Fig. 1.

Schematic representing the construction of noise datasets and independent component analysis. Resting state BOLD fMRI data are input to a generalised linear model (GLM) where noise confounds are modelled by nuisance regressors. Typically, the residuals from this fitting procedure are considered to be “de-noised” and used for further connectivity analysis. Here we study the fit of the data to the nuisance regressors, and decompose this “noise dataset” using independent component analysis (ICA). We examined 3, 6, 12, or 24 head motion regressors (translations, rotations, their derivatives, and quadratic terms) or 2 physiologic regressors (end-tidal CO₂ and cardiac rate), in addition to linear and quadratic detrending for signal drift removal, and fixed the dimensionality of the ICA output to be 20.