80
significantly greater for hand classification than for foot classification (62% vs. 53%, p <
0.001).
To avoid assumptions introduced by predefining regions of interest, we trained
additional classifiers with whole brain data. Because a linear SVM was used, the decision
boundary can be mapped directly to image space (LaConte, et al. 2007). This provides an
assumptions-free map (without predefined ROIs) of voxels that contain significant
information about the body site of stimulation. As shown in Fig. 4, voxels with high
feature space weights were found in SI, Slfoot, and S2, similar to the functional
activation maps obtained from the traditional univariate methods.
Figure 4. Support vector weight maps.
Map of the support vector weights ( ∣ weights ∣ >10 colored yellow) assigned to each voxel in an
ROI-free analysis, for the same subject as shown in Figure 2. Note the high weights for voxels
in Sl (left), Slfoot (middle) and S2 (right).
To study how classification performance changed with ROI size across all
subjects, classifiers were trained and tested with sub-ROIs consisting of from 1 to 70
voxels randomly selected from SI, S2 and MST∕STP (Figs. 5A, B, and C, respectively).
The accuracy with one voxel was low but performance increased as more voxels were
added to the ROI. The increase had a rapid initial phase followed by a slow, nearly linear