| |
Abstract:
We propose a multivariate approach that systematically
identifies the most typical and atypical subjects in a
multi-subject fMRI experiment. For this pilot study, we used data
from a study of multimodal selective attention, discussed in a
related poster (see V.P. Clark et al.). Multiple regression was
performed on individual subject's data to create Z-score maps of
significant activation in each of four visual attention conditions.
"Canonical subject analysis" compares subjects in two steps. First,
for each subject a similarity matrix is derived for the four
intra-subject attention conditions. This matrix describes an
"activation space" unique to each subject, relating each anatomical
pattern of activation to the remaining three. Second, the
similarity matrices derived in the first step are themselves
compared for similarity, resulting in a second-order similarity
matrix across subjects, a "subject space" for this experiment. From
this second-order matrix we determine which of the 14 subjects are
closest to the center, or "most typical", of the inter-subject
space, and which are multi-variate outliers. The application of
this technique to experiments with more task conditions and more
subjects accordingly enable us to close in on a prototypical
pattern of activation with general validity, without loss of
specificity.
|
