Support vector regression based multivariate lesion-symptom mapping.

Abstract

A novel multivariate lesion-symptom mapping (LSM) methodology was developed in this study. Lesion analysis is a classic model for studying brain functions. Using lesion data, focal brain-behavior associations have been widely assessed using the massive voxel-based lesion symptom mapping (VLSM) method. Assessing each voxel independently, VLSM suffers from low sensitivity after correcting for the enormous number of comparisons. It is also incapable for assessing a spatially distributed association pattern though the brain-behavior associations generally involve a collection of functionally related voxels. To solve these two outstanding problems, we carried out the first multivariate lesion symptom mapping (MLSM) in this study using support vector regression (SVR). In the so dubbed SVR-LSM, the symptom relation to the entire lesion map rather than each isolated voxel is modeled using a non-linear function, so the inter-voxel correlations are intrinsically considered, resulting in a potentially more sensitive way to examine lesion-symptom relationships. Evaluations using synthetic data and real data showed that SVR-LSM gained a much better performance (in terms of sensitivity and specificity) for detecting brain-behavior relations than VLSM. While the method was designed for lesion analysis, extending it to neuroimaging data will be straightforward.