Abstract Details

Title Support Vector Machine Classification of Stroke Using Resting State Functional Connectivity

Topic Cerebrovascular Disease and Interventional Neurology

Presentation(s) P03 - (-)

Poster/Presentation
Number 193

Objective

Background BACKGROUND: Resting-state functional MRI (rs-fMRI) is a powerful technique for studying whole brain neural connectivity. Pattern analysis methods can be used with rs-fMRI data to extract significant information and classify individual scans.

Design/Methods DESIGN/METHODS: Blood oxygen level-dependent (BOLD) time courses from rs-fMRI scans were extracted for 160 regions that cover most of the brain. All possible (12,720) region-to-region temporal correlations, or functional connections, were calculated for each subject. These correlation measures were input as features into a multivariate pattern recognition algorithm to build a model for discriminating between normal subjects (n=24, mean age = 47.4 years) and acute stroke patients (n=24, mean age = 58.8 years).

Results RESULTS: A linear kernel SVM classifier discriminated between stroke and normal subjects with 83% accuracy (p-value <1 [times] 10-6, sensitivity = 79%, specificity = 87.5%). The classifier was influenced most by the connectivity of the cingulo-opercular (30% relative weight) and sensorimotor (34% relative weight) networks.

Conclusions CONCLUSIONS: SVM methods can be applied to rs-fMRI derived brain connectivity measures to accurately discriminate between stroke and normal subjects. Connections between brain regions that most influentially drive the classifier can be extracted to gain insight into pathological aging (stroke) and healthy aging subject connectivity.

Authors/Disclosures
PRESENTER	No disclosure on file
Filippo Martinelli Boneschi	No disclosure on file
Veena A. Nair, PhD (DEPT. OF RADIOLOGY , UW MADISON)	The institution of Dr. Nair has received research support from NIH.
Matthew Jensen, MD (Concord Hospital)	No disclosure on file
Marcus Chacon, MD (University of Wisconsin Hospital)	No disclosure on file
Justin A. Sattin, MD (University of Wisconsin)	The institution of Dr. Sattin has received research support from NIH / NINDS.
Vivek Prabhakaran, MD (University of Wisconsin)	Dr. Prabhakaran has nothing to disclose.

��ɫ��

��ɫ��