Introduction

Extremely preterm birth (<27 weeks of gestation) has been linked to white matter (WM) abnormalities in infants¹ but the functional implications of these abnormalities are poorly understood. Thus, the long-term effect of these alterations on adolescents needs further investigation. Brain WM wiring can be represented by a network of all possible pairwise connections between brain regions. Such representation provides valuable insights about the structure of the brain-network. Specifically, in this work, we analyse the hub structure of extremely preterm (EP) born subjects compared to their full term (FT) born socioeconomically matched peers. Because of their role as focal points for the information transfer in the network, any dysfunction of hub regions may have disproportional effect on the integrity of the remaining network. In this analysis, we aim to investigate if the EP birth leads to altered hub brain structure.

Methods

Diffusion-weighted MRI volumes were corrected for thermal noise, Gibbs-ringing artefact, eddy current-induced distortion and subject movements artefacts. T1-weighted images were bias-corrected using N4ITK algorithm. Tissue parcellations of the corrected T1-weighted volumes were obtained using Geodesic Information Flow (GIF)². We applied Constrained Spherical Deconvolution³ to estimate the fiber orientation distribution in each voxel and performed Anatomically Constrained Tractography (ACT)⁴. From GIF template, we consider 121 cortical regions as nodes V of network G(V,E). The connection E between each node-region V is quantified after applying Spherical-Deconvolution Informed Filtering of Tracks⁵. Weighted structural networks were computed for 81 EP born adolescents (51 females, 30 males) along with 50 FT born adolescents (30 females, 20 males) as proposed by ⁶.
Different brain regions are identified as hubs depending on the specific centrality measure applied. To devise a more robust identification of these regions we propose a consensus-based approach by combining rankings across centrality measures as in⁷. We perform consensus classification for each average network of EP and FT born subjects. Hubs display a high level of connectivity strength S(j) and high betweenness centrality B(j) as well as high nodal efficiency E_nodal(j). Each node is assigned a score of one each time 1) the node is in the top 25% of nodes with stronger S(j), or 2) it is placed in the top 25% of nodes with highest B(j) or 3) highest E_nodal(j). With a maximum of 3 scores, nodes that scored 2 or higher are classified as hubs. Figure 1 shows the main steps of hubs identification.
Statistical analysis is performed to detect group differences regarding S(j), B(j) and E_nodal(j) in hubs. The analysis is performed using student t-test. Group, gender membership and brain volume may be confounding variables. There is difference in brain volume between females and males as well as between EP and FT born subjects as shown in figure 2. To account for this, we devise 4 nested GLM models, where we compare the full model (the model containing all regressors) to a set of restricted models. Specifically:

1. Y=β₀X₀+β₁X₁+β₂X₂+β₃X₃+ε
2. Y=β₀X₀+β₁X₁+β₂X₂+ε
3. Y=β₀X₀+β₁X₁+β₃X₃+ ε
4. Y=β₀X₀+β₁X₁+ε

where X₁, X₂ and X₃ are the regressors for the brain volume, the group membership and gender membership respectively, Y is the mean value of the graph metric under analysis. We compare the nested model fit to the graph metrics that showed a statistically significant difference between the two groups and investigate whether the effect on the graph metric is explained by volume difference, group or gender membership.

Results

The putative hubs identified for EP and FT born groups are shown in figure.3. There is an overall agreement between the two groups. The statistical analyses indicated that there is a significant decrease (p=2.6e^-8) in S(j) in EP (μ=8.9±1.1) with respect to FT (μ=10.3±1.3) born subjects. To a lesser extent, E_nodal(j) is reduced (p=1.7e^-4) in EP (μ=0.23±0.03) with respect to FT (μ=0.21±0.03). B(j) is comparable (p=0.3) between EP (μ=0.105±0.010) and FT (μ=0.104±0.009). The F-test analysis on nested GLM models revealed that brain volume and group membership (model.3) are the most accurate to describe the variance in S(j) and E_nodal(j) (top figure.4). The brain volumes describe 37% and 29.5% of the variance in S(j) and E_nodal(j) respectively. The group membership accounts for 12% of the variance in S(j) and 4.4% of the variance in E_nodal(j) (bottom figure.4).

Discussion/conclusions

We proposed a consensus-based method to identify hubs in EP and FT born network. The identified regions are comparable between the groups and consistent with previous studies⁷ on healthy subjects by identifying bilateral caudate, thalamus, putamen, precuneus, superior fontal gyrus and precuneus(Figure.3). The statistical analysis revealed that, although the EP and FT born subject have comparable hubs regions, the hubs in the EP group have a statistically significant reduced connectivity strength as well as nodal efficiency, while the betweenness centrality is comparable. The F-test on nested GLM models suggests that the observed difference is explained by brain volume difference and EP birth. The analysis of variance on model.3 suggests that although the brain volume influence is greater, the prematurity is significant to induce lower S(j) and E_nodal(j). The results suggest that while the core structure of the EP born brain is preserved, the connectivity and capacity of information flow is reduced and that is linked to reduced brain volume as well as EP birth.

Acknowledgements

We acknowledge the EPSRC-funded UCL Centre for Doc-toral Training in Medical Imaging (EP/L016478/1), the National Institute forHealth Research (NIHR) and the MRC (MR/J01107X/1)