Identifying associations in dense connectomes using structured kernel principal component regression

Abstract

A powerful and computationally efficient multivariate approach is proposed here, called structured kernel principal component regression (sKPCR), for the identification of associations in the voxel-level dense connectome. The method can identify voxel -phenotype associations based on the voxels' whole-brain connectivity pattern, which is applicable to detect linear and non-linear signals for both volume-based and surface-based functional magnetic resonance imaging (fMRI) data. For each voxel, our approach first extracts signals from the spatially smoothed connectivities by structured kernel principal component analysis, and then tests the voxel-phenotype associations via a general linear model. The method derives its power by appropriately modelling the spatial structure of the data. Simulations based on dense connectome data have shown that our method can accurately control the false-positive rate, and it is more powerful than many state-of-the-art approaches, such as the connectivity-wise general linear model (GLM) approach, multivariate distance matrix regression (MDMR), adaptive sum of powered score (aSPU) test, and least-square kernel machine (LSKM). To demonstrate the utility of our approach in real data analysis, we apply these methods to identify voxel-wise difference between schizophrenic patients and healthy controls in two independent resting-state fMRI datasets. The findings of our approach have a better between-sites reproducibility, and a larger proportion of overlap with existing schizophrenia findings. Code for our approach can be downloaded from https://github.com/weikanggong/vBWAS.