73
A finite impulse response model was used to fit the MR time series in the context
of the general linear model using the AFNI program 3dDeconvolve (Cox 1996). The
average response to each trial type in each voxel was determined with nine individual
tent functions that modeled the entire BOLD response from 0 seconds to 16 seconds
post stimulus, accounting for overlapping responses from consecutive trials without any
assumptions about the shape of the hemodynamic response (Glover 1999). An F-test
was used to find active voxels, defined as those in which the tent functions for the hand
and foot stimulation trials accounted for a significant fraction (p < 10^6) of the variance.
CIassiflerTrainingandTesting
Separate classifiers, as implemented in SVMIight (Joachims 1999) were
constructed for each subject. Complementary analyses with a different package, LibSVM
(Chang and Lin 2001), gave very similar results. Within each subject, the SVM was
trained using one set of data from the subject. Then, the SVM was tested on additional
data not used fortraining.
The input to the SVM consisted of a matrix of pattern vectors, Xy,/. X had N rows
corresponding to the number of active voxels, with y corresponding to the trial type and
/ corresponding to the trial index of that trial type. Since the feature dimension N was
high, a linear kernel was used to lower the computation time (LaConte, et al. 2005;
LaConte, et al. 2007). Separate classifiers were constructed for each pair of stimuli and
combined using a decision directed acyclic graph (Platt, et al. 2000).