Hierarchical subdivision and effect of ICA model dimensionality on the interoceptive task-derived brain networks

Abstract

Network analysis has become a tool of choice for studying functional Magnetic Resonance Imaging (fMRI) data. In this study, we used Independent Component Analysis (ICA) to characterize human brain networks related to interoception. Group-ICA was applied to blood oxygenation level-dependent (BOLD) fMRI data collected from 15 healthy subjects, who underwent intravesical stimulation, at four different ICA dimensionalities (i.e., K = 10, 20, 30, and 40 components) to assess hierarchical breakdown (subdivision) and the impact of ICA model dimensionality on the characteristics of intrinsic brain networks derived from imaging visceral interoception. The default mode network, the visual network, the network of inferior frontal/inferior temporal gyri, the limbic association network, the brainstem, and cerebellar networks were identified even at the lowest dimensionality of 10. The central executive network, the salience network, the self-referential network, the dorsal attention network, and the sensorimotor network appeared only with the ICA model dimensionality of at least 20. The ventral attention network, the thalamic network, and additional default mode, salience, limbic association, and brainstem networks identified with decomposition of 40 components suggesting that detection of some networks requires increasing model dimensionality. Furthermore, the lack of specific network subdivision might be related to the nature of the BOLD fMRI data upon which ICA was performed.