A PCA-based thresholding strategy for group studies of brain connectivity - with applications to resting state fMRI

Functional brain connectivity can be measured from BOLD fMRI recordings using graph representations incorporating inter-regional similarity of time courses. The resulting binary graphs are commonly constructed using thresholding of correlation values between pairwise time series anchored at their corresponding brain locations, representing the nodes in the graph. In this work we propose a new data driven approach to the challenge of selecting correlation threshold levels in group studies. Our approach addresses a collection of subject-specific graphs, thresholded at different levels. It combines information such that variability is preserved, and detection of sub-groups in the sample, having specific properties, is facilitated. The method is illustrated on simple synthetic graphs, and also tested on data from a resting state fMRI study of healthy elderly people with varying genetic risk of Alzheimers disease.