Individual patterns of functional connectivity in neonates as revealed by surface-based Bayesian modeling

Abstract

Resting-state functional connectivity is a widely used approach to study the functional brain network organization during early brain development. However, the estimation of functional connectivity networks in individual infants has been rather elusive due to the unique challenges involved with functional magnetic resonance imaging (fMRI) data from young populations. Here, we use fMRI data from the developing Human Connectome Project (dHCP) database to characterize individual variability in a large cohort of term-born infants (N = 289) using a novel data-driven Bayesian framework. To enhance alignment across individuals, the analysis was conducted exclusively on the cortical surface, employing surface-based registration guided by age-matched neonatal atlases. Using 10 minutes of resting-state fMRI data, we successfully estimated subject-level maps for fourteen brain networks/subnetworks along with individual functional parcellation maps that revealed differences between subjects. We also found a significant relationship between age and mean connectivity strength in all brain regions, including previously unreported findings in higher-order networks. These results illustrate the advantages of surface-based methods and Bayesian statistical approaches in uncovering individual variability within very young populations.

Keywords: Bayesian modeling; RSFC; brain development; fMRI; functional connectivity; neonatology; resting-state networks; statistical methods.

Figure 1:. Estimation of subject-level IC maps.

To estimate the posterior mean and variance of the subject-level ICs, along with the model parameters including the mixing matrix and residual variance, a computationally efficient expectation-maximization (EM) algorithm is used. An initial estimation of the model parameters $\left({\Omega }_i\right)$ is obtained from dimensionality reduction $\left(y_i\right)$ of the BOLD timeseries, while the mixing matrix ($A_i^{(0)}$) is obtained from a rough dual regression estimate of the individual IC maps. The empirical population prior parameters, namely, the mean $\left({\mu }_0(v)\right)$ and the between-subject variance ${\sigma }_0(v)$ of the group-level IC maps are calculated from a uniform sample of the cohort. After the EM algorithm converges, an estimate of the subject-level mean $\left(\widehat{{\mu }_{i,q}}(v)\right)$ and standard deviation $\left(\widehat{{\sigma }_{i,q}}\right)$ for each IC is obtained. A binary mask of significant engagement for each IC is obtained from a t-test corrected for multiple comparisons on the individual mean and variance IC maps.

Figure 2:. Subject-level maps obtained from 10 minutes of resting-state fMRI data.

(A) Five representative IC maps obtained from dual regression for a term-born infant scanned at 42.6 weeks PMA. (B) Posterior mean IC maps derived from the Bayesian model produced cleaner maps than the dual regression approach. White arrows highlight some areas of notable improvement in cluster convexity (e.g. default mode network) or reduction of spurious engagement (e.g. somatosensory, inferior parietal). All maps are projected onto the inflated 40-week atlas. A symmetrical scale around zero is defined by setting the maximum at the 98th percentile of FC (a.u.) for each IC map.

Figure 3:. Subject-level maps across different subjects.

Unthresholded t-statistic maps were computed from the mean and standard error maps derived from the Bayesian model estimation. The five ICs shown in Figure 2 are displayed for three term-born infants scanned at 40.6 weeks PMA, 41.7 weeks PMA, and 43.9 weeks PMA, respectively. Although a general spatial topology is preserved within each specific network, there are evident variations across individuals (as indicated by the white arrows).

Figure 4:. Group vs. individual parcellations

Functional parcellations were obtained using a winner-takes-all approach, dividing the cortex into eight distinct brain networks and subnetworks. The group parcellation was obtained from the Z-score maps derived from the group ICA analysis on 24 subjects. The individual parcellations were obtained from the subject-level t-maps derived from the Bayesian estimation on the subjects shown in Figure 3. White arrows highlight notable topographical differences between subjects. The results are displayed only in regions where the global signal-to-noise ratio (SNR) is greater than 17 dB (see Figure S3 for a whole depiction of the global SNR computed for this cohort). All parecellation are projected onto the 40-week inflated atlas.

Figure 5:. Effect of age at scan.

The relationship between age at scan and individual connectivity strength for each network was assessed by a Spearman’s rank correlation test, after controlling for sex and motion. All networks exhibited significant effects with age (p < 0.05) as determined by the corresponding Spearman’s p-values. A dashed blue line showing a linear trend is also included for illustration purposes.

Figure 6:. Frequency map of the entire cohort.

Each vertex represents the percentage of subjects that share the same parcel or network. Networks with higher spatial overlap such as the primary sensory RSNs are represented by warm colors. Note that boundaries between networks and areas with low SNR show higher variability (i.e., low spatial overlap) across subjects.