Multivariate Brain Functional Connectivity Through Regularized Estimators.

Functional connectivity analyses are typically based on matrices containing bivariate measures of covariability, such as correlations. Although this has been a fruitful approach, it may not be the optimal strategy to fully explore the complex associations underlying brain activity. Here, we propose extending connectivity to multivariate functions relating to the temporal dynamics of a region with the rest of the brain. The main technical challenges of such an approach are multidimensionality and its associated risk of overfitting or even the non-uniqueness of model solutions. To minimize these risks, and as an alternative to the more common dimensionality reduction methods, we propose using two regularized multivariate connectivity models. On the one hand, simple linear functions of all brain nodes were fitted with ridge regression. On the other hand, a more flexible approach to avoid linearity and additivity assumptions was implemented through random forest regression. Similarities and differences between both methods and with simple averages of bivariate correlations (i.e., weighted global brain connectivity) were evaluated on a resting state sample of N = 173 healthy subjects. Results revealed distinct connectivity patterns from the two proposed methods, which were especially relevant in the age-related analyses where both ridge and random forest regressions showed significant patterns of age-related disconnection, almost completely absent from the much less sensitive global brain connectivity maps. On the other hand, the greater flexibility provided by the random forest algorithm allowed detecting sex-specific differences. The generic framework of multivariate connectivity implemented here may be easily extended to other types of regularized models.