Autocorrelation with SPM FAST produces edge effects in time series data?


We are investigating the effects of FAST on our fMRI BOLD data. Our paradigm is deep phenotyping of many (10+) scans per subject, collected at a TR of 0.46 with a multiband factor of 8 on a 3T Siemens PRISMA machine. We believe that accounting for autocorrelation is important under such circumstances but were wondering if SPM FAST is applied properly. We are concerned that the W matrices (whitening) produced by FAST produces extreme edge effects in the time-series data that do not look intended.

Attached is a figure showing the W matrices, and another figure illustrating the effect of W x g (session-level grand-mean scaling) on the time series data, summarized by the dot-product of a multi-voxel neural pattern of interest (Neural Pain signature). Notice that in the orange line, the beginning and the end of the data become extreme relative to the original data. Run-3 also exhibits a very strange oscillatory pattern that was not present in the original data.

Can someone help us shed light on this issue? Thank you very much.