albeit they are extracted so as to be in-phase independent (by BSS definition). Our focus
on lagged communication between components effectively yields dependence measures unbiased by
volume conduction effects, which is a major concern about the validity of any dependence measures
issued by EEG measurements. We are able to show the organization of the extracted components in
two networks. Within each network components oscillate coherently with multiple-frequency
dynamics, whereas between networks they exchange information at non-random multiple time-lag
rates.
Keywords:
Blind Source Separation (BSS), Independent Component Analysis (ICA), Connectivity, In-Phase Coherence, Out-
Of-Phase Coherence.
3
Introduction
The functional organization of the brain at rest is currently conceived in terms of resting state
networks (RSNs), clusters of brain regions, mostly cortical, inter-connected anatomically and
functionally (Beckmann et al. 2005; Damoiseaux et al. 2006; Fox and Raichle 2007; Mantini et al.
2007; van den Heuvel et al. 2008). The study of RSNs has shifted the focus in neuroimaging from the
exact localization of specialized brain functions (looking for “things in a place”) to the understanding
of the interplay of widespread brain structures.
Regarding the study of RSNs by EEG, biophysical and neurophysiological studies suggest that
each RSN may exhibit complex dynamic associated with multiple frequencies simultaneously (Mantini
et al. 2007). Studying the distribution of scalp EEG power at rest, as in Chen et al. (2008), does not
allow the study of RSNs by EEG because scalp voltage is a mixing of underlying source activity
(volume conduction: see Nunez and Srinivasan 2006) and because scalp power is not a comprehensive
measure of widespread coherent synchronization. Instead, it appears more appropriate extracting robust
synchronizations all over the cortex and testing them altogether (the whole cohort) along the frequency
dimension. The multivariate tool permitting such investigation in large samples of individuals is group
independent component analysis (gICA) (Calhoun et al. 2001; Schmithorst and Holland 2004).
Whereas ICA relies on high-order statistics, here we use another blind source separation (BSS)
framework based on second-order statistics. We have previously argued that second-order BSS
methods suits well EEG data (Congedo et al. 2008). Importantly, it allows us to address the problem of
possible dependence among the extracted regions of coherent synchronization.
Group BSS extracts scalp spatial maps and associated EEG time-courses, referred together to as
components. We wish to extract components that can be found on a substantial proportion of normal
individuals at rest. Since there is no way to establish a-priori how many of such components can be
established, nor if they are reliable, we employ a test-retest strategy using two independent large
sample normative databases (N=57 and N=84) and retain as many components as we can replicate,
proceeding in descending order of explained variance; a similar strategy has been previously employed
(Damoiseaux et al. 2006) in an fMRI study. Once robust normative components are extracted, we
characterize the cortical structures involved in each component by a distributed source localization of
the spatial maps (Greenblatt et al. 2005) and their spectral profile. We are able to replicate on the two
databases seven components with nearly identical spatial and frequency distribution (detailed in a
separate paper). In this paper we study the out-of-phase (lagged) coherence (dependence) of the
extracted components using recent advances on connectivity measures adapted to EEG data. We are
able to show the organization of the components in two networks. Within each network components
oscillate coherently with multiple-frequency dynamics, whereas between networks they exchange
information at non-random multiple time-lag rates.
Material and Methods
It is well known that spectral EEG measures follow developmental equations, meaning that the
frequency composition of the EEG reflects the maturational status of the brain (John et al. 1980). In
order to avoid confounding age effects we consider in this study only adult individuals between 17 and
30 y.o. Two independent normative databases previously acquired were used for this study. One is a
subset of the normative database of the Brain Research Laboratory (BRL), New York University
(N=57; age range 17-30) and the other the adult normative database of Nova Tech EEG (NTE), Inc.,
Mesa, AZ (N=84; age range 18-30). Exclusion criteria for the BRL database were known psychiatric or
neurological illness, drug/alcohol abuse, current psychotropic/CNS active medications, history of head
injury (with loss of consciousness) or seizures. Exclusion criteria for the NTE database were a
psychiatric history in any relative and participant of drug/alcohol abuse, head injury (at any age, even
very mild), headache, physical disability and epilepsy.
Recording procedures and settings were very similar for the two databases. In both cases about 3-
5 minutes of EEG was continuously recorded while participant sat with the eye-closed on a
comfortable chair in a quiet and dimly lit room. EEG data were acquired at the 19 standard leads
4
prescribed by the 10-20 international system (FP1, FP2, F7, F3, FZ, F4, F8, T3, C3, CZ, C4, T4, T5,
P3, PZ, P4, T6, O1, O2) using both earlobes as reference and enabling a 60 Hz notch filter to suppress
power line contamination. The impedance of all electrodes was kept below 5K Ohms. Data of the NTE
database were acquired using the 12-bit A/D NeuroSearch-24 acquisition system (Lexicor Medical
technology, Inc., Boulder, CO) and sampled at 128 Hz, whereas data of the BRL database were
acquired using the 12-bit A/D BSA acquisition system (Neurometrics, Inc., New York, NY) and
sampled at 100 Hz. For consistency, we subsequently up-sampled the BRL database to 128 Hz using a
natural cubic spline interpolation routine (Congedo et al. 2002). In order to minimize inter-subject
variability we removed from all data any biological, instrumental and environmental artifacts, paying
particular attention to biological artifacts generated by the eyes, the hearth and the muscles of the neck,
face and jaw. EEG recordings were visually inspected on a high-resolution screen and epochs
containing visible artifacts were marked and ignored for ensuing analysis.
Frequency-domain statistics
All statistics used in this study are summarized in the complex hermitian Fourier cross-spectral
matrices
E E
f
⋅
∈
ℂ
S
, where f is the discrete frequency index and E the number of electrodes
(Bloomfield 2000). We can write
f f f
i
= +
S C Q
, where i=
1
−
. Symmetric
E E
f
⋅
∈ℜ
C
, referred to as
the cospectral matrix, holds in the main diagonal the power spectra and in the off-diagonal elements
the in-phase (or with a half cycle phase shift, i.e., opposite sign) dependency structure. Antisymmetric
E E
f
⋅
∈ℜ
Q
, referred to as the quadrature spectral matrix, holds in the off-diagonal elements the out-of-
phase (a quarter cycle in either direction) dependency structure. Both dependency structures are of
second order. The cospectral matrix is equivalent to the covariance matrix of the data band-pass filtered
for its discrete frequency: by Parseval’s theorem the sum of all co-spectral matrices is equivalent to the
data covariance matrix. We estimate individual cross-spectral matrices as the average obtained by Fast
Fourier Transform (FFT) on 50% sliding overlapping 2-seconds windows tapered using the function
introduced by Welch (1967). Finally, we obtain group average cospectral matrices as the grand average
across individuals.
Group Blind Source Separation
For E scalp sensors and M≤E EEG dipolar fields to be estimated, the BSS linear model employed
describes the superposition principle, i.e., we state
( ) ( )
t t
=
v As
, (1)
where
( )
E
t
∈ℜ
v
is the sensor measurement vector,
E M
⋅
∈ℜA
is a time-invariant non-singular
mixing
matrix and
( )
M
t
∈ℜ
s
holds the time-course of the source components. Note that model (1) describes
the instantaneous (in-phase) diffusion of current source over measurement sites, in fact describing the
effect of direct current and volume conduction (Congedo et al. 2008). Our source estimation is given
by
ˆ
( ) ( )
t t
=
s Bv
, (2)
where separating matrix
M E
⋅
∈ℜB
is what we wish to estimate by BSS assuming no knowledge of A
and weak assumptions on s.
A wide array of BSS/ICA methods exist. In this work we use a method based on the approximate
joint diagonalization (AJD) of Fourier cospectral matrices, which is a robust and computationally fast
approach (Congedo et al. 2008). We diagonalize grand-average cospectral matrices in the frequency
range 0.5-30 Hz only, which is the range with highest signal-to-noise ratio. This results in F=60
frequencies with 0.5 Hz resolution (0.5, 1, 1.5, …, 30 Hz). This approach is analogous to the averaging
group ICA approach described for fMRI by Schmithorst and Holland (2004). The weak assumption on
the source process we need to solve the BSS problem with this approach is that sources are
uncorrelated and have non-proportional power spectrum (Congedo et al. 2008).
In order to estimate M<E source components the matrix B is found with a classical two-stage
process, which allows the estimation of the M most energetic components while reducing the noise.