I'm happy to report that a collaborative project from Penn is now published, spearheaded by Phil Cook. In this paper we explored combining approaches in order to relate individual differences in gray and white matter to behavioral performance. Phil implemented two core steps in a group of participants that included frontotemporal dementia (FTD) patients and healthy older adults. For each participant we had T1- and diffusion-weighted structural images, providing cortical thickness and fractional anisotropy (FA) measurements. We also had a set of behavioral measures that included category fluency ("Name as many animals as you can in 30 seconds") and letter fluency ("Say as many words beginning with the letter F as you can in 30 seconds").
Regions of interest for cortical thickness (top row) and FA (bottom row) defined by eigenanatomy.
Phil first used eigenanatomy in order to define regions of interest (ROIs) for the gray matter images. Eigenanatomy is a dimensionality reduction scheme that identifies voxels that covary across individuals; ROIs are chosen that can maximally explain variance in the dataset.
The second step is my favorite aspect of this work, and can be implemented regardless of how ROIs are defined. Phil used a model selection procedure implemented in R to assess which combination of ROIs best predicted behavior. He used a combination of cross-validation and AIC to evaluate what predictors performed best. The elegant thing about this approach is that it incorporates both gray matter and white matter predictors in the same framework; thus, the model selection procedure can tell you whether gray matter alone, white matter alone, or some combination best explain the behavioral data.
Cross-validation model selection suggests the including both gray and white matter predictors (black circles) results in significantly better performance than any single modality, and that 4 regions provide the best predictions.
ROIs significantly associated with verbal fluency.
Perhaps not surprisingly, combining gray matter and white matter was consistently better than using either modality alone, as one might expect from a cortical system comprised of multiple regions connected with white matter tracts. It is encouraging that the regions identified are sensible in the context of semantic storage and retrieval during category fluency.
More importantly, the approach that Phil put together tackles the larger problem of how to combine data from multiple modalities in a quantitative, model-driven approach. I hope that we see more studies that follow a similar approach.
Reference:
Cook PA, McMillan CT, Avants BB, Peelle JE, Gee JC, Grossman M (2014) Relating brain anatomy and cognitive ability using a multivariate multimodal framework. NeuroImage 99:477-486. doi:10.1016/j.neuroimage.2014.05.008 (PDF)