Considerable evidence indicates that the functional connectome of the healthy human brain is highly stable, analogous to a fingerprint.
We investigated the stability of functional connectivity across tasks and sessions in a cohort of individuals with chronic stroke using a supervised machine learning approach.
Twelve individuals with chronic stroke underwent functional magnetic resonance imaging (fMRI) seven times over 18 weeks. The middle 6 weeks consisted of intensive aphasia therapy. We collected fMRI data during rest and performance of 2 tasks. We calculated functional connectivity metrics for each imaging run, then applied a support vector machine to classify data on the basis of participant, task, and time point (pre- or posttherapy). Permutation testing established statistical significance.
Whole brain functional connectivity matrices could be classified at levels significantly greater than chance on the basis of participant (87.1% accuracy; P < .0001), task (68.1% accuracy; P = .002), and time point (72.1% accuracy; P = .015). All significant effects were reproduced using only the contralesional right hemisphere; the left hemisphere revealed significant effects for participant and task, but not time point. Resting state data could also be used to classify task-based data according to subject (66.0%; P < .0001). While the strongest posttherapy changes occurred among regions outside putative language networks, connections with traditional language-associated regions were significantly more positively correlated with behavioral outcome measures, and other regions had more negative correlations and intrahemispheric connections.
Findings suggest the profound importance of considering interindividual variability when interpreting mechanisms of recovery in studies of functional connectivity in stroke.
]]>