Atypical Development of Functional Brain Networks in Neonates with Congenital Heart Disease
- Jung-Hoon Kim1,
- Josepheen De Asis-Cruz1,
- Nickie N. Andescavage1,
- Kushal Kapse1,
- Mary T. Donofrio2,
- Gilbert Vezina3,
- Adre du Plessis4, and
- Catherine Limperopoulos1
- 1Developing Brain Institute, Children’s National Hospital, Washington, DC 20010
- 2Division of Cardiology, Children’s National Hospital, Washington, DC 20010
- 3Division of Diagnostic Imaging and Radiology, Children’s National Hospital, Washington, DC 20010
- 4Prenatal Pediatrics Institute, Children’s National Hospital, Washington, DC 20010
-
Author contributions: J-H.K., J.D.A-C., N.N.A., A.d.P., and C.L. designed research; J-H.K. and J.D.A-C. performed research; J-H.K., J.D.A-C., K.K., G.V., A.d.P., and C.L. contributed unpublished reagents/analytic tools; J-H.K., J.D.A-C., and C.L. analyzed data; J.-H.K., J.D.A-C., N.N.A., K.K., M.T.D., G.V., A.d.P., and C.L. edited the paper; J-H.K., J.D.A-C., N.N.A., and C.L. wrote the paper.
Abstract
Congenital heart disease (CHD) affects ∼1% of live births in the United States and is the most prevalent congenital disorder. Despite advances in neonatal cardiovascular surgery improving survival, neurodevelopmental impairments remain prevalent, impacting motor skills, social behavior, and executive function. Motor deficits and long-term challenges in emotional regulation and memory are particularly common. Recent research using resting-state functional MRI (rs-fMRI) has revealed disorganized brain networks in newborns with CHD. However, those few prior rs-fMRI studies examining the impact of CHD have relied on predefined brain parcellations to compare group-level connectivity, limiting the ability to capture spatial alterations in neonatal brain networks in CHD. Understanding these network-level changes is critical for elucidating mechanisms of neurodevelopmental impairment and identifying early biomarkers of risk. To address these gaps, our study introduces two conceptual advances: (1) a data-driven approach to investigate atypical brain network development in high-risk CHD and (2) the use of a population-sized, independent dataset of healthy newborns to derive a normative set of neonatal brain networks. By analyzing a large rs-fMRI of human newborns (N = 448; 219 females and 229 males), we identify atypical brain activity in the sensorimotor and limbic networks of newborns with complex CHD. Notably, before cardiovascular surgery, these networks are split into left and right hemispheric subnetworks. Postoperatively, these components coalesce into a singular, symmetric pattern resembling networks observed in healthy neonates. Our study highlights the potential of rs-fMRI to detect subtle, early functional disruptions in CHD and may inform future biomarkers of neurodevelopmental risk.
Significance Statement
Congenital heart disease, the most common congenital disorder, affects 1% of live births and is associated with persistent neurodevelopmental impairments despite improved surgical survival. These deficits, including motor, socioemotional, and cognitive challenges, may stem from early brain network disruptions. Prior resting-state fMRI studies in CHD relied on predefined parcellations, limiting detection of subtle spatial alterations. In this study, we used a data-driven approach and leveraged an independent normative dataset to define resting-state networks. Comparing CHD patients and healthy controls against these independently derived networks, we reveal atypical sensorimotor and limbic network organization preoperatively, which normalizes post-surgery. These findings highlight the potential of rs-fMRI to identify early biomarkers of neurodevelopmental risk and guide targeted interventions in this high-risk population.
Introduction
Congenital heart disease (CHD) is the most prevalent birth defect and affects ∼1% of live births in the United States (Parker et al., 2024). Despite advances in neonatal cardiac surgery, neurodevelopmental disorders remain highly prevalent in this population (Mussatto et al., 2014; Gaynor et al., 2015; Derridj et al., 2021; Sanz et al., 2021; Loblein et al., 2023; Gelaw and Tilahun, 2025). CHD impacts multiple developmental domains, including motor skills, social behavior, and executive function (Limperopoulos et al., 2001; Holm et al., 2007; Brandlistuen et al., 2011; Calderon and Bellinger, 2015; Bolduc et al., 2020; Clancy et al., 2020; Turner et al., 2022; Vogel et al., 2025). While critical periods of risk for brain injury include pre-, intra-, and postoperative stages, growing evidence suggests that brain growth and development are disrupted prenatally (Masoller et al., 2016; Kelly et al., 2017). Many children with CHD later experience difficulties with emotional regulation and memory, linked to limbic system impairments (Bonthrone et al., 2022).
Recent resting-state fMRI (rs-fMRI) studies in newborns with CHD has revealed disorganized brain networks and autonomic dysconnectivity, which may underlie later sociocognitive deficits and autonomic dysfunction (De Asis-Cruz et al., 2018; De Asis-Cruz et al., 2023). Functional near-infrared spectroscopy (fNIRS) corroborates similar connectivity changes (Fourdain et al., 2023; Provost et al., 2023). Structural study also reported delayed maturation of the connectome in CHD neonates, characterized by lower integration and higher segregation compared with controls (Feldmann et al., 2020). Importantly, these early disturbances persist into adolescence and young adulthood (Enguix et al., 2022).
To date, most studies examining CHD-related connectivity in newborns have relied on group-average parcellations derived from adult (Enguix et al., 2022) or healthy infants (De Asis-Cruz et al., 2018). While previous parcel-based rs-fMRI studies provided important knowledge, showing altered connection strength within brain networks in CHD (De Asis-Cruz et al., 2018; Enguix et al., 2022), the inherent limitations in parcel-based analysis preclude detailed evaluation of spatial network patterns in CHD relative to healthy peers. Data-driven, parcel-free analytical techniques, independent component analysis (ICA), may offer higher-resolution comparisons. In adults, ICA exhibited network-like brain maps that better represent the brain system than predefined brain parcellations (Fan et al., 2021; Reeves et al., 2025) and has been widely applied to study neurological disorders, including Alzheimer’s disease (Meda et al., 2012) and schizophrenia (Wu et al., 2015; Wu et al., 2018), as well as to characterize task-specific network organizations engaged across various behavioral tasks (Duff et al., 2012). While prior studies have successfully used ICA in healthy infants (Fransson et al., 2007; Fransson et al., 2009; Doria et al., 2010; Toulmin et al., 2015), ICA has not been applied to CHD neonate, as ICA requires large datasets to ensure reliability.
To address this gap, we employed ICA approach combined with a large independent normative dataset to establish unbiased reference networks. Unlike conventional ICA applications that compare spatial weights or apply dual regression, our method leverages ICA-derived networks to characterize dynamic brain states, labeling each time point by its dominant state for subsequent analysis. Dynamic state-based approaches using ICA have been widely applied in older children and adults (Karahanoğlu and Van De Ville, 2015; Suk et al., 2016; Liu et al., 2019; Wang et al., 2021; Palmer et al., 2022). Prior infant studies have examined brain state dynamics using alternative methods such as dynamic functional connectivity, yet most relied on parcel-based analyses and lacked independent normative templates (Ma et al., 2020; Yates et al., 2022; França et al., 2024; Damera et al., 2025). To our knowledge, this is the first application of ICA-derived networks combined with an independent normative dataset in a neonatal CHD cohort. This framework enables novel evaluation of temporal variability in network organization during early development. We observed a notable hemispheric split in sensorimotor and limbic networks before surgery. Postoperatively, RSN patterns shifted toward more typical, unified configurations, possibly reflecting improved hemodynamics.
Materials and Methods
Subjects and MRI scans
We studied two large rs-fMRI datasets consisting of 664 human newborns (315 females; 349 males). The first dataset consisted of scans from 200 newborns (112 females; 88 males); of which 99 had a diagnosis of congenital heart disease (CHD) and 101 were healthy controls. All MRI studies were performed at the Developing Brain Institute (DBI) at Children's National Hospital. The second cohort consisted of 464 rs-fMRI scans from full-term human neonates (203 females; 261 males) that are publicly available as part of the developing human Connectome Project (dHCP) and referred to as the dHCP dataset. The dHCP dataset was utilized to derive the data-driven, detailed set of neonatal brain networks, and the DBI dataset was used to investigate group differences in CHD and control brain networks by comparing these to the normative brain networks defined by the dHCP dataset.
The Developing Brain Institute dataset: healthy control and congenital heart disease
We analyzed a cohort of 200 newborns with complete MRI datasets (including both structural and functional scans) as well as demographic and clinical information. Exclusion criteria for infants included maternal factors such as pregnancy complications (e.g., pregnancy-induced hypertension and hypothyroidism), multiple pregnancies, maternal medications, as well as infant-related factors such as metabolic disorders, genetic disorders, dysmorphic features, chromosomal abnormalities, and evidence of major structural malformation or congenital infections affecting brain development. All studies were conducted under the regulations and guidelines approved by the Institutional Review Board (IRB) of Children's National; written informed consent was obtained from each maternal-infant dyad who participated in the study.
MRI acquisition protocols were identical between controls and CHD groups. Specifically, structural and functional MR images (referred to sMRI and rs-fMRI, respectively) were acquired using 3T scanner (Discovery MR750, GE HealthCare). Structural (sMRI) scanning parameters were as follows: T1-weighted spoiled gradient echo; TR, 7.37 ms; TE, 2.48 ms; voxel size, 0.625 × 1 × 0.625 mm; and T2-weighted fast spin-echo, TR, 2,500 ms; TE, 84.59 ms, voxel size, 0.625 × 1 × 0.625 mm. For rs-fMRI, resting-state scans were acquired using the following parameters: TR, 2,000 ms; TE, 35 ms; voxel size, 3.125 × 3.125 × 3 mm; flip angle, 60°; field of view, 200 mm; and matrix size, 64 × 64. The scan duration was 6.7 min (200 volumes) for most scans, except for two older scans with a shorter duration (150 and 120 volumes). These two scans were excluded from the analysis due to their insufficient scan length after preprocessing. Aside from the scan duration, all other acquisition parameters were consistent across all rs-fMRI scans. Prior to the scan, babies were fed, swaddled in a warm blanket, immobilized using an infant vacuum pillow, and equipped with ear plugs/earmuffs. Heart rate and oxygen saturation were monitored by a nurse during the scan. Anatomical T1w and T2w MRIs obtained before/after cardiovascular surgery were reviewed by an expert pediatric neuroradiologist (G.V.) for injury or structural anomalies. No major abnormalities were detected; however, 33.7% of infants (among 77 neonates analyzed in CHD group) exhibited minor-to-mild injuries, including tiny foci of hemorrhage (n = 13), mild white matter injury (n = 7), mild ventriculomegaly (n = 3), mild ventricular prominence (n = 2), and a small subacute hematoma (n = 1). After cardiovascular surgery, 24.7% of infants were reported to have new brain findings including microhemorrhages (n = 8), mild white matter injury (n = 4), bilateral multifocal acute infarcts (n = 2), mild increased prominence of the ventricles (n = 2), focal encephalomalacia/gliosis (n = 1), small nonhemorrhagic punctate lesions (n = 1), and slight enlargement of the sulcal spaces (n = 1). Cardiac diagnoses were categorized as infants requiring corrective or palliative surgery based on CHD diagnosis and cardiovascular procedure.
Developing human brain mapping: independent cohort of healthy newborns
The developing Human Connectome Project (dHCP) is an open science project approved by the UK National Research Ethics Authority (Edwards et al., 2022). We utilized the third data release of the dHCP dataset, which was available at https://data.developingconnectome.org/. Of the 464 subjects (203 females; 261 males), we analyzed data from 294 subjects (131 females and 163 males); with good quality (radiology score = 1/2) rs-fMRI datasets.
MR scans were acquired using a 3T scanner (Philips Achieva system). High-resolution T1- and T2-weighted anatomical imaging was acquired with followings: T1-weighted image; spatial resolution, 0.8 mm isotropic; FOV, 145 × 122 × 100 mm; and TR, 4,795 ms; T2-weighted image; spatial resolution, 0.8 mm isotropic; FOV, 145 × 145 × 108 mm; TR, 12,000 ms; and TE, 156 ms. Functional MRI scans were acquired with acquisition specifications: multislice gradient-echo echoplanar imaging (EPI) sequence; multiband factor, 9. TR, 392 ms; TE, 38 ms; flip angle, 34; spatial resolution, 2.15 mm isotropic; and total volume, 2,300 (∼15 min). The downloaded dHCP rs-fMRI data was already preprocessed and registered into the volumetric brain template (Fitzgibbon et al., 2020). The image quality of rs-fMRI was reviewed using the summary report provided by dHCP, and we visually inspected each of downloaded dHCP data.
Preprocessing of rs-fMRI data
Preprocessing pipeline of rs-fMRI data of DBI dataset has been validated previously (De Asis-Cruz et al., 2015; De Asis-Cruz et al., 2018; Kim et al., 2023a,b). Briefly, the preprocessing steps consisted of (1) within-volume motion correction, (2) slice-time correction, (3) removal of first four volumes to stabilize the magnetization, (4) de-spiking of timeseries, (5) bias-field correction, (6) motion correction over timeseries, (7) coregistration of fMRI into structural MRI, (8) wrapping to a standard neonate brain template, (9) intensity normalization, (10) smoothing at 5 mm full-width half-maximum, (11) bandpass filtering at 0.009–0.08 Hz, and (12) nuisance regression. For the nuisance regression, regressors consisted of (1) six motion estimates, their first-derivates and quadratic from the rigid-body analysis of head motion, (2) signals averaged over localized white matter (Jo et al., 2010), and the (3) first three principal components of the ventricular CSF signal (Behzadi et al., 2007; Muschelli et al., 2014). Volumes with excessive head motion (volume-to-volume head motion > 0.2 mm) or >10% of voxels classified as intensity outliers were excluded in the analysis. After censoring motion-corrupted volumes, some rs-fMRI scans failed to pass preprocessing steps due to their lower degree of freedom and were subsequently excluded. Finally, preprocessed rs-fMRI data were projected into the standard cortical space using HCP workbench command -volume-to-surface-mapping.
For dHCP dataset, the downloaded dHCP fMRI data was already preprocessed, and no further preprocessing steps were applied. We simply projected preprocessed rs-fMRI data in the volumetric space to the cortical space, corresponding to DBI rs-fMRI data, using HCP workbench command -volume-to-surface-mapping. Then, we normalized the timeseries as zero mean and unitary variance, per vertex.
Establishing a normative set of resting-state brain networks
The dHCP dataset was used to define a normative set of resting-state brain networks (RSNs; Fig. 1A). RSNs were derived using multivariate exploratory linear optimized decomposition into independent components (MELODIC) algorithm (Beckmann and Smith, 2004). Each RSN (i.e., independent component, IC), defined by MELODIC, consisted of distinct brain regions representing various brain networks. Specifically, all of the dHCP datasets (# of subjects = 294, data length per dataset = 2,000) were fed as input to the MELODIC algorithm and yielded a total of 15 IC. The spatial pattern of each IC represents each RSN. The number of RSNs (=15) was heuristically determined given observations described in our previous studies (Kim et al., 2023a; Kim et al., 2024a).
Graphic illustration depicting the similarity to a normative set of resting-state brain networks (RSNs). A, RSNs from a large newborn cohort of 297 full-term newborns using group independent component analysis (GICA). B, Each time point of resting-state fMRI data from individuals either from control or CHD group is compared with the normative RSNs. Given its similarity to RSNs, each time point is assigned one of RSNs. Per subject, the similarity and prevalence are estimated. C, Correlational distance from exemplified subject. Given the estimated distance map, per volume, one of 15 RSNs is assigned, as highlighted as blue box. Finally, pattern similarity and prevalence are calculated.
We further derived group-specific RSNs. We independently applied MELODIC algorithm to data of controls and pre-op CHD group, each set # of ICs to 15. Then, the order of group-specific RSNs was set based on its similarity to the normative RSN set. To improve the reproducibility of our findings, we repeated the above analysis five times; each with random initialization. We reported similarity averaged over five repetitions. For visualization purposes, the group-specific IC set with the best similarity to the normative RSN set was presented in figures.
Prevalence and similarity of individual data to normative brain set
We evaluated the (dis)similarity of controls versus CHD groups with respect to the normative set of RSNs derived from the
dHCP dataset (Fig. 1B). Specifically, per scan of controls (n = 77) or pre-op CHD (n = 77), we calculated the distance of each time point to the set of RSNs, defined as <mml:math display="inline"><mml:mrow><mml:mi mathvariant="normal">Dis</mml:mi></mml:mrow><mml:msub><mml:mrow><mml:mi mathvariant="normal">t</mml:mi></mml:mrow><mml:mrow><mml:mi>t</mml:mi><mml:mo>,</mml:mo><mml:mi>k</mml:mi></mml:mrow></mml:msub><mml:mspace
width=".1em"></mml:mspace><mml:mo>=</mml:mo><mml:mspace width=".1em"></mml:mspace><mml:mn>1</mml:mn><mml:mo>−</mml:mo><mml:mrow><mml:mi
mathvariant="normal">abs</mml:mi></mml:mrow><mml:mo stretchy="false">[</mml:mo><mml:mrow><mml:mi>r</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mrow><mml:msub><mml:mi>P</mml:mi><mml:mi>t</mml:mi></mml:msub><mml:mo>,</mml:mo><mml:mrow><mml:mi
mathvariant="normal">RS</mml:mi></mml:mrow><mml:msub><mml:mrow><mml:mi mathvariant="normal">N</mml:mi></mml:mrow><mml:mi>k</mml:mi></mml:msub></mml:mrow><mml:mo
stretchy="false">)</mml:mo></mml:mrow><mml:mo stretchy="false">]</mml:mo><mml:mo>,</mml:mo></mml:math>Distt,k=1−abs[r(Pt,RSNk)],
where r stands for Pearson’s correlation coefficient, Pt for the cortical pattern at t-th timepoint, and RSNk for the cortical pattern of kth RSN, respectively (Fig. 1C for exemplified subject). Then, we defined two measures given the calculated correlation distance: prevalence and similarity.
The prevalence was estimated by (1) assigning each timepoint to one of 15 RSNs having the minimal distance, (2) summing the
number of time points for each RSN, and (3) dividing by the total timepoints. The similarity of each scan was calculated as
follows: (1) per timepoint, selecting the maximal Pearson’s r value (i.e., minimal distance) among 15 r values from 15 RSNs, (2) converting Pearson’s r values to Fisher’s z-score, and (3) averaging z-score over time points (Fig. 1C, right panel).
Postoperative change of brain networks
Finally, we evaluated whether and how brain networks in newborns with CHD changed after cardiovascular surgery. For this analysis, we examined a subset (n = 33) of CHD neonates with rs-fMRI data before and after surgery. Analogous to the previous analysis, we evaluated the change of similarity compared with the normative RSN set between pre- and post-op CHD using a paired t test.
Statistical test
We applied multiple statistical models, including an unpaired two-sample t test (control vs pre-op CHD), a paired t test (pre-op CHD vs post-op CHD), and a linear regression model. When appropriate, Bonferroni’s correction was used to adjust for multiple comparisons. The significance level (α) was determined based on the total number of tests (n). For example, for the linear regression model, n = 15 (corresponding to the number of RSNs), resulting in an α threshold of 0.0017. Results were considered statistically significant only if they met the Bonferroni-adjusted criterion (pBonferroni < 0.05). For tests that did not survive correction, the uncorrected p value was reported. All analyses were performed using MATLAB.
Results
In the DBI dataset, 200 newborns had complete MRI and demographic data. Of these, 46 were excluded due to rs-fMRI preprocessing failure, primarily resulting from insufficient degrees of freedom after censoring motion-corrupted volumes. The remaining 154 neonates included 77 diagnosed in utero with CHD and 77 controls. Within the CHD group, a subset of newborns (n = 33) underwent both preoperative and postoperative scans, which were used to examine brain network changes following surgery.
In total, we analyzed rs-fMRI datasets from 448 neonates: 294 from the publicly available dHCP and 154 from DBI. In DBI, control group (birth gestational age: 39.06 ± 1.52 weeks) underwent MRI between 38.57 and 47.57 postmenstrual age (PMA; mean and SD: 41.72 ± 1.93 weeks). The gestational age at birth of CHD group was 38.85 ± 1.27 weeks and the scan PMA was 40.21 ± 2.08 and 43.89 ± 3.92 weeks, for pre-op and post-op, respectively.
Here, the dHCP dataset served as the normative brain networks of neonates and the difference between controls and CHD was evaluated by comparing to this normative brain networks. The mean and standard deviation of their scan PMA were 41.18 and 1.70 weeks, respectively. Demographic details regarding the dHCP dataset can be found elsewhere at (Edwards et al., 2022). Detailed demographic information for controls and CHD group of DBI database is provided in Table 1. In CHD group, about one-third had single ventricle (1 V), and their subtypes were highly heterogeneous. Among the 77 newborns in the CHD group, 49 among 50 infants with 2V-CHD underwent 2 V repair (49/51 had 2V-CHD), while the remaining 28 underwent 1 V palliation.
Demographic information of the studied cohorts
Normative set of resting-state networks of neonates
We derived 15 RSNs representing the normative set of neonatal brain networks using the large-scale healthy newborn population (Fig. 2). The values at each vertex (scaled from –3 to 3) represent ICA loadings, indicating the degree (positive or negative) to which each component contributes to the observed fMRI signal patterns. As noted in previous studies, we also found that the pattern of neonatal RSNs was largely reminiscent of the brain networks of adults. For example, several RSNs (RSN1, 2, 3, 4, and 5) covered the sensorimotor network, each with a unique coverage of sensorimotor regions. RSN2 largely covered the primary motor area, RSN4 spanned the somatosensory region, RSN1 covered the premotor area, and RSN3 and 5 covered the supplementary motor cortex. Other sensory-related RSNs included 10 and 12—visual areas—and 7 and 9—auditory regions. We further observed higher-order networks, which integrate information from multiple networks, including 6, 8, 11, 13, and 15. Lastly, while harder to interpret, we found that the pattern of RSN14 resembled the topography of global signal—global synchrony within each hemisphere. To that end, we successfully derived the normative RSNs of healthy neonates at the voxel level via a data-driven approach.
The full set of normative resting-state networks (RSNs). Those RSNs are defined using the large datasets of healthy full-term newborns. The scale of each map is standardized, each with 0 mean and 1 standard deviation.
Atypical development of brain networks in newborns with CHD
We compared brain networks in neonates with CHD to the normative RSN set. Briefly, we compared the pattern similarity and prevalence (see Materials and Methods for the definition of measures) between controls and CHD groups. We found that the similarity of overall RSN increased over scan age in both cohorts (Fig. 3; r = 0.35 and 0.30, p = 0.002 and 0.008 in CHD and controls, respectively). To further examine whether this trend varied across different age ranges, we divided each cohort into younger and older subgroups based on their median ages (41.72 weeks for controls and 39.57 weeks for CHD). Subjects at the median age were included in both subgroups. In controls, younger neonates showed a significant increase in RSN similarity with age (r = 0.35; p = 0.026; n = 43), whereas no trend was observed in the older subgroup (r = 0.03; p = 0.879; n = 44). In contrast, CHD neonates exhibited the opposite pattern: no significant trend in the younger subgroup (r = −0.14; p = 0.357; n = 43) but a strong positive trend in the older subgroup (r = 0.47; p = 0.001; n = 44).
Evolution of newborn resting-state brain networks with advancing postmenstrual age (PMA). The scatterplot of similarity over aging. The similarity is estimated using the full set of RSNs. Different colors highlight different condition groups. Each dot stands for each subject. N of both groups = 77. Line stands for the linear fit.
We also found that CHD had lower similarity than controls even after controlling potential confounding factors, sex and birth weight (t = 2.02, uncorrected p, puncorr = 0.045), suggesting atypical brain development in CHD. By further extending the analysis to the level of each RSN, we demonstrated that the majority of RSNs increased their similarity over gestational age (Table 2; Bonferroni-corrected p, pBonferroni < 0.05) except RSN1, 4, 9, 11, and 15 (pBonferroni = 0.055, 0.065, 0.013, 0.074, and 0.096, respectively). Of these, two RSNs (RSN1 and RSN15) showed significantly lower similarity in CHD compared with controls (pBonferroni < 10−3 for both). Among all the derived RSNs, only RSN2 showed significant association to both variates, age (t = 2.99, puncorr = 0.003) and condition (i.e., CHD; t = 2.00, puncorr = 0.047) although neither remained significant after Bonferroni’s correction.
The association of pattern similarity of each RSN to scan age and condition
Next, we investigated whether the prevalence of each RSN also varied by age or CHD (Table 3). Unlike similarity, the prevalence was independent of age (puncorr > 0.05 for all RSNs except RSN13 with marginal significance, puncorr = 0.011, not survived after Bonferroni’s correction). Interestingly, RSN1, showing condition-wise difference in the similarity, also showed difference in its prevalence (higher prevalence in control, 11.55 ± 3.36% vs 8.03 ± 5.21% in CHD; t = 3.90, pBonferroni < 10−3). Altogether, our findings demonstrated the vulnerability of developing brain networks to CHD, and importantly, this vulnerability was selective to the type of brain networks.
The association of prevalence of each RSN to scan age and condition
Association between brain networks and CHD diagnostic categories
We compared pattern similarity between newborns with 1V-CHD (n = 27) and 2V-CHD (n = 50). After controlling for age, sex, and birth weight, we found no significant association between CHD type and overall brain network pattern similarity, with the exception of RSN11 (Table 4). In RSN11, resembling the default mode network, the pattern similarity was higher in the 1V-CHD group (0.13 ± 0.02) compared with the 2V-CHD group (0.11 ± 0.03), with an uncorrected p value of 0.029. However, this difference did not survive after correction for multiple comparisons (Bonferroni-adjusted).
The association of pattern similarity of RSN to CHD type
Altered pattern of brain networks in CHD
Given previous findings supporting the atypical development of brain networks in CHD, here we further investigated pattern difference of brain networks in CHD neonates. Briefly, we mapped group-specific brain networks by applying MELODIC to each cohort and compared group-specific networks to the normative brain maps (see Materials and Methods for details). Given findings from the previous section, we focused on two brain networks (RSN1 and 15), which showed significant group difference in their similarity (Fig. 4). We observed that RSN1 was represented by single brain network in controls (r = 0.65 ± 0.01 vs r = 0.39 ± 0.01 in second best) while, in CHD, it was distributed into two brain networks, each covering the left or right hemisphere (r = 0.37 ± 0.03 and 0.32 ± 0.01, respectively; Fig. 4A). We confirmed that the difference in the best versus second best similarity in controls (Δz = 0.34 ± 0.03) was significantly higher than CHD = 0.06 ± 0.04 (p < 10−6, one-tail two-sample t test). We found that RSN15 also showed a similar hemispheric split in CHD (Fig. 4B; r = 0.35 ± 0.01 and 0.32 ± 0.02), but not in controls (r = 0.55 ± 0.01 and 0.29 ± 0.01). We also confirmed that the difference in the best versus second best similarity in controls (Δz = 0.32 ± 0.03) was significantly higher than CHD, Δz = 0.03 ± 0.03 (p < 10−6, one-tail two-sample t test). We also found that pronounced activation in pars region (Fig. 4B; IC7 in controls). In sum, our findings suggested the adverse impact of CHD resulting in separation of specific RSNs (RSN1 and 15) into two hemispheres.
Altered pattern of resting-state networks in CHD. For two RSNs (IC1 and IC15 for A and B), group-specific RSNs (black, control; red, CHD) are selected based on their best similarity to each of the normative RSNs. R values in black and red boxes represent the estimated similarity to the normative RSN map (blue).
Changes in resting-state networks after cardiovascular surgery
In a subset of our cohort (n = 33), we acquired rs-fMRI after cardiovascular surgery. We tested whether cardiovascular surgery was associated with improved functional brain network. For this analysis, we only included the subset of neonates with both rs-fMRI data before/after surgery (pre/post-op). Analogous to previous analysis, we evaluated the similarity of each CHD group with respect to the normative set of RSNs derived from the dHCP dataset. Given our previous finding regarding the sensitivity of RSN1 and 15 to CHD condition, we focused on these two RSNs. For both RSN 1 and 15, the similarity of post-op images increased compared with pre-op (Fig. 5A,B; paired sample t test, t = 4.00 and 2.81, puncorr < 10−3 and =0.009, for RSN1 and 15, respectively). Given that similarity increased with age, this finding may have been due to advanced age of the post-op cohort (Table 3). To confirm that the increment in similarity in post-op CHD was not driven by the differences in age, we demonstrated that increment degree in pattern similarity was not correlated to the age gap—post-op versus pre-op (for RSN1 and 15; r = −0.19 and −0.08; puncorr = 0.33 and 0.70, respectively). Notably, after controlling for scan age, sex, and birth weight, the similarity in the post-op CHD group was comparable to that of controls, with no significant differences observed (t = 0.17 and 0.85; puncorr = 0.87 and 0.40, for RSN1 and 15, respectively). We also found that the network patterns of post-op CHD group resembled the pattern of controls, rather than pre-op CHD (Fig. 5C,D). Specifically, the hemispheric split observed in pre-op CHD was less pronounced in post-op CHD (for RSN1, r = 0.57 ± 0.06 and 0.33 ± 0.07; for RSN15, r = 0.51 ± 0.01 and 0.18 ± 0.01). We confirmed that, compared with controls, post-op CHD showed similar level of difference in the best versus second best similarity (p = 0.300 and 0.997, RSN1 and RSN15, respectively; one-tail two-sample t test). Notably, RSN15, which was derived from the dHCP dataset, exhibited some differences compared with the network patterns derived from the DBI dataset (which included all controls and CHD groups). These differences were particularly evident in the pars region and may be attributed to variations in rs-fMRI acquisition parameters and/or preprocessing pipelines used across datasets, highlighting the need for future studies to systematically investigate this aspect. Altogether, our findings from a subset of neonates with both pre- and postoperative rs-fMRI data demonstrated improved postoperative hemodynamics was associated with improved similarity, particularly within RSN1 and RSN15. This was independent of age-related effects and brought the postoperative CHD network patterns closer to those of healthy controls, with reduced hemispheric asymmetry.
RSNs after cardiovascular surgery resembles RSNs of healthy newborns. A, B, Similarity of group-specific resting-state networks (RSNs) to two normative RSNs, RSN1 and RSN15. Red bar stands for standard deviation. The difference between controls and Pre/Post-Op CHD was tested using linear regression model, by controlling the effect from scan age, sex, and birth weight. The difference between Pre- and Post-Op CHD was evaluated using paired sample t test.***Bonferroni-corrected p < 0.05. C, D, Group-specific RSNs (black, control; red, CHD) are selected based on its best similarity to each of normative RSN. R value in colored boxes stands for the estimated similarity to the normative FBN map (black box). HC: Healthy control.
Discussion
Using a data-driven approach, we identified atypical brain networks in newborns with complex CHD. Normative RSNs were derived from a large independent dataset (Fig. 2) and used as a reference to compare CHD and control groups, revealing significant alterations in CHD (Fig. 3; Tables 2, 3). No association was found between RSN patterns and CHD subtype (Table 4). Notably, sensorimotor (RSN1) and limbic (RSN15) networks showed hemispheric splits preoperatively (Fig. 4), which shifted toward more typical, unified patterns after surgery (Fig. 5), suggesting improved hemodynamics may support partial recovery of functional brain network.
We first established normative RSNs using the dHCP dataset and examined CHD versus control differences relative to these benchmarks. A key strength of our approach is reduced susceptibility to preprocessing-related biases, as the dHCP dataset underwent distinct processing steps. Prior work in adults shows that preprocessing methods like global signal regression can introduce spurious group differences (Saad et al., 2012). Despite neonatal rs-fMRI challenges such as motion artifacts (Kim et al., 2023b), our controls aligned closely with the normative set, supporting its validity. Leveraging independent datasets may offer an objective foundation for neonatal research and should be more widely adopted.
RSN similarity to the normative set increased with age in both groups (Fig. 3, Table 2), consistent with maturation toward the dHCP benchmark (mean age = 41.18 weeks). Controls showed increasing similarity in younger infants that plateaued with age, whereas CHD neonates exhibited minimal change early but a marked increase at older ages, suggesting delayed network maturation. This interpretation, however, should be viewed cautiously, as the age distribution of the CHD group differed from that of the dHCP dataset, and few CHD infants were scanned beyond the normative age. Larger cohorts spanning a wider age range are needed to determine whether CHD networks ultimately converge with normative trajectories or reflect persistent alterations.
Our findings indicate that CHD disrupts the developmental trajectory of specific brain networks. The sensorimotor network (RSN1) was most affected, consistent with prior reports of motor delays in children with complex CHD (Newburger et al., 2012; Gaynor et al., 2015; Naef et al., 2017; Huisenga et al., 2023). This suggests that early disruption of sensorimotor network formation may underlie long-term motor impairments. The limbic network (RSN15) was also vulnerable (Fig. 4), aligning with its role in emotion regulation (Heimer and Van Hoesen, 2006; Catani et al., 2013) and behavioral evidence of emotion regulation difficulties in CHD neonates (Clancy et al., 2020). Taken together, early disruptions in sensorimotor and limbic networks may contribute to the neurodevelopmental profiles observed in children with CHD following surgery.
Interestingly, these CHD-sensitive networks (sensorimotor and limbic) were not age dependent, unlike other RSNs. We speculate this may reflect the early emergence of those networks in utero (Prechtl, 1990; Vasung et al., 2010; Fagard et al., 2018), making them more vulnerable to CHD-related insults. However, this hypothesis requires confirmation, as prior fetal neuroimaging studies have also reported the emergence of other networks during this period (Jakab et al., 2014; van den Heuvel et al., 2018; Turk et al., 2019). The lack of age-dependency may indicate that maturation of sensorimotor and limbic networks pauses during the neonatal period or occurs later in infancy. Alternatively, maturation may progress toward subcortical regions, such as the cerebellum, which were not examined in this study (see limitations).
Overall, 2V-CHD is associated with better neurodevelopmental outcomes than 1V-CHD (Sommer et al., 2008; Hoskoppal et al., 2010; Huisenga et al., 2021; Nasr et al., 2023). However, some studies report no differences at school age, with both groups scoring below normal (Brosig et al., 2017). Consistent with this, our findings (Table 4) suggest CHD subtype may not strongly influence large-scale brain network organization. These results should be interpreted cautiously, as our ICA-based approach may lack sensitivity to subtle effects and our sample size may limit statistical power.
The hemispheric separation observed in RSN1 and RSN15 preoperatively was absent after surgery, with patterns resembling healthy controls. While age differences could partly explain this, two follow-up analyses suggest recovery was not solely age related. We propose that cardiovascular surgery improved cerebral oxygen delivery, facilitating network reorganization, consistent with prior reports (Smith, 1988; Taylor, 1998; Mahle et al., 2002). However, long-term neurodevelopmental challenges remain common despite early recovery (Bellinger et al., 1995; Marino et al., 2012). The absence of postoperative differences may reflect analytical limitations, including exclusion of cerebellar activity, a region critical for socioemotional development and highly vulnerable during rapid growth (Bobylova et al., 2007; Limperopoulos et al., 2007; BOLDUC et al., 2011; Koning et al., 2017; Kim et al., 2024b). Reduced cerebro-cerebellar plasticity could sustain lasting impairments and serve as a biomarker for outcomes. Alternatively, some network interactions (e.g., visual-auditory) emerge later and may not be detectable in neonates. Long-term follow-up is needed to clarify these early postoperative findings.
In our cohort, conventional structural MRI revealed no major brain malformations in either the CHD or controls groups. Among neonates with CHD, preoperative scans were reviewed by an experienced pediatric neuroradiologist (G.V.), and no cases of overt ischemic injury, major hemorrhage, or any significant structural abnormalities were identified, aside from some mild findings (see Materials and Methods for details). These results suggest that observed network differences likely reflect subtle functional disruptions rather than visible structural lesions (Fig. 4). This underscores the value of rs-fMRI for detecting early functional changes that may precede structural abnormalities and predict later neurodevelopmental outcomes. Specifically, we anticipate that deviations from normative network patterns, quantified by similarity measures, could serve as early biomarkers, a hypothesis we aim to validate in future longitudinal studies.
Here, we utilized two metrics, similarity and prevalence, to assess the vulnerability of the developing brain networks in newborns CHD. These captured distinct effects: similarity was more sensitive to CHD and age, while prevalence showed minimal differences. Conceptually, similarity reflects how closely brain organization matches normative patterns, whereas prevalence indicates time spent in specific RSN states. These measures likely represent different aspects of network architecture. Supporting this, an adult study found prevalence sensitive to psychotic experiences while RSN similarity remained stable (Wang et al., 2021), suggesting architecture and state stability vary across conditions. Our framework, leveraging an independent template to evaluate both metrics, offers a promising approach for studying neurological conditions and developmental trajectories in fetal-neonatal-infant cohorts.
This study has several limitations. First, the small sample size, especially in the postoperative group, may limit generalizability. Newborns with severe CHD were less likely to undergo post-op MRI (e.g., 40% of 1V-CHD in pre-op only vs ∼26% in both scans), introducing potential bias. We mitigated this by incorporating an independent dataset, but larger cohorts are needed. Second, our analysis focused on cortical surfaces, excluding deep brain activity. Voxel-level analysis, commonly applied in neonatal rs-fMRI (Smyser et al., 2011; De Asis-Cruz et al., 2015; Tortora et al., 2019; Eyre et al., 2021; Damera et al., 2024; Kim et al., 2024b), could address this but has lower specificity and sensitivity than surface-based methods (Grova et al., 2006; Brodoehl et al., 2020). Third, high variability in CHD severity limited our ability to assess its impact on brain network development. We observed no significant differences between 1V-CHD and 2V-CHD subtypes. As our center continues to collect data from newborns with various CHD subtypes, we hope to address this question in future studies when a larger sample pool becomes available. Fourth, due to data discontinuity, we could not investigate the transition between RSNs. Given prior studies exploring change of between-brain-state transitions (or between-state switches) across developmental stages (Karahanoğlu and Van De Ville, 2015; Liu et al., 2019; Ye et al., 2025) and in various clinical conditions (Suk et al., 2016; Wang et al., 2021; Palmer et al., 2022), it can be a promising direction for future work. Lastly, sleep state was not accounted for due to the lack of reliable sleep-stage classification. Although challenging to assess during MRI acquisition, brain state (e.g., active vs quite sleep; sleep vs naturalistic movie-watching) has been shown to influence neonatal brain network organization (Mitra et al., 2017; Yates et al., 2023; Mueller et al., 2025). Future studies incorporating robust approaches to characterize or control for brain state during neonatal fMRI acquisition, such as electrocorticogram-based sleep stage classification (Govindan et al., 2026), may help mitigate potential confounding effects arising from state-dependent neural variability.
In summary, we applied a novel data-driven approach to identify atypical resting-state networks in newborns with complex CHD, revealing disruptions, particularly in sensorimotor and limbic networks, that were not explained by structural abnormalities. These disruptions showed partial normalization after cardiovascular surgery, suggesting improved cerebral perfusion may support early functional recovery. Our findings highlight the value of rs-fMRI for detecting subtle network-level changes that may precede structural alterations and predict long-term neurodevelopmental outcomes. Future work, especially by translating those findings into the individual-level analysis, should validate whether deviations from normative connectivity can serve as early biomarkers of neurodevelopmental risk.
Footnotes
- Received September 30, 2025.
- Revision received January 25, 2026.
- Accepted January 28, 2026.
-
We thank the families who participated in this study. This work was funded by grant T32 5T32HD098066-03 from the National Institute of Child Health and Human Development (J-H.K.), grant R01 HL116585-01 from the National Heart, Lung, and Blood Institute, National Institutes of Health (C.L.), grant MOP-81116 from the Canadian Institute of Health Research (C.L.), and the A. James & Alice B. Clark Foundation’s Parent-Child Health Initiative (J-H.K., J.D.A-C., N.N.A., K.K., A.d.P., and C.L.).
-
The authors declare no competing financial interests.
- Correspondence should be addressed to Catherine Limperopoulos at climpero{at}childrensnational.org.
- Copyright © 2026 the authors
