Poster E118, Monday, March 26, 2:30-4:30 pm, Exhibit Hall C
Determining the functional anatomy of the human brain by using a combined VLSM and Bayesian network analysis approach
Audrey Arnoux1,2, Monica N. Toba1, Joel Daouk3, Jean-Marc Constans3, Laurent Puy1,2, Momar Diouf4, Mélanie Barbay1,2, Olivier Godefroy1,2; 1Laboratory of Functional Neurosciences, EA 4559, University of Picardy Jules Verne, Amiens, France, 2Department of Neurology, Amiens University Hospital, Amiens, France, 3Department of Imaging, Amiens University Hospital, Amiens, France, 4Department of Biostatistics, Amiens University Hospital, Amiens, France
Objectives: The ability of voxel-based lesion-symptom mapping (VLSM) to define the functional anatomy of the human brain has not been fully assessed. With a view to assessing VLSM’s validity, the present study analyzed the technique’s ability to confirm the known clinical-anatomic correlates of hemiparesis in stroke patients. Method: Lesions (in at least 5 patients) associated with transformed limb motor score (after adjustment on lesion volume) at 6 months were examined in 272 patients using VLSM. The value of additional multivariable linear, logistic and Bayesian analyses was examined. Results: We checked that motor hemiparesis was fully accounted for by corticospinal tract (CST) lesions. Conventional VLSM analysis flagged up 2 regions corresponding to the CST, but also 8 regions located outside the CST. All 10 brain regions achieving statistical significance in the VLSM analysis were submitted to additional analyses. The backward linear regression analysis selected 5 regions, one only corresponding to the CST. The logistic regression analysis selected correctly the CST. The Bayesian network analysis selected regions including the CST and identified the source of multicollinearity. These lesions evaluated by structural equation modeling resulted in an excellent fit. Analyses of confounding factors showed that conventional VLSM analyses were strongly influenced by lesion frequency and multicollinearity. Conclusions: Conventional VLSM analyses are sensitive but weakened by a type I error due to the combined effects of multicollinearity and lesion frequency. We demonstrate that the addition of a Bayesian network analysis, and to a lesser extent of logistic regression, controlled for this type I error.
Topic Area: METHODS: Neuroimaging