Electroencephalographic imaging of higher

brain function

Alan Gevins*, Michael E. Smith, Linda K. McEvoy, Harrison Leong and Jian Le
EEG Systems Laboratory and SAM Technology, 101 Spear Street, no. 204, San Francisco, CA 94105, USA

High temporal resolution is necessary to resolve the rapidly changing patterns of brain activity that
underlie mental function. Electroencephalography (EEG) provides temporal resolution in the millisecond
range. However, traditional EEG technology and practice provide insu¤cient spatial detail to identify
relationships between brain electrical events and structures and functions visualized by magnetic
resonance imaging or positron emission tomography. Recent advances help to overcome this problem by
recording EEGs from more electrodes, by registering EEG data with anatomical images, and by
correcting the distortion caused by volume conduction of EEG signals through the skull and scalp. In
addition, statistical measurements of sub-second interdependences between EEG time-series recorded
from di¡erent locations can help to generate hypotheses about the instantaneous functional networks that
form between di¡erent cortical regions during perception, thought and action. Example applications are
presented from studies of language, attention and working memory. Along with its unique ability to
monitor brain function as people perform everyday activities in the real world, these advances make
modern EEG an invaluable complement to other functional neuroimaging modalities.
Keywords: electroencephalography; MRI; attention; working memory; language

di¡erences between EEG and MEG are slight in compar-

1. INTRODUCTION
ison with their similarities.
The purpose of functional brain mapping is to measure The sensitivity of the EEG to changes in mental
patterns of neuronal activity associated with sensory, activity has been recognized since Berger (1929) reported
motor and cognitive functions, or with disease processes. a decrease in the amplitude of the dominant (alpha)
To be e¡ective in this, an imaging modality needs both rhythm of the EEG during mental arithmetic. In addition
near-millimetre precision in locating regions of activated to the type of tonic alterations in brain electrical activity
tissue and sub-second temporal precision for character- reported by Berger, EEG measurements of phasic,
izing changes in patterns of activation over time. stimulus-related changes in brain activity (such as evoked
Increasingly ¢ne anatomical resolution is available from potentials (EPs)) are well-suited to measuring sub-second
functional magnetic resonance imaging (fMRI) and component processes of sensory, motor and cognitive
positron emission tomography (PET) technologies. processes (Hillyard & Picton 1987; Regan 1989).
However, these technologies only provide indirect Measurements of the coherence, correlation or covariance
measures of neuronal electrical activity, and their of EEG time-series from di¡erent electrode sites help to
temporal resolution is too gross to resolve the rapidly generate hypotheses about the functional networks that
shifting patterns of activity that are characteristic of form between di¡erent cortical regions during these
actual neurophysiological processes. processes. The temporal resolution and sensitivity of the
The electroencephalogram (EEG), in contrast, has a EEG would seem to make it an ideal complement to
temporal resolution as ¢ne as the analogue-to-digital fMRI and PET. Unfortunately, the spatial detail provided
sampling rate used to record it (typically in the range by conventional EEG recordings has been so coarse that
1^5 ms). For simplicity we use the term EEG in a general it has only been possible to interpret EEGs meaningfully
sense to refer both to recordings of brain electrical with respect to underlying functional neuroanatomy at
activity and, except where noted, to recordings of brain the level of entire cortical lobes, if at all. Although the
magnetic activity called magnetoencephalograms or ability to infer the three-dimensional (3D) distribution of
MEGs. The nature of MEG recording technology and the electrical sources in the brain from scalp EEG recordings
relative strengths and weaknesses of the EEG and MEG has fundamental physical limits, the amount of spatial
approaches have been reviewed elsewhere (Cohen & information that can be recovered from the scalp-
Cu¤n 1991; Leahy et al. 1998; Williamson & Kaufman recorded EEG is frequently underestimated. Indeed, the
1987). Here it su¤ces to say that, from a broad perspec- very low level of spatial resolution obtained from conven-
tive that considers all neuroimaging modalities, the tional EEGs re£ects the fact that activity is measured at
only a small number of scalp sites during most routine
recordings; moreover, modern spatial signal-enhancing
*
Author for correspondence (alan@eeg.com). methods have not often been applied. This is rapidly

Phil. Trans. R. Soc. Lond. B (1999) 354, 1125^1134 1125 & 1999 The Royal Society
1126 A. Gevins and others Electroencephalographic imaging

Figure 1. High-resolution EEG mapping of somatosensory cortex activation. The image on the right is a 122-channel,
computer-enhanced topographic map of the potentials evoked by repetitive electrical stimulation of the left middle and right
index ¢ngers. The scalp-recorded EEG data have been mathematically deblurred and projected down to the external surface of
the subject's brain, which was reconstructed from his MRI. It is clear that localized areas of the left and right cerebral cortex
have been activated, corresponding to the stimulation of the ¢ngers on the right and left hands. The image on the left shows a
conventional, 18-channel topographic EEG map (viewed from above the head, with the nose at the top) evoked by the same
stimuli; there is an obvious lack of useful spatial information.

changing: an increasing number of laboratories have been

recording EEGs (and MEGs), from more than 100 loca-
tions, post-processing data to increase spatial detail, and
in some cases registering these data with other images of
brain structure and function.

2. ENHANCING THE SPATIAL RESOLUTION OF

THE EEG
Better spatial sampling is the ¢rst requirement for
extracting more detailed information about higher brain
function from scalp-recorded EEGs. The 19-channel
`10/20' montage of electrode placement (Jasper 1958)
commonly used in clinical and research EEG recordings
has an inter-electrode distance of ca. 6 cm on a typical
adult head. This spacing might be dense enough for
detecting signs of gross pathology or of di¡erentiating the
gross topography of EP components, but it is insu¤cient
for resolving the ¢ner topographical di¡erences that are
important in studying higher brain functions. By
increasing the number of electrodes to over 100, average
inter-electrode distances of ca. 2.5 cm can be obtained on

Figure 2. (Cont.) blurring e¡ect of the scalp and skull were

mathematically removed by using the deblurring method;
the resulting spatially sharpened data were projected on to
the cortical surface, which was constructed from the subject's
MRI. Activity at two instants before the button press (top)
and two instants after the button press (bottom) are plotted.
Activation both before and after the response is strongly
lateralized to the left hemisphere (the hemisphere
contralateral to the hand used to make the response). This
Figure 2. Deblurred movement-related potentials. EPs, time ¢gure shows lateralized activation of the precentral gyrus
locked to the onset of a button press response made with the before and immediately after a button press response is made;
right middle ¢nger, were recorded from a high-density (128 ca. 250 ms after the response, the focus of activation moves to
channel) electrode montage attached to the scalp. The the postcentral gyrus.

Phil. Trans. R. Soc. Lond. B (1999)

 Electroencephalographic imaging A. Gevins and others 1127

Figure 3. Validation of the

EEG deblurring method by
comparison with data recorded
directly from the cortical
surface. A 64-channel grid was
implanted in this epilepsy
patient as part of clinical care;
the thumb, middle ¢nger and
little ¢nger of the right hand
were stimulated separately.
Average evoked cortical
potentials recorded with the
intracerebral grid are shown at
the top; the computed cortical
potentials produced by the
deblurring method are at the
bottom. The main maxima and
minima are similarly located.
Improvements to deblurring
are being made to reduce the
noise ampli¢cation evident at
the edge of the recording array.

Figure 4. Deblurred EPs during reading. The processing of syntactic (grammatical) words di¡ers from the processing of content
(semantic) words. The image on the right shows the result of deblurring an EP that reached peak amplitude ca. 445 ms after the
presentation of pronouns that were the subject of a short sentence. It had a focus of activity over the left lateral frontal
operculum, and a duration of ca. 200 ms. The image on the left illustrates the deblurred EP elicited by presentation of content
words at the same point in time. Both classes of words show common activation over the left parietal region and over the
calcarine cortex during this interval. Unlike PET or fMRI, with EEG it is usually unnecessary either to combine data from
several subjects or to subtract an experimental condition from a presumed control so as to measure the neuronal patterns of
mental processes.

an average adult head. This is within the range of the tion amounts to a spatial low-pass ¢ltering, which causes
typical cortex-to-scalp point spread function (i.e. the size a blurring of the potential distribution at the scalp. In
of the scalp representation of a small, discrete cortical recent years a number of spatial enhancement methods
source) (Gevins et al. 1990). have been developed for reducing this distortion.
For electrical (but not magnetic) recordings, the useful- The simplest and most widely accessible of these
ness of such increased spatial sampling remains limited methods is the spatial laplacian operator, usually referred
by the distortion of neuronal potentials as they are to as the laplacian derivation (LD). It is computed as the
passively conducted through the highly resistive skull second derivative in space of the potential ¢eld at each
(Gevins et al. 1991; Van den Broek et al. 1998). This distor- electrode. The LD is thought to be proportional to the

Phil. Trans. R. Soc. Lond. B (1999)

1128 A. Gevins and others Electroencephalographic imaging

current entering and exiting from the scalp at each problems of locating generator sources are as severe for
electrode site (Nunez 1981; Nunez & Pilgreen 1991), and is MEG as they are for EEG (see below). Further, the cost
independent of the location of the reference electrode of MEG technology is at least an order of magnitude
used for recording. It is relatively insensitive to signals greater than that required for EEG studies, and the asso-
that are common to the local group of electrodes used in ciated infrastructure required to perform MEG studies is
the computation, and is thus more sensitive to high more complex and in£exible. Thus, for most laboratories,
spatial frequency local cortical potentials. A simple and for some applications (particularly those in which a
method of computing the LD assumes that electrodes are subject's head cannot be immobilized, such as long-term
equidistant and at right angles to each other, an approxi- monitoring or ambulatory recordings), MEG does not
mation that is only reasonable at a few scalp locations provide a viable alternative to EEG recordings.
such as the vertex. A more accurate approach is based on
measuring the actual 3D position of the electrodes and
3. VERIFICATION AND APPLICATION OF HIGH-
using 3D spline functions to compute the LD over the
RESOLUTION EEG IN EXPERIMENTAL STUDIES
actual shape of a subject's head (Le et al. 1994). The main
shortcoming of the LD is that it unrealistically assumes Exploratory studies of deblurring and other high-
that the skull has the same thickness and conductivity resolution EEG techniques focused on the spatial
everywhere, which limits the improvement in spatial enhancement of sensory EPs, in which a great deal of a
detail that the method can achieve. priori knowledge exists concerning their underlying
This shortcoming of the LD can be ameliorated by neural generators (Gevins et al. 1994b). For example,
using a realistic model of each subject's head to correct ¢gure 1 contrasts the spatial detail obtained for the somato-
the EEG potential ¢eld locally for distortion resulting sensory stimulation of ¢ngers on each hand by using
from conduction to the scalp. One such method is called conventional 18-electrode EEG mapping methods with
¢nite element deblurring. It provides a computational that obtained using 124 electrodes and deblurring. The
estimate of the electrical potential ¢eld near the cortical deblurred somatosensory responses clearly isolate activity
surface by using a realistic mathematical model of volume to the region of the central sulcus in each hemisphere.
conduction through the skull and scalp to project scalp- Similar localization has been obtained with movement-
recorded signals downwards (Gevins et al. 1991, 1994b; Le related potentials. Figure 2 illustrates the results of
& Gevins 1993). Each subject's magnetic resonance image deblurring potentials locked in time to a button-press
(MRI) is used to construct a realistic model of his or her response made with the right hand. The major foci of
head in the form of many small tetrahedral elements activity occur in the contralateral (left hemisphere) in the
representing the tissues of scalp, skull and brain. By somatomotor region of the pre-central and post-central
assigning each tissue a conductivity value, it is possible to gyri. Demonstrations such as these serve to verify the
calculate the potential at all ¢nite element vertices by reasonableness of the approach, but a better validation is
using Poisson's equation. Given that the actual conduc- obtained by a comparison of the deblurred potentials
tivity value of each of these ¢nite elements is unknown, a with subdural grid recordings in epileptic patients under-
constant value is used for the ratio of scalp to skull going evaluation for ablative surgery. So far these
conductivity; the conductivity of each ¢nite element is set validation studies have produced a reasonable degree of
by multiplying this constant by the local tissue thickness agreement between the deblurred potentials and those
as determined from the MRI. Thus, even though true measured directly at the cortical surface (¢gure 3).
local conductivity is unknown, the procedure is well- Recent developments suggest that high-resolution EEG
behaved with respect to this source of uncertainty, methods are useful tools in the experimental analysis of
because it accounts successfully for relative conductivity higher-order brain functions. For example, spatial
variation due to regional di¡erences in scalp and skull enhancement of EEGs related to component processes in
thicknesses. reading has yielded results that are highly consistent with
In initial applications, the deblurring method has been current knowledge of the functional neuroanatomy
shown to be reliable and more accurate than the LD thought to be involved with visual pattern recognition
(Gevins et al. 1991, 1994b). This improvement occurs at and language functions (Gevins et al. 1995a). In one study,
the expense of obtaining and processing each subject's EPs were elicited during a simple cued matching task
structural MRI. Although deblurring can substantially requiring one of four types of matching judgement:
improve the spatial detail provided by scalp recorded graphic (visual identity of unfamiliar non-letter character
EEGs, it does not provide conclusive information about strings), phonemic (homophonic pseudo-words), semantic
the location of generating sources. Nevertheless, the (antonymy) and grammatical (noun ^ verb agreement).
improved spatial detail facilitates the formation of more Each trial of the task began with a cue that indicated
speci¢c hypotheses about the distribution of active which of the four conditions to expect. After 1s this cue
cortical areas during perceptual, cognitive and motor was followed by the ¢rst stimulus, which in turn was
tasks. followed 1s later with the comparison stimulus. Several
The issue of blurring of brain signals by the skull can striking between-condition di¡erences were evident in
be largely avoided by recording the magnetic rather than highly localized EP patterns. For example, larger ampli-
the electrical ¢elds of the brain, because the skull has no tude EP waves occurred in the grammatical condition
e¡ect on magnetic ¢eld topography. However, this trans- (relative to the other language tasks) at 445 ms after the
parency does not eliminate the need for using a high ¢rst stimulus; the duration of the wave was only
density of sensors to make an accurate map of the spatial ca. 200 ms. These potentials were largest at scalp locations
topography of brain magnetic ¢elds. Furthermore, the near the presumed region of Broca's language area in the

Phil. Trans. R. Soc. Lond. B (1999)

 Electroencephalographic imaging A. Gevins and others 1129

frontal cortex of the left hemisphere (¢gure 4), and their

4. IDENTIFYING THE SOURCES OF EEG PHENOMENA
task correlates are consistent with the postulated
functional neuroanatomy of this region. Thus, the Neither the LD, nor more-advanced EEG spatial
functional localization of cognitive processes inferred enhancement algorithms such as deblurring, nor MEG
from spatially enhanced and anatomically registered recordings, provide any conclusive 3D information about
electrophysiological measurements can be compared with where the source of a scalp-recorded signal lies in the
the results of lesion studies and other neuroimaging brain. In some cases, such as when healthy subjects
approaches. As a complement to these approaches, the perform di¤cult cognitive tasks and strong signals are
¢ne-grained temporal resolution of EP measurements, in recorded over areas of association cortex (i.e. dorsolateral
combination with improved topographic detail, adds prefrontal, superior and inferior parietal, inferotemporal
valuable insights gained by characterizing both the and lateral temporal), the hypothesis that EEG potentials
regionalization of functions and the sub-second dynamics are generated in these areas is the most plausible.
of their engagement. However, counter-examples can always be presented. In
Modern EEG methods have also been used to study addition to a visual examination of the potential ¢eld
sub-second and multi-second distributed neural processes distribution, another procedure for generating hypotheses
associated with working memory, the cognitive function about the neuroanatomical loci responsible for generating
of creating a temporary internal representation of infor- neuroelectric events measured at the scalp is called dipole
mation during focused thought (Gevins et al. 1996; modelling (Fender 1987; Scherg & Von Cramon 1985).
McEvoy et al. 1998). In task conditions that placed a high Dipole modelling uses iterative numerical methods to ¢t a
load on working memory functions, subjects were asked to mathematical representation of a focal, dipolar current
decide whether the stimulus on each trial matched either source, or collection of such sources, to an observed scalp-
the verbal identity or the spatial location of a stimulus recorded EEG or MEG ¢eld.
occurring three trials previously (ca. 13.5 s ago). This Source modelling does not, in general, provide a
required subjects to concentrate on maintaining a unique and physically correct answer about where in the
sequence of three letter names or three spatial locations brain activity recorded at the scalp is generated, because
concurrently; they had to update that sequence on each solving for the source of an EEG or MEG distribution
trial by remembering the most recent stimulus and could recorded at the scalp is a mathematically ill-conditioned
drop the stimulus from four trials back. In two `inverse problem' that has no unique solution; additional
corresponding control conditions, only the verbal identity information and/or assumptions are required to permit
or spatial location of the ¢rst stimulus had to be remem- choice between candidate source models. Although some
bered. Both spatial and verbal working memory tasks of this a priori information is obvious and harmless (i.e.
produced highly localized momentary modulation of EPs that the potentials must arise from the space occupied by
over prefrontal cortical areas relative to control the brain), other assumptions border on presupposing
conditions, with deblurred voltage maxima approxi- unknown information (i.e. that the potentials arise only
mately over Brodmann's areas 9, 45 and 46 (¢gure 5). from the cortex, or that the number of active cortical
These brief (ca. 50 ms and 200 ms) events occurred in areas is known).
parallel with a sustained EP wave and were maximal One simple, convenient, and potentially clinically
over the superior parietal lobe and the supramarginal useful approach for potentials elicited by simple sensory
gyrus, with a slight right-hemisphere predominance. It stimulation is to assume that the scalp potential pattern
began ca. 200 ms after stimulus onset, returned to near arises from a single point dipole source, as shown in
the baseline by ca. 600 ms after stimulus in control ¢gure 6. Although not anatomically or physiologically
conditions, and was sustained up to ca. 1s or longer in the realistic, such simple models can sometimes be useful for
working memory conditions. The sub-second EP e¡ects locating the centre of mass of primary sensory cortex and
occurred in conjunction with multi-second changes in the hence major functional landmarks such as the central
ongoing EEG, of which the theta-band power focused sulcus. When justi¢ed by simple voltage topography
over the midline frontal cortex is shown in ¢gure 5 (¢gure 7), models of this sort can also be useful for
(Gevins et al. 1997; Smith et al. 1999). These EEG ¢ndings generating initial hypotheses about the possible sources
might provide the ¢rst direct evidence in a single experi- underlying other phenomena.
ment to support the idea that the various types of atten- Most complex scalp-recorded electrophysiological
tion are associated with neural processes with distinct phenomena are poorly approximated by a single dipole
time-courses in distinct neuronal populations. The source model, and obtaining estimates of the strength and
increased theta-band power might be a marker of the 3D locations of the underlying neuronal generators when
continuous focused attention required to perform the task, there are multiple, time-overlapped active sources has
and might re£ect the engagement of the anterior cingulate widely recognized practical and theoretical di¤culties
gyrus, a conjecture supported by dipole modelling (Miltner et al. 1994; Mosher et al. 1992). Intense e¡ort is
(Gevins et al. 1997). In contrast, the momentary attention being allocated to the development of improved methods
required for scanning and updating the representations of for source analysis for electrical phenomenon that are
working memory might be re£ected in increased EP likely to arise from multiple and/or distributed sources
peaks over lateralized regions of dorsolateral prefrontal (Gorodnitsky et al. 1995; Grave de Peralta-Menendez &
cortex, whereas the maintenance of a representation of Gonzalez-Andino 1998; Koles 1998; Mosher et al. 1992;
the stimuli being remembered might be re£ected in the Tesche et al. 1995; Wang et al. 1993). Even so, regardless of
parietally maximal EP wave and other concomitant which method is used to formulate them, such source
changes in the EEG (Gevins et al. 1996, 1997). generator hypotheses must ultimately be veri¢ed

Phil. Trans. R. Soc. Lond. B (1999)

1130 A. Gevins and others Electroencephalographic imaging

the neural dynamics of thought processes. Even the

simplest cognitive tasks require the functional coordina-
tion of a large number of widely distributed specialized
brain systems. A simple response to a sensory stimulus
involves the coordination of sensory, association and other
areas that prepare for, register and analyse the stimulus,
the motor systems that prepare for and execute the
response, and other distributed neuronal networks. These
distributed networks serve to allocate and direct
attentional resources to the stimulus, to relate the
stimulus to internal representations of self and environ-
ment to decide what action to take, to initiate or inhibit
the behavioural response, and to update internal
representations after receiving feedback about the result
of the action. In the ongoing EEG, hypotheses about
functional interactions between cortical regions are some-
times drawn from measurements of statistical inter-
relationships between time-series recorded at di¡erent
sites. These can be quanti¢ed by various measures of
spectral, wave shape or information-theory similarity,
including spectral coherence (Walter 1963), correlation
(Brazier & Casby 1952; Gevins et al. 1981, 1983; Livanov
1977), covariance (Gevins et al. 1987, 1989a,b), information
measures (Callaway & Harris 1974; Mars & Lopes da
Silva 1987), nonlinear regression (Lopes da Silva et al.
1989) and multichannel time-varying autoregressive
Figure 5. Deblurred EPs and ongoing EEG related to modelling (Gersch 1987).
sustained focused attention. High-resolution EEG methods Some of the above methods can be used to characterize
have made it possible to measure both sub-second phasic and the spatio-temporal relationships between sub-second EP
multi-second tonic regional brain activity simultaneously components. Because the EP waveform delineates the
during the performance of cognitive tasks. In this experiment time-course of event-related mass neural activity of a
a sequence of increased sub-second EP peaks and waves was
neuronal population, the coordination of two or more
observed over frontal (¢rst and second columns) and parietal
(third column) cortices during a di¤cult working memory
populations during task performance should be signalled
task, in comparison with control conditions with lower by a consistent relationship between the morphology of
working memory requirements. These sub-second changes the EP waveforms emitted by these populations, with
in the working memory tasks were accompanied by consistent time-delay (Gevins & Bressler 1988). If the
longer-lasting (4 s) increases in ongoing EEG theta-band relationships are linear, as they often seem to be, this
power (fourth column). These EEG ¢ndings suggest coordinated activity might be measured by the lagged
that various types of attention are associated with neural correlation or covariance between the EPs, or segments of
processes that have distinct time-courses in distinct neuronal EPs, from di¡erent regions (Gevins et al. 1987, 1989a,b).
populations. The amplitude scale is constant across One such measure of this type of process is referred to as
experimental conditions within each column; the EP scale is an EP covariance (EPC). (Of course, a signi¢cant
voltage, the EEG scale is z-scored spectral power.
covariance of this type is only a measure of statistical
association and does not map the actual neuronal
independently. In rare cases this might be done in patient pathways of interaction between functionally related
populations in the context of invasive recordings populations.) Studies of the neurogenesis of EPCs are still
performed for clinical diagnostic purposes. More in their infancy (Bressler et al. 1993; Gevins et al. 1994a),
commonly, another type of imaging modality, such as and any interpretations of EPCs in terms of the under-
PETor fMRI, has to be employed. Indeed, one promising lying neural processes that generate them must therefore
approach to this issue is to use information about the be made very cautiously. (It is noted, however, that EPC
cortical regions activated by a task as mapped by 3D results have so far been highly consistent with the known
functional neuroimaging methods such as fMRI or PET large-scale functional neuroanatomy of frontal, parietal
to constrain source models, and to derive information and temporal association cortices.) Meanwhile, EPCs
about the spatio-temporal dynamics of those sources from have provided fascinating glimpses of the complex,
EP measurements (George et al. 1995; Heinze et al. 1994; rapidly shifting, distributed neuronal processes that
Mangun et al. 1998; Sereno 1998; Simpson et al. 1995). underlie simple cognitive tasks.
The EPC technique has yielded its most interesting
results as a tool for studying preparatory attentional
5. DISTRIBUTED FUNCTIONAL NETWORKS OF
networks, the changes in brain activity associated with
SIMPLE COGNITIVE TASKS
readiness for an impending event or action. For example,
Independently of whether de¢nitive knowledge of subjects in one experiment (Gevins et al. 1987, 1989a,b)
source con¢gurations exists, changes in the spatial distri- performed a task that required graded ¢nger-pressure
bution of EEG phenomena can be used to characterize responses with either the right hand or the left hand

Phil. Trans. R. Soc. Lond. B (1999)

 Electroencephalographic imaging A. Gevins and others 1131

Figure 6. Localization of an EEG

dipole model in the somatosensory
cortex of the right hemisphere from
scalp-recorded data evoked in
response to transient electrical
stimulation of the left index ¢nger.
This popular type of source
generator localization modelling
produces anatomically plausible
results in the case of simple sensory
stimulation.

Figure 8. Preparatory EPC patterns preceding accurate and

inaccurate responses. EPCs involving left frontal, midline
precentral and left central and parietal electrode sites are
prominent in patterns preceding accurate responses (by
0.5^1 s) (left). The number and magnitude of EPCs that
precede inaccurate responses are smaller (right).

patterns di¡ered according to the hand that subjects

expected to use. Figure 8 shows right-hand preparatory
EPCs for seven subjects for those trials for which the
response (ca. 0.5^1s later) was subsequently either accurate
Figure 7. Deblurred frontal midline theta EEG activity
and localization of the corresponding source model in the
or inaccurate. The set of subsequently accurate trials are
region of the anterior cingulate cortex. Topographic data characterized by covariances of the left prefrontal elec-
correspond to the di¤cult working memory task condition trode with electrodes overlying the same motor, somato-
depicted in ¢gure 5 (Gevins et al. 1997). The data were sensory and parietal areas that were involved in actual
processed with the deblurring method, and the spatially response execution. (Simultaneous measurement of £exor
sharpened results were projected on to the cortical surface, digitori muscle activity showed that the ¢nger that would
which was constructed from the subject's MRI. The subsequently respond was not active during the prepara-
upwards-orientated arrow superimposed on the midline tory interval.) The preparatory patterns preceding inac-
sagittal image depicts the localization of a point dipole curate responses di¡ered markedly from those preceding
source model for these data. accurate responses, with fewer EPCs between the left
frontal site and other electrodes. Such results suggest that
proportional to visual numeric stimuli from one to nine. one important role of frontal lobe integrative mechanisms
The hand to be used was cued 1s before the stimulus. A is the anticipatory scheduling and coordination of the acti-
375 ms EPC analysis window spanned the interval vation of those specialized brain regions that will partici-
preceding the stimulus number so as to measure how EP pate in a future cognitive event.

Phil. Trans. R. Soc. Lond. B (1999)

1132 A. Gevins and others Electroencephalographic imaging

With respect to measuring spatial interrelationships territory of how brains think when performing everyday
between measurements of the ongoing human EEG, there activities in the real world (Gevins et al. 1995b).
is currently a renewed interest in coherent brain activity
in the 40 Hz range or in a broader high-frequency We thank our many colleagues at EEG Systems Laboratory and
`gamma' band (ca. 35^100 Hz). This interest derives SAM Technology, past and present, for their contributions to
largely from studies of neuronal activity in animal the work described here. This research was supported by grants
from The Air Force O¤ce of Scienti¢c Research, The National
preparations that suggest that 40 Hz or gamma-band Institute of Mental Health, The National Institute of Neuro-
oscillations might represent the action of an integrative logical Disorders and Stroke, The National Science Foundation,
mechanism for binding the distributed neural populations The National Aeronautics and Space Administration, The Air
activated by a sensory experience into an integrated Force Research Laboratory, The O¤ce of Naval Research, and
percept (Freeman 1975; Freeman & Skarda 1985; Gray et The National Institute of Alcoholism and Alcohol Abuse of the
al. 1989). It has been suggested that such coherent activity United States Federal Government.
is measurable over distributed regions of the human scalp
with EEG (Sheer 1989) or MEG (Llinäs & Ribary 1993). REFERENCES
However, in studies of brain electrical activity recorded
directly from the cortical surface in human neurosurgical Berger, H. 1929 Ûber das Elektroenzephalogramm des
Menschen. Arch. Psychiat. 87 (Nervenkrankheiten), 527^570.
patients while they performed somatosensory discrimina-
Brazier, M. A. B. & Casby, J. U. 1952 Crosscorrelation and
tion tasks, no spatial modulation of gamma-band activity autocorrelation studies of electroencephalographic potentials.
was found to be related to stimulus type in a simple Electroencephalogr. Clin. Neurophysiol. 4, 201^211.
somatosensory discrimination task (Menon et al. 1995), Bressler, S. L., Coppola, R. & Nakamura, R. 1993 Episodic
although such evidence was abundant in the lower- multiregional cortical coherence at multiple frequencies
frequency EPs (Gevins et al. 1994a; Menon et al. 1995). during visual task performance. Nature 366, 153^155.
Callaway, E. & Harris, P. 1974 Coupling between cortical poten-
tials from di¡erent areas. Science 183, 873^875.
6. CONCLUSIONS Cohen, D. & Cu¤n, B. N. 1991 EEG versus MEG localization
The neurophysiology of mentation involves the rapid accuracy: theory and experiment. Brain Topogr. 4, 95^103.
coordination of processes in widely distributed cortical Fender, D. H. 1987 Source localization of brain electrical
activity. In Methods of analysis of brain electrical and magnetic
and subcortical areas. The electrical signals that accom-
signals, vol. 1 (ed. A. S. Gevins & A. Rëmond), pp. 355^403.
pany higher cognitive functions are subtle, spatially Amsterdam: Elsevier.
complex, and change both in a tonic multi-second fashion Freeman, W. J. 1975 Mass action in the nervous system. New York:
and phasically in sub-second intervals in response to Academic Press.
environmental demands and internal representations of Freeman, W. J. & Skarda, C. A. 1985 Spatial EEG patterns,
environment and self. No one brain-imaging technology non-linear dynamics and perception: the neo-Sherringtonian
is currently capable of providing both near-millimetre view. Brain Res. Rev. 10, 147^175.
precision in localizing regions of activated tissue and sub- George, J. S., Aine, C. J., Mosher, J. C., Schmidt, D. M.,
second temporal precision for characterizing changes in Ranken, D. M., Schlitt, H. A., Wood, C. C., Lewine, J. D.,
patterns of activation over time. However, by combining Sanders, J. A. & Belliveau, J. W. 1995 Mapping function in
the human brain with magnetoencephalography, anatomical
several technologies it seems possible to achieve this
magnetic resonance imaging, and functional magnetic
degree of ¢ne spatio-temporal resolution. Modern high- resonance imaging. J. Clin. Neurophysiol. 12, 406^431.
resolution EEG is especially well-suited for monitoring Gersch, W. 1987 Non-stationary multichannel time series
rapidly changing regional patterns of neuronal activation analysis. In Methods of analysis of brain electrical and magnetic
accompanying purposive behaviours, whereas fMRI signals, vol. 1 (ed. A. S. Gevins & A. Rëmond), pp. 261^296.
seems ideal for precisely determining their 3D localiza- Amsterdam: Elsevier.
tion and distribution. It is a topic of current research to Gevins, A. S. & Bressler, S. L. 1988 Functional topography of
determine how to combine EEG and fMRI data from the the human brain. In Functional brain imaging (ed. G.
same subjects doing the same tasks. Pfurtscheller), pp. 99^116. Toronto: Hans Huber.
Finally, an obvious but often unconsidered feature of Gevins, A. S., Doyle, J. C., Cutillo, B., Scha¡er, R. E.,
EEG technology is worth mentioning: its extreme Tannehill, R. S., Ghannam, J. H., Gilcrease, V. A. & Yeager,
C. L. 1981 Electrical potentials in human brain during
compactness and simplicity. This fact has important
cognition: new method reveals dynamic patterns of correla-
practical considerations that are frequently not considered tion. Science 213, 918^922.
in scienti¢c discussions of brain-mapping technology. For Gevins, A. S., Scha¡er, R. E., Doyle, J. C., Cutillo, B.,
example, EEG has the potential to serve as a sensitive, Tannehill, R. S. & Bressler, S. L. 1983 Shadows of thought:
low-cost, portable monitor for clinical assessment and shifting lateralization of human brain electrical patterns
other applications (Gevins et al. 1998; A. Gevins and during brief visuomotor task. Science 220, 97^99.
M. E. Smith, unpublished data). The compactness of Gevins, A. S., Morgan, N. H., Bressler, S. L., Cutillo, B. A.,
EEG technology also means that, unlike all other White, R. M., Illes, J., Greer, D. S., Doyle, J. C. & Zeitlin,
functional neuroimaging modalities (which require G. M. 1987 Human neuroelectric patterns predict perfor-
massive machinery, large teams of technicians and mance accuracy. Science 235, 580^585.
Gevins, A. S., Bressler, S. L., Morgan, N. H., Cutillo, B. A.,
complete immobilization of the subject), EEGs can be
White, R. M., Greer, D. & Illes, J. 1989a Event-related
collected from an ambulatory subject who is literally covariances during a bimanual visuomotor task. I. Methods
wearing the entire recording apparatus. This feature of and analysis of stimulus and response-locked data.
EEGs will facilitate research into the still uncharted Electroencephalogr. Clin. Neurophysiol. 74, 58^75.

Phil. Trans. R. Soc. Lond. B (1999)

 Electroencephalographic imaging A. Gevins and others 1133

Gevins, A. S., Bressler, S. L., Morgan, N. H., Cutillo, B. A., Le, J. & Gevins, A. S. 1993 Method to reduce blur distortion
White, R. M., Greer, D. & Illes, J. 1989b Event-related from EEGs using a realistic head model. IEEE Trans. Biomed.
covariances during a bimanual visuomotor task. II. Engng 40, 517^528.
Preparation and feedback. Electroencephalogr. Clin. Neurophysiol. Le, J., Menon, V. & Gevins, A. 1994 Local estimate of the
74, 147^160. surface Laplacian derivation on a realistically shaped scalp
Gevins, A. S., Brickett, P., Costales, B., Le, J. & Reutter, B. surface and its performance on noisy data. Electroencephalogr.
1990 Beyond topographic mapping: towards functional ^ Clin. Neurophysiol. 92, 433^441.
anatomical imaging with 124-channel EEGs and 3-D MRIs. Leahy, R. M., Mosher, J. C., Spencer, M. E., Huang, M. X. &
Brain Topogr. 3, 53^64. Lewine, J. D. 1998 A study of dipole localization accuracy for
Gevins, A. S., Le, J., Brickett, P., Reutter, B. & Desmond, J. MEG and EEG using a human skull phantom.
1991 Seeing through the skull: advanced EEGs use MRIs to Electroencephalogr. Clin. Neurophysiol. 107, 159^173.
accurately measure cortical activity from the scalp. Brain Livanov, M. N. 1977 Spatial organization of cerebral processes. New
Topogr. 4, 125^131. York: Wiley.
Gevins, A. S., Cutillo, B. A., Desmond, J. E., Ward, M. & Llinäs, R. & Ribary, U. 1993 Coherent 40-Hz oscillations
Bressler, S. L. 1994a Subdural grid recordings of distributed characterize dream state in humans. Proc. Natl Acad. Sci. USA
neocortical networks involved with somatosensory discrimi- 90, 2078^2081.
nation. Electroencephalogr. Clin. Neurophysiol. 92, 282^290. Lopes da Silva, F., Pijn, J. P. & Boeijinga, P. 1989
Gevins, A. S., Le, J., Martin, N. K., Reutter, B., Desmond, J. & Interdependence of EEG signals: linear vs. nonlinear associa-
Brickett, P. 1994b High resolution EEG: 124-channel tions and the signi¢cance of time delays and phase shifts.
recording, spatial deblurring and MRI integration methods. BrainTopogr. 2, 9^18.
Electroencephalogr. Clin. Neurophysiol. 90, 337^358. McEvoy, L. K., Smith, M. E. & Gevins, A. 1998 Dynamic cortical
Gevins, A. S., Cutillo, B. A. & Smith, M. E. 1995a Regional networks of verbal and spatial working memory: e¡ects of
modulation of high resolution evoked potentials during verbal memory load and task practice. Cerebr. Cortex 8, 563^574.
and nonverbal matching tasks. Electroencephalogr. Clin. Mangun, G. R., Buonocore, M. H., Girelli, M. & Jha, A. P.
Neurophysiol. 94, 129^147. 1998 ERP and fMRI measures of visual spatial selective
Gevins, A. S., Leong, H., Du, R., Smith, M. E., Le, J., attention. Hum. Brain Mapp. 6, 383^389.
DuRousseau, D., Zhang, J. & Libove, J. 1995b Towards Mars, N. J. & Lopes da Silva, F. H. 1987 EEG analysis methods
measurement of brain function in operational environments. based on information theory. In Methods of analysis of brain
Biol. Psychol. 40, 169^186. electrical and magnetic signals, vol. 1 (ed. A. S. Gevins & A.
Gevins, A. S., Smith, M. E., Le, J., Leong, H., Bennett, J., Rëmond), pp. 297^307. Amsterdam: Elsevier.
Martin, N., McEvoy, L., Du, R. & Whit¢eld, S. 1996 High Menon, V., Freeman, W. J., Cutillo, B. A., Desmond, J. E.,
resolution evoked potential imaging of the cortical dynamics Ward, M., Bressler, S. L., Laxer, K., Barbero, N. & Gevins,
of human working memory. Electroencephalogr. Clin. A. S. 1995 Spatiotemporal correlations in human gamma
Neurophysiol. 98, 327^348. band electrocorticograms. Electroencephalogr. Clin. Neurophysiol.
Gevins, A., Smith, M. E., McEvoy, L. & Yu, D. 1997 High 98, 89^102.
resolution EEG mapping of cortical activation related to Miltner, W., Braun, C., Johnson Jr, R., Simpson, G. V. &
working memory: e¡ects of task di¤culty, type of processing, Ruchkin, D. S. 1994 A test of brain electrical source analysis
and practice. Cerebr. Cortex 7, 374^385. (BESA): a simulation study. Electroencephalogr. Clin.
Gevins, A., Smith, M. E., Leong, H., McEvoy, L., Whit¢eld, S., Neurophysiol. 91, 295^310.
Du, R. & Rush, G. 1998 Monitoring working memory load Mosher, J., Lewis, P. & Leahy, R. 1992 Multiple dipole
during computer-based tasks with EEG pattern recognition modeling and localization from spatio-temporal MEG data.
methods. Hum. Factors 40, 79^91. IEEE Trans. Biomed. Engng 39, 541^557.
Gorodnitsky, I. F., George, J. S. & Rao, B. D. 1995 Nunez, P. L. 1981 Electric ¢elds in the brain: the neurophysics of EEG.
Neuromagnetic source imaging with FOCUSS: a recursive New York: Oxford University Press.
weighted minimum norm algorithm. Electroencephalogr. Clin. Nunez, P. L. & Pilgreen, K. L. 1991 The spline-Laplacian in
Neurophysiol. 95, 231^251. clinical neurophysiology: a method to improve EEG spatial
Grave de Peralta-Menendez, R. & Gonzalez-Andino, S. L. 1998 resolution. J. Clin. Neurophysiol. 8, 397^413.
A critical analysis of linear inverse solutions to the Regan, D. 1989 Human brain electrophysiology. New York: Elsevier.
neuroelectric inverse problem. IEEE Trans. Biomed. Engng 45, Scherg, M. & Von Cramon, D. 1985 Two bilateral sources of the
440^448. late AEP as identi¢ed by a spatio-temporal dipole model.
Gray, C. M., Engel, A. K. & Singer, W. 1989 Oscillatory Electroencephalogr. Clin. Neurophysiol. 62, 32^44.
responses in cat visual cortex exhibit inter-columnar synchro- Sereno, M. I. 1998 Brain mapping in animals and humans. Curr.
nization which re£ects global stimulus properties. Nature 338, Opin. Neurobiol. 8, 188^194.
334^337. Sheer, D. E. 1989 Sensory and cognitive 40 Hz event-related
Heinze, H. J., Mangun, G. R., Burchert, W., Hinrichs, H., potentials. In Brain dynamics, vol. 2 (ed. E. Basar & T. H.
Scholz, M., Munte, T. F., Gos, A., Scherg, M., Johannes, S. & Bullock), pp. 339^374. Berlin: Springer.
Hundeshagen, H. 1994 Combined spatial and temporal Simpson, G. V., P£ieger, M. E., Foxe, J. J., Ahlfors, S. P.,
imaging of brain activity during visual selective attention in Vaughan, J., H.G., Hrabe, J., Ilmoniemi, R. J. & Lantos, G.
humans. Nature 372, 543^546. 1995 Dynamic neuroimaging of brain function. J. Clin.
Hillyard, S. A. & Picton, T. W. 1987 Electrophysiology of cogni- Neurophysiol. 12, 432^449.
tion. In Handbook of physiology, vol. 5, 2nd edn (ed V. B. Smith, M. E., McEvoy, L. K. & Gevins, A. 1999
Mountcastle), pp. 519^584. Bethesda, MD: American Neurophysiological indices of strategy development and skill
Physiological Society. acquisition. Cogn. Brain Res. 7, 389^404.
Jasper, H. H. 1958 The ten ^ twenty electrode system of the Tesche, C. D., Uusitalo, M. A., Ilmoniemi, R. J., Huotilainen,
International Federation. Electroencephalogr. Clin. Neurophysiol. M., Kajola, M. & Salonen, O. 1995 Signal ^ space projections
10, 371^375. of MEG data characterize both distributed and well-
Koles, Z. J. 1998 Trends in EEG source localization. localized neuronal sources. Electroencephalogr. Clin. Neurophysiol.
Electroencephalogr. Clin. Neurophysiol. 106, 127^137. 95, 189^200.

Phil. Trans. R. Soc. Lond. B (1999)

1134 A. Gevins and others Electroencephalographic imaging

Van den Broek, S. P., Reindeers, F., Donderwinkel, M. & Wang, J. Z., Williamson, S. J. & Kaufman, I. 1993 Magnetic
Peters, M. J. 1998 Volume conductions e¡ects in EEG and source imaging based on the minimum-norm least-squares
MEG. Electroencephalogr. Clin. Neurophysiol. 106, 522^534. inverse. BrainTopogr. 5, 365^371.
Walter, D. O. 1963 Spectral analysis for electroencephalograms: Williamson, S. J. & Kaufman, L. 1987 Analysis of neuromag-
mathematical determination of neurophysiological relation- netic signals. In Handbook of electroencephalography and clinical
ships from records of limited duration. Exp. Neurol. 8, neurophysiology, vol. 1 (ed. A. S. Gevins & A. Rëmond),
155^181. pp. 405^448. Amsterdam: Elsevier.

Phil. Trans. R. Soc. Lond. B (1999)