Spine-Prints: Transposing Brain Fingerprints to the Spinal Cord

Abstract

Functional connectivity (FC) patterns in the human brain form a reproducible, individual-specific "fingerprint" that allows reliable identification of the same participant across scans acquired over different sessions. While brain fingerprinting is robust across healthy individuals and neuroimaging modalities, little is known about whether the fingerprinting principle extends beyond the brain. Here, we used multiple spinal functional magnetic resonance imaging (fMRI) datasets acquired at different sites to examine whether a fingerprint can be revealed from FCs of the cervical region of the human spinal cord. Our results demonstrate that the functional organisation of the cervical spinal cord also exhibits individual-specific properties, suggesting the potential existence of a spine-print within the same acquisition session. This study provides the first evidence of a spinal cord connectivity fingerprint, underscoring the importance of considering a more comprehensive view of the entire central nervous system. Eventually, these spine-specific signatures could contribute to identifying individualized biomarkers of neuronal connectivity, with potential clinical applications in neurology and neurosurgery.

Keywords: brain fingerprinting; fMRI; functional connectivity; inter-subject variability; spinal cord imaging.

Figure 1.. Datasets overview.

Left panel: coronal view of the PAM50 template (y=70) with spinal level annotations (C2-C8) and the gray matter probability mask on the axial view. The three datasets are shown sequentially with the mean motion-corrected fMRI and the tSNR maps. Dataset 3 additionally includes whole-brain coverage, with the brain data shown alongside the MNI template of 2mm resolution in the sagittal plane (x=45).

Figure 2.

Overview of functional connectivity (FC) matrices and identifiability results for the three datasets. A. FC matrices of the first run from each dataset. The most-left matrix is a zoomed-in view of the spinal level C4 for both Dataset 1 and 2 to show the labeling and sorting of the spinal cord ROIs used for all the spinal levels in all matrices. On the right Dataset 3 FC displays the brain ROIs with Yeo’s sorting of the 7 functional networks (VIS, SM, DA, VA, L, FP, DMN), followed by the subcortical regions (SUB) and the 7 spinal levels. The brain functional networks have the left hemisphere first (Lh) and the right (Rh) after. B. The second row shows the corresponding identifiability matrices for each dataset, along with the reached scores and accuracies. C. On the left, a plot illustrates participants identifiability accuracy as a function of the top-K (from 1 to 5) highest correlation values across the three datasets. On the right, a detailed plot for Dataset 3 shows identifiability accuracy as a function of top-K values, with separate curves for brain-only and spine-only data.

Figure 3.. ICC results.

(A) Spinal cord ROIs displayed in a cross section of the spinal cord (gray matter (GM) regions: dh = dorsal horns, iz = intermediate zone, vh = ventral horns, white matter (WM) regions: sl = spinal lemniscus, cst = cortico-spinal tract, fc = fasciculus cuneatus, fg = fasciculus gracilis). (B) Average ICC matrix across all spinal levels and datasets, indicating the 95th percentile in the colorbar (0.526). (C-D-E), namely, Datasets 1,2, and 3. (E) For Dataset 3, brain ROIs are sorted according to the Yeo functional networks (Yeo et al. 2011) (VIS = visual, SM = somatomotor, DA = dorsal attention, VA = ventral attention, L = limbic, FP = fronto-parietal, DMN = default mode network, SUB = subcortical). The color bar values are adjusted according to the distribution of the values, reporting the 95th percentiles of each dataset’s ICC as minimum to show a filtered ICC matrix.

Figure 4.

Dataset 3 fingerprinting using brain-only, spine-only, and brain–spine interaction as inputs. The identifiability matrices reflect the correlations across the participants’ runs, with participants’ IDs sorted by highest correlation in each case, leading to differing axis labels across matrices. Identification accuracies (chance level = 8.3%) are 93.3% for brain-only, 53.3% for spine-only, and 40% for the interaction.