75
ensuring that there were no systematic order effects that could help classification
performance.
The image intensity used for classification was obtained from the adjusted MR
time series, which had regressors of no interest removed in order to reduce temporal
variation. The regressors of no interest included a mean and linear trend for each scan
series (accounting for slow drifts in the MR time series); head motion estimates from
the image registration algorithm; and the estimated hip trial responses. For two-way
classification, estimates of the response to the unclassified trial types were also
considered to be regressors of no interest. To the extent that noise from the regressors
of no interest remained in the MR time series, classification performance will be
impaired, resulting in an underestimate of classifier performance.
Additional Analyses
To identify voxels for classification, a leave-one-out procedure was used to
identify voxels that responded to somatosensory stimulation (p < IO'6) in the five scan
series used for training each classifier. Data in the left-out scan series was not used to
construct the activation map for the corresponding classifier to avoid introducing bias.
For further analysis, the active voxels were grouped into different regions of interest
(ROIs) based on anatomical and functional criteria using the same leave-one-out
procedure (Fig. 2). The primary somatosensory cortex (Sl) ROI was created from all
active voxels in and hear the central sulcus, postcentral gyrus and postcentral sulcus.
The secondary somatosensory cortex (S2) ROI was created from all active voxels in and
near the parietal operculum. A visual association ROI was created from all active voxels