Report

Dynamic Brain Sources of Visual Evoked Responses

See allHide authors and affiliations

Science  25 Jan 2002:
Vol. 295, Issue 5555, pp. 690-694
DOI: 10.1126/science.1066168

This article has a correction. Please see:

Abstract

It has been long debated whether averaged electrical responses recorded from the scalp result from stimulus-evoked brain events or stimulus-induced changes in ongoing brain dynamics. In a human visual selective attention task, we show that nontarget event-related potentials were mainly generated by partial stimulus-induced phase resetting of multiple electroencephalographic processes. Independent component analysis applied to the single-trial data identified at least eight classes of contributing components, including those producing central and lateral posterior alpha, left and right mu, and frontal midline theta rhythms. Scalp topographies of these components were consistent with their generation in compact cortical domains.

Increasing evidence suggests that correlated activity in neural populations is important for brain function and may be coupled to field potential oscillations over a wide range of frequencies (1). Most electroencephalographic (EEG) studies of human visual perception, however, have assumed the averaged event-related potential (ERP) evoked in humans by brief visual stimuli reflects neural activity within discrete, functionally defined visual cortical processing regions. In this view, response averaging removes background EEG activity (considered to be noise), whose time course is presumed to be independent of experimental events, as well as most artifacts produced by eye and muscle activity. Other researchers, by contrast, suggest that ERP features arise from alterations in the dynamics of ongoing neural synchrony generating the scalp EEG (2–4). By these accounts, ERP features are produced through stimulus-induced resetting of the phase of ongoing field potential oscillations, a phenomenon observed in vitro (5).

Averaged ERPs evoked by sudden onsets of simple visual stimuli contain a prominent negative peak (N1) at 150 to 200 ms, modulated by attention (6), whose time course varies across scalp locations. The N1 is typically followed by a train of positive and negative deflections, termed “alpha ringing” from its resemblance to ongoing (8 to 12 Hz) alpha band EEG activity, whose time course also varies across scalp locations (7) consistent with observations that individual subjects may have multiple alpha rhythms with overlapping scalp topographies (8, 9).

To distinguish between the two hypotheses for the genesis of the average ERP, we compared the averaged ERPs and unaveraged EEG epochs time-locked to onsets of nontarget visual stimuli presented to the left visual field of 15 adult subjects in a spatial visual selective attention experiment [Web fig. 1 (10)] (11,12). The frequency dependence and the scalp topography of the mean power in the single-subject average-ERP waveforms (Fig. 1A) resembled that of the unaveraged EEG (Fig. 1B). However, below 20 Hz the mean spectral amplitude of the poststimulus averaged ERP (Fig. 1A) was up to five times larger than that expected (Web fig. 3A) from the EEG spectrum of the single-trial responses (Fig. 1B), assuming that EEG phase (measured relative to stimulus onset) in the single trials averaged to form the ERP was randomly distributed (13). Further, at all scalp channels except the left occipital, the near 15 decibel (dB) increase in alpha power in the ERP waveforms beginning during the N1 interval (Fig. 1C) was not accompanied by any significant increase in alpha power in the single trials (Web fig. 3B) (14, 15).

Figure 1

Power spectra of unaveraged EEG and averaged ERP data have similar topographies. (A) Mean power spectra of 1-s ERPs averaging epochs time-locked to presentations of nontarget stimuli in the left visual field. Each trace represents one scalp channel (mean of 15 single-subject ERP spectra). Heads above, scalp distributions of power at four frequencies. (B) Mean power spectra, at each scalp channel, of the on-average 922 1-s EEG trial epochs averaged to create the single-subject ERPs. (C) Event-related spectral perturbation (ERSP) plot (15) showing mean poststimulus increases in spectral power of the single-subject ERPs, averaged across 15 subjects. Nongreen areas in the time/frequency plane that show significant (P < 0.02) poststimulus increases or decreases (see color scale) in log spectral power in the averaged ERP waveform at a central parietal electrode site (POz) relative to mean power in the averaged 1-s prestimulus ERP. Topographic scalp maps show topography of the poststimulus power increases in the ERP across all 29 scalp channels at three indicated points in the time/frequency plane. (Event-related increases and decreases in spectral power in the unaveraged single trials are shown in Web. fig. 3B). (D) Percentage of subjects with (P < 0.02) significant poststimulus intertrial coherence (ITC) (16), measuring the degree of phase resetting (i.e., phase consistency or phase locking) of EEG activity in single trials, at site POz.

Event-related intertrial coherence (ITC), a measure of the consistency across trials of EEG spectral phase at each frequency and latency window (16), revealed that during the N1 period, the uniform phase distribution across trials, existing before and during stimulus presentation, was replaced by a phase distribution weighted toward a dominant phase. This partial but significant “phase resetting” occurred following stimulus onsets in all scalp channels and EEG frequencies below 20 Hz (Fig. 1D). At 10 Hz in central posterior channels, ITC remained significant for 700 ms. Mathematically, the poststimulus ERP could be accounted for by the phase resetting of the EEG activity, as shown previously for auditory responses (17).

To test the implication of these results that the visual ERPs were not sums of a sequence of brief fixed-latency, fixed-polarity potential events, as often assumed, we sorted single-trial EEG epochs from each subject according to their poststimulus (0 to 293 ms) power at the peak alpha frequency (10.25 Hz) and separated the bottom and top 10% into two trial subsets. We then sorted trials in each subset according to their phase at the same frequency and time window. Plotting the phase-sorted single trials as color-coded horizontal lines in a rectangular image allowed visualization of the relationship between single-trial rhythmic EEG activity and its ERP average (Fig. 2, arrows). The ERP average of trials with strongest poststimulus alpha power (Fig. 2, upper panel) was large (>10 μV). Alpha phase of these trials was unevenly distributed from 200 ms before stimulus onset to 700 ms after stimulus onset. The prestimulus phase bias occurred because, consistent with previous reports by Brandt and others (18–21), poststimulus alpha activity in single trials tended to be largest in trials in which stimuli were presented at scalp-negative alpha phase.

Figure 2

The average ERP is produced by stimulus-induced phase resetting of ongoing EEG activity. Rectangular images, alpha phase-sorted “ERP-image” plots in which each horizontal line in the rectangular image represents a (color-coded) single trial, here at posterior central scalp site POz. Note the color μV scale on the right. The two ERP images show two (>1200) trial subsets consisting of the 10% of trials drawn from each of the 15 subjects having the highest or lowest power, respectively, at the peak alpha frequency (10.25 Hz) in the indicated poststimulus time window (0 to 293 ms, dotted lines). Before plotting, each trial subset was sorted (top-to-bottom) by its relative alpha phase in the same time window. The sigmoidal shape of the phase-sorted poststimulus alpha wave fronts (red and blue) in the highest-alpha trial subset (upper image) indicate the uneven distribution of poststimulus alpha phase. Upper traces below show the averaged ERPs for the high-alpha (brown trace) and low-alpha (blue trace) trial subsets. Lower traces show the time course of intertrial coherence (ITC) (11) at 10.25 Hz for the same trial subsets, together with (red line) the (P = 0.02) ITC significance level. The prominent negative (N1) peak in the highest-alpha ERP waveform (lower orange arrow) is the sum of more negative than positive single-trial values (between upper orange arrows) at the same response latency.

For the subset of trials with weakest poststimulus alpha power (Fig. 2, lower panel), however, the ERP average was small (<1 μV), inconsistent with the assumption that the ERP sums fixed-latency, fixed-polarity potentials that are generated independent of the rest of the EEG. Figure 2 suggests that the fixed-latency, fixed-polarity components of early visual ERP records, if they exist at all, must be small (≪1 μV) (2).

If the N1 and other features of the visual ERPs arise primarily from phase resetting of ongoing EEG processes, which processes contribute to their generation? Because of their distance from the cortex, electrical potentials recorded at any scalp electrode sum the projected activities of multiple brain (and often nonbrain) sources. Under favorable circumstances, these can be separated by independent component analysis (ICA) (22–25). Applied to EEG data, ICA finds spatial filters that separate the recorded activity into the sum of spatially fixed and distinct, temporally maximally independent component processes. To find sources of the ongoing EEG that contribute to the much smaller averaged ERP, we applied infomax ICA (26), separately for each subject, to concatenated 31-channel, 200-ms poststimulus EEG epochs (between 50 ms and 250 ms after stimulus onset) from ∼3000 target and nontarget stimuli presented in the left, center, or right visual field (27).

ICA linearly decomposed each subject's EEG data in the N1 interval into 31 maximally independent components, each characterized by a different, fixed scalp map, showing the spatial projection of the component to each scalp channel, and a time course of activation in each trial. To determine which independent components were common across subjects, we performed cluster analysis on the component maps and activity spectra (28). Eight of the resulting clusters (Fig. 3) contained multiple components that were among the six largest contributors for each subject.

Figure 3

Eight clusters of independent components of the poststimulus single-trial EEG data, derived by infomax ICA (15), accounted for most of the grand mean ERP. Each cluster contained components among the top six contributors to the N1 interval in the single-subject ERP and is represented by a mean scalp map and normalized power spectrum (mean ± 1 SD). Each cluster comprised 9 to 23 independent components from 8 to 13 of the subjects. Together, the eight clusters comprised 110 independent EEG components drawn from all 15 subjects. Unselected components had unique scalp maps or spectra, or represented nonbrain artifacts.

Components in the two lateral posterior clusters (labeled αLP and αRP in Fig. 3) accounted for the early lateral-occipital positivity (P1) as well as for part of the early part of the broad frontal N1 (Fig. 4, middle right). Their scalp maps generally resembled projections of single equivalent current dipoles in lateral occipital cortex (29). Concurrent, and more highly rhythmic, central-posterior alpha components (αCP), found in nine subjects, made dominant contributions to alpha ringing without accompanying increases in alpha power (Fig. 4, left panel). In six subjects, the αCP scalp map was best fitted by a source model comprising two synchronous dipoles located in left and right calcarine cortices. Two more component clusters (μLC, μRC) with centrolateral scalp maxima and ∼10- and ∼20-Hz spectral peaks (Fig. 4, right panel) accounted for separate left and right μ rhythms (30–32), and also contributed to the late N1. Scalp maps of these components were generally consistent with single compact cortical sources in the hand representation area of sensorimotor cortex. Another cluster of frontocentral components (FC) exhibited poststimulus intertrial phase coherence both in the theta and alpha bands that accounted for half the N1 variance at anterior channels (Fig. 4, center left). Single-dipole inverse source models of these components were concentrated in or near left dorsal anterior cingulate cortex (Web fig. 4). Together, the eight component clusters accounted for 77% of variance in the grand-mean ERP N1 as well as 79% of ERP variance in the subsequent alpha ringing period (Web fig. 5).

Figure 4

Characteristics of four of the contributing component clusters (left to right): αCP, central posterior alpha; FC, left frontocentral; αLP, lateral posterior alpha; μRC, right central mu. Top row: Normalized mean (± SD) log power spectra and cluster-mean scalp maps. Second row: ERP image plots of cluster activity in all single trials, sorted by phase in a three-cycle poststimulus time window at the indicated frequency. Solid vertical lines, stimulus onset; dotted lines, N1 interval. Middle row: Black traces, the envelope (most positive and most negative values at each time point) of the grand-mean ERP for the subjects contributing to each cluster. Blue filled outlines, envelope of the component cluster contribution. Fourth row: For each time and frequency, the number, among all the subjects contributing to the cluster (indicated above), whose cluster component(s) exhibited significant (P < 0.02) phase locking (ITC) (16) to the stimuli. Green areas, significant ITC in fewer than four subjects. Bottom row: Mean event-related spectral perturbations (ERSPs) (15), showing cluster mean differences in log spectral EEG power relative to log power in the 1-s prestimulus EEG baseline.

In the traditional model of ERP generation, averaging of sensory event-related EEG data is assumed to reveal a succession of reliably evoked ERP components (often identified with single response peaks) produced in sensory processing areas with fixed latencies and polarities, and to eliminate the (background) oscillatory EEG processes that are assumed to be unaffected by stimulation and irrelevant to brain stimulus processing. Our results extend previous conclusions (2–4) that many ERP features are instead produced by partial EEG phase resetting by suggesting a parsimonious explanation for the observed spatiotemporal complexity of both the N1 and the subsequent alpha ringing in these data as arising from stimulus-induced phase resetting of ongoing activity within multiple, maximally independent EEG domains. Although we do not suggest all features of averaged ERPs are necessarily generated by partial phase resetting of EEG processes without concurrent energy increases, for these data, phase resetting explains (i) why the ERP average of epochs with low single-trial alpha energy is so small (Fig. 2), (ii) why the strong, ∼10-Hz (“alpha-ringing”) peak in the power spectrum of the ERP (Fig. 1C) need not be accompanied by an event-related increase in alpha band EEG power in single trials, and (iii) why the latencies of the resulting ERP peaks need not match the (50 to 100 ms) latency of initial neural activation in visual areas.

That the scalp topographies of the EEG processes separated by ICA can be mostly fit by single equivalent current dipoles is consistent with their generation in compact cortical domains. These results suggest that the scalp EEG is largely the sum of a limited number of such processes, and that the location, orientation, and spectral character of these domains may be similar across subjects. Separation of EEG data by ICA into independent domains appears to reveal a new system different from the hierarchical organization of cortical areas involved in the representation of sensory information, a system involving synchronous electromagnetic field activity within relatively large independent EEG domains. Some physiological results suggest that the EEG domains that are partially phase reset to produce the ERP components (Fig. 3) may extend into more than one visual processing area (33). Neuromodulatory processes may be involved in regulating synchrony in large thalamocortical populations (34), whereas synchrony in small cortical domains is supported by networks of local inhibitory interneurons (1,35). It is possible that the spatial extent of the EEG domains identified by ICA may be identified by concurrent EEG and functional magnetic resonance imaging (fMRI) (36, 37).

After a visual event, spatially distinct and otherwise independent EEG domains may also exhibit transient frequency-domain coherence that may not depend on stimulus-induced phase resetting. For example, during the N1 interval following presentation of face-image stimuli, local field potential signals recorded directly in fusiform gyrus and in several other human brain areas exhibit transient coherence in the alpha band (38, 39). This transient coherence may organize top-down brain responses and focus further processing of stimuli (40,41). The functional relationship of phase resetting, underlying ERP components, to these more general brain dynamic events remains to be explored. Finally, the analysis approach used here, if applied to other types of event-related brain data that are typically averaged, such as optical recordings, neural spike trains, and functional magnetic imaging signals, might reveal additional facts about cortical dynamics (42).

  • * To whom correspondence should be addressed. E-mail: smakeig{at}ucsd.edu

REFERENCES AND NOTES

View Abstract

Stay Connected to Science

Navigate This Article