Fig. 1: Framework for analyzing multiway multiscale interactions within brain networks.

The preprocessed infant resting-state fMRI data were input into Multivariate Objective Optimization ICA with Reference (MOO-ICAR) using the spatially constrained NeuroMark2.1 Template, derived from over 100K subjects³¹. This process yielded subject-specific estimates of 105 intrinsic connectivity networks (ICNs) and their corresponding time courses in infant rsfMRI. These 105 ICNs were then grouped into six brain domains: visual (VI), cerebellar (CB), temporal (TP/TMN), subcortical (SC), somatomotor (SM), and higher cognitive (HC), as depicted in A. As we know, the network in the human brain is not isolated; interactions occur both pairwise and beyond pairwise for information exchange. In this study, we estimated pairwise (interaction order = 2) and triple interactions (interaction order = 3) among intrinsic connectivity networks (ICNs), shedding light on multiway interactions in the infant brain, as shown in B. We first estimated the pairwise interactions for each subject and then concatenated these pairwise interactions (105 × 105) across subjects (F × Subjs, where F represents the total number of features). For triple interactions, we flattened the 3D interaction tensors (105 × 105 × 105) after estimating them and then concatenated them across subjects. Due to the high dimensionality of triple interactions and to explore the latent space underlying these complex pairwise and triple interactions, we applied group independent component analysis (GICA) to decompose the intricate patterns of both pairwise and triple interaction components. Subsequent analyses then focused on exploring the relationship between age and the decomposed brain network interaction components, as illustrated in C.