82
foot). In order to determine whether MVPA could also be successful in a more difficult
somatosensory classification task, in experiment 2 touches were delivered to closely
spaced locations on the body surface (different digits on the same hand and the foot).
In every one of 8 subjects, decoding performance was much greater than
chance, with a mean of 56% ± 4% SEM across subjects (chance performance of 25% for
4-way decoding). As a more rigorous test, we measured decoding performance for 3-
way decoding of different fingers on the same hand. Performance was good for finger
decoding, with a mean of 68% ± 3% SEM across subjects (chance performance of 33%).
Subdividing the active voxels revealed significant differences in decoding performance
between ROIs (Fig. 3D; F(4,39)=7.3, p = 0.0002). The best performance was found in Sl
(61% ±4% for the Sl ROI) and S2 (50% ± 3%).
TheMvpAanalysisusedmultivariateinformationfrommanyVoxeIsto
Successfully classify individual stimulation trials. Traditional univariate methods examine
the average BOLD response to different trial types averaged across voxels in an ROI.
Could classification be performed with the BOLD response to individual trials in an ROI?
First, we examined the easier classification task of experiment 1. Figure 6A shows the
average response to left hand and right hand touches in left hemisphere S2 of a single
subject. The response, averaged across all trials, was significantly greater for right hand
than left hand touches. Next, we measured the response to each individual trial in left
hemisphere S2 (Fig. 6B). While on average the BOLD signal change was greater in right
hand than left hand trials, the distributions of the signal changes were largely
overlapping. The optimal classification boundary was calculated as the average of the