Global signal (GS) regression is a commonly used preprocessing approach in the analysis of resting state fMRI data. However GSR should be used with caution as it can not only induce spurious anti-correlations,¹ but may also remove signal of neural origin. A prior animal study found strong temporal correlations between the local field potential at a single location and the fMRI signal across a widespread region of the cortex.² Here we used simultaneously acquired EEG/fMRI data in human subjects to study the relation between the fMRI GS time course and an EEG-based measure of time-varying vigilance during both the eyes open (EO) and eyes closed (EC) conditions at rest. EEG-fMRI data were simultaneously acquired on 9 healthy subjects (3 females) during three resting-state sessions using a 3T GE MR750 system and a 64 channel EEG system (Brain Products). Each session included one EC and one EO scan (5 minutes each). EEG data processing included: 1) MR gradient artifacts removal, 2) low pass filtering (fc=30Hz) and then down-sampling to 250Hz, 3) removing cardio-ballistic and residual artifacts using OBS-ICA.^3,4 A spectrogram was created using a short-time Fourier transform with a 1311 point 4-term Blackman-Harris window and 65.7% overlap, resulting in 1.8s temporal resolution. For each channel in the spectrogram matrix, the value in each frequency bin was divided by the root mean square (rms) of the bin values across frequencies. A relative EEG amplitude spectrum was then calculated by taking the rms of the normalized spectrogram entries across time and channels. At each time point, a measure of vigilance (Vig) was then defined as the rms of relative alpha amplitude (7-13Hz) divided by the rms of the relative delta (1-4Hz) and theta (4-7Hz) amplitudes.⁵ In order to account for the hemodynamic delay, the vigilance time series was convolved with a hemodynamic response function. Outlier detection was applied to the mean of all EEG amplitude time courses to remove motion-contaminated time segments from both the spectrogram and fMRI time series. fMRI data were acquired with the following parameters: echo planar imaging with 166 volumes, 30 slices, 3.4×3.4×5mm³ voxel size, 64×64 matrix size, TR=1.8s, TE=30ms. For each session, high resolution anatomical data were acquired using a magnetization prepared 3D fast spoiled gradient (FSPGR) sequence. Each anatomical volume was registered with the functional data. Nuisance regressors (1^st +2^nd order Legendre, 6 motion regressors and their first derivatives, mean BOLD signals from the WM and CSF voxels and their first derivatives, RETROICOR⁶ and RVHRCOR⁷ noise terms) were removed from the raw data through linear regression. For each voxel, a percent change BOLD time series was obtained by subtracting the mean value and then dividing the resulting difference by the mean value. The global signal was formed by averaging the percent change time series across all brain voxels. For each scan the correlation between the GS and Vig was calculated after excluding the censored time points. The correlation coefficients were then normalized to Fisher-z scores for group comparison.Figure 1 displays the correlation coefficients between the GS and Vig time series for each subject. We found significant (p<0.05, uncorrected) negative correlation between the GS and Vig in 70% and 63% of the EC and EO scans, respectively. Performing t-tests on the z-transformed correlations also confirmed the overall negative relations in both EC (t=-8.47, p=5.9×10^-9) and EO conditions (t= -8.43, p=6.4×10^-9). We did not find a significant difference between the EC and EO conditions (EO-EC: t= 1.5, p=0.14). Figure 2 shows the GS and Vig time series from a representative subject. For display purposes, the vigilance is multiplied by -1. Figure 3 illustrates the mean of the percent BOLD images averaged over the bottom 10% and the top 10% time points of the vigilance time series for the same subject. Consistent with the negative correlation between the Vig and GS, periods of high vigilance are characterized by a widespread reduction in the BOLD signal. Our results are consistent with prior studies that have found negative correlations on a regional basis between the BOLD and EEG alpha power ^8-10, and also relatively higher theta power during alpha deactivation.¹⁰ Our results contribute to the understanding that the GS has a significant neuronal component and further emphasizes the need to exercise caution when regressing out the GS.