An empirical evaluation of multivariate lesion behaviour mapping using support vector regression

Multivariate lesion behaviour mapping based on machine learning algorithms has recently been suggested to complement the methods of anatomo-behavioural approaches in cognitive neuroscience. Several studies applied and validated support vector regression-based lesion symptom mapping (SVR-LSM) to map anatomo-behavioural relations. However, this promising method, as well as the multivariate approach per se, still bears many open questions. By using large lesion samples in three simulation experiments, the present study empirically tested the validity of several methodological aspects. We found that i) correction for multiple comparisons is required in the current implementation of SVR-LSM, ii) that sample sizes of at least 100 to 120 subjects are required to optimally model voxel-wise lesion location in SVR-LSM, and iii) that SVR-LSM is susceptible to misplacement of statistical topographies along the brain’s vasculature to a similar extent as mass-univariate analyses.