Altered Functional Brain Network Integration, Segregation and Modularity in Very Preterm Infants at Term Equivalent Age

Abstract

Objectives:

To determine the functional network organization of the brain in infants born very preterm at term equivalent age and to relate network alterations to known clinical risk factors for poor neurologic outcomes in prematurity.

Study design:

Resting state functional MRI data from 66 VPT infants (gestational age < 32wks and birth weight <1500g) and 66 healthy full term neonates, acquired as part of a prospective, cross-sectional study, were compared at term age using graph theory. Features of resting state networks, including integration, segregation and modularity, were derived from correlated hemodynamic activity arising from 93 cortical and subcortical regions of interest and compared between groups.

Results:

Despite preserved small world topology and modular organization, resting state networks of VPT infants at TEA were less segregated and less integrated than those of full term infants. Chronic respiratory illness (ie, bronchopulmonary dysplasia and the length of oxygen support) was associated with decreased global efficiency and increased path lengths (p-value < 0.05). In both cohorts, four functional modules with similar composition were observed (parietal/temporal, frontal, subcortical/limbic and occipital). The density of connections in three out of the four modules was decreased in the VPT network (P value < .01), however in the occipital/visual cortex module, connectivity was increased in VPT relative to controls (p-value < 0.0001).

Conclusions:

Early exposure to the ex utero environment is associated with altered resting state network functional organization in VPT infants at TEA likely reflecting disrupted brain maturational processes.

Resting state functional MRI (rs-fMRI) is an imaging technique that reveals, without any explicit external stimulation, consistent and reproducible patterns of interacting brain regions called resting state networks that correspond to known functionally relevant brain systems (i.e., visual, sensorimotor, and auditory, among others) [1,2]. In preterm infants, rs-fMRI has allowed researchers to begin to characterize the effects of early exposure to the extrauterine environment on the neurodevelopment of preterm infants [3–6]. Resting state fMRI estimates neuronal activity by measuring spontaneous, low-frequency fluctuations in blood oxygen level dependent signals (BOLD). Because rs-fMRI is non-invasive, rapidly acquired, requires no input from the participant, and investigates multiple brain systems simultaneously [2], it has emerged as a promising tool for evaluating pediatric brain function. In preterm infants, rs-fMRI has shown intact resting state network patterns but with reduced complexity and altered connectivity strengths in some neural networks [4,6–8]. Connectivity changes in preterm brains have also been shown to correlate with long-term neurologic outcomes [9,10].

Studies of preterm resting state networks using a mathematical approach known as graph theory [11–13] showed that the organization of preterm brains for specialized and distributed processing of information differed from healthy, full term controls. Building on these network studies, we compared the functional organization of resting state neural networks in a sample of unsedated preterm infants without or with mild brain injury and healthy full term newborns. We investigated the effects of clinical risk factors on network connectivity and hypothesized that the effect of prematurity would be associated with measurable alterations in strength and organization of the resting state networks.

Methods

VPT infants and healthy full term controls were enrolled in two prospective observational studies [14]. For the VPT cohort, inclusion criteria were: gestational age (GA) at birth <32 weeks and birth weight < 1500 grams. Exclusion criteria included: chromosomal anomalies, dysmorphic features, congenital brain malformations, central nervous system infection, and metabolic disorders. Only VPT infants without brain injury or with mild brain injury (e.g., low grade intraventricular hemorrhage, or mild white matter injury; Figure 1 [available at www.jpeds.com]) based on Kidokoro’s brain MRI scoring system were included [15]. All healthy full-term control infants were born without delivery complications and had normal brain MRI scans. Exclusion criteria for the control group were: maternal drug use and disease (e.g., gestational diabetes, psychiatric disorders), congenital infections, chromosomal abnormalities, and dysmorphic features.

Figure 1 Online. Preterm inclusion and exclusion criteria.

Demographic and clinical information was collected from medical record reviews. In the preterm cohort, several risk factors were chosen a priori based on previous literature for their potential adverse effect on brain development, these included: the presence of mild brain injury (yes/no), moderate to severe bronchopulmonary dysplasia (BPD), the length of supplemental oxygen requirement, postnatal steroid treatment, necrotizing enterocolitis diagnosis, need for cardiac vasopressor support, and need for patent ductus arteriosus surgical ligation [16–25].

Image acquisition and processing

All infants were scanned without sedation using the same 3T GE scanner. Acquisition parameters are detailed in Appendix 1 (available at www.jpeds.com). Resting state data were preprocessed using a previously validated pipeline that employed a series of temporal and spatial denoising strategies to minimize the influence of noise from head motion, scanner drift, and subject physiology (i.e., respiration), among others, on the measured BOLD signal [26]. BOLD signals were measured from 93 regions of interest (ROIs): 90 cortical and subcortical ROIs defined using an automated anatomical labelling atlas [27] mapped to a neonatal brain [28] and 3 infratentorial ROIs derived from DrawEM segmentation [29]. Correlation (Pearson r) between BOLD signals from all possible ROI pairs were then computed and used for network analyses. Appendix 1 includes additional information.

Network analyses

We compared network functional organization between the 2 groups using the following graph metrics: 1) clustering coefficient (Cc), 2) local efficiency (lE), 3) characteristic path length (L), 4) global efficiency (GE), 5) small world architecture, and 6) modular organization [30]. Clustering coefficient and local efficiency are segregation parameters. Segregation reflects the extent to which densely connected regions of the brain (modules or clusters) perform specialized functions [31]. Segregated networks have high Cc and high lE. Network integration refers to the ability of resting state networks to share information globally and is captured by the metrics characteristic path length and global efficiency. Integrated networks have short L and high GE. Small-world topology, quantified with the small world index (σ) that relates Cc and L, refers to the balance between network specialization and integration observed in complex networks like the brain [11,32–36]; if σ > 1, meaning Cc is high and L is low, then the network is considered small-world. Modularity is related to segregation and reflects the tendency of networks to subdivide into functionally meaningful clusters. A module refers to a group of nodes that are highly connected to each other but are sparsely connected to others outside of their group [37]. This relationship is reflected by the modularity index (Q; [38]). For a mathematical description of these measures, please refer to the work of Rubinov and Sporns [30].

In the preterm cohort, we studied the relationship between clinical risk factors and global properties of the resting state networks. We computed the averaged global network metrics (Cc, L, gE, lE) over the range of correlation thresholds and assessed the relationship between risk factors associated with prematurity and global resting state networks properties using ANCOVA.

Statistical Analyses

Statistical analyses were performed using Matlab 2017a and SAS 9.3. Between the preterm cohort and healthy full-term group, we compared segregation and integration metrics from individual network using permutation testing (100,000 iterations) on the residuals adjusted for postmenstrual age at MRI, sex, and motion. Modularity was evaluated on group-averaged connectomes at a density of 15% (sparsest density with the least probability of spurious connections). We then identified modular communities using 10,000 iterations of the Louvain algorithm [40]. Between the preterm cohort and healthy full-term group, we compared the modularity index and the intra-modular functional connectivity using two-sided two samples t-test, and the intra-modular density using Chi-square test. Finally, the relationship between risk factors associated with prematurity and global resting state networks properties was assessed using ANCOVA analysis controlling for sex, GA at birth, day of life at MRI, and motion.

Results

We studied 132 infants: 66 VPT infants at TEA and 66 healthy full-term control infants. Of the 66 VPT infants, 21 had mild brain injury: seven with low grade intraventricular hemorrhage (grade I-II), 13 with mild white matter injury, and one with small cerebellar punctate hemorrhage. Postmenstrual age at MRI was higher in the full-term group; this was used as a covariate in the rest of the analysis. Table 1 summarizes the demographic and clinical characteristics of the cohorts.

Table 1.

Clinical characteristics of preterm infants and full-term control infants.

	Preterm Infants, n=66	Healthy Full-term Infants, n=66	P-value
Perinatal Characteristics
Birth gestational age, wk, mean ± SD [range]	27.36 ± 2.68 [22.4 - 32]	39.48 ± 0.99 [37.3-41.3]	<0.0001
Birthweight, g, mean ± SD [range]	931 ± 305 [480-1500]	3373 ± 356 [2590-4011]	<0.0001
Small for gestational age1, n (%)	9 (14)	6 (9)	0.41
Male, n (%)	29 (44)	36 (55)	0.76
Native American; Hispanic; White Asian; Black; Multiethnic, n (%)	0; 15 (23); 9 (14) 0; 39 (59); 3 (4)	2 (3); 6 (9); 16 (24) 5 (8); 24 (51); 3 (5)	0.021
Vaginal delivery, n (%)	27 (40)	45 (68)	0.0017
Apgar score at 5 min, median [range]	8 [1-9]	9 [8-10]	<0.001
Maternal age, y, mean ± SD	27.91 ± 5.9	29.2 ± 7.18	0.26
MRI Characteristics
Postmenstrual age at MRI, wk, mean ± SD [range]	40.21 ± 1.56 [37.4-44.4]	41.1 ± 1.1 [38.4-44.3]	0.0002
Day of life at MRI, d, mean ± SD [range]	89.97 ± 21.4 [41-137]	11.21 ± 4.88 [4-20]	<0.0001
Head circumference at MRI, cm, mean ± SD [range]	33.26 ± 2 [27-37]	35.71 ± 1.09 [32-39]	<0.0001
Weight at MRI, g, mean ± SD [range]	2818 ± 681 [1870-5400]	3596 ± 422 [2475-4470]	<0.0001

¹SGA if BW less than the 10% percentile for sex and GA, based on the Fenton growth chart

Network segregation, integration, and small-world architecture

Resting state networks of VPT and full-term infants exhibited small-world topology (σ > 1). This balance between network integration and segregation suggested that even in the VPT infants, efficient global processing of information was preserved. However, segregation and integration measures per se were reduced in VPT infants compared with healthy newborns (Figure 2). Segregation parameters clustering coefficient (Cc) and local efficiency (lE) and integration metrics path length (L) and global efficiency (gE) were all significantly decreased in VPT infants.

Figure 2. Clustering coefficient, path length, global and local efficiency in VPT (blue) and full term infants (black).

Metrics were significantly different between groups at all tested thresholds using permutation testing (p-value < 0.05); tests controlled for sex, postmenstrual age at MRI, and motion.

Modular organization

Preterm and full-term infant resting state networks showed modular organization (Figure 3). The modularity index, however, was significantly lower in the VPT infants compared with the full-term infants suggesting reduced density of connections within modules in VPT. Although modules in VPT resting state networks tend to be less connected, their composition was highly similar to full-term newborns (Figure 3). In both cohorts, resting state networks were divided into four modules (Table 2 and Figure 3), namely: (1) Module 1 - primary somatosensory, motor, and auditory cortices and association areas within the parietal/temporal cortices; (2) Module 2 - regions in the frontal cortex; (3) Module 3 - subcortical, limbic, and paralimbic brain regions; and (4) Module 4 - primary and association visual cortices/occipital cortex.

Figure 3. Modular brain organization in preterm and full-term infants.

Spheres are nodes, lines are edges; color represent module membership. Note reduced number of connection between nodes in modules 1-3 (1 – red, 2 – green, and 3 – yellow) and increased density in module 4 (blue) in VPT compared with full-term infants.

Table 2.

Modular brain organization differences between preterm and full-term infants

		Preterm Infants, n=66	Full-term Control Infants, n=66	P-value
Modularity (Q)		0.49 ± 0.005	0.51 ± 0.002	<.001
Module 1 -Parietal/Temporal	Node	26	24
	Density	42.2%	52.5%	0.01
	Connectivity strength	0.49 ± 0.2	0.54 ± 0.22	0.0086
Module 2 -Frontal	Node	22	22
	Density	36.4%	49.4%	0.0048
	Connectivity strength	0.55 ± 0.25	0.55 ± 0.25	0.68
Module 3 -Subcortical/Limbic	Node	22	21
	Density	69.7%	80.5%	0.0092
	Connectivity strength	0.51 ± 0.21	0.56 ± 0.2	0.0023
Module 4 - Occipital	Node	23	26
	Density	41.5%	23.7%	<0.0001
	Connectivity strength	0.51 ± 0.25	0.535 ± 0.22	0.239
Between Module Connections	Density	4.8%	4.2 %	0.44
	Connectivity strength	0.39 ± 0.11	0.42 ± 0.13	0.0256

When we compared the number of connections within each module at a density threshold of 15 % (i.e., same total number of connections in VPT and full-term group-averaged functional connectome), density and connectivity strength within each module were perturbed in VPT infants (Table 2). VPT infants had significant lower density in all modules except Module 4 (i.e., occipital areas) where connection density was higher. Connectivity strength was significantly reduced in VPT infants in the parieto-temporal and subcortical/limbic modules (Table 2).

Clinical risk factors

In the preterm cohort, we showed that both moderate to severe BPD and prolonged need for supplemental oxygen were associated with increased path length (β>0) and decreased global efficiency (β<0) (Table 3), suggesting reduced integrative abilities in VPT infants. In addition, longer period of oxygen support was also associated with decreased local efficiency (β<0) suggesting decreased specialization in VPT resting state networks (Table 3).

Table 3.

Relationship between clinical risk factors and global network properties in the preterm cohort.

Risk Factors Studied	Preterm infants (n=66)	Global network metric	Model p-value	Association with global network metrics
Mild brain injury	21 (32%)	cc	0.002	β=0.001; p=0.94
		L	0.002	β=0.01; p=0.85
		gE	0.011	β=−0.001; p=0.86
		lE	0.0008	β=0.002; p=0.85
Moderate to severe bronchopulmonary dysplasia*	24 (37%)	cc	0.0005	β=−0.02 ;p=0.15
		L	<0.0001	β=+0.17; p=0.01*
		gE	<0.0001	β=−0.03; p= 0.02*
		lE	0.0001	β=−0.02; p=0.05
Length of oxygen support (days)	67 ± 41 [0-172]	cc	0.0006	β=−0.0004 ; p=0.36
		L	0.0001	L: β=+0.0027; p=0.03*
		gE	0.0001	gE: β=−0.004; p=0.01*
		lE	<0.0001	lE: β =−0.0004; p=0.026*
Postnatal steroid treatment	21 (32%)	cc	0.0013	β=−0.001; p=0.92
		L	0.0012	β=−0.038; p=0.58
		gE	0.0006	β=2.2×10”6; p=0.99
		lE	0.0006	β=0.005; p=0.69
Necrotizing enterocolitis	19 (29%)	cc	0.0019	β=−0.004; p=0.78
		L	0.0017	β=−0.05; p=0.42
		gE	0.0011	β=0.002; p=0.84
		lE	0.0007	β=0.006; p=0.63
Cardiac vasopressor treatment	19 (30%)	cc	0.0006	β=0.021; p=0.18
		L	0.0014	β=−0.02; p=0.76
		gE	0.0006	β=−9.9 ×10−6; p=0.99
		lE	0.0003	β=0.016; p=0.18
Surgical ligation of patent ductus arteriosus	12 (18%)	cc	0.002	β=0.0018; p=0.92
		L	0.0023	β=3.6χ10−4 ; p=0.99
		gE	0.0011	β=−0.002 ; p=0.86
		lE	0.0008	β=7χ10−4 ; p=0.96

^*Severity of BPD in VPT infants: treatment with oxygen for at least 28 days and need for more than 21% oxygen support (moderate: less than 30% oxygen; severe: more than 30% oxygen) at discharge or at 36 weeks postmenstrual age (PMA) whichever comes first [30]

Discussion

Our study revealed aberrant functional network organization in a large sample of very preterm infants scanned at term equivalent age. Compared with healthy term newborn infants, VPT infants with no or mild brain injury had reduced resting state network segregation and integration measures and altered modular connectivity. Among the risk factors analyzed, only the clinical indicators of severity of respiratory illness (i.e., moderate to severe BPD and prolonged supplemental oxygen support) were associated with network alterations in VPT infants. These findings underscore the potential detrimental link between respiratory illness severity and functional brain development in VPT infants.

We showed small world organization in VPT and full-term infants consistent with previous studies using rs-fMRI [3,11–13,39]. This organization, reflective of efficient, specialized and integrated neuronal communication [41], likely emerges before the 31^st week of gestation and continues to mature as gestation progresses [13]. Although small world architecture was preserved, resting state networks of VPT infants were less segregated (i.e., lower clustering coefficients and reduced local efficiency) and integrated (i.e., longer path lengths and reduced global efficiency). Reduced clustering was previously reported by Scheinost et al in a smaller cohort (n = 12 preterm infants) [12]; they, however, did not observe differences in path lengths between premature and term infants, possibly due to the smaller sample size. Our results suggest that preterm birth somehow reconfigures connections among brain regions. In preterm infants, longitudinal studies have reported age-dependent increases in functional connectivity strength [4,6,7], network complexity and magnitude [7], integration [11,13], and modularity [11,13] during the third trimester suggesting a vulnerability window. Thus, it is not surprising that in some networks at TEA, preterm infants exhibit reductions in connectivity strength [7], decreased interhemispheric connectivity [5,6], altered lateralization of language areas [42], altered thalamocortical connectivity [8,43], and impaired basal ganglia – frontal cortex connectivity [44]; and it provides additional support to the likely effects of preterm birth on the organization of emerging functional networks. Our understanding of how functional connectivity in VPT infants evolves over time remains incomplete, but available follow-up studies already suggest that functional connectivity remains compromised in adults born very preterm [45–49].

Interestingly, only the clinical risk factors linked to respiratory disturbances showed a negative association with resting state network organization: moderate to severe BPD and length of oxygen support were linked to reduced global and/or local efficiency of the resting state networks. Respiratory issues such as BPD and prolonged need for oxygen support have been associated with impaired brain development and adverse cognitive functioning [19,20,50,51]. Hypoxic episodes resulting from respiratory disturbances are thought to underlie the pathophysiological mechanisms responsible for these cerebral alterations [52–56]. Hypoxic-ischemic events are a major cause of prematurity related brain injury [57–59], and disturbances in cerebral oxygenation have also been associated with subtler delays in structural brain maturation in high-risk neonatal population [60,61]. Taken together, these findings emphasize the crucial role of adequate oxygen supply in the early establishment of the neural circuitry.

Mild brain injury was not related to alterations in global properties of the resting state networks in our study. Because white matter injury has been associated with altered functional connectivity in preterm infants [5,43,62], we hypothesize that the degree of brain injury severity plays a crucial role thus the potential functional connectivity disturbances following mild brain injury still remains to be elucidated.

Small-world topology of networks relate closely to its modular organization [32,63]. Modular systems with functionally specialized subsets of brain regions that are sparsely connected to other modules, such as what we observed in resting state networks of VPTs at TEA and full-term newborns, tend to be small world (i.e., high segregation balanced with high integration; [64]). Modular organization of the developing brain has been described in third trimester fetuses [65], preterm infants [11,13], and healthy neonates [11,39]. We also reported modular organization of resting state networks in VPT infants at TEA, but with reduced modularity compared with full-term infants. Four functional modules with similar composition to full-term infants were identified. Consistent with previous work, modules in infants were mostly anatomically constrained [3,39,66]; adjacent areas/regions belonging to the same lobe tend to belong to the same module. In contrast, adult modules are composed of neighboring (i.e. subcortical module), as well as spatially distant, but functionally associated regions such as the default-mode and attention modules (fronto-parietal) [37,64,67,68]. Cognitive networks are immature at birth [69] and the default-mode network only starts to exhibit adult-like properties during the first years of life [69–71], thus it is not surprising that the modular organization at term age does not include cognitive module yet. Due to the postnatal development of cognitive processes and establishment of long-range connections during early childhood, regional organization of the resting state networks evolve from a local organization during childhood to a distributed organization in adulthood [63,72,73].

Preterm networks, while still modular, were less segregated (i.e. lower modularity index) compared with full term resting state networks. Previous work has shown that networks become more segregated with increasing age in the first two years of life [74]. This reduced specialization in VPT infants may suggest less mature brain networks compared with their full term counterparts. Examining the density of connections within each module, however, suggests a more nuanced picture where, depending on the neural network, both delayed and accelerated maturation seem to occur. In preterm infants, connection density was reduced in three out of the four modules – parieto-temporal, frontal, and subcortical/limbic. Interestingly, VPT infants had an increased number of connections in the occipital/visual module. This finding converges with the results of Padilla et al that demonstrated increased volumes of areas involved in visual processing in extremely preterm infants at TEA compared with healthy controls.[75] This increased connection density in the visual cortex is likely due to experience-dependent processes. Early extrauterine exposure is likely associated with increased visual stimulation during a critical period of development for the visual system [76–78]. Synaptic density of the visual cortex dramatically increases from mid-gestation to the first months of postnatal life [76,77,79,80]. Increased visual inputs during this critical period might have lasting consequences on visual processing [78,81,82]. Additional studies are needed to elucidate the relationship between early visual experiences, functional organization of the visual cortex, and impairments in cortical visual processing. It would be intriguing to explore if other sensory areas of the brain (i.e., auditory cortex) that are prematurely exposed to ex utero stimulation would also present with increased connection density. The modular analysis performed was not suited to elucidate the possible link between prematurity and the functional development of other sensory areas sensitive to exogenous stimulation during the third trimester.

Our study limitations deserve mention. First, our understanding of the biological mechanisms underlying BOLD responses remains incomplete [83], even more so in newborn infants than in adults [84]. As such, interpreting and comparing findings must be done carefully. Having said that, there is a high degree of consistency between our findings and other published reports in newborn infants. In addition, observed emergence of resting state networks coincide with the timing of established developmental processes (i.e. development of somatosensory and motor before higher cognitive/association networks). Second, recent work has shown that the disruption of resting state networks in preterm infants may begin in utero [85] with an additive negative effect of fetal growth restriction [86]. Thus, prenatal, in addition to postnatal risk factors associated with preterm birth may also impact the early development of the resting state networks. Next, similar to other studies [34], only positive correlations were included in our network analyses because most existing graph methods were optimized for networks with positive correlations [87–89]. To evaluate negatively correlated brain activity, newer approaches accounting for anti-correlated activity need to be used [30,89,90]. Last, newborn imaging poses numerous technical challenges one of which is motion correction. To minimize the influence of motion, we removed (‘scrubbed’) high motion volumes, and used motion parameters and their derivatives as regressors in our statistical analyses. Correcting geometric distortion in EPI images is another technical issue. We did not perform distortion correction in our study, however, rigorous visual evaluation of our EPI images show minimal shape distortion. We have now included field map acquisition in our newborn MRI sequences so we can systematically evaluate distortion effects on connectivity measures.

To summarize, we report reduced segregation and integration in resting state networks of VPT infants at TEA and their association with prematurity related respiratory illness severity. Notably, these network alterations were present despite intact small-world topology and modularity. Longitudinal studies covering the prenatal (i.e., healthy fetal controls) and early school age periods would be necessary to provide additional insights into the role of premature birth on the development of functional brain networks, including possibly identifying the onset of these alterations, and to understand its long-term impact on neurobehavior.