Brainstem and Deep Gray Nuclei Modulate Brain Network Efficiency in Low-risk Term Newborns

Abstract

In low-risk term newborns, network analysis of source EEG has identified the hub regions that play an important role in information transfer in the network. However, the network efficiency changes after removing particular hub regions are not yet clear. The resting state 124-channel high-density electroencephalography (HD-EEG) was collected from low-risk term newborns within 72 hours after birth. The functional connectivity (spectral coherence) matrix was built using source EEG data in the delta band. The Small-World Propensity (SWP) metric was computed after removing each hub region individually and four hubs simultaneously. In the individual analysis, the SWP was significantly reduced after removing the right caudate, left thalamus, right thalamus, and brainstem among 18 hub regions from the functional brain network (P < 0.05). A significant reduction in the SWP was observed when the same four hubs were removed simultaneously. Our data provide evidence for reduced network efficiency after removing hubs in low-risk newborns. These findings highlight the critical role of specific hubs and demonstrate diminished network efficiency after their removal in low-risk term newborns.

I. Introduction

The human functional brain architecture is governed by two principles, namely, functional segregation and functional integration [1]. The segregation principle states that some functional processes involve local brain regions [2]. In contrast, the integration principle highlights that coordinated information flow among distributed brain regions are required even for simple behaviors. It has been proposed that optimal brain function must balance local specialization and global integration for flexible communication across the brain. The balance between segregation and integration in a network can be quantified by measuring the small-worldness of a network [3]. Previous studies have shown that newborn’s functional brain networks possess small world network architecture at birth,[4, 5] indicating that their brain networks are sufficiently organized to maintain a balance between network segregation and integration [6]. Hubs are the brain regions with higher connections with other brain regions in the network. Due to having more connections, hubs play a critical role in integrating information across different brain regions [7]. Functional brain network studies in newborns have shown the existence of rich club hub regions, i.e., regions that are more highly interconnected with each other than expected by chance [5]. Recent studies show that source imaging techniques can be applied to high-density electroencephalography (HD-EEG) recordings to study newborn’s functional brain networks at the bedside [5, 8]. The high temporal resolution of EEG in milliseconds allows the characterization of frequency-specific functional connectivity. Our previous HD-EEG study showed the small worldness architecture in newborns and identified the frequency-specific hubs [5]. However, the effects of hub removal on network efficiency are not yet clear in low-risk term newborns.

In this study, we evaluated the effect of removing hubs on small-world architecture in the newborn’s functional brain networks. The functional brain networks were constructed at the source level for the brain regions specified by the automated anatomical labeling (AAL) atlas for each individual. The hubs were identified based on the strength metric in our previous study [5]. The network efficiency was calculated using the Small World Propensity (SWP) metric. However, the role of specific hubs on overall network efficiency remains unknown. Since hubs play a critical role in information transfer in the network, we hypothesized that the removal of certain hubs would reduce the efficiency of functional brain networks.

II. MATERIALS AND METHODS

A. Subjects

We recruited low-risk term newborns from singleton pregnancies delivered at 37–41 weeks of gestation age (GA) at the Inova Women’s Hospital in Fairfax, Virginia. Inclusion criteria were 1-minute Apgar scores greater than or equal to 7, absence of perinatal or neonatal complications, and birth weight between 10th to 90th percentile for gestational age. Newborns from pregnancies complicated by conditions such as diabetes, substance abuse, and hypertension were excluded from the study. All newborns were studied by 124-channel EEG recordings within 72 hours of age. Parents provided written informed consent for the procedures involved in this study. This study was approved by the Institutional Review Boards of Inova Fairfax Hospital, Falls Church, Virginia, and Children’s National Hospital, Washington, DC.

B. EEG acquisition and preprocessing

The scalp EEG signals were recorded in the newborns within 72 hours after birth using the 124-channel HD-EEG system (Electrical Geodesics Inc., Eugene, OR, USA). A 250 Hz or 1000 Hz sampling rate was used to record the EEG data. The newborns were swaddled and placed in a bassinet during the entire EEG recording, and the recording duration varied between 21 and 60 minutes. In addition to the electrical brain activity, continuous electrocardiogram (ECG) and video recordings were made. EEG data were preprocessed using the MATLAB 2020b (Mathworks Inc., Natick, MA, USA) software [9]. EEG data recorded at 1000 Hz were down-sampled to 250 Hz, and a fourth-order Butterworth high-pass filter with a cutoff frequency of 0.1 Hz was applied. The volume conduction and EKG interference on EEG signals were identified and attenuated using the previously described method [10]. The EEG artifact detection and removal procedure is described in our previous study [11]. A certified pediatric neurophysiologist identified the sleep/wake states (active sleep, quiet sleep, awake, and artifact) in EEG [11].

C. Source-level analysis

The source analysis steps were implemented using the functions in the FieldTrip toolbox and custom in-house scripts [12] (Fig. 1). In this study, the realistic head model was built using the T2-weighted magnetic resonance images (MRI) from a full-term newborn at 41 weeks of GA. T2-weighted MRI data were acquired using the 3T MRI scanner (General Electric Discovery MR750) with an 8-channel high-resolution brain array receive only coil (T2-weighted 3D CUBE sequence; flip angle = 90°; voxel size = 0.625 × 1 × 0.625 mm³; echo time = 64.7 ms; repetition time = 2500 ms). The MRI was segmented into brain, skull, and scalp tissues using the Brain Extraction Tool (BET) toolbox in the FMRIB Software Library (FSL, v6.0) [13]. The conduction model was built with three layers (brain, skull, and scalp) using the Boundary Element Method (BEM) with conductivities of 1.79 S/m, 0.2 S/m, and 0.43 S/m for brain, skull, and scalp, respectively [14]. The source model was created with a rectangular grid of 1 mm spacing enclosed by the brain surface. The scalp EEG electrode positions were co-registered to the scalp surface, and the lead field matrix was computed as implemented in FieldTrip. The source positions according to the regions in the automated anatomical labeling (AAL) atlas were identified [15]. Source time series were extracted separately for every subject from the voxel corresponding to the medoid of each region for the duration of artifact-free 3-minute EEG using the linearly constrained minimum variance (LCMV) beamformer [16].

Figure. 1.

The T2-weighted magnetic resonance imaging (MRI) data were acquired from full-term born at 41 weeks of gestational age (GA). The realistic head model was built using the boundary element method (BEM). The automated anatomical labeling (AAL) atlas was used to define the brain regions, and the leadfield matrix was computed. The linear constrained minimum variance (LCMV) beamformer was used to extract the time series from different brain regions. The functional connectivity matrix was calculated using the spectral coherence in the delta band and small-world propensity (SWP) was calculated for the entire matrix and after the removal of hub regions.

D. Functional brain connectivity based on source-level time series

The functional brain connectivity was calculated in the source space by applying spectral coherence to the source-level EEG data. The spectral coherence was calculated using the Welch periodogram method. In this method, the data from each pair of regions was divided into three-second windows. In every window, power spectra and cross-spectrum were computed. The power spectrum of a signal was calculated as the square of the magnitude of the Fourier transform of the signal. The cross-spectrum between two signals was computed as the multiplication of the Fourier transform of the first signal and the complex conjugate of the Fourier transform of the second signal. The power spectra and cross-spectrum were averaged across all windows to obtain their estimates. The coherence between two brain regions was calculated as the ratio of the square of the magnitude of the estimate of the cross-spectrum to the product of the estimates of auto-spectra of the two signals. The coherence was calculated for the delta frequency band (0.5 to 4 Hz). The maximum coherence was calculated for every pair of regions in the delta band and used in further analysis. The functional connectivity matrix was constructed for each subject in which rows and columns represent the brain regions, and the value represents the functional connectivity (coherence) between the corresponding brain regions. SWP measures the small-world characteristic of the network. The SWP value ranges from 0 to 1. The higher value indicates that the given network has a stronger small-world architecture. Our previous study identified 18 brain regions (left olfactory, left caudate, left insula, left amygdala, left putamen, left pallidum, left parahippocampal gyrus, left hippocampus, left thalamus, right olfactory, right caudate, right amygdala, right putamen, right pallidum, right parahippocampal gyrus, right hippocampus, right thalamus, and brainstem) as hubs in the delta frequency band based on the strength metric [5]. A node was considered a hub if its strength was larger than the average network strength by more than one standard deviation. In this study, we calculated the SWP metric after removing each hub region as well as four tuple hubs simultaneously. A tuple is an ordered collection of elements.

E. Statistical analysis

The continuous and categorical data were summarized using mean (standard deviation) or counts (percentage). All the statistical analyses were implemented in Matrix Laboratory (MATLAB). The difference in SWP before and after the removal of the hub was assessed using the paired-sample t-test. P < 0.05 was considered significant.

III. Results

In this study, we analyzed the data from 112 newborns who met the inclusion criteria and were born between 37 and 41 weeks of GA with a mean (standard deviation) of 39.1 (0.7) weeks. Among 112 newborns, 57 (50.8%) were male, and 55 (49.1%) were female. 59 (52.6%) newborns were delivered vaginally, and 53 (47.3%) by cesarean section. The mean Apgar (standard deviation) scores measured at 1 and 5 minutes were 8.2 (0.4) and 8.9 (0.2), respectively. The normality of predictor variables was checked using the Shapiro-Wilk test. The SWP metric before and after the removal of each hub region was normally distributed (alpha = 0.05). The SWP metric analysis showed that newborns functional brain networks had a small world architecture immediately after birth (SWP > 0.6). The median SWP value was 0.8527, while the minimum and maximum values were 0.7395 and 0.8925 respectively. SWP was reduced after the removal of each hub region (Figure. 2), and the decrease was shown in Figure 3. However, the reduction in SWP was significant only after the removal of four hub regions, namely the right caudate, left thalamus, right thalamus, and brainstem (P < 0.05).

Figure. 2.

Each boxplot shows the distribution of small world propensity (SWP) values for term newborns after removing a specific hub region shown in the x-axis.

Figure. 3.

Each boxplot shows the distribution of the reduction in SWP after removing each hub shown in the x-axis. The asterisk (*) denotes p < 0.05.

We generated all possible 4-tuple hub permutations for the 18 hubs, which resulted in 3060 tuples. The SWP was calculated separately after removing each 4-tuple. The vector norm was calculated for the SWP values after the removal of every 4-tuple for 112 newborns. The norm obtained for every 4-tuple is shown in Figure. 4. This result showed that the lowest norm for a tuple (number: 3025) that consisted of hub combinations, right caudate, left thalamus, right thalamus, and brainstem.

Figure. 4.

The norm of the small-world propensities for all 112 newborns for every 4-tuple removal.

IV. Discussion and conclusions

Hub brain regions play an important role in the human brain as they are highly connected and facilitate the communication among different brain regions. Our previous study identified 18 hub brain regions in the delta band in healthy, low-risk term newborns [5]. In this study, we investigated the effect of hub removal on small-world architecture. Individually removing the brainstem, left thalamus, right thalamus, and right caudate hubs from the newborn functional network led to a decrease in small-worldness. Notably, eliminating these four hubs simultaneously resulted in the most significant reduction compared to the original network. This emphasizes the pivotal role of these hubs in sustaining efficient neural information flow in the newborn brain. Since these brain regions mature very early in development[17], the impairment in these regions might cause neurodevelopmental disorders or other cognitive, language, motor, or behavioral problems. The specific mechanisms through which these hubs contribute to network function remain unclear. However, ongoing causal analysis aims to unravel the distinctive features of information flows among these hubs, shedding light on their intricate roles in the network. There are some strengths and limitations to this study. A significant strength of this study includes using HD-EEG recordings from low-risk term newborns for the source analysis. In addition, we also implemented the graph theoretical approaches and characterized the functional brain network. Some limitations of our study include the use of the term MRI to construct the head model and the use of standard electrode locations for co-registration in source analysis. In newborn studies, an MRI from a newborn or template can be used to construct a reliable head model to address the lack of individual MRI [18]. The mean error between the standard and actual EEG electrode positions could be 0 to 15 mm. This error might reduce the signal-to-noise ratio of reconstructed source signals, which could lead to underestimating the results [19]. In future studies, we plan to acquire and use individual MRI and electrode locations in source analysis.