August 18, 2002
Reply to Technical Comment on Makeig et al.,
Science, Jan. 25, 2002
In our recent Science paper(1),
we argued that most features of the averaged event-related potential (ERP)
following brief non-target visual stimuli in our data were best explained as
arising from partial phase resetting of multiple ongoing EEG processes rather
than from an (overlapping) series of brief, monophasic potentials (with
fixed-latency and polarity), independent of ongoing EEG processes and
concomitant to bursts of stimulus-evoked neural activity in sensory (and other)
cortical areas. We demonstrated, first, a robust post-stimulus partial
phase-consistency in our single trials, and a 15-dB pre-to-post stimulus power
increase in the ERP averages in the absence (at nearly all scalp sites) of
significant related power increases in the single-trial data. [The commenters]
suggest these facts could be explained by the presence of what we might call
'true' ERP activity that, while quite small relative to the ongoing (but
phase-random) EEG signals, dominated the averaged ERP by appearing at the same
latency and polarity following all (or most) stimulus onsets. They use this
model to construct simulated data superficially resembling our key figure ((1), Fig. 2). However, close inspection
of our Fig. 2 shows that the key assumption they use to build their simulation
is not relevant to our data.
They assume
that, as we demonstrated (Fig. 2, upper panel), the 10% of trials with highest
post-stimulus alpha power were dominated by trials in which the posterior alpha
was phase-negative at stimulus onset, the 10% of trials with lowest
post-stimulus alpha (Fig. 2, lower panel) must conversely have been dominated
by trials in which posterior alpha at stimulus onset was phase-positive. If so,
these trials might have actually subtracted out the 'true-ERP' features from
the average of these trials by phase cancellation. Unfortunately for this
suggestion, the lowest (blue) trace in our Fig. 2 clearly shows this was not
the case for our lowest-alpha data. Inter-trial coherence (ITC), a measure of
non-uniform phase distribution shown in the bottom traces of the figure, was
not significant at stimulus onset, meaning that the phase of whatever alpha
activity was present in the lowest-alpha power trials was randomly distributed
and did not, therefore, decrease the 'true-ERP' activity through phase
cancellation (2). Our key point stands, that
in this 15-subject data set 'true' ERP activity, as estimated from the trials
with lowest alpha activity, was very small at best.
A fall-back
position suggested to us by another ERP researcher might be that the 'true' ERP
was generated only when alpha activity appeared in single trials. This proposal
is weakened, however, by another result we presented(3). At frequencies across the ERP
frequency range, the scalp topography of spectral power in the mean ERPs
strongly resembled the mean topography of power in the whole EEG at the same
frequencies. Thus, the 'true' ERP activity in our data unrelated to ongoing EEG
processes must have (a) appeared only during alpha activity, and (b) with its
same scalp distribution.
Further, such a
'true' ERP model assumes (c) that ongoing EEG processes did not
undergo partial phase resetting following stimulation. This assumption seems to
us improbable, since a large number of theoretic and experimental
demonstrations have shown that strong inputs to nonlinear systems that produce
oscillatory activity very often produce reproducible phase perturbations or
'phase resetting' (3). Thus, and in the absence of independent evidence to
support it, a true-ERP plus unperturbed-EEG model for our data seems to us both
unwieldy and implausible.
Instead, several
convergent facts about our data lead us to tentatively conclude that the
averages of single-trial EEG data we showed were dominated by partial
phase-resetting of processes composing the whole EEG activity. Motivated by
this conclusion, we presented results of decomposition by Independent Component
Analysis (ICA)(4)
of the ongoing single-trial EEG activity recorded during 100-ms post-stimulus
(N1) intervals that identified several EEG sources of phase-reset activity in
the ERP. We reported that such decomposition produced classes of EEG components
that strongly resembled traditionally labeled EEG processes, and that these
components had scalp maps compatible with synchronous activity occurring within
compact regions of cortex. These findings, since largely replicated by us for
other data sets, suggest that ICA,
applied to suitable collections of single-trial data from EEG (and/or MEG)
experiments, may give accurate information about the event-related dynamics of
ongoing brain electromagnetic processes.
In general, the
problem of determining the dynamics and spatial distribution of electromagnetic
brain activity from observations on the scalp is mathematically
underdetermined. Synchronous activity within the cortical neuropile creates
far-field potentials on the scalp, and scalp electrodes or magnetic sensors
each record the sum of all such activities generated in any suitably-oriented
region of cortex. The inverse problem of identifying the cortical contributors
and their signals from the recorded EEG and/or MEG data can be separated into
two parts: first, decomposing the recorded data into sums of spatially
separable (and often overlapping) field maps generated by activity within
spatially or functionally separable cortical regions, and second determining
the spatial distribution of each identified cortical area. Unfortunately,
neither part of the problem is uniquely determined.
The first part,
linearly unmixing the putative contributions of active generator regions to the
recorded scalp signals, may be accomplished in any number of ways (just as 4 =
2+2 = 3+1 = -3+7, etc.). Deciding on a unique, functionally meaningful solution
requires making additional assumptions. The standard ERP model uses one such
linear decomposition (Recorded Data = EEG + ERP); our ICA
model uses another (Recorded Data = Sum of independent EEG processes). The
intimate relationships between average ERP and single-trial EEG activity we
demonstrated in our data lead us to believe that for these data our ICA
model is more plausible though final judgment, possibly including elements of
both models, will require independent physiological evidence.
Though the
results we presented, and our interpretation of them, pose fundamental
challenges to the dominant interpretation of event-related potential (ERP) and
magnetic–field (ERF) data, similar objections have in fact been raised over
nearly three decades by others (5-7).
In our paper, we were careful to point out that partial phase resetting is most
probably not a sufficient or parsimonious explanation for all ERP phenomena – for
example small, early potentials associated with primary sensory cortices, and
near-DC ERP features. We do firmly believe, however, that exploring dynamic
properties of single event-related EEG (and MEG) epochs will lead to new
discoveries about event-related brain dynamics. Related non-averaging
approaches to analyzing event-related brain imaging and other biological data
may also prove fruitful (8).