Abstract
Post-stroke neuropsychological evaluation can take a long time to assess motor, cognitive, language impairments, etc. in subjects with no overt clinical deficits. A ten-minute non-invasive diffusion tensor imaging (DTI) can potentially pick subtle impairments like subclinical language deficit. Machine learning can supplement neuropsychological test to detect language deficit from white matter tracts in post-stroke DTI with high accuracy. Thirty subjects were categorized into three groups based on stroke status and language neuropsychological test (verbal fluency task): ten healthy controls, ten ischemic stroke subjects with language deficit (normed verbal fluency score < -1.5) and ten ischemic stroke subjects with no language deficit (normed fluency score >= -1.5). Subjects were age and gender matched across groups. DTI with 56 directions acquired on 3T GE MRI 750 scanners were processed using standard steps to generate fractional anisotropy (FA) maps that measure microstructural integrity of the brain. Individual 48 white matter tracts were extracted using Johns Hopkins University (JHU) ICBM-DTI-81 white matter atlas. Mean FA per tract per subject was computed. A support vector machine (SVM) used mean FA to classify among groups. With a Gaussian kernel and leave-one-out cross validation, SVM ranked tracts predicting language deficit. Groups were age and gender matched resulting in no significant difference in the age across groups ( p > 0.05). While a three-class SVM yields a 57% classification accuracy, binary SVM on pairs of groups provides accuracy of 80% (healthy controls vs. stroke with language deficit), 85% (healthy controls vs. stroke with no language deficit) and 75% (stroke with language deficit vs. stroke with no language deficit). Top independent white matter tract with highest weight given by classifier was: splenium of corpus callosum, right superior corona radiata, right external capsule respectively for the three classifications above. In conclusion, SVM classifier can predict whether a given subject has language deficit from post-stroke DTI with accuracy significantly higher than random chance. Weights obtained from SVM indicate importance of specific tracts driving the classifier and distinguishing the groups.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have