Then, we trained a machine-learning classifier (a Support Vector Machine, SVM) to automatically predict coma outcome (good poor) based on whole-brain FC of comatose patients.
Multivariate pattern classification, as described in Klöppel et al. [10], was performed with a linear kernel by identifying a separating hyperplane that maximizes the distance between different clinical groups based on whole-brain or ROIs information.
In the present study, we subdivided the insula based on whole-brain functional connectivity maps, using a data-driven clustering technique.
For this reason, our first imaging results are based on whole-brain contrasts, at a relatively liberal statistical threshold (initial cluster forming threshold Z>2, family wise error corrected cluster extent threshold p<0.05).
In previous work, based on whole-brain MTR histograms, it was found that SLE patients with active neuropsychiatric symptoms, past neuropsychiatric symptoms, and SLE patients without neuropsychiatric symptoms could be distinguished, suggesting diagnostic potential for these parameters [ 6- 8].
Conversely, pattern recognition approaches (e.g. SVM) are multivariate techniques (i.e. they take into account spatial correlations in the data), and if the discrimination is based on the whole brain pattern, all voxels within the brain will contribute to the classification and no local inferences based on these approaches should be made.
Much better temporal resolution can be obtained by identifying regions of interest (ROIs) based on the whole brain analyses, then reconstructing virtual electrode outputs targeted specifically at each ROI.
These conclusions are based on whole-cell recordings from GRF2 knockout hippocampal brain slices using low frequency stimulation coupled with depolarization of the post-synaptic cell (LFS pairing).
