Rank-adaptive covariance changepoint detection for estimating dynamic functional connectivity from fMRI data
The analysis of functional connectivity (FC) networks in resting-state functional magnetic resonance imaging (rs-fMRI) has recently evolved to a dynamic FC approach, where the functional networks are presumed to vary throughout a scanning session. Central challenges in dFC analysis involve partitioning rs-fMRI into segments of static FC and achieving high replicability while controlling for false positives. In this work we propose Rank-Adapative Covariance Changepoint detection (RACC), a changepoint detection method to address these challenges. RACC utilizes a binary segmentation procedure with novel test statistics able to detect changes in covariances driven by low-rank latent factors, which are useful for understanding changes occurring within and between functional networks. A permutation scheme is used to address the high dimensionality of the data and to provide false positive control. RACC improves upon existing rs-fMRI changepoint detection methods by explicitly controlling Type 1 error and improving sensitivity in estimating dFC at the whole-brain level. We conducted extensive simulation studies across a variety of data generating scenarios, and applied RACC to a rs-fMRI dataset of subjects with schizophrenia spectrum disorder and healthy controls to highlight our findings.
READ FULL TEXT