Figure 1. Overview of the analysis pipeline.

We used the same Shen parcellation for diffusion-weighted MRI (DWI) and functional magnetic resonance imaging (fMRI) data. Time-resolved functional connectivity was estimated using a metric for phase connectivity. A proxy measure for metastability was derived from the phase information. Time-resolved networks were subsequently extracted from the concatenated data from all subjects using non-negative tensor factorization (NNTF). Dwell times and nonstationarity (excursions from the median) were retrieved for each spatial pattern of functional connectivity (six resting-state networks and one ‘residual’ network). At the same time, time-resolved connectivity was predicted on a sample-by-sample basis based on a linear combination of eigenmodes (hidden patterns in the anatomical network). Measures were used for classification and feature ranking.