Multivariate lesion-symptom mapping using support vector regression

Abstract

Lesion analysis is a classic approach to study brain functions. Because brain function is a result of coherent activations of a collection of functionally related voxels, lesion-symptom relations are generally contributed by multiple voxels simultaneously. Although voxel-based lesion symptom mapping (VLSM) has made substantial contributions to the understanding of brain-behavior relationships, a better understanding of the brain-behavior relationship contributed by multiple brain regions needs a multivariate lesion symptom mapping (MLSM). The purpose of this paper was to develop an MLSM using a machine learning-based multivariate regression algorithm: support vector regression (SVR). In the proposed SVR-LSM, the symptom relation to the entire lesion map as opposed to each isolated voxel is modeled using a non-linear function, so the intervoxel correlations are intrinsically considered, resulting in a potentially more sensitive way to examine lesion-symptom relationships. To explore the relative merits of VLSM and SVR-LSM we used both approaches in the analysis of a synthetic dataset. SVR-LSM showed much higher sensitivity and specificity for detecting the synthetic lesion-behavior relations than VLSM. When applied to lesion data and language measures from patients with brain damages, SVR-LSM reproduced the essential pattern of previous findings identified by VLSM and showed higher sensitivity than VLSM for identifying the lesion-behavior relations. Our data also showed the possibility of using lesion data to predict continuous behavior scores.

The ins and outs of meaning: Behavioral and neuroanatomical dissociation of semantically-driven word retrieval and multimodal semantic recognition in aphasia

“…We (Zhang et al, 2014) developed SVR-LSM as a method for multivariate lesion-symptom mapping. In a typical whole-brain VLSM analysis the number of voxels (hundreds of thousands) is much greater than the number of patients (usually < 1000).…”

“…Because the nonlinear transform contains interactive operations involving different voxels, it automatically takes the spatial correlations among voxels into account during model training, resulting in higher sensitivity for lesion-symptom relation detection as compared to a linear SVR-LSM as well as to VLSM (Zhang et al 2014). To locate the symptom-associated brain regions, the multivariate parametric map of the nonlinear SVR-LSM is projected back to the original brain space.…”

“…In addition, because the LSM betas are identified simultaneously for all voxels rather than from each voxel independently, SVR-LSM is less sensitive to differences in statistical power across voxels that result from differences in the proportion of participants with lesions in each voxel. These differences in power can have substantial effects on results of traditional VLSM analyses but are mitigated in SVR-LSM (this is discussed in more detail in Zhang et al, 2014). There is no single agreed-upon method for statistical thresholding in SVR-LSM (for some discussion see Zhang et al, 2014); however, our present goal is to evaluate the convergence between SVR-LSM and our previous VLSM analyses, so we thresholded the SVR-LSM maps to have the same number of supra-threshold voxels as the VLSM analyses, where the threshold was set based on false discovery rate (FDR) correction and a permutation-based cluster size threshold.…”

“…Limitations of mass-univariate VLSM also motivated recent work by our group to develop and test a multivariate lesion-symptom method based on support vector regression (SVR-LSM, Zhang et al, 2014). Analyses conducted with real and simulated data revealed that SVR-LSM has higher sensitivity and specificity for detecting lesion-behavior relations, particularly when multiple regions contribute to behavior and when the lesion proportion differs substantially across voxels.…”

Multivariate Connectome-Based Symptom Mapping in Post-Stroke Patients: Networks Supporting Language and Speech

“…This organization sug-gests that brain-behavior relationships could be better elucidated by applying multivariate, rather than univariate, analysis (Schmah et al, 2010; as supported by several fMRI studies: Misaki et al, 2010;. Lesion studies could also benefit from a multivariate approach (Smith et al, 2013;Mah et al, 2014;Yourganov et al, 2015;Zhang et al, 2014); instead of analyzing the effect of lesions in different locations independently, as is done in voxelwise lesion symptom mapping (VLSM; Bates et al, 2003), a multivariate method accounts for the interactions between spatial locations.…”