Prenatal social disadvantage is associated with alterations in functional networks at birth Authors: Ashley N. Nielsen, Regina L. Triplett, Lourdes M. Bernardez, Ursula A. Tooley, Max P. Herzberg, Rachel E. Lean, Sydney Kaplan, Dominique Meyer, Jeanette K. Kenley, Dimitrios Alexopoulos, David Losielle, Aidan Latham, Tara A. Smyser, Arpana Agrawal, Josh S. Shimony, Joshua J. Jackson, J. Philip Miller, Marcus E. Raichle, Barbara B. Warner, Cynthia E. Rogers, Chad M. Sylvester, Deanna M. Barch, Joan L. Luby, Christopher D. Smyser Published: 2024 ABSTRACT SignificanceSocial disadvantage relates to variations in early trajectories of brain development. Yet, the role and significance of exposure during the prenatal period, when brain systems begin to rapidly emerge, are poorly understood. Here, we demonstrate that prenatal social disadvantage is associated with alterations in functional connectivity that are present at birth and are distributed across functional networks and subcortical structures key to executive function and behavioral adaptation. This evidence that deviations in brain development associated with social disadvantage begin in utero suggests that interventions to promote adaptive brain development and reduce psychiatric risk may need to begin prenatally and underscore the importance of social supports for expectant parents that reduce exposures associated with social disadvantage. Social disadvantage, including reduced household income and neighborhood disadvantage, is associated with higher risk for poor developmental, psychiatric, educational, and occupational outcomes across the lifespan (1). Social disadvantage is particularly impactful when experienced in early life (2), presumably because adverse exposures in early childhood alter or disrupt complex neural processes that occur during typical early brain development (3). Further still, indirect prenatal exposure through social disadvantage experienced by mothers during pregnancy may alter fetal brain development and yield early deviations in brain structure, function, metabolism, and connectivity through hormonal, microbiome, genetic, epigenetic, and immune factors that cross the placenta and fetal blood–brain barrier (4). These early neural alterations may, in turn, contribute to predisposition toward risk or resilient developmental trajectories (2–4). Recent work from our group demonstrated that the relation between prenatal exposure to social disadvantage and brain development is evident in the first weeks of life, with greater exposure associated with smaller brain volumes (5) and aberrant white matter microstructure (6). Herein, we investigate whether prenatal exposure to social disadvantage is associated with variation in infant functional networks to contribute to the characterization of alterations in the emerging organization of large-scale brain systems that underlie elevated risk for poor outcomes. Intrinsic activity across large-scale brain systems is spatiotemporally structured (7) such that activity of distant, but functionally related, brain areas are correlated at rest (i.e., functional connectivity [FC]), and these brain areas coordinate to form functional networks (8). Functional network organization supports complex cognitive, social, and emotional behavior (9); constrains learning and plasticity (10); and can vary across individuals (11) and in disease (12). These networks can be observed in the first weeks of life (13) and serve as foundational precursors that are refined over development (14). Alterations in FC associated with social disadvantage have been identified in children and adolescents (15–19). Further, some alterations have clinical significance and mediate the links between social disadvantage and subsequent behavioral and socioemotional outcomes (15–17, 19). The amygdala has been consistently implicated (15–17), but recent work has delineated the broader impact of social disadvantage, including brain-wide alterations in FC across both association networks important for executive function and sensorimotor networks key to adaptively interacting with the environment (18, 19). One hypothesized link between FC and social disadvantage is that the quality, predictability, and nature of a child’s experiences in their environment during early development alter the coactivation and coordination of brain areas over time (20). Another possibility is that prenatal exposure to social disadvantage initiates these changes in emerging functional networks, with alterations in FC beginning in utero and persisting into childhood and adolescence rather than being driven by postnatal experiences alone. Given the limited social mobility in the United States and countries around the world (21), childhood social disadvantage is frequently preceded by maternal social disadvantage during pregnancy, making it difficult to disentangle the effects of pre- and postnatal exposure. Without measuring FC at birth, extant studies linking social disadvantage and FC in children have been unable to capture the contribution of prenatal social disadvantage to observed differences in functional brain organization. Further, characterizing the effects of prenatal social disadvantage on FC across the brain at birth may illuminate adaptations that confer early vulnerability and increase adversity-linked risk for psychopathology. The current study includes a cohort of 399 mother–infant dyads participating in the Early Life Adversity, Biological Embedding, and Risk for Developmental Precursors of Mental Disorders (eLABE) study. Initially recruited through a larger investigation (22), mothers were identified in the first trimester of pregnancy and oversampled for low income. Prenatal exposure to social disadvantage was assessed during all three trimesters using demographic and survey measures and analyzed using factor analysis to derive a latent measure (encompassing parental education, health insurance, income, neighborhood disadvantage, nutrition) (23). Of the 385 infants imaged during natural sleep within the first weeks of life, a cohort of 261 healthy, full-term infants had at least 10 min of low head-motion resting-state functional magnetic resonance imaging (fMRI) data acquired to measure FC across the cortex, subcortex, and cerebellum. Multivariate machine learning techniques were implemented to extract patterns of infant FC that vary according to the latent measure of social disadvantage experienced by mothers during pregnancy. Critically, a separate validation sample of 151 infants with high rates of maternal social disadvantage participating in the Prenatal Cannabis Use and Development of Offspring Brain and Behavior During Early Life (CUDDEL) study were used to independently validate the findings identified in the eLABE sample. Multiple strategies were employed to interrogate the importance of individual functional networks and subcortical structures for the detection of variation in infant FC related to prenatal social disadvantage. Last, we examined whether either prenatal psychosocial stress (i.e., depression, anxiety, racial discrimination, stressful life events) or differences in brain maturity may contribute to the relationship between prenatal social disadvantage and variation in infant FC. Results Prenatal exposures were assessed using demographic and survey measures during all three trimesters and combined to derive a latent measure of prenatal social disadvantage (23) and align with previous eLABE studies (5, 6). Maternal health insurance status (private, public, underinsured), education (7 d (n = 32), low birth weight (<2,000 g n = 1), and identification of congenital anomaly in infancy (n = 1). Of the remaining 284 neonates, 23 infants had fewer than 10 min of low motion fMRI data, leaving a sample of 261 neonates. We report the framewise displacement in mm (mean = 0.19; SD = 0.16; range = 0.04 to 1.2) and the amount of retained low-motion data in minutes (mean = 16.6; SD = 4.4; range = 10.1 to 41.9) for the included sample. Image preprocessing and FC processing. The processing pipelines utilized here have been previously reported in Nielsen et al. (14) and Sylvester et al. (13) and a brief description is provided below. Both the eLABE and CUDDEL samples were processed in an identical manner to limit potential differences due to analysis across cohorts, incorporating surface-based strategies to account for individual and developmental differences in cortical folding (55). fMRI preprocessing included correction of intensity differences attributable to interleaved acquisition, bias field correction, intensity normalization of each run to a whole brain mode value of 1,000, linear realignment within and across runs to compensate for rigid body motion, and linear registration of BOLD images to the adult Talairach isotropic atlas. Neonates were registered: BOLD to individual T2 to group-average T2 from the eLABE cohort to 711-2N Talairach atlas. Field distortion correction was performed, using the FSL TOPUP toolbox (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/TOPUP) and applied with FSL’s applytopup. Following initial processing, fMRI data were censored at FD > 0.25 mm, with the additional restriction that only epochs of at least three consecutive frames FD < 0.25 mm were included. This FD threshold was selected after taking into account the smaller radius of infants’ heads (56) and reviewing motion traces in multiple subjects (57). In order to be included in the study, a minimum of 10 min (750 frames) of data were required. We calculated the number of frames retained after censoring and the average FD across all frames (precensoring) for each individual subject and evaluated a covariate of interest. Next, we used the Melbourne Children’s Regional Brain Atlases [MCRIB (55)] as a surface-based neonatal parcellation approach to generate surfaces for each subject. Volumetric resting-state BOLD timeseries were mapped to subject-specific surfaces using established procedures adapted from the Human Connectome Project as implemented in Connectome Workbench 1.2.3. MCRIB was also used to parcellate the subcortex for each subject to generate the volumetric resting-state BOLD timeseries for left and right amygdala, hippocampus, putamen, pallidum, caudate, thalamus, and cerebellum. After mapping to the surface, each dataset underwent rs-fc preprocessing. In the initial iteration, the data were processed with the following steps: a) demean and detrend within run, band-pass filtering (0.005 Hz < f < 0.1), interpolation of censored frames, b) multiple regression with nuisance timeseries including white matter, ventricles, extra-axial cerebrospinal fluid, and whole brain, as well as 24-parameter Friston expansion regressors derived from head motion. fMRI data were aligned across subjects into the “fs_LR32k” surface space using spherical registration. Time courses for surface data were smoothed with geodesic 2D Gaussian kernels (σ = 2.25 mm). FC network construction. For each participant, whole-brain FC was generated by extracting the timeseries from 333 previously defined surface parcels (58) (Fig. 1) and 14 subcortical regions and calculating the Fisher z-transformed Pearson correlation between each pair of parcels. For each individual, this yields a 347 × 347 correlation matrix. We used definitions of functional networks derived in healthy adults (58) that include processing networks like the auditory, visual, somatomotor-hand, somatomotor-mouth, retrosplenial, and control and other association networks like the frontoparietal, cingulo-opercular, dorsal attention, ventral attention, salience, default mode, cingulo-parietal. Multivariate Machine Learning Procedures. In alignment with the procedures reported in Nielsen et al. (14), SVR with a linear kernel was implemented to identify multivariate patterns of FC that carry information about prenatal exposure to social disadvantage. SVR is a sensitive and powerful tool to extract the multivariate relationship between a set of many features (here, functional connections) and labels (here, prenatal social disadvantage) in a training sample which can subsequently be applied to make predictions about unseen test individuals. Ten-fold cross-validation was used to identify patterns of FC in the training set (i.e., 90% of the eLABE sample) and test whether these patterns could accurately discriminate differences in social disadvantage in the remaining individuals in the test set (i.e., 10% of the eLABE sample). To minimize the impact of confounding covariates with the potential to influence FC or the relationship between social disadvantage and FC, we controlled for PMA at scan, sex, amount of data, mean FD (prescrubbing), maternal medical risk, tobacco use, and marijuana use. Within each fold of 10-fold cross-validation, linear models were fit to best capture the relationship between the regressors and FC from each functional connection (60,031) and the relationship between the regressor and social disadvantage, separately in both the training set and testing set. The residuals from these models were then used as the features and labels for training and testing with SVR. Ten-fold cross-validation was repeated with 100 random partitions of the dataset into training and test sets to ensure that performance was not due to serendipitous random seeding (SI Appendix, Fig. S7). Deidentified data used for training (FC matrices, social disadvantage, covariates) are publicly available (59). Validation in the CUDDEL Sample. Predictive approaches like SVR are designed to optimize prediction and could potentially identify multivariate patterns of FC that do not generalize to describe the impact of early adversity outside of the original training sample (27). Cross-validation alone is not a sufficient test of the robustness of data-driven, multivariate predictive models (60). Because the CUDDEL study did not collect all measures used in the eLABE sample, training to identify and evaluate patterns of FC related to prenatal social disadvantage was slightly modified. First, because the CUDDEL sample did not incorporate some of the measures used as covariates (e.g., maternal medical risk), the SVR model was generated without regressing covariates. Second, the CUDDEL sample did not acquire all measures that comprise the latent factor composite social disadvantage variable. Thus, only the Income-to-Needs Ratio was used to evaluate whether estimates of social disadvantage generated from infant FC were related to true exposure to social disadvantage. For each fold of cross-validation, the multivariate patterns of FC identified in the eLABE sample were also applied to estimate the social disadvantage of the individuals in the validation sample (CUDDEL). This generates a single estimate of social disadvantage based on FC for each individual in the eLABE sample and 10 estimates of social disadvantage based on FC (i.e., for each fold of 10-fold cross-validation) for each individual in the validation sample. Performance was quantified by calculating the correlation between the estimated social disadvantage and actual Income-to-Needs Ratio (i.e., R2) in the eLABE and validation samples separately. To evaluate whether this performance was better than expected by chance, null models were generated by identifying patterns of FC related to social disadvantage that was permuted before training in the eLABE sample. Estimates of social disadvantage based on permuted data were then correlated with prenatal Income-to-Needs. This process was repeated 500 times and the observed correlation between estimates of social disadvantage based on FC and prenatal Income-to-Needs was compared to the distribution of correlations generated with null models to produce a P-value. Because the Income-to-Needs Ratio was acquired in both samples, a separate validation of the patterns of infant FC related to Income-to-Needs was also conducted. For each fold of cross-validation, the multivariate patterns of FC identified in the eLABE sample were also applied to estimate the prenatal Income-to-Needs Ratio of the individuals in the validation sample (CUDDEL). This generates a single estimate of Income-to-Needs Ratio based on FC for each individual in the eLABE sample and 10 estimates of the Income-to-Needs Ratio based on FC (i.e., for each fold of 10-fold cross-validation) for each individual in the validation sample. Performance was quantified by calculating the correlation between the estimated and actual Income-to-Needs Ratio (i.e., R2) in the eLABE and validation samples separately. Detection of Social Disadvantage with FC from Individual Functional Networks. Feature selection (i.e., restricting the number of features used for training), as detailed in Nielsen et al. (14), was used to test the extent to which the FC from a single network might capture the variance in FC related to social disadvantage. For each network, all connections associated with parcels from a single functional network (e.g., within the FP and between the FP and the rest of the brain) were subselected and then only those features were used to detect evidence of prenatal exposure to social disadvantage. False discovery rate was used to correct for the multiple comparisons (12 networks). To further disentangle the relationship between evidence of prenatal exposure to social disadvantage, network size, and network identity, we compared the detection of social disadvantage from FC when the same number of parcels were selected from each functional network. A number of parcels was randomly subselected from a single functional network, from a single parcel to the maximum number of parcels in a network (e.g., for DMN = 41). Only the FC from the connections associated with those sampled parcels was then used as features for estimation of social disadvantage from FC (number of features: 333-13,653). This process was repeated 25 times for each network and each parcel number. The average variance explained as a function of sampled parcel number was generated for each functional network. Detection of Social Disadvantage with Permuted FC from Individual Networks and Subcortical Structures. Feature permutation (i.e., manipulating a restricted set of features used for testing), as detailed in Nielsen et al. (14), was used to investigate whether a given set of features was necessary to capture the brain-wide patterns of FC that facilitate the detection of differences in prenatal exposure to social disadvantage. Permutation sensitivity analysis disrupts the true relationship between social disadvantage and FC present in a limited set of features by permuting these features across the subjects in the test set. The model trained to detect social disadvantage using all the features is then applied to the test set containing permuted features. Permuting these features will negatively impact the estimation of social disadvantage with brain-wide FC if a set of features is uniquely important for detection of social disadvantage. This approach was applied for functional networks and subcortical structures as described below. Networks. For each network, all connections associated with parcels from a single functional network (e.g., within the CO and between the CO and the rest of the brain) were subselected and these features were then permuted across subjects in the test set as described above. To ensure that the impact of permuting the FC of a given network was due to network identity and not network size, the impact of permuting a matched set of randomly selected connections was also examined. For each network, we generated a matched set of functional connections randomly selected from the matrix and these features were then permuted across subjects in the test set. Fifty matched randomly selected feature sets were generated for each network. P-values were calculated for each network by comparing the detection of social disadvantage (R2) when features from a single network were permuted and when randomly selected features were permuted. False discovery rate was used to correct for the multiple comparisons (12 networks). Subcortical structures. The connections associated with a single subcortical structure (left and right) were also selected and these features were then permuted in the test set as described above. The impact of permuting a matched set of randomly selected connections was also examined. One thousand matched randomly selected feature sets were generated as the null for all subcortical structures. P-values were calculated for each subcortical structure by comparing the detection of social disadvantage (R2) when features from a single subcortical structure are permuted and when randomly selected features are permuted. False discovery rate was used to correct for the multiple comparisons (seven subcortical structures). Benchmarking Variation in Infant FC Related Social Disadvantage. Individual components of social disadvantage. SVR was used to determine the extent to which multivariate patterns of infant FC carried information about the individual components of the Social Disadvantage latent factor in isolation. SVR was used for the continuous variables (Income-to-Needs, Area Deprivation Index, Healthy Eating Index) and support vector classification as used for the binary and categorical variables (Insurance: public vs. private; Education: less than college graduate vs. college graduate or more). Regression models were generated with and without the covariates used above (PMA at scan, sex, amount of data, mean FD (prescrubbing), maternal medical risk, tobacco use, and marijuana use). Composite of social disadvantage. SVR was used to build additional models to identify patterns of FC related to equally weighted composites of the components of the latent factor of social disadvantage. Using only the subjects that did not have any missing data, a composite of social disadvantage was created using all components (n = 184) and another composite was created excluding the Healthy Eating Index (n = 223). To generate a composite, the components were z-scored and averaged together. To avoid potential leakage across the training and test sets, the individuals in the test set were excluded to calculate the mean and SD used to z-score the components. Regression models were generated with and without the covariates used above (PMA at scan, sex, amount of data, mean FD (prescrubbing), maternal medical risk, tobacco use, and marijuana use). Brain maturity vs. social disadvantage. Because an infant’s age at scan has previously been shown to be related to infant FC, further analyses were conducted to demonstrate that differences in brain maturity alone did not yield patterns of FC that can estimate social disadvantage. First, we built models to estimate measures of brain maturity (PMA at scan, gestational age at birth, chronological age since birth) using SVR in this sample for context. Though the models that identify patterns of FC related to social disadvantage above control for linear relationships with PMA at scan, it is possible that nonlinear or more complex relationships with brain maturity could inflate performance. Thus, we next expanded the maturity covariates (gestational age at birth, chronological age since birth, PMA at scan, gestational age at birth^2, chronological age since birth^2, PMA at scan^) to account for these potential relationships and examine the effect on estimation of social disadvantage from infant FC. Because greater social disadvantage was associated with earlier delivery, a younger gestational age at birth, and a younger PMA at scan (SI Appendix, Fig. S4A), we also determined whether social disadvantage could be estimated from infant FC in an age-restricted sample that only included infants born between 39 to 40 wk gestation (n = 152). As above, we used SVR to identify patterns of FC related to prenatal social disadvantage. Last, permutation sensitivity analyses were performed at the level of parcels for both the detection of social disadvantage and PMA at scan from variation in infant FC. As above, the connections associated with a single parcel (analogous to a seed map) were selected and these features were permuted across subjects in the test set, removing the information related to either social disadvantage or PMA at Scan. Five hundred matched randomly selected features were generated as the null for all parcels. P-values were calculated for each parcel separately for social disadvantage and PMA at scan by comparing the variance explained when features from a single parcel are permuted and when randomly selected features are permuted. False discovery rate was used to correct for the multiple comparisons. The distribution of parcels important to detect social disadvantage and PMA at scan from FC and their overlap are depicted in SI Appendix, Fig. S5. Maternal psychosocial stress vs. social disadvantage. Maternal psychosocial stress is potentially related to social disadvantage and/or infant FC. Thus, we examined how this variable contributed to the observed detection of prenatal social disadvantage from infant FC. SVR was used to determine the extent to which multivariate patterns of FC carried information about maternal psychosocial stress. Supplementary Material Appendix 01 (PDF)