Kernel machine tests of association between brain networks and phenotypes

PLoS One. 2019 Mar 21;14(3):e0199340. doi: 10.1371/journal.pone.0199340. eCollection 2019.


Applications of quantitative network analysis to functional brain connectivity have become popular in the last decade due to their ability to describe the general topological principles of brain networks. However, many issues arise when standard statistical analysis techniques are applied to functional magnetic resonance imaging (fMRI) connectivity maps. Frequently, summary measures of these maps, such as global efficiency and clustering coefficients, collapse the changing structures of graph topology from many scales to one. This can result in a loss of whole-brain spatio-temporal pattern information that may be significant in association and prediction analyses. Drawing from the electrical engineering field, the resistance perturbation distance is a quantification of similarity between graphs on the same vertex set that has been shown to identify changes in dynamic graphs, such as those from fMRI, while not being computationally expensive or result in a loss of information. This work proposes a novel kernel-based regression scheme that incorporates the resistance perturbation distance to better understand the association with biological phenotypes from fMRI using both simulated and real datasets.

MeSH terms

  • Adult
  • Algorithms
  • Brain / anatomy & histology*
  • Brain / diagnostic imaging
  • Brain / physiology*
  • Brain Mapping
  • Computer Simulation
  • Connectome / methods*
  • Connectome / statistics & numerical data
  • Female
  • Functional Neuroimaging
  • Humans
  • Machine Learning
  • Magnetic Resonance Imaging
  • Male
  • Middle Aged
  • Models, Neurological
  • Nerve Net / anatomy & histology
  • Nerve Net / diagnostic imaging
  • Nerve Net / physiology
  • Phenotype
  • Regression Analysis
  • Statistics, Nonparametric