The spatiospectral characterization of brain networks: fusing concurrent EEG spectra and fMRI maps

Abstract

Different imaging modalities capture different aspects of brain activity. Functional magnetic resonance imaging (fMRI) reveals intrinsic networks whose BOLD signals have periods from 100 s (0.01 Hz) to about 10s (0.1 Hz). Electroencephalographic (EEG) recordings, in contrast, commonly reflect cortical electrical fluctuations with periods up to 20 ms (50 Hz) or above. We examined the correspondence between intrinsic fMRI and EEG network activity at rest in order to characterize brain networks both spatially (with fMRI) and spectrally (with EEG). Brain networks were separately identified within the concurrently recorded fMRI and EEG at the aggregate group level with group independent component analysis and the association between spatial fMRI and frequency by spatial EEG sources was examined by deconvolving their component time courses. The two modalities are considered linked if the estimated impulse response function (IRF) is significantly non-zero at biologically plausible delays. We found that negative associations were primarily present within two of five alpha components, which highlights the importance of considering multiple alpha sources in EEG-fMRI. Positive associations were primarily present within the lower (e.g. delta and theta) and higher (e.g. upper beta and lower gamma) spectral regions, sometimes within the same fMRI components. Collectively, the results demonstrate a promising approach to characterize brain networks spatially and spectrally, and reveal that positive and negative associations appear within partially distinct regions of the EEG spectrum.

Figure 1

Fusing EEG and fMRI by deconvolving group component timecourses. Group ICA is conducted separately on the frequency by spatial EEG (a) and spatial BOLD fMRI (d) data. The 2D EEG (a) or fMRI (b) data is converted into an [frequency × channel] amplitude spectra vector or voxel magnitude vector, respectively (represented by the red lines in a and b). The vectors are concatenated across epochs forming a 2D data matrix for each individual (represented by the boxes in a and b). The separate group ICA (implemented with GIFT software) generates an individual time course and source matrix for the EEG (b and c) and fMRI recording (e and f). The first level group components are fused by deconvolving the individual EEG source time course and the individual fMRI source time course (g). Deconvolution (as implemented here) treats the BOLD fMRI component time course as the output of the EEG component time course convolved with an unknown (estimated) impulse response function (IRF). Linked EEG-fMRI sources are identified when the estimated IRF significantly deviates from zero. The two time courses shown (in g) correspond to the time course from a segment of a recording for the EEG group component representing the upper alpha band (e.g. 10–11.5 Hz) (in c) and a group fMRI component with a peak response over occipital cortical regions (in f). The negative peak in the estimated IRF (bottom right) demonstrates that increases in the upper alpha band are associated with a subsequent reduction in the occipital BOLD response.

Figure 2

Frequency by channel EEG sources. The 2D frequency (y-axis) by channel (x-axis EEG sources are demonstrated in a for the 10 selected components. The components are organized from lowest peak frequency (upper left) to the highest peak frequency (lower right). The x axis denotes the electrodes corresponding to occipital (O), parietal (P), temporal (T), central (C), and frontal (F) locations. The characteristic EEG frequency bands are indicated by the horizontal bars at 4, 8, 10, 12, 20, and 30 Hz. The values represent the normalized (i.e. z-scored) source values. The peak response and topography of the 2D sources is represented in b. The upper plot indicates the average normalized spectrum of the four channels (indicated in white) with the largest normalized (i.e. z-scored) peak amplitude. The topographic plot indicates the average normalized spectrum for each channel across frequencies.

Figure 3

Spatial fMRI components and their relationship to frequency by spatial EEG. The 56 selected BOLD fMRI components are z-scored, thresholded at z > 4 and displayed in (a). The components are organized in the composite maps according to the approximate spatial location of their peak response. The relationship between each BOLD fMRI component and the 10 EEG components (shown in Fig 2) is indicated in (b). Significant positive associations are indicated along the white grayscale axis and significant negative associations are indicated along the black grayscale axis. White and black pixels indicate the regions with the largest positive or negative peak in the IRF, respectively. Non-significant relationships are in gray. The solid lines in (b) separate the components according to their grouping in (a) and the colors in (a) correspond to the corresponding colors in (b). The anatomical labels are abbreviated in (b) according to the BrainInfo portal (http://braininfo.rprc.washington.edu/).

Figure 4

Impulse response functions (IRFs). The average IRF (n = 50) is plotted for each of the 10 EEG group components (shown in Fig. 2) for a group fMRI component with a peak response over occipital cortical regions (a). The error bars represent the range of 97.5% of the values of the bootstrap resampled IRFs. Significant deviations are indicated by the diamonds (see the main text for details). The location of deviations overlaps well with the location of modulation expected based upon the canonical hemodynamic response function (HRF) (e.g. with a peak between 4–10 seconds). The occipital fMRI component is related to modulations in multiple EEG components (representing different frequencies and/or spatial locations). Increased responses within the “low alpha / high alpha” “high alpha” and “low beta” EEG components are related to reduced fMRI component responses, while increases in the “delta”, “theta”, “upper beta”, and “lower gamma” components are related to increased responses within the occipital fMRI component.

Figure 5

Surface rendering of spatiotemporal networks. The BOLD fMRI components with negative relationships with the EEG components are rendered on the surface composite plots in a (blue). The components with positive relationships are rendered on the surface composite plots in b (red). The components were z-scored and thresholded at z > 4 within each plot. Different shades of each color within each plot were used to differentiate the different components. The peak response of the EEG component is indicated by the labels on top. The spatial topography and spectral characteristics of each EEG component is shown in Fig 2.