Abstract
For forty years, much human electrophysiologic
thinking has been based on the concept that EEG data recorded from the scalp
following sensory stimulation are dominated by successive far-field correlates
of bottom-up brain sensory processing, as represented in evoked potential (EP)
averages. I will present evidence for an
alternate view that human EEG data are dominated by by
oscillatory processes relating to time-varying, top-down control of cortical
dynamics and attention. This view suggests a reorientation of scientific and
engineering focus towards modeling brain dynamics of humans as active operators
rather than as passive perceivers and programmed responders. Such research presents new engineering
challenges. There is a need, first, to understand and model the process of
partial phase resetting of ongoing and intermittent nonlinear oscillatory
processes, and more generally, of inter-process synchronization. It is probable
that real time signal processing capable of separating EEG processes from
non-brain signals and of monitoring brain synchronization events may allow
high-level cognitive monitoring that could be used in man-machine interfaces
and for neuropsychological training.
1. Introduction
Historically,
electroencephalographic (EEG) data analysis has been a technically
underdeveloped area of neuroscience. Typically, psychologists using
Event-Related Potential (ERP) methods prefer to use simple but limited measures
of peak amplitudes and latencies in ERP averages at single scalp channels. Our
recent work, together with very recent neurophysiologic evidence from humans
and animals, suggests that ERPs averaging data from a
number of experimental trials conceal rather than reveal the essential nature
of event-related brain dynamics [1]. These results show that active brain
responses to significant events or external stimuli involve synchronized
oscillations in local field potentials in a number of brain regions [2]. These
brain dynamic events appear to begin in the frontal cortex, implying they carry
or channel top-down information about intention, including attentional
focus, to sensorimotor brain areas [3] triggering
other dynamic events that carry or channel bottom-up information from sensory
to response-selection areas [4]. Methods we have developed or integrated for
EEG research, including Independent Component Analysis (ICA) and time/frequency
analysis, now allow detection and modeling of such brain dynamic events from
high-dimensional electrical data collected noninvasively
from the human scalp.
1.1 Limitations of ERP standard methods.
Contrary
to the common assumption that event-related EEG data are comprised of an
average ERP plus unrelated EEG activity, our recent analyses show that ERP are
generated chiefly by imbalances in the distribution of spectral phase in EEG
processes whose frequency contents and spatial projections to the scalp vary
across individuals. The averaged event-related potential (ERP) evoked by brief
unattended visual stimuli, for example, consists of a sequence of positive and
negative peaks that are generally assumed to reflect activity in individual
visual cortical processing regions. In this view, response averaging removes
'background' electroencephalographic (EEG) activity, whose time course is
presumed to be independent of experimental events, as well as artifactual potentials produced by eye and muscle activity,
sparing scalp potentials produced by information processing within cortical
areas. Other researchers have claimed that some ERP features may also arise
from perturbations in the ongoing neural synchrony generating the scalp EEG [5][6].
1.2
Visualizing brain dynamics in single trials
Recently, we introduced a new method of
visualizing the relation of single-trial EEG dynamics to time locking events:
the 'ERP-image' [7][8][9]. ERP-image visualization
allows relationships between dependent variables (e.g., reaction time, alpha
component phase, gamma band amplitude, etc.) and time-and-phase locked response
dynamics to be seen clearly. Combining ICA with
ERP-image visualization allows temporally and functionally independent
components of EEG dynamics to be separated and visualized. For example, we [1]
have recently investigated which model of ERP generation more accurately
characterizes dynamic EEG changes occurring during the prominent (N1)
negativity in the ERP peaking near 200 ms after sudden onsets of visual
stimuli. The results contradict the standard model of event-related brain
dynamics that underlies almost all ERP research. This traditional model holds
that event-related EEG data are the sum of an averaged ERP plus `background'
EEG, based on two assumptions: (1) Each response-generating brain area, silent
in the EEG recording prior to the experimental events of interest, becomes
briefly and reliably active with a similar time course and polarity following
each event, (2) While ongoing EEG activity remains wholly unaffected.
While details of the biophysics of the synchronous
neural activity that produces the scalp-recorded EEG are not yet well
understood, there is no reason to expect that cortical areas contributing to ERPs may not also exhibit synchronized activity before as
well as after experimental events, thus contributing to the ongoing EEG, and
that EEG processes may not be perturbed by stimuli or cognitive events. Conclusions from data presented in [1] cast
strong doubt on the basic assumption that ERP peaks are produced within just
those cortical processing areas that show post-stimulus neural firing rate
increases. Instead, they suggest that the visual evoked responses are largely
produced by event-related changes in the statistical distribution of phase,
relative to the inducing events, of oscillatory activity within spatial domains
of neural synchrony that generate the ongoing EEG.
1.3. EEG spectral dynamics
In the last
decade, a new wave of EEG analysis approaches has emerged in which the object
of study is not solely the amplitudes and latencies of peaks in ERP averages,
but event-related changes in EEG dynamics in single event-related data epochs.
An early effort in this direction was by Pfurtscheller
and Araniber [10], who first reported a method for
quantifying the average transient suppression of alpha band (circa 10-Hz)
activity following stimulation. In the
last decade, researchers studying Pfurtscheller's
event-related desynchronization (ERD, spectral
amplitude decreases), and event-related synchronization (ERS, spectral
amplitude increases) in a variety of narrow frequency bands (4-40 Hz) have
reported on their systematic dependencies on task and cognitive state variables
as well as on stimulus parameters [11]. For example, Williamson et al. [12]
reported that, given a visually presented arithmetic problem to compute
mentally, the resulting subject alpha-band ERD resolved only when the
calculation was complete.
Makeig
[13] reported event-related changes in the full EEG spectrum,
yielding a 2-D time/frequency measure he called the event-related spectral
perturbation (ERSP). This method avoided problems associated with analysis of a priori narrow frequency bands, since
bands of interest for the analysis could be based on significant features of
the complete time/frequency transform. The ERSP method was adapted under
different names by Bertrand and colleagues, and by other European groups
studying event-related gamma band (30-100 Hz) EEG phenomena [14].
Its adoption coincided with the new availability to EEG laboratories of scientific workstations with
sufficient computational power to complete these analyses quickly.
1.4. Event-related coherence
In 1994, Rappelsburger
and colleagues introduced event-related coherence (ERCOH) elaborating on
creative studies of EEG coherence by Petsche and colleagues
[15][16]. The application of spectral coherence to EEG
analysis had had a long history, though it had been used infrequently because
of its relative computational complexity and difficulty in interpretation. Petsche showed, for example, that in creative men imagining
how to draw a picture representing an abstract concept (for example,
"liberty"), the distinguishing EEG characteristic was increased
coherence between electrodes located over homologous left and right hemispheric
locations, whereas the distinguishing characteristic in creative women was exactly
the opposite [17]. Another application of coherence to ERP analysis was
introduced by the Bertrand group [14] who demonstrated a measure of the
consistency of phase consistency at each time surrounding an experimental event
class, dubbing this a "phase locking factor." Makeig and colleagues
[8] proposed the term "inter-trial coherence" (ITC) for this measure.
1.5. Independent Component Analysis
In 1995, Bell and Sejnowski [18]
published an iterative algorithm based on information theory for decomposing
linearly mixed signals into temporally independent by minimizing their mutual
information. The problem of blind separation of recorded multi-channel signals
into sums of temporally independent sources had been posed some years earlier.
First approaches to blind source separation minimized third and fourth-order
correlations among the observed variables and achieved limited success in
simulations [19] and approach extended by Cardoso and
Laheld [2] in their JADE algorithm. Bell and Sejnowski [21]
generalized this approach, demonstrating a simple neural network algorithm that
used joint information maximization or 'infomax' as a
training criterion. By using a compressive nonlinearity to transform the data
and then following the entropy gradient of the resulting mixtures, they were
able to demonstrate unmixing of ten recorded voice
and music sound sources that had been mixed with different weights in ten
simulated microphone channels. Their algorithm used only minimal assumptions
about the nature of the sources to be separated. Mixing weights (and thus scalp
projections) of individual components were assumed to be fixed, and the time
courses of the sources mutually independent. Bell & Sejnowski
demonstrated a similar approach for performing blind deconvolution,
and later applied their 'infomax' method to
decomposition of visual scenes [22].
1.6. Applications to EEG Analysis
The
first applications of blind decomposition to biomedical time series analysis
were presented by Makeig et al. [23][24], who applied
the infomax ICA algorithm to decomposition of EEG and event-related
potential (ERP) data and reported the use of ICA to monitor alertness. This first report demonstrated
segregation of eye movements from brain EEG phenomena, and separation of EEG
data into constituent components defined by spatial stability and temporal
independence. Subsequent technical reports [25][26] demonstrated successful
separation of six simulated EEG sources mixed into six simulated EEG channels
using a realistic three-shell head model. Unmixing
performance of the ICA algorithm was shown to degrade gracefully in the
presence of noise added to simulate sensor noise or additional small EEG
sources. A range of biomedical applications of ICA have recently been reviewed by Jung et al., [27].
2. COMBINING ICA WITH
TIME-FREQUENCY ANALYSIS
In general, oscillatory event-related
time/frequency dynamics appear to be more functionally relevant indices of
event-related brain dynamics than peaks in the average ERP. However, they are
largely invisible to analysis of averaged evoked responses. Our analyses
clearly show that independent component brain networks are rapidly linearly
coupled and decoupled at a wide range of specific frequency bands in accord
with task events and subject behavior. The picture that emerges from these results is that a
forced-choice response demand induces coherent theta
activity in a distributed network of frontal and parietal cortical regions.
Often, the theta activity and coherence replaces pre-existing coherences at
higher EEG frequencies. We have seen similar patterns of
induced coherent theta bursts following both nontarget
and target stimuli in a visual selective attention task, in response to targets
stimuli in an auditory response task, and in response to adverse target
trajectories in a compensatory tracking task. Thus, we believe that frontal-parietal theta network dynamics,
well represented in event-related scalp EEG data, indexes the brain's
monitoring of and response to momentary cognitive challenges.
6. APPLICATIONS
It
appears good measures of performance or capacity to perform can be derived from brain synchronization events supporting
cognitive engagement. The same technology that can find nonlinear
correlations between physiological measures and moment-to-moment fluctuations
in performance in monitoring tasks should also be demonstrable in other task
situations and may find applications in cognitive monitoring.
11. References
[1] Makeig S, Westerfield, M.,
Jung T-P, Townsend, Courchesne, E. and
Sejnowski TJ, Electroencephalographic sources of visual evoked responses . Science 295:690-4, 2002.
[2]
Klopp, J., Marinkovic, K., Chauvel, P., Nenov, V., and Halgren, E. Early widespread cortical
distribution of coherent fusiform face selective
activity. Human
Brain Mapping, 11, 286-293, 2000.
[3] von Stein, A., Chiang, C., and Konig, P. Top-down processing mediated by interareal synchronization. Proceedings of the National Academy of Sciences USA, 97(26), 14748-14753, 2000.
[4]
Fries, P., Reynolds, J., Rorie, A., and Desimone, R. Modulation of Oscillatory Neuronal
Synchronization by Selective Visual Attention. Science, 291(5508), 1560-63, 2001.
[5]
Basar, E. EEG-brain
dynamics : relation between EEG and Brain evoked
potentials, Elsevier/North-Holland Biomedical Press, New York, N.Y., 1980.
[6]
Brandt, M. E. Visual and auditory evoked phase resetting of the alpha EEG. International Journal of Psychophysiology,
26(1-3), 285-98, 1997
[7]
Jung T-P, Makeig S, Westerfield M, Townsend J, Courchesne E, and Sejnowski TJ. Analyzing
and visualizing single-trial event-related potentials, In: Advances in Neural Information Processing Systems, 11:118-24, 1996.
[8]
Makeig, S., Westerfield, M., Jung, T.-P., Covington, J., Townsend, J., Sejnowski,
T. J., and Courchesne, E. Independent components of
the late positive event-related potential in a visual spatial attention task.
Journal of Neuroscience,
19(7), 2665-2680, 1999.
[9] Jung T-P, Makeig S, Humphries C, Lee
T-W, McKeown MJ, Iragui V, Sejnowski TJ. Removing
electroencephalographic artifacts by blind source separation, Psychophysiology, 37:163-78, 2001.
[10] Pfurtscheller, G., and Aranibar,
A.. Event-related cortical desynchronization
detected by power measurements of scalp EEG. Electroencephalography and Clinical Neurophysiology, 42, 817-826.1977.
[11]
Pfurtscheller, G., and Andrew, C. Event-Related changes
of band power and coherence: methodology and interpretation. Journal of Clinical Neurophysiology,
16(6), 512-519, 1999.
[12]
Williamson, S. J., Kaufman, L., Curtis, S., Lu, Z. L., Michel, C. M., and Wang,
J. Z. Neural substrates of working memories are revealed magnetically by the
local suppression of alpha rhythm. Electroencephalography and Clinical Neurophysiology. Supplement, 47, 163-80, 1996.
[13] Makeig S. Auditory event-related dynamics of the EEG spectrum and
effects of exposure to tones. Electroencephalog. clin.
Neurophysiolog., 86:283-93, 1993.
[14]
Tallon-Baudry, C., Bertrand, O., Delpuech,
C., and Pernier, J. Stimulus specificity of
phase-locked and non-phase-locked 40 Hz visual responses in human. Journal of Neuroscience,
16(13), 4240-9, 1996.
[15]
Petsche, H., Kaplan, S., von Stein, A., and Filz, O. The possible meaning of the upper and lower alpha
frequency ranges for cognitive and creative tasks. International Journal of Psychophysiology,
26, 77-97, 1997.
[16]
Sarnthein, J., Petsche, H.,
Rappelsberger, P., Shaw, G. L., and von Stein, A.Synchronization between prefrontal and posterior
association cortex during human working memory. Proceedings of the National Academy
of Sciences of the United States of America, 95(12), 7092-6, 1998.
[17]
Petsche, H. Approaches to verbal, visual and musical
creativity by EEG coherence analysis. International Journal of Psychophysiology, 24(1-2), 145-59, 1996.
[18]
Bell, A. J., and Sejnowski, T.
J. An information-maximization approach to blind separation
and blind deconvolution. Neural Computation, 7(6), 1129-59, 1995.
[19]
Comon, P. Independent component analysis, a new
concept? Signal Processing, 36(3),
287-314, 1994.
[20]
Cardoso, J. F., and Laheld,
B. H. Equivariant adaptive source separation. IEEE Transactions on Signal
Processing, 44(12), 3017-30, 1996.
[21]
Bell, A. J., and Sejnowski, T.
J. Learning the higher-order structure of a natural sound, Network: Computation in Neural Systems 7:261-266. 1996.
[22]
Bell, A., J., and Sejnowski, T.
J. The 'Independent components' of natural scenes are edge filters, Vision Research, 37(23) 3327-3338,
1997.
[23] Makeig S, Bell AJ,
Jung T-P, Sejnowski TJ. Independent Component
Analysis of Electroencephalographic Data, In: Advances in Neural Information Processing Systems 8:145-51, 1996.
[24] Makeig S, Jung T-P, Bell AJ, Ghahremani
D, Sejnowski TJ. Blind separation
of event-related brain responses into independent components, Proc. Natl. Acad. Sci.
USA, 94:10979-84, 1997.
[25] Ghahremani D,
Makeig S, Jung T-P, Bell AJ, Sejnowski TJ. Independent component analysis of simulated EEG using a three-shell
spherical head model, Tech Rep. INC-9601, Institute for Neural
Computation, University of California, San Diego, 1996.
[26]
Makeig S, Jung T-P, Ghahremani D and Sejnowski TJ, Independent
Component Analysis of Simulated ERP Data, Tech Rep. INC-9606,
Institute for Neural Computation, University of California, San Diego, 1996.
[27] Jung T-P, Makeig S, McKeown
M.J., Bell, A.J. , Lee T-W, and Sejnowski TJ. Imaging Brain Dynamics
Using Independent Component Analysis, Proceedings
of the IEEE. 89, 1107-22, 2001.