74
In each subject, six scan series were collected, each containing a random
sequence of somatosensory stimuli. This allowed the use of leave-one-out cross-
validation to assess classification performance. Within each subject, six different SVMs
were constructed, each trained on a different set of five scan series collected from the
subject. Then, each SVM was tested on the single left-out scan series not used for
training. Arranging the samples in this way avoids splitting samples from one run into
both training and test sets which may be problematic due to dependency among
successive samples within each run (Ha×by, et al. 2001).
Because the BOLD response to brief somatosensory stimulation was relatively
punctate (Fig. 6A), in order to estimate thé response to individual trials we made the
simplifying assumption that the image intensity in a voxel at a given time reflected only
the somatosensory stimulus delivered two TRs (4 seconds) previously; this meant that
the estimated response to a single trial contained small contributions from previous
trials. This did not introduce bias into the classifier for two reasons. Most importantly,
all training trials were from different five-minute scan series (separated by 30 seconds -
30 minutes) from the trial being classified, preventing BOLD spillover between testing
and training trials. Any BOLD response spillover could only hurt classification
performance (by providing a less accurate estimate of the true response), and not help
classification performance (by introducing a classification signal into neighboring trials,
as would occur if training and testing was performed within a single scan series).
Second, first order counterbalancing was used when designing the stimulus sequence,
and the stimulus sequence for each scan series was randomized independently,