The independent component analysis (ICA) method provides a complete decomposition of single-trial (or continuous) EEG data, separating the data into distinct information sources. As the results presented below show, the amount of information about cortical dynamics provided by this method is large. Here, we detail for the first time dynamics occurring within single trials of the classes of maximally independent EEG processes whose event-related activities contribute to 31-channel visual target responses recorded during a test of spatial selective visual attention ().
Response Dynamics of the Scalp Recordings: ERPs and ERP Images
To orient readers used to analyses of the raw scalp-channel data, we first present conventional ERP results at selected channels. As we previously reported, in these experiments performance level was high; more than 95% of targets were followed by a button press within the allotted response window (150–1000 ms). Mean subject-median reaction time (RT) was 352 ms. The average ERP time locked to onset of the target stimulus followed by a subject button press contained the expected late positive complex or P300 positivity following early stimulus-locked peaks conventionally termed P1, N1, P2, and N2 (A). The scalp topography of the late positive complex varied continuously through its extent (A, scalp maps).
Grand Mean and RT-Sorted Single-Trial Responses at Sites Fz and Pz
In the grand average of the same epochs, each time-locked to the subject response (B), the early response-locked peaks became smeared out, and the P300 and succeeding negative dip more concentrated. In two-dimensional “ERP-image” plots of the 8,413 single trials from all 15 subjects (C–F), potential fluctuations in single trials are shown as color-coded horizontal lines, here normalized by channel baseline variability then sorted (across all trials) by RT and smoothed (vertically) with a 300-trial moving average. The ERP images clearly show that the early visual response peaks at central posterior site Pz (E) were time-locked to stimulus onset, while the late positivity at Pz immediately followed the button press (compare E and F) except in the trials with the quickest RTs. Over half of these were contributed by two fast-responding subjects whose responses, unlike those of the other 13 subjects, preceded P300 onset.
At frontocentral channel Fz, however, the late positivity in the stimulus-locked grand average (C, bottom) was largely composed of two response-locked positive peaks, separated by 200 ms, that, together with intervening and flanking negativities, could be partially modeled by a two-cycle, 5-Hz wavelet (D). The single P300 peak at Fz in the stimulus-locked ERP (A) “smears out” the two-cycle pattern that is captured clearly in the response-locked average (B and D), while highlighting a concurrent, broader, and possibly stimulus-locked positivity in faster-RT trials (C).
Event-related spectral perturbations
summarizes the grand mean time course of changes from prestimulus baseline in log spectral EEG power at all the scalp channels time-locked to button presses (solid vertical line) across the EEG frequency range. During and just prior to the button press, an approximately 3-dB increase in low-theta-band power peaked (red area) near 4 Hz in bilateral central and posterior cortex. This increase remained significant (p < 0.01) for 14 of the 15 subjects even after the subject-mean ERP was subtracted from each trial (data not shown).
Changes in Mean Scalp Spectral Power Time-Locked to the Subject Response
A concurrent but weaker theta power increase near 6 Hz (B) was maximal at frontocentral and parietal scalp sites. The theta increase at these sites was accompanied by a blocking of mu activity around 10-Hz and 22-Hz, maximal at the left and right central scalp but also widespread over posterior scalp (C, D, and F). Following the button press, a late central bilateral increase in beta activity (maximal at 16–18 Hz) appeared (E). was computed prior to performing ICA and removing eye movement artifacts. The diffuse, far frontal increase in 3–10 Hz activity that began 500 ms after the button press (G) doubtless reflects increased subject eye activity following the target response.
Conventionally, characterizing the sources of ERP () or event-related spectral perturbation (ERSP) () processes is thought to be difficult because the scalp sensors are relatively far from the actual brain sources and therefore each sums the volume-conducted activities of several source areas. Moreover, the biophysical inverse problem of determining the potential source distribution from a given scalp map is in general severely underconstrained, with many mathematically correct but physiologically different solutions. Nonetheless, infomax ICA, applied to the concatenated single trials for each subject, after removing trials containing out-of-bounds or uncharacteristic artifacts, decomposed the whole set of concatenated EEG signals into 31 spatially fixed, maximally temporally independent component processes, and the scalp maps associated with many of these processes resembled the scalp projections of synchronous activity in either one or sometimes two nearly bilaterally symmetric cortical patches
Component contributions to the single-trial EEG signals
shows two single trials at site Pz (black traces) from one subject after removal of six eye and muscle artifact components. Projected activities of the three independent components most strongly contributing to each trial are shown as thin colored traces and accompanying scalp maps. Since infomax ICA provides a complete linear decomposition, the observed data (black traces) are in each case the sum of the remaining 25 (31 minus six) component projections, including the three component projections shown. In the upper trial, and typically, the single-trial P300 at Pz was accounted by ICA as summing contributions from at least two independent EEG processes. Component IC1 for this subject (ranked first by amount of total EEG variance accounted for) was later included in the parietal “P3b” component cluster (described below) on the basis of its scalp map and activity power spectrum.
Independent Component Decompositions for Two Single Trials
While component IC1 accounted for the largest part of the P300 peak in the upper trial, in the lower trial the same IC1 process showed mixed low alpha and beta activity with a smaller postmotor response positivity. Note that the positive postresponse contribution of IC1 in this trial (thin blue trace) was sometimes larger than the observed positivity in the whole-channel data (thick black trace). At these times, some of the other 24 components contributed negative potentials to the signal at this scalp channel, partially canceling the IC1 positivity in the recorded data. Thus ICA, applied to the continuous or concatenated single-trial data, may actually recover more of the actual projected signals than are available in the single-channel data.
Independent component clusters
Cluster analysis, applied to the normalized scalp topographies and power spectra of all 465 components from the 15 subjects (see Materials and Methods
), identified at least 15 clusters of components having similar power spectra and scalp projections. These component clusters also showed functionally distinct activity patterns. Six distinct component clusters (data not shown) accounted for eye blinks, horizontal eye movements, and left and right temporal muscle noises, respectively. These were effectively removed from the activity of the other component clusters by the ICA decomposition and are not further considered here.
Equivalent dipole locations
shows the results of modeling the grand mean scalp maps for each of the nine independent component clusters as the projection of an equivalent dipole. Residual scalp map variances unaccounted for by these models were relatively small (range: 0.87% to 9.55%; mean: 4.93%), though the equivalent dipole locations for the individual clustered components were not all tightly clustered, as shown by the spatial standard deviation ellipses.
Mean Component Cluster Equivalent Dipole Locations
The location of the equivalent dipole for a radially oriented cortical source patch (or here, effective sum of patches) is typically deeper than the cortical patch itself (Baillet et al. 2001
). The equivalent dipoles for individual components in the P3b cluster (data not shown) were scattered across parietal and central cortex (as indicated in by P3b's larger spatial standard deviation). Therefore, the equivalent dipole for the mean scalp map of the P3b cluster was unnaturally deep and represented the center of the active cortical source distribution only symbolically. The mean equivalent dipole location for the cluster designated “P3f” was estimated chiefly from the two periocular electrodes. Moreover, the complicated geometry of the frontal skull cannot be well fit to the spherical head model used here. Thus, the mislocalization of the P3f cluster equivalent dipole below the orbitofrontal brain surface should not be taken literally.
The spatial standard deviations of the other component cluster dipoles were smaller. Their equivalent dipoles indicate the respective dominant cortical regions of their source domains. Though the mean cluster scalp map for the central occipital alpha cluster (Cα) could be fit satisfactorily by a single equivalent dipole located in the central occiput, for several of the cluster components a better model of the component scalp map was obtained from a symmetric dipole pair in left and right pericalcarine cortex (data not shown).
Component Cluster Dynamics
Mean dynamics properties of nine nonartifact component clusters are summarized in –, each of which shows the mean scalp map and response-locked ERP image, activity and ERP spectra, and ERSP for one or more component clusters. Because of the complexity of the results, we report and interpret the nine component clusters in four groups based on shared dynamic features.
Far-Frontal and Parietal Component Clusters Contributing to the P300
Three Posterior Alpha-Rhythm Component Clusters
Two clusters contributing to the late positive response (P3f and P3b)
Two component clusters made distinct contributions to the late positive complex of the target ERP. After subtracting the larger back-projected scalp-data contributions of components accounting for blinks and saccadic eye movements, the response-locked ERP at both periocular channels contained a broad, approximately 2-μV positive-going scalp potential peaking on average 39 ms before the recorded button press. A–D shows the mean scalp map and dynamic properties of a cluster of ten independent components from ten subjects that together largely accounted for an ERP feature whose time course was highly similar to the peak we labeled P3f (for P3-frontal) in an earlier report on decomposing the matrix of 25 condition ERPs from these experiments (Makeig et al. 1999a
). Note, in the ERP image (C), the absence of sharp excursions not regularly time-locked to experimental events, which would mark blinks or lateral eye movements. Such activity at far frontal and periocular channels was effectively separated out by ICA into artifact components (data not shown). Instead, as shown in B, the P3f component cluster accounted for nearly all the positivity occurring before the button press (designated P3f by Makeig et al. 1999a
), particularly in shorter latency-response trials (C).
The P3f cluster-mean response-locked positivity began near 150 ms, consistent with direct neurophysiological evidence that by 150 ms after stimulus onset, visual information is spread throughout the brain by a complex web of afferent and efferent connections (Klopp et al. 2000
; Hupe et al. 2001
). Subtracting the button travel time (approximately 25 ms, roughly estimated from electromyographic recording during one experimental session) and the neuromuscular conduction time (approximately 15 ms) suggested that the P3f peak at 39 ms before the button press occurred very near the moment of the subcortical motor command. It is thus tempting to speculate that the P3f process should originate in frontal structures involved in motivated decision making and response selection, such as orbitofrontal cortex (Ikeda et al. 1996
), though the sparse spatial sampling of the present data does not allow more specific conclusions.
The far-frontal (P3f) component cluster response appears similar to the 280-ms “P2a” peak noted in responses to (foveal) visual “oddball” stimuli by Potts et al. (1998)
. Potts and Tucker (2001)
reported that P2a was maximal near the eyes but can be recorded over most of the face and may also be found in attention conditions involving no subject button press. The scalp map of the ERP-derived P3f component derived by ICA applied to the 25 condition ERPs (Makeig et al. 1999a
) also included bilateral parietal features not seen here in the P3f component cluster. Evidently, the temporal and far-frontal projections joined in the ERP-derived P3f component map did not cohere in the much more extensive single-trial data and so were separated by ICA applied to the concatenated single-trial data. This detail points to the advantage of decomposing a sufficient number of unaveraged data trials over decomposing even a relatively large set of averaged responses.
E–H shows the mean scalp map and activity patterns of a bilaterally distributed cluster of 15 components (from nine subjects) that projected most strongly to posterior and central scalp sites and made a substantial contribution to the slow postmotor P300 or P3b positivity. Examination of the raw ERP waveforms of the six subjects not contributing to this cluster suggested the absence of a typical central parietal positivity in their target responses. The mean cluster scalp map (E) resembled that of the response-locked parietal ERP peak itself (see B). The ERP image of the normalized single-trial component activity (G) includes an early series of small, positive and negative, wave fronts following the (sorted) time of stimulus delivery (dashed curve) by fixed delays. These are followed by a large response-locked positivity (red area) accounting for 62% of postre-sponse ERP variance at 300 ms for site Pz. The P3b cluster positivity is clearly smaller in late-response trials (G, top), consistent with the Pz data (F). The mean cluster ERSP (H) reveals a significant (3-dB) post-response low-theta power increase.
The significance of the stimulus-locked P300 (or P3b) peak over central parietal cortex has long been debated. Our results clearly show (E) that in these experiments, this peak was time-locked to and predominantly followed the motor response. P3b onset occurred at about the moment of the motor command, coincident with the P3f peak. It could not, therefore, index activity involved in making the motor decision or action, as this is seemingly associated with the P3f process. The equivalent dipole distribution of the P3b cluster was broad, the strongest commonality being dipole orientation toward the central parietal scalp. Between-subject variability in locations of P3b generators have also been reported by researchers using other source localization methods (Moores et al. 2003
). It is possible that more advanced three-dimensional component clustering methods, applied to decompositions from more subjects and more data channels, might allow further distinctions among processes in this cluster.
Central midline clusters
and show mean properties of four classes of components producing the two-cycle postresponse evoked-response pattern seen clearly in the response-locked data at site Fz (see D). Principal among these were two component clusters () projecting maximally to the frontal midline (FM) and central midline (CM) scalp, respectively.
Two Mediofrontal Independent Component Clusters Showing a Postmotor Theta Response Pattern
Two Mu Rhythm Component Clusters Also Showing the Postmotor Theta Response Pattern
In the RT-sorted FM cluster ERP image (C), the two negative wave fronts follow the curving trace marking stimulus onsets, the second of these merging with the earlier RT-locked negativity. Though the vertex-maximum CM component cluster (E–H) also exhibited the postmotor theta feature (F and G, red arrows) with the mid frontal and mu rhythm clusters (see below), it contributed little to the broader (P3b) positivity produced mainly by the central parietal (see E) and central occipital () component clusters.
Though the equivalent dipole locations of components in the two midline clusters were somewhat overlapping (see ), their mean equivalent dipole locations were generally consistent with sources in or near the dorsal anterior cingulate and cingulate motor areas, respectively (Ullsberger and von Cramon 2003
). These areas are implicated by fMRI and neurophysiological experiments as participating in motor response selection and anticipation of the consequences of events, including those involving self-perceived errors (Shima and Tanji 1998
; Luu and Tucker 2001
; Manthey et al. 2003
; Ullsperger and von Cramon 2003
). The phase/latency of the postmotor theta burst appeared to be consistent across quicker and slower responses.
The theta burst appears to resemble other reported FM EEG activity patterns: theta bursts or trains (fmθ) appearing during mental concentration (Mizuki et al. 1980
; Gevins et al. 1997
; Uchida et al. 2003
) and brief bursts of theta activity linked to and following the error-related negativity (Luu and Tucker 2001
; Luu et al. 2004
), an ERP peak whose latency matches the first negativity in the FM cluster postresponse ERP (at approximately 60 ms). Inverse source modeling has placed the generating cortical domain of the ERN and fmθ in or near the dorsal anterior cingulate. In this ICA decomposition, however, the two-cycle postmotor theta burst pattern appeared not only in the FM cluster, but also in the CM, mu, and parietal clusters (see below).
The scalp map of the CM cluster (E) resembles scalp maps of the “P3a” or “P3novel” ERP peaks seen, e.g., when unique and unexpected stimuli are included in a randomly alternating sequence of target and nontarget stimuli (Courchesne et al. 1975
; Polich and Comerchero 2003
). Here, however, the CM cluster made only a small contribution to the stimulus-locked target ERP.
Mu rhythm clusters
The left and right mu rhythm component clusters (Lμ and Rμ in ) exhibited the defining feature of mu rhythms—distinct spectral peaks near 10 Hz and 22 Hz that are strongly blocked following movements, with equivalent dipoles located roughly over hand motor cortex (and/or adjacent postcentral somatosensory areas), and oriented roughly orthogonal to the directions of the central sulci. Both the ERP and ERSP peaks were larger in the left mu cluster (contralateral to the response hand) than in the right. In common with the midline clusters, the mu component clusters contained the two-cycle postmotor theta pattern (D) concurrent with a mean theta power increase (H). They also made slower, positive-going contributions to the parietal ERP, particularly to the late “slow wave” phase of the stimulus-locked P300 complex that, unlike the main (P3b) peak, exhibits a polarity reversal over the central scalp (Simson et al. 1977
In a previous ICA analysis of ERPs from these experiments, the late slow wave phase of the stimulus-locked P300 complex was confined to a single component (Makeig et al. 1999a
), but here it was separated into distinct left and right mu rhythm processes in at least ten of the subjects. More detailed source analysis of magnetoencephalographic mu rhythms has assigned their source mainly to somatosensory cortex (Forss and Silen 2001
). Thus, the postresponse slow positivity (C) of the Lμ cluster, larger following slower responses, might index tactile feedback from the hand and button surface (Makeig et al. 1999a
Posterior alpha clusters
shows the dynamics of three clusters of components projecting to the posterior scalp. Each had a distinct near 10-Hz alpha frequency peak in its activity spectrum, most pronounced in components of the central cluster (F, inset). The stimulus-locked ERP contributions of the two lateral posterior alpha clusters, shown as sloping wave fronts in C and K, included an early stimulus-locked peak accounting for most of the P1 ERP peak (near 145 ms) and for part of the succeeding N1 peak, which summed contributions from several clusters. In the central alpha component cluster, the initial stimulus-locked response feature was followed by a train of approximately 10-Hz stimulus-locked waves. These can be said to be produced by partial phase resetting of the intermittent alpha activity of these components following stimulus onsets, since they were accompanied by no mean increase in alpha power. The central alpha cluster also made an appreciably broad, triangular contribution (F and G) to the P300 positivity, while the contributions of the lateral clusters to the response-locked ERP, beginning just before the button press, were small and narrow. The mean response-related ERSPs for these three clusters were weak (D, H, and L), and their postresponse alpha and beta blocking were brief and weak, compared to the two mu clusters (see ). The lateral clusters, but not the central cluster, exhibited a low beta increase above baseline (near 14 Hz) beginning near the button press.
While the existence of multiple alpha rhythms has long been noted, ICA here neatly separated their activities and identified their complete, overlapping scalp maps based on the relative independence of their activity patterns in the unaveraged data. The central posterior alpha processes had a stronger alpha-band peak than lateral posterior alpha components, and showed longer-lasting phase resetting following visual stimulus onsets (G). The longer phase memory implied by the prolonged phase resetting is compatible with longer bursts of alpha activity in these components. Possibly, the distinct dynamics of the central and lateral posterior clusters may serve different though still unknown purposes.
The “trapezoidal” signature of several of the central posterior component scalp maps (E) is compatible with a model comprising two equivalent dipoles located symmetrically in left and right pericalcarine cortex. To be fused into a single infomax ICA component, activity in both hemispheres must have been largely synchronous with negligible phase delay. Alpha-band activity in two cortical areas can indeed be synchronous if the two areas are densely connected, here most likely via corpus callosum. Synchronization of bilateral generator regions via dense callosal coupling might also support the observed sharp (around 10-dB) alpha peak in the activity spectra of these components.
In these data, the lateral posterior alpha components were always unilateral and never bilateral. Possibly this may reflect the lower density of direct connections between these areas. Early ERP features in these experiments appeared predominantly contralateral to the stimulus locations. No doubt this was because these stimuli were presented above and usually lateral to fixation. In other data, we have noted that visual ERPs time-locked to foveally presented stimuli usually contain a bilateral posterior P1/N1/P2 complex (Makeig et al. 2004
Component ERP Contributions
Together, the nine component clusters accounted for 91.1% of the variance of the response-locked grand mean ERP at all channels in the 1000 ms following stimulus onset, as well as for 90.8% of the variance of the stimulus-locked grand mean ERP. A and B shows the envelopes (most positive and negative channel values, across all channels, at each time point) of the stimulus-locked and response-locked grand mean ERPs (black) and the envelope (red fill) of the summed back-projections to the scalp of the components comprising the nine clusters. The normalized grand average activity time courses for the nine clusters are shown in C–H, for comparison with the time courses of the grand mean ERP (A and B).
Component Time Courses and Summed Scalp Projections
Note that stimulus-locked component cluster ERP activity first appeared in the lateral posterior alpha clusters (at 100 ms). Onset of the stimulus-locked ERP of the P3b cluster at about the same time was soon followed by the far-frontal P3f cluster onset (near 120 ms, C). The stimulus-locked ERP deflection began at the same moment in the four postmotor theta clusters (E). Six of the nine clusters had a negative peak in their stimulus-locked ERP average near 200 ms, confirming the spatial complexity of the N1 peak, as indicated by invasive measures (Klopp et al. 2000
) and comparable to previous analysis of nontarget epochs from this data set (Makeig et al. 2002
In the response-locked cluster ERPs, note that the P3f cluster activity appeared to begin early, while response-related activity in the P3b cluster ERP diverged from baseline 10–20 ms before the P3f peak, concurrent with a posterior-negative peak in the left mu cluster ERP. The posterior-positive peaks in the response-locked ERPs of both mu clusters, the early shoulder of the P3b cluster ERP peak, the central cluster ERP slow wave, and the negative-going peak of the FM cluster ERP all occurred together, about 100 ms after the P3f peak.
shows the individual and summed independent component cluster contributions to the grand mean ERP at sites Fz and Pz. At Pz, no component cluster contributed more than a third of peak parietal P300 amplitude in either the stimulus-locked or response-locked ERPs (C and D). The largest cluster contribution to the peak at Fz was also from the P3b cluster, which contributed about half its peak amplitude (A and B). The P3f cluster contributed at best a third. These results cast doubt on claims that the target-response P300 peak at Fz predominantly indexes frontal activity.
Cluster Projections to the Scalp ERP
Post-Response Theta Synchronization
Phase coherence analysis of consistent phase relationships between the FM and CM clusters, and between the FM and left mu clusters, time-locked to and following the motor response showed that significant theta phase coherence appeared in the data, even after the respective component ERPs were subtracted from each trial, indicating a transient postresponse phase linkage between these otherwise maximally independent processes. illustrates this phenomenon in the time domain using phase-sorted ERP-image plots. Sorting trials by the phase (with respect to the button press) of the postmotor theta activity of the FM cluster (A) and then imaging the single-trial activities of the CM cluster in the same trial order (B) induced partial theta phase ordering on the CM data (i.e., the slightly diagonal wave fronts in B). The converse procedure (C) gave a similar result (D).
Phase Coupling of Theta Components: Time-Domain View
is an animation representing the joint response-related theta-band dynamics occurring in and between all nine component clusters (Delorme et al. 2002
). The figure shows an analysis window centered 89 ms after the button press. Note that the transient theta phase coherence between the mu, parietal, and midline components was selective: phase coherence between FM and CM, FM and Lμ, and Rμ and P3b clusters (indicated by linking cylinders) were significant, whereas no significant phase linkage occurred in this time period between the FM and P3b clusters, nor between the CM and Lμ clusters. This selectivity diminishes the possibility that the observed transient phase linkages were produced by appearance of postresponse EEG activity not separated out by ICA into a separate component and therefore misattributed by the ICA model to nearby independent components. The results presented here, however, do not allow us to completely discount this possibility.