Paper detail

Effects of Scan Length and Shrinkage on Reliability of Resting-State Functional Connectivity in the Human Connectome Project

In this paper, we use data from the Human Connectome Project (N=461) to investigate the effect of scan length on reliability of resting-state functional connectivity (rsFC) estimates produced from resting-state functional magnetic resonance imaging (rsfMRI). Additionally, we study the benefits of empirical Bayes shrinkage, in which subject-level estimates borrow strength from the population average by trading a small increase in bias for a greater reduction in variance. For each subject, we compute raw and shrinkage estimates of rsFC between 300 regions identified through independent components analysis (ICA) based on rsfMRI scans varying from 3 to 30 minutes in length. The time course for each region is determined using dual regression, and rsFC is estimated as the Pearson correlation between each pair of time courses. Shrinkage estimates for each subject are computed as a weighted average between the raw subject-level estimate and the population average estimate, where the weight is determined for each connection by the relationship of within-subject variance to between-subject variance. We find that shrinkage estimates exhibit greater reliability than raw estimates for most connections, with 30-40% improvement using scans less than 10 minutes in length and 10-20% improvement using scans of 20-30 minutes. We also observe significant spatial variability in reliability of both raw and shrinkage estimates, with connections within the default mode and motor networks exhibiting the greatest reliability and between-network connections exhibiting the poorest reliability. We conclude that the scan length required for reliable estimation of rsFC depends on the specific connections of interest, and shrinkage can be used to increase reliability of rsFC, even when produced from long, high-quality rsfMRI scans.