Gross, W.L., Zhou, Y.Q., Binder, J.R. (2014) Global signal regression improves fMRI prediction of language outcome after left anterior temporal lobectomy. Presented at the annual meeting of the Society for Neuroscience, Washington DC
Noise segregation is a critical component of fMRI analysis. Because many noise sources are distributed evenly across the brain image, global signal regression (GSR; removing spatially averaged signal components) has become a popular method of noise removal. Although this can be done to improve SNR in any fMRI analysis, it is particularly relevant to functional connectivity analyses, where spatial correlations are key.
Current literature on GSR cites multiple theoretical issues with this methodology. Because calculation of global signal (GS) includes voxels of interest, it is weakly correlated with signals of interest (Murphy et al., 2009) and seems to reduce the magnitude and extent of activation (Aguirre et al., 1998). It produces negative correlations in connectivity analyses, which some observers propose are factitious (Murphy et al., 2009). Because of these effects, many have argued against using GSR. However, it is not clear if these effects are purely artifactual, as there are no external standards, and most arguments are based on theoretical conjecture. A few studies have compared fMRI to electrical recordings in the brain (Scholvinck et al., 2010; Keller et al., 2013) and found that, although the GS is present in neural activity, GSR improves the correspondence of fMRI with high gamma power.
To test the validity of GSR, we applied it to an fMRI analysis that predicts neuropsychological outcome after left anterior temporal lobe surgery (Sabsevitz et al., 2003; Binder et al., 2008). We took the ability to predict outcome as an externally valid standard. If predictions improve after GSR, SNR must be increasing in the data. This analysis included 36 patients with intractable epilepsy who underwent left anterior temporal lobectomy and preoperative fMRI mapping. The maps were used to produce lateralization indices of language, which were correlated with outcome. GS was calculated by taking the spatial average across all voxels in the brain, and was applied to the analysis either as a regressor, or through proportional scaling (Gavrilescu et al., 2002). Voxel counts were also calculated within ROIs known to be associated with the task, along with the rest of the brain.
On all models tested, adding GSR (either as a regressor, or through proportional scaling) increased the R2 of the outcome prediction model. Interestingly, it also consistently reduced the voxel count in every ROI examined. It appears that GSR decreases noise in the data, and improves information content, although this is not reflected in the raw voxel counts. These data provide support for the use of GSR and caution against making strong conclusions about data quality based only on voxel counts.