5 EEG Source Imaging and Multimodal Neuroimaging
111
a standard approach, symmetrical methods have yet to present singular methods for
discussion. It may be better to explore this topic in a historical manner, then, reviewing major landmark developments. Once more, however, we can simplify the topic
by splitting it into two major categories: data driven approaches and model-based
methods.
Data-Driven Integration
Data-driven methods are those that focus predominately on integrating the datasets
from EEG and fMRI. Under these paradigms, little attention is given to the actual
structures and dynamics of neural populations or how they interact with cortical blood
flow. It is instead assumed that the numerical methods will provide the necessary
context for application. This means that the methods are generally straightforward
and computationally accessible.
Perhaps the earliest effort to appear as a symmetrical data-driven approach was
presented by Martinez-Montes et al. in 2004 [55]. Under their approach, a multiway
Partial Least-Squares (PLS) method was used to decompose EEG and fMRI data
into a sum of elements, which were called “atoms.” Following this method, EEG
atoms contained three main properties—their spatial, temporal, and spectral signatures—following a method of Parallel Factor Analysis, where the EEG Signal was
represented as a trilinear model of these parts:
S dwt
N k
k1
a dk b wk c tk + ε dwt
(5.45)
where S is the EEG signal; a, b, and c are normalized vectors influenced by d, w, and
t, which respectively represent electrode (spatial), frequency (spectral), and time
(temporal) components; N k is the number of components; k specifies the current
component; and ε is the error. fMRI is similarly modeled considering time and voxel
as their key dimensions:
F st
N k
k1
u sk v tk + ε st
(5.46)
where F st is the fMRI signal; u and v represent the two signatures defined by their
voxel (spatial component represented by s) and time (temporal component represented by t). Once more, N k and k are the number of components and the current
component, and ε is the error. These representations can be seen in Fig. 5.10.
The method then sought to maximize the covariance between the temporal properties of the EEG and fMRI signals (the c and v vectors in Fig. 5.10), matching
the signals. EEG signals were further broken down into frequency bands of interest,
focusing on only one band at a time and identifying which frequencies presented a
significant correlation with the fMRI signal and how they were distributed throughout the brain. As in EEG-informed fMRI, this integration required that EEG data
be downsampled and convolved with an appropriate HRF to exist within the same
111
a standard approach, symmetrical methods have yet to present singular methods for
discussion. It may be better to explore this topic in a historical manner, then, reviewing major landmark developments. Once more, however, we can simplify the topic
by splitting it into two major categories: data driven approaches and model-based
methods.
Data-Driven Integration
Data-driven methods are those that focus predominately on integrating the datasets
from EEG and fMRI. Under these paradigms, little attention is given to the actual
structures and dynamics of neural populations or how they interact with cortical blood
flow. It is instead assumed that the numerical methods will provide the necessary
context for application. This means that the methods are generally straightforward
and computationally accessible.
Perhaps the earliest effort to appear as a symmetrical data-driven approach was
presented by Martinez-Montes et al. in 2004 [55]. Under their approach, a multiway
Partial Least-Squares (PLS) method was used to decompose EEG and fMRI data
into a sum of elements, which were called “atoms.” Following this method, EEG
atoms contained three main properties—their spatial, temporal, and spectral signatures—following a method of Parallel Factor Analysis, where the EEG Signal was
represented as a trilinear model of these parts:
S dwt
N k
k1
a dk b wk c tk + ε dwt
(5.45)
where S is the EEG signal; a, b, and c are normalized vectors influenced by d, w, and
t, which respectively represent electrode (spatial), frequency (spectral), and time
(temporal) components; N k is the number of components; k specifies the current
component; and ε is the error. fMRI is similarly modeled considering time and voxel
as their key dimensions:
F st
N k
k1
u sk v tk + ε st
(5.46)
where F st is the fMRI signal; u and v represent the two signatures defined by their
voxel (spatial component represented by s) and time (temporal component represented by t). Once more, N k and k are the number of components and the current
component, and ε is the error. These representations can be seen in Fig. 5.10.
The method then sought to maximize the covariance between the temporal properties of the EEG and fMRI signals (the c and v vectors in Fig. 5.10), matching
the signals. EEG signals were further broken down into frequency bands of interest,
focusing on only one band at a time and identifying which frequencies presented a
significant correlation with the fMRI signal and how they were distributed throughout the brain. As in EEG-informed fMRI, this integration required that EEG data
be downsampled and convolved with an appropriate HRF to exist within the same