Home

Initial SVR-LSM implementation

The application of support vector regression to lesion-symptom mapping was developed by Zhang et al. (2014). The approach treats voxel values of lesion maps as features (1/0) and the behavioral vector as the target of training. After the model is trained, the predictive hyperplane is back-projected into brain space to visualize the findings (SVR-β map). Voxelwise reliability is determined by a voxelwise threshold determined via permutation testing.

This Implementation

This implementation is described in DeMarco & Turkeltaub (2017), which is an adaptation of the initial implementation of support vector regression to lesion-symptom mapping (Zhang et al., 2014).

New Features

• Clusterwise FWE-correction via permutation testing
• Allows computation completely within MATLAB software (no libSVM to compile)
• Flexible handling of covariates via nuisance models
• Load/save configuration files
• Parallelization
• Batch run pre-saved analysis configurations from command line or graphic interface
• Invoke from command line with custom scripts
• Basic output summary and overview
• Model diagnostics

Details of new features

Cluster-level correction

Cluster level family-wise error correction is conducted via permutation testing.

Flexible handling of covariates

Arbitrary covariates are handled via nuisance model with a statistical partialing out strategy.

Lesion-volume control methods

Five options for lesion volume control are provided, including:

• No correction
• Direct Total Lesion Volume Control (dTLVC; Zhang et al., 2014): This approach divides the 1-values in lesioned voxels of lesion tracings by the square root of the lesion volume of each participant. Importantly, this does not transform spared 0-valued voxels, as zero divided by any value beside zero is also zero. In DeMarco & Turkeltaub (2017) we suggest that this remaining bias in unlesioned voxels is problematic when the behavior of interest is correlated with lesion volume.
• Regress on Behavior: This method covaries lesion volume out of the behavioral vector a la a semi-partial correlation.
• Regress on Lesion Data: This method covaries lesion volume out the raw 1/0 lesion map data on a voxel by voxel basis, a la a semi-partial correlation.
• Regress on Both: This method covaries lesion volume out of both lesion data and the behavior vector, a la a partial correlation. It is conceptually equivalent to an additional predictor in a multiple regression