2026-10-03
NASCAR splits four large-scale networks into 15 overlapping subnetworks on 1,000 HCP subjects; brainstem atlas similarity predicts CRS-R in 28 TBI patients at r=0.46.
Cortical large-scale networks can already be split into subnetworks. Subcortical mapping has not. Two barriers sit on top of each other. Subcortical structures are small, so distributed cortical signals get funneled into the same nuclei: networks overlap in space and correlate in time. ICA assumes statistical independence. Seed-based correlation assumes a linear point-to-point link. Neither can unmix that overlap. Subcortical BOLD is also weak, and physiological noise is worse in deep tissue. There is still no consensus atlas of subcortical functional connectivity, which leaves neuromodulation targets talking to a whole thalamus or a whole brainstem.
This paper runs NASCAR tensor decomposition on resting-state fMRI from 1,000 Human Connectome Project young adults, splits the default mode, sensorimotor, visual, and higher-order cognitive networks into 15 subnetworks, and releases a normative subcortical atlas.
NASCAR (Nadam-Accelerated SCAlable and Robust) is an unconstrained Canonical Polyadic decomposition. BrainSync first rotates each subject's time series onto a shared temporal axis. The data then form a tensor of about 36,000 spatial locations × 1,200 time points × 1,000 subjects. Each component is an outer product of a spatial map, a time course, and a subject-loading. ICA and PCA force independence or orthogonality. NASCAR does not, so two maps can occupy the same thalamic voxels and two time courses can correlate. Subcortex has few samples, and a low-rank fit would otherwise be pulled toward the cortex, so the cerebellum is dropped and cortex is spatially downsampled by three. Rank is set to 30.
Stability comes from bootstrap: 100 draws with replacement over the 1,000 subjects, 30 components each, 3,000 maps clustered with spatial cross-correlation plus UMAP and HDBSCAN. Primary visual and sensorimotor networks show up in nearly every run. DMN, salience, executive control, and attention networks sit at moderate-to-high reproducibility. The 15 subnetworks are then assigned by hand to four families using anatomy, prior maps, and spatial, temporal, subject, and spectral correlation matrices. The subcortical piece of each subnetwork is the last third of the same spatial vector, recovered jointly with cortex, not by planting a seed in the thalamus after the fact.
The HCP data used here are 3T, LR phase-encode only, first session only, 2 mm, TR 720 ms, 15 minutes. Maps are thresholded at the top 5% for figures.
The 15 subnetworks are: three DMN (medial temporal, dorsomedial prefrontal, and a new ventral posterior cingulate component, DMN-vPCC), three sensorimotor (face, hand, foot), five visual (primary, visuo-motor, ventro-dorsal stream, dorso-dorsal stream, ventral stream), and four higher-order cognitive (salience, executive control, dorsal attention, ventral attention).
The heaviest overlap sits in caudate, putamen, hippocampus, and thalamus. DMN-vPCC is the DMN branch that reaches brainstem arousal nuclei, including locus coeruleus and midbrain reticular formation. The three sensorimotor subnetworks overlap heavily in thalamus and separate on cortex along the face-hand-foot map.
On the clinical side, 28 patients with acute severe traumatic brain injury from the RESPONSE cohort get personalized subnetwork maps. Similarity between each patient's brainstem map and the 1,000-subject atlas, plus age and sex, is used in leave-one-out OLS to predict Coma Recovery Scale-Revised (CRS-R) total scores.
| Region | Pearson r | p |
| Brainstem | 0.46 | 0.01 |
| Thalamus | 0.44 | 0.02 |
| All subcortex | 0.08 | 0.69 |
| Cortex | 0.03 | 0.88 |
| Whole brain | -0.06 | 0.75 |
In a single OLS fit on all 28, the significant predictors are age (p=0.014), ventral attention (p=0.012), salience (p=0.021), and face sensorimotor (p=0.042). The consciousness signal sits in brainstem and thalamus. It does not sit on the whole cortical map.
For network analysts, this is a complete demonstration that allowing overlap can be pushed from cortex into subcortex, with atlases planned for OpenNeuro and Lead-DBS. For DBS, ultrasound, and drug targeting, the caudate, putamen, hippocampus, and thalamus hubs are a finer grain than the whole thalamus. The TBI analysis is a proof of principle: a healthy atlas can be used as a ruler on patients. The sample is small.
This is a method-driven atlas paper. The analysis problem is real, the resource will be public, and the clinical numbers do not yet support a treatment claim.
The authors list several: there is no ground truth for large-scale networks; mesoscale fMRI overlap is not microcircuit overlap; resolution, SNR, and registration blur boundaries; auditory and language networks are missing; HCP young adults will not match other ages or pathologies; NASCAR is computationally heavy; research-grade scans have not been tested in the clinic.
The clinical numbers are the weaker part. Twenty-eight patients, 17 predictors, uncorrected p-values, OLS leave-one-out, r=0.46 is a moderate correlation with a real overfitting risk. Subnetwork labels still come from post-hoc visual inspection, the same human step as ICA. Only one LR run per subject was used, so half of the HCP resting data never entered the tensor.