We have previously shown that hierarchical clustering can also be used to extract functional connectivity networks, and that it can be used to investigate the intra-network organization of resting-state networks through stratifying data into a hierarchical structure naturally. In this study, we devise a data-driven scheme to analyze resting-state networks using hierarchical clustering and applied this scheme to analyze the sensory-motor network.The purpose of this study is to apply the data-driven hierarchical clustering analysis scheme to the sensory-motor network to analyze its functional sub-division, and intra-network functional organization.The sensory-motor was previously extracted using full-brain hierarchical clustering at voxel level from 86 health subjects¹. The dendrogram for the network, as obtained though hierarchical clustering, was cut at increasing cut levels (k). At each cut, the new clusters were evaluated for significance (p<0.01). The significance criterion is based on both intra-cluster correlation and cluster size, obtained through group-level statistics from intra-cluster correlation values and Monte-Carlo simulations of cluster formation from Gaussian noise (fig. 1). Clusters that are determined to be not meaningful in terms of functional subdivision were removed from the dendrogram. This process was repeated up to cut level k=80 whereupon enough sub-networks (19) were obtained to warrant figure display. A cluster stands out to be the largest among the resulting sub-networks, and appears to be the central hub of the network². It was chosen for further investigation to illustrate the frameworks capability to analyze resting-state networks greater detail. The same procedure was carried out as previously. This further sub-division corresponds to cutting the dendrogram at nodes corresponding to a plateau in the inconsistency coefficient, indicating that this is a natural termination point for further sub-divisions.

The intra-network hierarchy of the extracted sub-networks and their neuroanatomical locations are depicted in figure 2. The extracted sub-networks include 4 groups of bilateral sub-networks: Insula-auditory sub-networks (fig. 2, clusters 1-3); paracentral lobule and cingulate-motor sub-networks (fig. 2, clusters 4 and 5); facial expression control (fig. 2, clusters 11-13); and hand movement control (fig. 2, clusters 6, 7, 9, and 10). There exist also two groups of more loosely connected unilateral sub-networks: The right parietal sub-network group (fig. 2, clusters 15 and 16) and the left insular-STG (superior temporal gyrus) sub-network group (fig. 2, clusters 17-19). This intra-network organization of intrinsic functional connectivity derived from spontaneous activity of the brain at rest reflects consistently the functional and neural anatomic connectivity topography network, which consists of S1, M1, and some pre/post- central gyrus areas divided into dorsal and ventral subgroups in addition to the parietal operculum and the auditory cortex³.

The functional sub-division of the paracentral lobule effectively distinguishes the supplementary motor area (SMA) (fig. 3, cluster 4), areas in the somatosensory cortex responsible for the legs and feet⁴ (fig. 3, clusters 1-3), and higher order motor-learning areas⁵ bordering the cingulate cortex (fig.3, cluster 5).

The implemented hierarchical clustering scheme is a conscientious method where certain internal characteristics of the data, such as cluster size and intra-cluster correlation. Another characteristic of the data, the inconsistency coefficient, created a natural end-point to the sub-division. Further sub-divisions (up to k=200) yielded no more significant clusters given our cluster evaluation criteria. Hierarchical clustering nevertheless produces a large amount of information that can be difficult to survey even under the current scheme. This visualization problem is overcome by presenting only manageable sized sub-sets at a time. To demonstrate this, we further sub-divided the cluster corresponding to the paracentral lobule as an illustrative example. While our choice is in some ways natural due to its comparatively larger size, and central role in functional modulation within the network, it should be noted that this can be done equally well for any other sub-networks

We have shown that hierarchical clustering is capable of analyzing the functional organization of resting-state networks in different levels of detail. The functional hierarchy constructed through voxel-wise hierarchical clustering corroborate consistently with the known modular organizations from previous clinical and experimental studies.