Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_6027_Библиотеки_им_академика_М_И_Перельмана.pdf
Скачиваний:
0
Добавлен:
31.08.2026
Размер:
29 Мб
Скачать
19 Event-Related Potentials 253

19.6 Conclusion

This chapter introduced the concept of ERPs, their characteristic waveform, and a selection of commonly investigated components. ERPs have been studied exten­sively for more than half a century and have proven to be a powerful tool to explore sensory and cognitive processing. At the same time, the logic behind the analysis is easy to grasp: cutting the data into segments around the events, removing offsets, and averaging over segments. The results can be communicated through a variety of output measures and compelling gures. Thus, ERPs provide a simple yet elegant solution to investigate evoked brain responses with high temporal resolution. The ERP method is still widely popular in EEG research as reected by the prolic publication of new studies. Hence, ERPs should be an essential part of the EEG researchers analysis toolkit.

References

Bruin, K. J., & Wijers, A. A. (2002). Inhibition, response mode, and stimulus probability: A
comparative event-related potential study. Clinical Neurophysiology, 113, 1172–1182.
Brunia, C. H. M., Van Boxtel, G. J. M., & Böcker, K. B. E. (2012). Negative slow waves as indices
of
anticipation: The Bereitschaftspotential, the contingent negative variation, and the stimulus­preceding negativity. In E. S. Kappenman & S. J. Luck (Eds.), The Oxford handbook of event- related potential components. Oxford University Press.
Eimer, M. (2011). The face-sensitive N170 component of the event-related brain potential. In The
Gehring, W. J., Liu, Y., Orr, J. M., & Carp, J. (2012). The error-related negativity (Ern/Ne). In E. S.
Handy, T. C. (2005). Event-related potentials: A methods handbook. MIT Press. Haxby, J. V., Hoffman, E. A., & Gobbini, M. I. (2000). The distributed human neural system for
Hernandez-Pavon, J. C., Veniero, D., Bergmann, T. O., Belardinelli, P., Bortoletto, M., Casarotto,
Hutzler, F., Braun, M., Võ, M. L. H., Engl, V., Hofmann, M., Dambacher, M., Leder, H., & Jacobs,
Jeffreys, D. A., & Axford, J. G. (1972). Source locations of pattern-specic components of human
Keil, A., Debener, S., Gratton, G., Junghofer, M., Kappenman, E. S., Luck, S. J., Luu, P., Miller,
Kutas, M.,
handbook of face perception (Vol. 28, pp. 329–344). Oxford University Press.
Oxford
Kappenman Oxford University Press.
face
S.,
Casula, E. P., Farzan, F., Fecchio, M., Julkunen, P., Kallioniemi, E., Lioumis, P., Metsomaa, J., Miniussi, C., Mutanen, T. P., Rocchi, L., Rogasch, N. C., Sha, M. M., Siebner, H. R., Thut, G., Zrenner, C., Ziemann, U., & Ilmoniemi, R. J. (2023). TMS combined with EEG: Recom­mendations and open issues for data collection and analysis. Brain Stimulation, 16, 567–593.
A.
M. (2007). Welcome to the real world: Validating xation-related brain potentials for
ecologically valid settings. Brain Research, 1172, 124–129.
visual evoked potentials. I. Component of striate cortical origin. Experimental Brain Research, 16, 1– 21.
A., & Yee, C. M. (2014). Committee report: Publication guidelines and recommendations for
G. studies using electroencephalography and magnetoencephalography. Psychophysiology, 51, 1–21.
language comprehension. Trends in Cognitive Sciences, 4, 463–470.
& S. J. Luck (Eds.), The Oxford handbook of event-related potential components.
perception. Trends in Cognitive Sciences, 4, 223–233.
& Federmeier, K. D. (2000). Electrophysiology reveals semantic memory use in
254 M. Hoppstädter
Liu, Y., Huang, H., McGinnis-Deweese, M., Keil, A., & Ding, M. (2012). Neural substrate of the
late
positive potential in emotional processing. The Journal of Neuroscience, 32, 14563–14572. Luck, S. J. (2014). An introduction to Luck, S. J., & Kappenman, E. S. (2011). The Oxford handbook of event-related potential compo-
nents
. Oxford University Press.
Maris, E., & Oostenveld, R. (2007). Nonparametric statistical testing of EEG- and MEG-data.
Journal of Neuroscience Methods, 164, 177–190.
Näätänen, R., & Kreegipuu, K. (2012). The mismatch negativity (MMN). In E. S. Kappenman &
J. Luck (Eds.), The Oxford handbook of event-related potential components. Oxford Uni-
S. versity Press.
Nordstrom, H., & Wiens, S. (2012). Emotional event-related potentials are larger to gures than
scenes
Polich, J. (2004). Clinical application of the P300 event-related brain potential. Physical Medicine
Pratt, H. (2012). Sensory ERP components. In E. S. Kappenman & S. J. Luck (Eds.), The Oxford
Rossion, B., & Jacques, C. (2012). The N170: Understanding the time course of face perception in
Sauseng, P., Klimesch, W., Gruber, W. R., Hanslmayr, S., Freunberger, R., & Doppelmayr,
Schandry, R., Sparrer, B., & Weitkunat, R. (1986). From the heart to the brain: A study of heartbeat
Schechter, I., Butler, P. D., Zemon, V. M., Revheim, N., Saperstein, A. M., Jalbrzikowski, M.,
Smulders, F. T. Y., & Miller, J. O. (2012). The lateralized readiness potential. In E. S. Kappenman
Sutton, S., Braren, M., Zubin, J., & John, E. R. (1965). Evoked-potential correlates of stimulus
Tanaka, J. W., & Curran, T. (2001). A neural basis for expert object recognition. Psychological
Walter, W. G., Cooper, R., Aldridge, V., McCallum, W., & Winter, A. (1964). Contingent negative
Wilding, E.
but are similarly reduced by inattention. BMC Neuroscience, 13, 49.
Rehabilitation Clinics of North America, 15, 133–161.
and
handbook of event-related potential components. Oxford University Press.
the
human brain. In E. S. Kappenman & S. J. Luck (Eds.), The Oxford handbook of event-
related potential components. Oxford University Press.
M.
(2007). Are event-related potential components generated by phase resetting of brain
oscillations? A critical discussion. Neuroscience, 146, 1435–1444.
contingent
Pasternak, transient visual evoked potentials to magno- and parvocellular-selective stimuli in schizophre­nia. Clinical Neurophysiology, 116, 2204–2215.
& University Press.
uncertainty.
Science,
variation: Nature, 203, 380–384.
processes. In E. S. Kappenman & S. J. Luck (Eds.), The Oxford handbook of event-related potential components. Oxford University Press.
scalp potentials. The International Journal of Neuroscience, 30, 261–275.
R., Silipo, G., & Javitt, D. C. (2005). Impairments in generation of early-stage
S. J. Luck (Eds.), The Oxford handbook of event-related potential components. Oxford
Science, 150, 1187–1188.
12, 43– 47.
An electric sign of sensori-motor association and expectancy in the human brain.
L., & Ranganath, C. (2012). Electrophysiological correlates of episodic memory
the event-related potential technique. MIT Press.
Chapter 20
EEG Source Analysis
Alejandro Ojeda
Abstract EEG source analysis enables the study of cognitive and physiological
ses closer to their origin within the brain. To that end, it combines anatomical
proces information, usually derived from participant-specic or template MRI segmenta­tions of the head and inner layers of tissue, with EEG signals using mathematical methods that transform the latter into a functional brain imaging modality. This chapter will touch on the fundamentals of EEG source analysis, including its modeling assumptions, representative source localization algorithms, and source­level statistical inference and connectivity.
Keywords Source imaging · Head model · Forward and inverse problems · Regular space · Multiple comparison correction · Source summarization
ization · Constraints · Statistical analysis and connectivity in the source

20.1 Introduction

As mentioned in Chap. 2, What is EEG?, the EEG is a voltage signal that can be measured on the scalp surface and originates from the electrical activity of cortical neurons. The layer of pyramidal neurons in the cortex is thought to be the primary contributor to the EEG (and MEG) signal because the spatial orientation of these neurons resembles equivalent current dipoles. When enough localized populations of these neurons re synchronously, they can produce extracellular currents large enough to spread across different layers of conductive tissue (e.g., cerebrospinal uid (CSF), bone, and skin) and can be measured with voltage sensors on the scalp. These conductive layers of tissue form the volume conductor, which has the effect of mixing currents produced by different neural populations. As a result, an EEG sensor on the scalp picks up a voltage signal composed of contributions from multiple
A. Ojeda (*) Brain Vision LLC, Garner, NC, USA e-mail:
alejandro.ojeda@brainvision.com
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026 T.
Warbrick (ed.), The EEG Handbook,
https://doi.org/10.1007/978-3-032-20450-9_20
255
256 A. Ojeda
neural sources scattered over the cortical surface (and not only those directly underneath the sensor).
1
Typically, we use electrical models of the head to describe how neural signals propagate and mix on their way from their source to the EEG sensors. Depending on the assumptions, these models can furnish different variants of Maxwells equations, which we can solve for simulated, unitary source activities (see Table 20.1). This is know
n as solving the forward problem of the EEG. The solution of which is the so-called lead eld matrix, which encodes the contribution of a unitary current dipole normal to the cortical surface at the jth location to the ith scalp sensor. If source direction is included in the model, the contributions along the x, y, and z axes are also solved for and included in the lead eld.
Terminology
Volume conductor Electrically conductive tissue within the head through which
Volume conductor effect
Head model Electrical model of the head that describes how neural currents
Sensor space Component of the head model that denes the sensor positions
Source space Component of the head model where neural sources are dened.
Forward problem Given the equations derived from the head model, one nds the
Lead eld matrix Matrix that contains the solution to the forward problem for a
neural currents propagate (e.g., brain, cerebrospinal uid, bone, and skin).
It is the mixing of extracellular currents produced by distant as well
as nearby neural populations; therefore, a voltage sensor on the scalp picks up a signal composed of contributions from multiple neural sources.
propagate includes information and assumptions about the geometry and conductivity of the tissue.
with digitized and co-registered with a skin segmentation derived from the participants structural MRI.
For segmentation derived from structural MRI to constrain sources to the grey matter. Another constraint could be to conne the
sources to the cortical surface only.
voltage unitary current dipoles placed on the source space.
dense contribution of every unitary source to the given sensor. For
example, the element k
unitary current dipole onto the ith sensor.
from their sources to the sensors. The model usually
respect to the head. For example, sensor positions could be
example, in realistic head models, one can use grey matter
projection on the sensor space produced by simulated
array of sources. Each row in this matrix represents the
represents the voltage induced by the jth
i, j
1
In addition to cortical sources, there is evidence that subcortical sources contributed to the EEG
signals (Seeber et al., 2019).
20 EEG Source Analysis 257
Inverse problem Estimate the sources that generated a set of scalp voltage
EEG source imaging (ESI)
Equivalent current
(ECD)
dipole
Constraints Set of mathematical restrictions used to bias the solution of an
Regularization parameter
Euclidean norm Square root of the sum of squares of a vectors components. This
Frobenius norm It is an extension of the Euclidean norm to matrices, and it is used
Type I error When statistically testing a hypothesis, this is the error of
Multiple comparisons problem
Anatomical connectivity
Functional connectivity
Effective connectivity Indicates a causal relationship between brain regions.
measurements. The solution of the inverse problem of the EEG in a dense source
space.
In this case, the output of the inverse algorithm is a map of estimated current source density values. The resulting map can be rendered into an image (electrical tomography) or used for sub­sequent statistical analyses.
A dipole model that summarizes the electrical activity of an underlying source conguration of unknown spatial extent. Dipole tting algorithms utilize this model to determine the location of a few equivalent sources that best explain a given scalp topography.
inverse
algorithm toward biologically plausible source
congurations. A parameter used to obtain a unique inverse solution. It can be
interpreted weight assigned to the constraints.
norm is typically used to represent the reconstruction error of an inverse algorithm when solving for a single source vector.
to solving for multiple source vectors simultaneously.
rejecting a false positive.
When performing inference on the source space, the larger the number therefore, incurring a Type I error.
Indicates connectivity at the structural level (axons and bundles of axons
Indicates an instantaneous relationship (correlation) between brain
as a trade-off between the reconstruction error and the
represent the reconstruction error of an inverse algorithm when
the null hypothesis when it is true. This is also known as
of sources, the more likely it is to nd erroneous effects,
connecting different brain regions).
regions.
As long as the geometrical and electrical properties of the participants brain dont change signicantly (due to aging, injury, illness, etc.), the lead eld matrix can be calculated once and then used for estimating the neural sources that most likely generated an observed brain topography for an experimental condition of interest. This is known as solving the inverse problem of the EEG. The solution to this problem turns the EEG into a functional brain imaging modality by mapping the signals measured on the scalp onto anatomically relevant brain structures (see Fig.
Thereby allowing researchers to study behavioral, physiological, and
20.1).
cognitive responses linked to brain function and anatomy.
258 A. Ojeda
Table 20.1 Common modeling approaches to solve the forward problem of the EEG
Method Summary Pros Cons References Spherical The sources are dis-
BEM Boundary Element
FEM Finite Element
tributed over a sphere, typically tted to the cortical surface extracted from partic­ipants MRI segmen­tation. The layers representing the skull and scalp are modeled accordingly, increas­ing the radius param­eter and using a ratio of xed conductivities for each layer.
(BEM): MRI
Method segmentations are used to extract cortex, skull, and scalp sur­faces and dene com­partments of homogeneous isotro­pic conductivity. The sources are modeled as dipoles distributed over the cortical sur­face. Potential distri­butions are solved numerically on the boundary of each compartment.
Method (FEM): MRI is used to discretize the entire volume into small elements. The anisotropy and inho­mogeneity of the tis­sue can be modeled if Diffusion Tensor imaging (DTI) data is available. The poten­tials are solved numerically for each element.
Has analytical solu­tion.
Serves as benchmark for more sophisticated methods.
Captures the geome­try
of the head better than the spherical model.
Can accurately model
any head shape and manage heterogeneous and anisotropic conductivities.
Inaccurate. Hallez
Does not model inhomogeneous and anisotropic conduc­tivities and can be challenging to apply on surfaces with holes.
The method has sin­gularities that need to be handled by the numeric solver.
et al. (
2007)
Stenroos and Sarvas (
2012) and
Gramfort
al.
et (
2010)
Wolters et al. (
2007)
Having an EEG source signal representation in a well-dened anatomical space opens the door to statistical analyses of individuals and groups in the source space. Furthermore, ERP, frequency-domain, correlation, and causality studies can be
20 EEG Source Analysis 259
EEG voltage topography
arg min
Fig. 20.1 Schematic representation of EEG forward and inverse problems. The right-to-left arrow represents the EEG voltage generation model; source dipole currents on the right spread throughout the tissue, producing a voltage pattern on the left. The left-to-right arrow indicates the inverse problem: the estimation of the most likely current source pattern that could have generated the observed EEG voltage scalp pattern
Current source
map
carried out entirely in the source space. In the rest of the chapter, we will delve further into (1) the fundamentals of EEG source imaging, (2) statistical inference in the source space, and (3) source connectivity.

20.2 Fundamentals of EEG Source Imaging

In this chapter, we distingu ish between source imaging and source localization. Both families of algorithms start with a scalp topography typically representing EEG voltages (or EEG voltage components if a decomposition such as ICA was previ­ously applied (Debener et al.,
ithms output is an image (or map) of current source density values that can be
algor rendered in the source space. By source localization, however, we mean that the output is a discrete set of points indicating the center of the equivalent current dipoles that generated the topography. The key difference between the two is that for imaging, the source model is dened using a dense set of points distributed on a cortical surface or 3D grid. While for localization, the source model is sparse, and we identify the location of a handful of equivalent current dipoles that provide a plausible explanation for the scalp topography. See Fig.
The discrete
source model has been successfully applied to localize auditory and visual ERPs (Makeig et al., 1997, 2002; Debener et al., 2008). However, despite its simpli
city, the sparse nature of this source space co mplicates further statistical analysis because equivalent dipoles may not be found in the same locations across conditions or subjects. Usually, it requires working on clusters of sources and other techniques (Bigdely-Shamlo et al., beyond
the scope of this chapter, so from now on we will focus on distributed
EEG source Imaging (ESI).
2010)). Then, by source imaging, we mean that the
20.2.
2013). Dealing with this type of analysis is
260 A. Ojeda
Distributed
Forward solver
Spherical
BEM
FEM
Other
Inverse solver
LORETA
Beamformer
LASSO
Elastic-Net
SBL
Other
Fig. 20.2 Typical EEG source analysis owchart. The source model determines the subsequent pipeline. If the model is discrete, we need to go the route of dipole modeling and clustering and decide how to perform subsequent single-participant and group analyses on dipoles or clusters of dipoles. If the source model is distributed, we build a participant-speci c or template head model and calculate the lead eld using a forward solver (see Table estimate the source maps for selected time points of interest in the EEG (see Table can perform connectivity analysis on the source time series. Statistical analyses can be performed on sources or connectivity features at the participant level or in groups. An important caveat for group analyses is that unless we use the same template head model for all participants, we need to co-register participant source spaces to ensure that the features on which we perform statistics correspond to the same anatomical structures
Calculate lead
Estimate sources
Connectivity
estimation
field
Statistical analysis
Source model
20.1). Next, we use an inverse solver to
Discrete
Dipole fitting
Dipole clustering
Connectivity, spectral, and
statistical analyses
20.2). Then, we

20.3 Forward Problem

The typical ESI pipeline starts by dening a source space. If participant-specic MRI segmentations are available, a personalized head model is constructed. Otherwise, a generic one can be used for all participants based on a population template MRI. We will revisit the distinction between participant-specic and template-based head models later when we discuss the statistical analysis in the source space. The head model is built according to the forward solver of choice. For BEM, we need to provide the cortex, in-skull, out-skull, and skin surfaces, and we solve for sources placed on the vertices of the cortical surface. For FEM, grey matter (usually cortex only), CSF, and head segmentations are needed, and we solve for a 3D grid of sources placed on the grey matter. A spherical model solver can work on a surface
20 EEG Source Analysis 261
and a 3D grid source space. Next, we need to co-register the MRI-derived head model with the position of the sensors where voltages were measured. Notwith­standing that every approximation can introduce errors, if participant-specic sensor positions are missing, we can use a standard montage (that includes our sensors as a subset) placed on the head surface. Then we proceed to calculate the lead eld matrix.

20.4 Source Estimation

The relationship between scalp voltages and current sources in the brain is referred to as the measurement model, and can be expressed by a linear equation (Dale & Sereno,
where v represents a vector of sensor voltages at a given time point, K is the lead eld matrix, measurement noise usually assum ed normally distributed. The normality of the sensor noise rests on the assumption that gross artifacts were removed in pre-processing. So, to estimate the source vector j, we need to invert this equation (solve the inverse problem of the EEG, see Fig. probl assumptions, there are innite source congurations that can render the same scalp topography. To tackle this problem, we seek to estimate a source vector explains the measurements subject to additional constraints,
1993),
v ¼ Kj þ e ð20:1Þ
j is a vector of current density values on the source space, and e is the
20.1). A key aspect of this inverse
em is that there are many more sources than sensors, so without further
j that best
2
þ λP jðÞ
2
ð20:2 Þ
where the rst term
j ¼ arg min
( v
kk
2
) is the data t, taken as the square of the l2-norm
Kj
2
v Kj
kk
j
(Euclidean norm) of the reconstruction error, the second is a penalty function that biases the solution toward a source conguration that exhibits some desired features, and λ is a regularization parameter that regulates the weight given to the penalty with respect to the data t term. The regularization parameter can be determined by minimizing the generalized cross-validation curve (Golub et al., L-curve class
method (Hansen, 1992), or by grid search. Equation 20.2 belongs to a
of problems called Regularized Least Squares (RLS), and many popular
1979), the
inverse algorithms can be written in this framework, some of which we summarize in Table source
20.2. See Michel and Brunet (2019) for an extended list of popular EEG
analysis tools.
Thanks to
the normality of the measurement noise, even many Bayesian inversion methods can be represented in the RLS framework and vice versa. In the Bayesian framework, the error term is derived from a data likelihood function, the penalty
262 A. Ojeda
Table 20.2 Shows a non-exhaustive collection of popular inverse methods from the RLS family
Method Summary Pros Cons References MNE Minimum norm estimation
is obtained with
(MNE)
P j j
ðÞ¼
kk
2
.
2
wMNE Weighted MNE is
obtained
P jðÞ¼ Wjkk
matrix W acts as a depth
with
2 2
, where the
normalization operator.
LORETA Low-resolution electro-
magnetic (LORETA) is obtained
with P jðÞ L is a discrete Laplacian
tomography
2
, where
¼ Lj
kk
2
operator that enforces cor­relation among nearby sources.
It has a closed­form
solution.
It has a closed­form
solution and can recover supercial and deeper sources.
It has a closed-
solution and
form can recover supercial and deeper sources. Enforces smooth continuous solu­tions. It is robust
It is biased toward supercial sources and sensitive to noise.
It is sensitive to noise.
It is less sensitive
focal sources,
to and it tends to overestimate the spatial extent of the source activations.
Hamalainen and Ilmoniemi (1994)
Lin et al. (2006)
Pascual­Marqui et al. (1994)
to noise.
LASSO Least absolute shrinkage
selection
operator (LASSO) is obtained with P( j )
¼kjk
, where the l1-
1
norm enforces sparse solutions.
It is good for recovering and localized solutions.
focal
It has an iterative solution, thus usually more expensive to compute than other closed-form
Tibshirani (1996)
solution algo­rithms. It is sensi­tive to noise.
ENET Elastic net (ENET) is
obtained
P jðÞ¼ M
where M
with
2
j
þ M2j
kk
1
2
and M2 are lin-
1
k
ear operator. ENET is also found in the literature as MxNE (Mixed-norm Estimate).
Combines l1 and
2 norms to pro-
l
mote solutions
,
k
1
that are smooth and non-overlapping, therefore limiting the excessive
Like Lasso, it is solved
iteratively.
Vega­Hernández et al. (
2008)
and Gramfort
al. (
2012)
et
smoothness sometimes exhibit by LORETA while enforcing local correlations.
See Paz-Linares et al.
(2017) and Vega-Hernández et al. (2008) for a more in-depth exposition
comes from a prior density, and the regularization parameters, known as hyperparameters, are estimated within the same probabilistic framework. Although an exposition of Bayesian approaches is beyond the scope of this chapter, it is worth briey mentioning that the Sparse Bayesian Learning (SBL) family of algorithms