High-frequency Broadband Modulations of Electroencephalographic Spectra

Abstract

High-frequency cortical potentials in electroencephalographic (EEG) scalp recordings have low amplitudes and may be confounded with scalp muscle activities. EEG data from an eyes-closed emotion imagination task were linearly decomposed using independent component analysis (ICA) into maximally independent component (IC) processes. Joint decomposition of IC log spectrograms into source- and frequency-independent modulator (IM) processes revealed three distinct classes of IMs that separately modulated broadband high-frequency ( approximately 15-200 Hz) power of brain, scalp muscle, and likely ocular motor IC processes. Multi-dimensional scaling revealed significant but spatially complex relationships between mean broadband brain IM effects and the valence of the imagined emotions. Thus, contrary to prevalent assumption, unitary modes of spectral modulation of frequencies encompassing the beta, gamma, and high gamma frequency ranges can be isolated from scalp-recorded EEG data and may be differentially associated with brain sources and cognitive activities.

Keywords: ECoG; EEG; EMG; ICA; emotion; gamma; neuromodulator; ocular motor tremor.

Figure 1

Independent spectral modulators of scalp EEG signals. ICA, applied to EEG data recorded at a large number of scalp electrodes, identifies (A) temporally distinct (independent) signals generated by partial synchronization of local field potentials within cortical patches (B), the resulting far-field potentials summed (Σ), in differing linear combinations, at each electrode depending on the distance and orientation of each cortical patch generator relative to the (A) recording and (C) reference electrodes. On average, power in the cortical IC signals decrease monotonically with frequency, but also exhibit continual, marked, and complex variations across time. Rather than viewing these variations as occurring independently at each frequency, spectral modulations may be modeled as exponentially weighted influences of several distinct but possibly overlapping modulator (IM) processes (D) that independently modulate via multiplicatively scaling (Π) the activity spectra of one or more independent component (IC) signals. On converting the IC spectra to log power, combined IM influences on IC spectra are converted to log-linear weighted sums of IM influences, allowing a linear ICA decomposition of the IC log-power spectra to separate the effects of the individual IM processes (D) on power at selected frequencies of IC sources (B).

Figure 2

Typical single-subject decomposition of log-spectral power modulations across an hour-long experimental session. The 14 IMs visualized here (14 rows) represent major classes of spectral modulation of 10 of the 16 ICs (rightmost 10 columns) entered into the log-spectral decomposition for this subject. The leftmost column shows histograms of the time-window weights for each IM. The top four IMs (IMs 1–3, 6) are examples of broadband modulators indexing EMG activity. ICs (IMs 1–3) and a putative ocular motor IC (IM6); note the EMG IMs’ multi-modal weight histograms (left). The other IMs (below) affect only brain ICs, either with a broadband pattern (IMs 8, 9, 13) else either predominantly in the theta (IM24), alpha (IMs 5, 20, 39, 43), or beta (IMs 6, 12) frequency range.

Figure 3

Effects of brain and muscle modulators on independent component spectra. Maximal effects of three IMs (columns) on the power spectra of three ICs (rows) are shown via their maximal (red traces), minimal (blue traces), and mean (black traces) log-power spectra. Outer light grey limits represent the 1st and 99th percentiles of spectral variation across all the 2-s windows during the session. Dark grey areas represent the 1st and 99th percentiles of the PCA-reduced spectral data. Note the much larger broadband IM modulation of an electromyographic IC (IM1 on IC58, upper left) compared to the separate but smaller broadband IM effects on brain ICs (IM8 on ICs 5 and 2, middle column). The effects of the upper alpha rhythm modulator (IM20) include shifting the peak alpha frequencies of IC5 and IC2 (blue versus red traces, right column).

Figure 4

Broadband independent modulators of brain and scalp muscle components. Brain and scalp muscle ICs are separately modulated by IMs with similar broadband high-frequency templates (upper rows). The left column shows broadband templates for each IC category (black trace is the mean). The right three columns show equivalent-dipole locations of the affected ICs. A distinct cluster of putative ocular motor IMs, shown in the bottom row, have a peak effect near 50 Hz on ICs many of whose bilaterally symmetric equivalent-dipole models (bottom right panels) are located near the eyes. (ICs whose best-fit equivalent-dipole model comprised two bilaterally symmetrical dipoles are represented with a dotted yellow line connecting the dipole pair). Dipole locations for scalp muscle ICs are outside the brain volume (middle row). Green lines in dipole plots connect ICs co-modulated by the same IM and the colors of the dipole spheres (yellow to red) indicate the relative strength of modulation (yellow = 50%, to red = 100% of maximal). Purple spheres indicate individually modulated ICs.

Figure 5

Broadband modulators of a representative data set with higher sampling rate. This decomposition of a data set acquired with a sampling rate of 512 Hz (from a different subject than Figure 2) allowed an upper frequency analysis limit of 256 Hz, allowing examination of broadband patterns between 128 Hz and 256 Hz. Negative spikes at 60 Hz and 180 Hz in some templates are residual effects of 60-Hz line noise. Note that IM9 (bottom row) has no effect above 150 Hz, while the modulatory effect of IM3 (middle row) is still increasing at 256 Hz.

Figure 6

Value-sorted time course correlations of all within-subject IM pairs for each pair of IM clusters. Traces represent sorted correlation coefficients between time weights from pair-wise comparisons of 11 IM clusters, each point representing a within-subject correlation of two IM time courses over 14 emotion imagination periods (excepting ‘compassion,’ see main text). IM clusters affected spectral changes in Delta, low Theta1, high Theta2, below-peak Alpha1, at-peak Alpha2, above-peak Alpha3, low to high Beta1-4 bands, and Broadband high-frequency activity, respectively. See Table 1 legend for frequency-band limits. Most time course correlations were quite weak, but were typically positive between all lower-frequency IM clusters. Correlations for broadband versus broadband IM pairs (arrow) were more often positive than for any other IM cluster pairs. Correlations of broadband IM time courses with lower-frequency IMs (ellipse) tended to be negatively correlated, though nearly all IM time course correlations were weak (|r| < 0.4).

Figure 7

Mean behavioral ratings of the 15 emotion labels used in the experiment. Subjects rated each word on two scales: ‘Valence (negative-positive)’ and ‘Arousal (calm-active)’ labeled ‘Very negative’ (0) to ‘Very positive’ (10), and ‘Low activity’ (0) to ‘Stimulating’ (10), respectively, with the midpoint (5) indicated as ‘Neutral’ on both scales. Rating data were collected from 100 subjects via an anonymous on-line survey. Each emotion point represents the mean z-score and the error bars the standard deviation. Colors are applied from a continuous color spectrum and used simply to differentiate emotions from one another and do not reflect any objective metric or emotion grouping.

Figure 8

Multi-dimensional scaling of median time weights of broadband IMs for each emotion. Similarities between median IM weights in the 15 emotion periods, drawn from log-spectral decompositions for each subject of brain source ICs only, as represented in the best-fitting two-dimensional space by non-metric multi-dimensional scaling (MDS). Colors of the balls represent the mean behavioral ratings of (positive or negative) valence of the 15 emotion terms by a separate subject cohort. The solid line shows the best-fit regression direction (r = 0.96) predicting mean rated valence for each emotion term from its location in the 2-D MDS space solely based on IM weights after neglecting compassion (see text). The dashed line orthogonal to this cleanly separates positive-valence emotions terms (warm color balls) from negative-valence terms (cool color balls).

Figure 9

Equivalent-dipole density of ICs affected by broadband IMs. Spatial density of equivalent dipoles (in IC equivalent dipoles/cm³, for 154 broadband IMs from 32 subjects), obtained by convolving each dipole location with a 3-D Gaussian blur (1-cm SD) and then summing after normalizing for boundary effects. White integers above and to the left of each slice image give their standard MNI brain z-axis coordinates, yellow text the nearest Talairach z-axis coordinates. Here, figure left is brain left.

Figure 10

Difference between positive and negative correlation-weighted IC equivalent-dipole densities of IMs whose median weights, across 14 emotion imagination periods, were positively or negatively correlated with behaviorally rated emotion valence. Regions of non-significant density differences were masked using permutation statistics (p > 0.003, uncorrected). Areas of significant density difference between positive and negative correlation densities are colored yellow/red, indicating broadband power increases during positive-valence emotions, or cyan/blue, indicating broadband power increases during negative-valence emotions. White integers near each slice image give the MNI z-axis coordinates; yellow text, the nearest Talairach z-axis values. In these images, left is left. Weights for ‘compassion’ were not included (see text).