https://orcid.org/0000-0002-1932-9557
Abstract
Obstructive sleep apnea (OSA) is characterized by recurrent airway collapse during sleep, resulting in intermittent hypoxia and sleep fragmentation that may contribute to alternations in brain structure and function. We hypothesized that OSA in children reorganizes and alters cortical structure, which can cause changes in cortical thickness correlation between brain regions across subjects.
We constructed cortical structure networks based on cortical thickness measurements from 41 controls (age 15.54 ± 1.66 years, male 19) and 50 children with OSA (age 15.32 ± 1.65 years, male 29). The global (clustering coefficient [CC], path length, and small-worldness) and regional (nodal betweenness centrality, NBC) network properties and hub region distributions were examined between groups.
We found increased CCs in OSA compared to controls across a wide range of network densities (p-value < .05) and lower NBC area under the curve in left caudal anterior cingulate, left caudal middle frontal, left fusiform, left transverse temporal, right pars opercularis, and right precentral gyri (p-value < .05). In addition, while most of the hub regions were the same between groups, the OSA group had fewer hub regions and a different hub distribution compared to controls.
Our findings suggest that children with OSA exhibit altered global and regional network characteristics compared to healthy controls. Our approach to the investigation of cortical structure in children with OSA could prove useful in understanding the etiology of OSA-related brain functional disorders.
To characterize associations of cortical thickness between brain regions, we constructed cortical structure networks based on cortical thickness from 41 controls and 50 children with obstructive sleep apnea (OSA). The global and regional network properties and the hub region distributions were examined between groups to assess the effect of OSA on cortical structure. We observed that the OSA group had increased clustering coefficients and lower nodal betweenness centrality in multiple brain regions. In addition, children with OSA had fewer hub regions and different hub distributions compared to controls. This study provides new evidence of disrupted cortical structure network that may further our understanding of structural brain alterations in children with OSA.
Introduction
Obstructive sleep apnea (OSA) is characterized by recurrent events of partial or complete airway collapse during sleep and occurs in both adults and children [1]. OSA causes sleep fragmentation and excessive daytime sleepiness with harmful systemic and brain responses, including blood-brain barrier dysfunctions, brain edema, and cerebral hypoxemia due to a mismatch between O2 demand and supply in the brain [2–4]. In children, these brain insults may lead to brain-related functional impairments, including emotional lability, disruptive behaviors, anxiety, hyperactivity, and cognitive dysfunction [5].
While OSA may cause accelerated brain aging in adults [6], OSA in children may stunt brain development [7]. Since the process of synaptic remodeling occurs almost exclusively during sleep, the effect of OSA on brain structures and neurodevelopment in children is likely to be more profound [8, 9]. Magnetic resonance imaging (MRI) can provide objective insights into the pathophysiological consequences of OSA in children by exploring abnormalities of brain structure that normally facilitate development and execution of a range of cognitive functions and behaviors. Previous studies have found that children with OSA show significant brain tissue change [7, 10–13]. Reduced gray matter volumes in prefrontal and temporal regions in OSA patients have been associated with attention deficits and limited visual-motor processing [11] and decrease in dentate gyrus mean diffusivity has been associated with poor verbal learning and memory [10]. In particular, one group found that children with OSA suffered acute and/or chronic brain tissue damage of multiple brain areas associated with autonomic, cognitive, and neuropsychologic control [12, 13] as well as correlation between disease severity and extent of tissue change [12]. Previous findings suggest that OSA could contribute to alterations in gray matter characteristics, including changes to the architecture of glia, vasculature disruptions, and disordered neuron development. In addition, in the central nervous system, intermittent hypoxia resulting from OSA is related to oxygen neuronal apoptosis, necrosis, and reactive oxygen species (ROS) production [14]. The increase of ROS production and neuronal apoptosis is associated with microglial mitochondria dysfunction (i.e. O2−/H2O2 formation) in cortical neurons [14, 15]. This, in turn, might cause excessive or insufficient synaptic pruning. Synaptic pruning is a natural process in the brain and progresses throughout the first 20 years of life to preserve strong brain connections and eliminate unused connections for cost-effective function [16, 17]. However, it is unclear how OSA in children impacts brain connections across the whole brain, which facilitates heterogeneous regional functions. Moreover, previous studies which reported that OSA caused alterations of functional or structural connections highlighted the need to address OSA as a network disorder [18–20]. Therefore, investigating the effect of OSA on brain networks in children is important for understanding potential associations of brain function disorders as well as the possibility of adverse effects on the neurodevelopment of these children.
The human brain has been modeled as a network of brain regions connected by anatomical tracts and anatomical or functional associations [21, 22]. Network analysis based on a graph theoretical approach has become increasingly popular in the neuroimaging field and has been in widespread use as a method to describe the properties of network organization [23]. The quantitative measurement of network organization can inform investigations of abnormal brain connections related to neurological and psychiatric disorders. Previous studies using brain network analysis in children reported a disruption of brain networks in neurological and psychiatric disorders including autism spectrum disorder [24], attention-deficit/hyperactivity disorder [25], and language disorders [26]. Therefore, a brain network analysis has the potential to be an effective tool for exploring the effects of OSA on children’s brain structure.
In the present study, we constructed a cortical structure network based on the association between brain regions of cortical thickness variations across individuals and applied a network analysis to investigate the differences in global and regional network properties as well as hub region distribution between children with OSA prior to initiation of treatment and matched controls. The primary hypothesis was that children with OSA would exhibit altered inter-region associations in cortical thickness compared to healthy, matched controls.
Methods
Subjects
Inclusion criteria
(1) Obese children (body mass index [BMI] > 95th percentile for age and sex) 12–18 years old undergoing evaluation of OSA were recruited from the sleep, adolescent, endocrine, and general pediatric obesity clinics at the Children’s Hospital at Montefiore Medical Center (CHAM) in Bronx, NY, (2) normal motor, cognitive, behavioral, and linguistic milestones during childhood, (3) intact adenoid and tonsils.
Exclusion criteria
(1) Treatment for OSA (e.g.: continuous positive airway pressure [CPAP]), (2) pre-existing neurological impairments including: seizures, developmental delay, brain tumor, degenerative brain diseases, (3) craniofacial and genetic syndromes, (4) history of head trauma, (5) cyanotic heart disease, (6) chronic lung disease with hypoxemia, and (7) MRI contraindications.
Subjects underwent T1-weighted structural brain MRI (Philips Achieva 3T scanner, 3D MPRAGE sequence, echo time (TE)/repetition time (TR)/inversion time (TI) = 3.8/8.3/900 ms, with 1 mm isotropic resolution, 338 s acquisition) during wakefulness at the Gruss Magnetic Resonance Research Center at the Albert Einstein College of Medicine. Following data acquisition, all MRIs were reviewed by an attending neuroradiologist. Determination of OSA or control was made by overnight polysomnography. The study was approved by the Institutional Review Board at the Albert Einstein College of Medicine. Informed consent was obtained from the participant and his/her parents of minors prior to enrollment in the study.
Overnight polysomnography
Overnight polysomnography (Natus/Xltek, Oakville, ON, Canada) was performed at the Sleep Disorders Center at CHAM. Sleep staging and scoring of arousals were performed per standard criteria by a registered polysomnographic technologist [27]. Obstructive apnea was defined as at least a 90% drop in peak signal excursion of the oronasal thermal sensor in the presence of continued respiratory effort lasting longer than two respiratory cycles. Hypopnea was defined as a 30% decrease in amplitude of the nasal pressure transducer associated with either a decrease in basal SpO2 by ≥ 3% or arousal [28]. Children were classified as OSA if they had an obstructive apnea index > 1 events/h, an obstructive apnea-hypopnea index (AHI) > 5 events/h, or both. We further stratified OSA subjects based on AHI as mild (AHI 5–14 events/h), moderate (AHI 15–29 events/h), or severe (AHI > 30 events/h).
T1-weighted image preprocessing
T1-weighted images were preprocessed using FreeSurfer v7.1.1 (https://surfer.nmr.mgh.harvard.edu/). The preprocessing pipeline was composed of volume- and surface-based analyses. The volume-based analysis applied a neuroanatomical label to each voxel while the surface-based analysis reconstructed the white matter and pial surfaces to extract local measures of cortical thickness [29, 30]. Briefly, the volume-based analysis involves the following steps: (1) brain extraction, and (2) normalization of brain tissue types. The surface-based analysis additionally involves (3) segmentation of white matter and gray matter, and (4) tessellation of white matter and pial surfaces. This analysis resulted in a triangular mesh consisting of approximately 160 000 vertices for each hemisphere. After the volume- and surface-based analyses, we performed inflation and a spherical transformation for the identification of gyral and sulcal regions and registration to standard FreeSurfer space (i.e. “fsaverage”, a template brain surface reconstruction, based on the combination of MRI scans from 40 healthy adults) [31].
Cortical thickness was obtained by analyzing the reconstructed surface of white matter and pial surfaces and calculating the closest distance between pairs of vertices on two different cortical surfaces [32]. This method of measuring cortical thickness generates a more accurate cortical morphology by using the intensity and continuity information from the whole MR volume throughout the segmentation and deformation processes [32]. More details on the procedures described in this section can be found in previous publications [29, 30].
Construction of structural cortical networks
Ideally, the human brain can be modeled as a network of neurons connected by synapses. However, it is difficult to model this network given an estimated 100 billion neurons with each having an average of 7000 synaptic connections [33]. Large-scale activation and coactivation of neuronal populations are necessary to fully capture brain functions. Thus, modeling the integration of these populations is considered the most reliable approach [34, 35]. In the present study, we constructed structural cortical networks based on cortical thickness measurements from MRI [21]. A regional cortical thickness for each region of interest (ROI) was defined as the average thickness of all vertices defined as belonging to that region parcellated using Desikan-Killiany atlas [36], which parcellates the cortical surface into 34 regions per hemisphere.
The present study utilized a sparse inverse covariance estimation (SICE) [37] approach to quantify the presence of structural interconnections in the framework of a network approach. Briefly, cortical thickness obtained from each region of an individual subject constructed a 2-D measurement matrix (i.e. Subject × ROI). Before the network construction, a linear regression was performed at every cortical region to remove the effects of age, sex, BMI, and mean cortical thickness across the entire cortex. The resulting residuals were used to substitute for the raw cortical thickness values [38]. The SICE algorithm was then used to find a robust estimate of the binary structure of the 2D cortical connectivity matrix (i.e. 0, 1 for the absence or presence of cortical correlation between regions, i.e. a 2D, ROI × ROI matrix) by imposing a sparsity constraint on the maximum likelihood estimate of the inverse covariance matrix [37]. The sparsity is controlled by the regularization parameter, λ. Since there is currently no gold standard to select a single network density defined as the ratio of the number of existing connections divided by the maximum possible number of connections [38], we generated binary connectivity matrices over a wide range of network densities (10%–35% in 1% increments) which met the small-worldness (SW) criterion. The network properties are greatly influenced by the total number of ROIs and brain connections and the difference in the number of brain connections between groups can cause undesirable differences in between-group network properties [23]. To reduce these differences, we used the stability approach to regularization selection which allows generatation of binary connectivity matrices with the same network density across groups [39]. These binary matrices capture the underlying topological organization of the human structural cortical networks.
Global cortical structure network characteristics
Clustering coefficient (CC), path length (PL), and SW were used to characterize the global cortical structure network, and test for differences in this structure between OSA patients and controls. Brain functions are supported not only by a single neuron or brain region, but also by clusters of neurons, or a group of interactions between brain regions [40]. In brain networks, a measure of specialized function and segregated processing is based on the concept of separate clusters or connected triangles [23, 41]. The CC is the average of the CC overall ROIs in a network, where the CCi of an ROI i is defined as the number of existing connections among the connected ROIs divided by all their possible connections. Namely, the CC quantifies the abundance of connected triangles in a network. In the brain network, a measure of the ability to combine pieces of specialized information from multiple brain regions is based on the concept of shortest paths [23, 41]. The PL is the minimum number of connections that link any two nodes of the network, averaged over all nodes. The SW measurement captures the balance between functional integration and segregation within the network [23]. A small world network meets the following criteria: normalized CC (CCreal/CCrand) >> 1, normalized PL (PLreal/PLrand) ≈ 1, and SW ((CCreal/CCrand)/(PLreal/PLrand)) > 1, where the CCrand and PLrand are the mean CC and PL, averaged over 100 matched random networks that preserve the number of ROIs, connections, and degree distribution of the actual network [42, 43].
Regional cortical structure network characteristics
Nodal betweenness centrality (NBC) was used to characterize the regional characteristics of cortical structure network, and to test for differences in this structure between OSA and controls. The NBCi of an ROI i is the number of shortest paths between any two ROIs that pass through ROI i [44]. NBC captures the relative importance of a brain region and the influence of a brain region over information flow between other regions [45]. In addition, a brain region was considered a hub if the NBC was greater than one SD above the average NBC of the network (NBCi > mean + SD). Hubs are highly connected brain regions, which facilitate the integration of information from different specialized or disparate subnetworks [46].
Statistical analysis
Demographic and clinical characteristics including sleep variables were compared between the groups with either two-sample t-tests or chi-square tests.
To determine statistical significance of group differences in network parameters, a nonparametric permutation test method was applied [47]. First, CC, PL, SW, and NBC at a given network density were computed separately for the control and OSA groups. To test the null hypothesis that the observed group differences could occur by chance, the set of regional cortical thickness measures for each subject was randomly reallocated (i.e. permuted) to control or OSA groups and we then reconstructed the cortical structure network for each permuted group. Then, the network parameters for each reallocated group were calculated and their differences between the permuted groups were calculated. This randomization procedure was repeated 5000 times and the 2.5 and 97.5 percentile points of distribution for each measure were used as the critical values for a two-tailed test of the null hypothesis. This procedure was repeated at every network density. Multiple comparisons for both global and regional properties were corrected by false discovery rate (FDR) correction [48]. We also assessed between-group differences in the areas under the curve (AUCs) for each measure. AUC was computed using trapezoidal rule formula for the values evaluated at each network density. We compared the three global measures between groups and the one regional measure for 68 ROIs. For each measure, the global results were FDR-corrected among 26 densities and the regional results were FDR-corrected among the 68 regions. In addition, AUCs were FDR-corrected among the 68 regions, and among the three measures for global results.
Results
Participants
The study enrolled 96 subjects. Among these participants, MRI’s for two participants did not have full brain coverage, and three participants were not able to complete the overnight sleep study after MRI scanning. Of the 91 participants (43 female and 48 male) included in final structural connectivity analysis, four were noted to have incidental brain findings upon review by an attending neuroradiologist. One control had a midline posterior fossa arachnoid cyst, one control had a 5 mm hyperintense lesion in the pituitary gland, one OSA subject had a small lacunar infarct in thalamus, and one OSA subject had a Chiari type I malformation of 5–7 mm below the foramen magnum. The data of these subjects were included in analyses since they did not involve cortical ROIs and all subjects were asymptomatic.
Fifty children had a diagnosis of OSA (mean age of 15.32 ± 1.65 years, mean BMI Z-score of 2.35 ± 0.35, mean height of 165.94 ± 9.71 cm, mean AHI of 15.19 ± 8.90 events/h, 29 (58%) male), while 41 were control participants (mean age of 15.54 ± 1.66 years, mean BMI Z-score of 1.97 ± 0.70, mean height of 163.34 ± 7.83 cm, mean AHI of 1.85 ± 1.38 events/h, 19 (46%) male). The control group did not differ significantly from the OSA group with respect to age, sex, or height (Table 1). In addition, the mean cortical thickness across the entire cortex of control group (2.52 ± 0.93 mm) did not differ from the OSA group (2.54 ± 0.1 mm, p-value = .20). There was a significant difference in BMI and BMI Z-score between the groups (p-value < .01). As mentioned previously, prior to network construction, we used linear regression to remove any effects of age, sex, BMI, and mean cortical thickness.
Demographic information of participants
| Controls (n = 41) | OSA (n = 50) | P | |
|---|---|---|---|
| Age (year) | 15.54 ± 1.66 | 15.32 ± 1.65 | .52 |
| Age range (year) | 12.39–18.96 | 11.88–18.33 | |
| Height (cm) | 163.34 ± 7.83 | 165.94 ± 9.71 | .17 |
| Weight (kg) | 87.49 ± 17.91 | 100.08 ± 24.50 | .01 |
| BMI (kg/cm2) | 32.64 ± 5.58 | 37.12 ± 7.37 | <.01 |
| BMI Z-score | 1.97 ± 0.70 | 2.35 ± 0.35 | <.01 |
| male, n (%) | 19 (46.4) | 29 (58.0) | .57 |
| Controls (n = 41) | OSA (n = 50) | P | |
|---|---|---|---|
| Age (year) | 15.54 ± 1.66 | 15.32 ± 1.65 | .52 |
| Age range (year) | 12.39–18.96 | 11.88–18.33 | |
| Height (cm) | 163.34 ± 7.83 | 165.94 ± 9.71 | .17 |
| Weight (kg) | 87.49 ± 17.91 | 100.08 ± 24.50 | .01 |
| BMI (kg/cm2) | 32.64 ± 5.58 | 37.12 ± 7.37 | <.01 |
| BMI Z-score | 1.97 ± 0.70 | 2.35 ± 0.35 | <.01 |
| male, n (%) | 19 (46.4) | 29 (58.0) | .57 |
Values are presented as n (%) or mean ± SD.
OSA, obstructive sleep apnea; BMI, body mass index.
Demographic information of participants
| Controls (n = 41) | OSA (n = 50) | P | |
|---|---|---|---|
| Age (year) | 15.54 ± 1.66 | 15.32 ± 1.65 | .52 |
| Age range (year) | 12.39–18.96 | 11.88–18.33 | |
| Height (cm) | 163.34 ± 7.83 | 165.94 ± 9.71 | .17 |
| Weight (kg) | 87.49 ± 17.91 | 100.08 ± 24.50 | .01 |
| BMI (kg/cm2) | 32.64 ± 5.58 | 37.12 ± 7.37 | <.01 |
| BMI Z-score | 1.97 ± 0.70 | 2.35 ± 0.35 | <.01 |
| male, n (%) | 19 (46.4) | 29 (58.0) | .57 |
| Controls (n = 41) | OSA (n = 50) | P | |
|---|---|---|---|
| Age (year) | 15.54 ± 1.66 | 15.32 ± 1.65 | .52 |
| Age range (year) | 12.39–18.96 | 11.88–18.33 | |
| Height (cm) | 163.34 ± 7.83 | 165.94 ± 9.71 | .17 |
| Weight (kg) | 87.49 ± 17.91 | 100.08 ± 24.50 | .01 |
| BMI (kg/cm2) | 32.64 ± 5.58 | 37.12 ± 7.37 | <.01 |
| BMI Z-score | 1.97 ± 0.70 | 2.35 ± 0.35 | <.01 |
| male, n (%) | 19 (46.4) | 29 (58.0) | .57 |
Values are presented as n (%) or mean ± SD.
OSA, obstructive sleep apnea; BMI, body mass index.
Polysomnography
Polysomnographic parameters of the two groups are shown in Table 2. As expected, there were significant differences in the control versus OSA groups with respect to the apnea index (0.11 ± 0.35 vs. 1.83 ± 2.79 events/h, p-value < 0.01), AHI (1.85 ± 1.38 vs. 15.19 ± 8.90 events/h, p-value < 0.01), arousal index (8.88 ± 5.11 vs. 20.97 ± 11.85 events/h, p-value < 0.01), and SpO2 nadir (92.78 ± 2.40 vs. 86.18 ± 7.33 %, p-value < 0.01).
Sleep variables of participants
| Controls (n = 41) | OSA (n = 50) | P | |
|---|---|---|---|
| Total sleep time (min) | 348.44 ± 65.23 | 347.19 ± 65.68 | .93 |
| Sleep efficiency (%) | 84.23 ± 11.93 | 80.05 ± 13.04 | .12 |
| AI (events/h) | 0.11 ± 0.35 | 1.83 ± 2.79 | <.01 |
| AHI (events/h) | 1.85 ± 1.38 | 15.19 ± 8.90 | <.01 |
| Mild (5 ≤ AHI < 15) | 30 (60.0) | ||
| Moderate (15 ≤ AHI < 30) | 16 (32.0) | ||
| Severe (AHI ≥ 30) | 4 (8.0) | ||
| Central sleep apnea (events/h) | 0.30 ± 0.39 | 0.86 ± 1.50 | .04 |
| Baseline SpO2 (%) | 99.05 ± 1.18 | 99.24 ± 1.24 | .46 |
| SpO2 nadir (%) | 92.78 ± 2.40 | 86.18 ± 7.33 | <.01 |
| Baseline ETCO2 (mmHg) | 40.77 ± 4.25 | 38.95 ± 5.43 | .13 |
| Peak ETCO2 (mmHg) | 48.18 ± 4.82 | 48.35 ± 6.21 | .89 |
| Arousal index (events/h) | 8.88 ± 5.11 | 20.97 ± 11.85 | <.01 |
| Controls (n = 41) | OSA (n = 50) | P | |
|---|---|---|---|
| Total sleep time (min) | 348.44 ± 65.23 | 347.19 ± 65.68 | .93 |
| Sleep efficiency (%) | 84.23 ± 11.93 | 80.05 ± 13.04 | .12 |
| AI (events/h) | 0.11 ± 0.35 | 1.83 ± 2.79 | <.01 |
| AHI (events/h) | 1.85 ± 1.38 | 15.19 ± 8.90 | <.01 |
| Mild (5 ≤ AHI < 15) | 30 (60.0) | ||
| Moderate (15 ≤ AHI < 30) | 16 (32.0) | ||
| Severe (AHI ≥ 30) | 4 (8.0) | ||
| Central sleep apnea (events/h) | 0.30 ± 0.39 | 0.86 ± 1.50 | .04 |
| Baseline SpO2 (%) | 99.05 ± 1.18 | 99.24 ± 1.24 | .46 |
| SpO2 nadir (%) | 92.78 ± 2.40 | 86.18 ± 7.33 | <.01 |
| Baseline ETCO2 (mmHg) | 40.77 ± 4.25 | 38.95 ± 5.43 | .13 |
| Peak ETCO2 (mmHg) | 48.18 ± 4.82 | 48.35 ± 6.21 | .89 |
| Arousal index (events/h) | 8.88 ± 5.11 | 20.97 ± 11.85 | <.01 |
Values are presented as n (%) or mean ± SD.
OSA, obstructive sleep apnea; AI, apnea index; AHI, apnea hypopnea index; SpO2, oxygen saturation; ETCO2, end-tidal carbon dioxide.
Sleep variables of participants
| Controls (n = 41) | OSA (n = 50) | P | |
|---|---|---|---|
| Total sleep time (min) | 348.44 ± 65.23 | 347.19 ± 65.68 | .93 |
| Sleep efficiency (%) | 84.23 ± 11.93 | 80.05 ± 13.04 | .12 |
| AI (events/h) | 0.11 ± 0.35 | 1.83 ± 2.79 | <.01 |
| AHI (events/h) | 1.85 ± 1.38 | 15.19 ± 8.90 | <.01 |
| Mild (5 ≤ AHI < 15) | 30 (60.0) | ||
| Moderate (15 ≤ AHI < 30) | 16 (32.0) | ||
| Severe (AHI ≥ 30) | 4 (8.0) | ||
| Central sleep apnea (events/h) | 0.30 ± 0.39 | 0.86 ± 1.50 | .04 |
| Baseline SpO2 (%) | 99.05 ± 1.18 | 99.24 ± 1.24 | .46 |
| SpO2 nadir (%) | 92.78 ± 2.40 | 86.18 ± 7.33 | <.01 |
| Baseline ETCO2 (mmHg) | 40.77 ± 4.25 | 38.95 ± 5.43 | .13 |
| Peak ETCO2 (mmHg) | 48.18 ± 4.82 | 48.35 ± 6.21 | .89 |
| Arousal index (events/h) | 8.88 ± 5.11 | 20.97 ± 11.85 | <.01 |
| Controls (n = 41) | OSA (n = 50) | P | |
|---|---|---|---|
| Total sleep time (min) | 348.44 ± 65.23 | 347.19 ± 65.68 | .93 |
| Sleep efficiency (%) | 84.23 ± 11.93 | 80.05 ± 13.04 | .12 |
| AI (events/h) | 0.11 ± 0.35 | 1.83 ± 2.79 | <.01 |
| AHI (events/h) | 1.85 ± 1.38 | 15.19 ± 8.90 | <.01 |
| Mild (5 ≤ AHI < 15) | 30 (60.0) | ||
| Moderate (15 ≤ AHI < 30) | 16 (32.0) | ||
| Severe (AHI ≥ 30) | 4 (8.0) | ||
| Central sleep apnea (events/h) | 0.30 ± 0.39 | 0.86 ± 1.50 | .04 |
| Baseline SpO2 (%) | 99.05 ± 1.18 | 99.24 ± 1.24 | .46 |
| SpO2 nadir (%) | 92.78 ± 2.40 | 86.18 ± 7.33 | <.01 |
| Baseline ETCO2 (mmHg) | 40.77 ± 4.25 | 38.95 ± 5.43 | .13 |
| Peak ETCO2 (mmHg) | 48.18 ± 4.82 | 48.35 ± 6.21 | .89 |
| Arousal index (events/h) | 8.88 ± 5.11 | 20.97 ± 11.85 | <.01 |
Values are presented as n (%) or mean ± SD.
OSA, obstructive sleep apnea; AI, apnea index; AHI, apnea hypopnea index; SpO2, oxygen saturation; ETCO2, end-tidal carbon dioxide.
Differences in global characteristics between children with OSA and controls
Cortical structure networks of both OSA and controls met the criteria of small-world network [((CCreal/CCrand)/(PLreal/PLrand)) > 1] across all network densities (10%–35% in 1% increments).
Permutation tests revealed significantly increased CC in OSA at network densities of 17%–20% (FDR corrected p-value < .05, Figure 1) and significantly higher AUC for CC in OSA compared to controls (FDR corrected p-value = .03, Figure 2). PL and SW were not significantly different between OSA and controls. The results remained similar when we performed the permutation tests stratified by age, BMI, or sex.
Comparison of properties of cortical structure network between children with obstructive sleep apnea (OSA) and controls with varying network densities. The gray shade shows the group differences obtained from 5000 permutations, and the real group differences were presented in blue. Dotted red lines are the 2.5% and 97.5% critical values. Asterisk denotes the significant difference at a given network density.
The difference in area under curve of network properties between children with obstructive sleep apnea (OSA) and controls. The clustering coefficient of children with OSA significantly increased relative to the controls (FDR corrected p-value = .03). Blackline and dotted red lines are the mean and 95% confidence interval of the group difference, respectively, obtained from 5000 permutations. Blue circles indicate the observed difference. Asterisk denotes the significant difference. Dotted gray line is the reference line (i.e. zero).
Differences in regional characteristics between children with OSA and controls
Compared with controls, children with OSA showed decreased AUC for NBC in six brain regions including left caudal anterior cingulate, left caudal middle frontal, left fusiform, left transverse temporal, right pars opercularis, and right precentral gyri (FDR corrected p-value < .05, Figure 3). There were no brain regions showing higher AUC of NBC in children with OSA than controls. The results remained similar when we performed the permutation tests stratified by age, BMI, or sex.
The difference in area under curve of nodal betweenness centrality (NBC) between children with obstructive sleep apnea (OSA) and controls. The NBC in six brain regions of children with OSA significantly decreased relative to the controls (FDR corrected p-value < .05). Blackline and dotted red lines are the mean and 95% confidence interval of the group difference, respectively, obtained from 5000 permutations. Blue circles indicate the observed difference. Dotted gray line is the reference line (i.e. zero).
Distribution of hub regions
Eight and six hub regions were identified for controls and children with OSA respectively (see Figure 4). Most of these hub regions were the same for OSA and controls. Three regions, left caudal anterior cingulate, left transverse temporal, and right frontal pole were hub regions in controls but not children with OSA. In contrast, left rostral anterior cingulate was a hub region in OSA but not in controls.
Spatial distribution of network hub regions based on nodal betweenness centrality (NBC). Network hub regions are highlighted by the red circles. The network hub regions were identified when the network regions were greater than one standard deviation above the mean of NBC. The size of each circle indicates the value of NBC.
Differences in network characteristics according to OSA severity
To examine the potential impact of OSA severity on cortical structure network properties, children with OSA were divided into two groups: (1) mild OSA (AHI 5–14 events/h, n = 30) and (2) moderate to severe OSA (AHI ≥ 15 events/h, n = 20). Similar to the whole group analysis, children with either mild OSA or moderate to severe OSA had increased AUC for CC (FDR corrected p-value < .05) and decreased AUC for NBC (FDR corrected p-value < .05) in left fusiform, right pars opercularis, and right precentral gyri compared to controls. OSA severity significantly affected the cortical structure network properties.
Children with moderate to severe OSA had increased AUC for CC (FDR corrected p-value = .037) and decreased AUC for NBC in right precentral gyrus (FDR corrected p-value = .003) compared to children with mild OSA. In addition, six and seven hub regions were identified respectively for children with mild OSA and moderate to severe OSA. Five regions, bilateral entorhinal, bilateral temporal pole, and left frontal pole were the same for children with either mild OSA or moderate to severe OSA. Right parahippocampal area was a hub region in children with mild OSA but not children with moderate to severe OSA. In contrast, left caudal anterior cingulate and left parahippocampal were hub regions in children with moderate to severe OSA but not in children with mild OSA.
Interaction between OSA and age, sex, or BMI
To test interactions between OSA and age, sex, or BMI, we stratified the participants into two categories of age (early puberty, age < 14 years vs. late puberty, age ≥ 14 years), sex (man vs. woman), or BMI (BMI z-score < 2.0 vs. BMI z-score ≥ 2.0). For nonparametric tests, we permuted participants stratified by each of the confounders separately. Then, we generated the null distribution of interaction terms as difference in difference. The OSA by age, sex, or BMI interaction effects were not significant for global and regional network properties (p-value > .05).
Discussion
OSA is known to affect brain structure; in children, such changes may lead to behavioral and cognitive deficits, especially those who do not receive treatment or are inadequately treated [7,49]. However, whether or not these changes involve cortical development as it relates to cortical thickness and the structural network of the cortex has not been investigated. Therefore, this study was motivated by a desire to address OSA in children as a network disorder underpinned by changes in cortical structure connectivity. Toward this goal, we modeled the brain as a network of brain regions, where we estimated the network based on inter-regional associations of cortical thickness across subject cohorts [21] and applied network analysis to describe the network organization properties [23]. Three major findings in the present study support our hypothesis. First, children with OSA had significantly increased CCs, a global measure of network integration and efficiency. Secondly, AUC for NBC, a local measure of network segregation, discriminated between regional characteristics of OSA and controls in left caudal anterior cingulate, left caudal middle frontal, left fusiform, left transverse temporal, right pars opercularis, and right precentral gyri. Finally, while most of the hub regions were the same for OSA and controls, there were four different hub distributions between controls and OSA. Microglial mitochondria dysfunction with increase of ROS production caused by intermittent hypoxia [14, 15] might induce the failure of eliminating unused connections. Residual connections may consequently lead to higher CC and lower centrality of affected brain regions in children with OSA than normal controls. Our findings suggest that untreated OSA in children is associated with a reorganization of cortical structure that affects global and regional characteristics.
The cortical structure networks of both children with OSA and controls followed a small-world network across a wide range of network densities in line with previous brain network studies in children [24, 25]. Brain maturation and development are impacted by alterations of global network properties [50]. Previous studies have suggested that brain network development or maturation was characterized by a shift from higher functional segregation in children to higher functional integration in adults [51]. In particular, decreased CCs suggest decreased functional segregation and maturation of interregional pathways [50, 52]. In the present study, we found significantly higher CCs in children with OSA compared to controls, a finding similar to previous network studies of disorders known to influence brain development [53–55]. In a dense network, there are more connected triangles than there are in a sparse network [56]. An abundance of triangles is thought to reflect increased residual connections associated with a failure of elimination of unused connections. Our finding thus supports the notion that OSA influences cerebral cortical development, including a gradual pruning of excess synaptic connections and change in neuronal size, glial cell density, and vasculature during childhood years [49] which is suggestive of unrefined functional segregation and immature interregional pathways.
Previous OSA studies have shown cortical thickening in left fusiform and cortical thinning in left caudal anterior cingulate, left caudal middle frontal, right pars opercularis, and right precentral gyri of children with OSA [49, 57]. These regional cortical changes can lead to altered association of cortical thickness between brain regions, which affects regional characteristics of the cortical structure network. To investigate change in regional characteristics, the present study tested differences in NBC between children with OSA and controls. The left caudal anterior cingulate, left caudal middle frontal, left fusiform, left transverse temporal, right pars opercularis, and right precentral gyri of children with OSA showed significant lower NBC. A brain region with high NBC indicates that the brain region is highly interactive with the other regions [45, 46]. The significant decrease of NBC in these regions shown in the present study is thus indicative of reduced structural interaction between these regions in children with OSA, and may be induced by hypoxia-related disruption of the structural pathways.
Structural impairment of the caudal anterior cingulate would affect its related pathways, such as those related to sensorimotor, affective, and cognitive processing [58, 59]. The caudal middle frontal gyrus intersects the precentral gyrus, which controls saccadic, rapid, and conjugate eye movements. It allows the central vision to scan numerous details within a scene or image [60]. In addition, the left middle frontal gyrus plays a key role in the development of literacy [61]. The fusiform gyrus is the region known to be involved in face processing [62] and the transverse temporal gyrus is related to the early processing of incoming auditory stimuli for the understanding of speech [63]. Thus, normal structural development of the temporal cortex, including fusiform and transverse temporal gyri, is important for social cognition and behavior in childhood [59, 63, 64]. The right pars opercularis, one of the subregions of Broca’s complex, is functionally correlated with bilateral precentral, bilateral middle frontal, left superior temporal gyri, and left Broca’s complex for phonological processing and syntactic processing [65]. The decreased centrality in these regions may indicate altered information transport and integration [66] and it may cause affective or cognitive deficits as well as academic difficulties in children with OSA, as has been documented previously [7, 13, 49].
The efficient integration of networks within the brain is supported by hub regions with high centrality that facilitate connecting distant brain regions [67]. In the present study, while most of the hub regions were the same for children with OSA and controls, we found that the children with OSA had fewer hub regions than controls and a different hub distribution. The number of hub regions increases during childhood [68] and a disruption of hub regions may contribute to neural dysfunction related to various psychiatric and neurological disorders [69]. Our findings of atypical hub distributions in children with OSA may be an indication of abnormal cortical development or damage to brain structures, either of which may result from intermittent hypoxia and sleep fragmentation, as have been thoroughly documented in OSA. Although this is beyond the scope of the current study, it remains an important line of inquiry for future research.
Previous studies found that OSA severity is associated with brain network characteristics [19, 20, 70]. Chen et al. reported that CC, PL, and SW were significantly correlated with AHI and Epworth Sleepiness Scale in adult men with severe OSA [19]. Park et al. Found that the regional topological values in multiple brain regions were significantly correlated with OSA severity in adults with OSA [20]. In the present study, we found that OSA severity was significantly correlated with the cortical structure network properties. Children with moderate to severe OSA had increased AUC for CC and decreased AUC for NBC in right precentral gyrus compared to children with mild OSA. In addition, there was a difference in hub distribution. Our findings support previous studies, suggesting that OSA severity worsens brain dysfunction [19, 20, 70]. Nevertheless, since the average population age in the most of previous studies was older than our study population, comparisons across studies are somewhat limited.
The main strengths of our study are (1) the large size sample compared to previous studies of children with OSA, and (2) the availability of multimodal data such as MRI imaging and PSG data in children, and (3) the application of a network analysis approach. On the other hand, the present study has several limitations. First, although FreeSurfer provides cortical and subcortical segmentations, some regions, including hippocampus, amygdala, and cerebellum are not available in the thickness pipeline due to resolution constraints or challenging anatomic shapes [71]. Recent studies have proposed a segmentation and parcellation method based on a multiatlas label fusion technique, which would allow estimates of cerebellar cortical thickness [72]. Future studies of OSA should aim to investigate the association of cortical thickness between cerebellar regions that are highly susceptible to hypoxia. In addition, we did not acquire data on mood or cognition in this dataset and therefore were not able to examine their association with our findings. Future studies should include behavioral correlates of these brain networks to more fully explore their relationships in children with OSA.
In summary, this study provides additional data that further explain the anatomical blueprint of atypically developed cortical structure networks in children with OSA. Our approach contributes to an understanding of the neurodevelopmental impact of untreated OSA in children. In addition, it stimulates future research that examines the role of network disorders in the development of learning difficulties, attention deficits, disruptive behaviors, and cognitive dysfunction in children with OSA.
Funding
This work was supported by the National Center for Advancing Translational Sciences (NCATS), components of the National Institutes of Health (NIH), through CTSA (Clinical and Translational Science Awards) grants UL1TR001073, KL2TR001071 and TL1TR001072, and HL-130468-A1.
Disclosure Statement
None declared.
Data Availability
The data underlying this article will be shared on reasonable request to the corresponding author.
Comments