Fetal Frontolimbic Connectivity Prospectively Associates With Aggression in Toddlers

Abstract

Background

Aggression is a major public health concern that emerges early in development and lacks optimized treatment, highlighting need for improved mechanistic understanding regarding the etiology of aggression. The present study leveraged fetal resting-state functional magnetic resonance imaging to identify candidate neurocircuitry for the onset of aggressive behaviors before symptom emergence.

Methods

Pregnant mothers were recruited during the third trimester of pregnancy to complete a fetal resting-state functional magnetic resonance imaging scan. Mothers subsequently completed the Child Behavior Checklist to assess child aggression at 3 years postpartum (n = 79). Independent component analysis was used to define frontal and limbic regions of interest.

Results

Child aggression was not related to within-network connectivity of subcortical limbic regions or within–medial prefrontal network connectivity in fetuses. However, weaker functional coupling between the subcortical limbic network and medial prefrontal network in fetuses was prospectively associated with greater maternal-rated child aggression at 3 years of age even after controlling for maternal emotion dysregulation and toddler language ability. We observed similar, but weaker, associations between fetal frontolimbic functional connectivity and toddler internalizing symptoms.

Conclusions

Neural correlates of aggressive behavior may be detectable in utero, well before the onset of aggression symptoms. These preliminary results highlight frontolimbic connections as potential candidate neurocircuitry that should be further investigated in relation to the unfolding of child behavior and psychiatric risk.

Aggression, defined as verbal or physical behaviors intended to cause harm, emerges early in human life. Aggressive acts peak during the preschool period, with more than 90% of preschoolers engaging in at least occasional aggressive acts (1). Although the mean frequency of aggressive behavior steadily decreases after 4 years of age, stability of aggression is typically high such that young children with the highest levels of aggression are most likely to be aggressive adults as well (2,3). When aggression does extend into adolescence and adulthood, it is a strong predictor of negative life outcomes, including psychosocial problems, psychopathology, physical violence, delinquency, and school dropout (4, 5, 6). By the time early life aggression morphs into moderate to severe antisocial behaviors in adolescence, these behaviors have become increasingly entrenched by social feedback loops and are difficult to treat (7, 8, 9). Indeed, current multimodal treatments for aggression and delinquency that are implemented during adolescence, such as multisystemic therapy, have inconsistent success rates that are often not superior to community treatment as usual (10,11). Improved mechanistic understanding of the etiology of aggression, along with earlier intervention, is a critical public health goal.

Basic neuroscience studies that illuminate candidate neurocircuitry before the entrenchment of pathological and maladaptive levels of aggression are lacking. However, translational preclinical research converges with functional magnetic resonance imaging (fMRI) experiments conducted in human adults to isolate dysregulated coactivation of limbic regions, including the amygdala, and the medial prefrontal cortex (mPFC) in the expression of aggressive behaviors. Among individuals who display more aggressive behaviors, dysregulated frontolimbic circuitry manifests both as increased excitatory signals from limbic regions (12, 13, 14) and decreased mPFC-initiated inhibition of limbic activity (15, 16, 17). Yet frontolimbic circuitry undergoes significant changes across development that may alter the nature of these brain-behavior associations. For instance, the amygdala and mPFC demonstrate positive functional coupling during infancy and childhood (18, 19, 20, 21) but become negatively coupled during adolescence and adulthood (20,22). It is imperative to examine brain-behavior associations across the spectrum of development to identify consistent and diverging patterns. Several studies have examined the relevance of infant frontolimbic circuitry for the development of internalizing symptoms (21,23, 24, 25), but none have examined whether early life frontolimbic connectivity is associated with externalizing symptoms. Determining the extent to which individual differences in neural circuitry are related to aggression early in human development represents a foundational gap in our knowledge.

Although it is often assumed that the brain causes behavior, work conducted in adolescents reveals that externalizing behaviors can precede certain neural alterations (26,27), suggesting bidirectional associations between disease states, behavioral phenotypes, and neural development. A particularly rigorous method of parsing the temporal unfolding of biobehavioral risk is to measure brain development before birth and before both the onset of aggression and the postnatal exposures that enhance risk for, or buffer against, this behavioral phenotype. Establishing this temporal unfolding is an essential, although not sufficient, step toward establishing a causal association between brain and behavior.

Frontolimbic brain structures are an especially promising candidate for such work, as they develop early in human life, rendering them feasible targets for examination in utero. Histological work reveals that the amygdala is differentiated by 17 to 18 weeks of gestation (28) and has reached a high degree of structural maturity by 36 weeks of gestational age (GA) (29). Although the PFC develops later (30), early anatomical connections from the amygdala to the forebrain form between the 24th and 26th weeks of gestation (28). Available literature thus suggests that although it continues to be refined by prenatal and postnatal experiences, the core scaffolding for frontolimbic circuitry is in place during the third trimester of pregnancy. Indeed, one study confirms that fetal resting-state fMRI (rsfMRI) demonstrates the capacity to noninvasively quantify individual differences in this circuit before birth (31). Fetal rsfMRI presents the unique opportunity to examine prospective associations between individual differences in the development of frontal and limbic neurocircuitry in utero and subsequent aggressive behavior.

In the current study, we sought to determine whether and how frontolimbic circuitry in utero was associated with childhood aggression using a longitudinal dataset including fetal rsfMRI scans. We specifically tested the novel hypothesis that interindividual variation in frontolimbic functional circuitry before the child’s birth will be associated with aggressive behavior at age 3 years. We also explored whether select comorbid risk factors modified these brain-behavior associations. In this study, we examined 2 risk factors that have been tied to heightened offspring aggression: child language development and maternal emotion dysregulation (32, 33, 34). By providing needed exploration of a core neurodevelopmental pathway that may contribute to heightened aggression early in life, these findings have potential to deepen our understanding of aggression etiology and illuminate sensitive periods for intervention timing.

Methods and Materials

Participants and Procedures

Healthy mothers with singleton pregnancies were recruited from Hutzel Women’s Hospital in Detroit, Michigan, during routine obstetrical appointments. Exclusion criteria included 1) non-native English speaker, 2) less than 18 years of age, or 3) presence of anatomical fetal brain abnormalities identified during ultrasound and/or MRI examination. Fetal MRI scans were obtained at Wayne State University when fetuses were between 22 and 39 weeks GA. Manually segmented and quality assured fMRI data were available for 165 fetuses at the time of this analysis. From these quality-assured data, fetuses were excluded if they were scanned before 25 weeks GA (n = 9), had low birth weight, or were born very preterm (<1800 g or <33 weeks GA) (n = 14). We also excluded fetuses with high average motion (>1.5 mm maximum excursion, >0.5 mm mean, rotational >2°) or had fewer than 100 functional volumes after scrubbing (n = 22). When children were 3 years old (mean [SD] = 36.43 [1.83] months), 79 of these mothers completed follow-up questionnaires about their child’s behavioral and socioemotional development (Figure 1A). The final sample consisted of 79 mother-fetus dyads who were predominantly from low- to middle-income households. Consistent with our exclusion criteria, included dyads had an older GA at the fetal MRI scan and at birth as well as less motion during the fetal MRI scan (Table 1). The Wayne State University Institutional Review Board approved all study procedures, and informed written consent was provided by participating mothers. Data from this sample have previously been used to examine the impact of prenatal stress (31,35,36) and cannabis use (37) on the developing brain, examine associations between preterm birth and functional network development (38), and optimize fetal fMRI preprocessing (39,40).

Figure 1.

Participant ages, attrition, and endorsement of aggressive toddler behavior. (A) A total of 120 mothers completed a resting-state functional magnetic resonance imaging (MRI) scan between 26 and 39 weeks of gestational age that passed quality checks. A subset of these mothers (n = 79) additionally completed a follow-up visit when their child was 3 years of age. (B) Mothers rated children’s aggressive behaviors using the preschool version of the Child Behavior Checklist (CBCL). The number of toddlers displaying each aggressive behavior is plotted. (C) Distribution of total aggression scores in the sample. In the present sample, aggression scores ranged from 0 to 27 (mean [SD] = 8.76 [6.43], median = 8.00). Ninety-one percent of mothers endorsed at least 1 aggressive behavior in their child.

Table 1.

Participant Sociodemographic Variables

	Final Sample, n = 79, Mean (SD) or n (%)	Excluded Sample, n = 86, Mean (SD) or n (%)	Statistical Differences Between Included and Excluded Dyads
Sociodemographic Variables
Maternal Age at Fetal MRI, Years	25.54 (4.66)	25.71 (5.49)	t163 = 0.21, p = .83
GA at Fetal MRI, Weeks	32.96 (3.71)	31.02 (4.50)	t163 = −3.00, p = .003a
Maternal Race/Ethnicity			χ24 (n = 158) = 3.01, p = .56
Asian American	1 (1%)	1 (1%)
Biracial	3 (4%)	4 (5%)
Black/African American	64 (86%)	71 (85%)
Hispanic/Latina	0 (0%)	0 (0%)
White/Caucasian	7 (9%)	7 (8%)
Maternal Education, HS Diploma/GED or Less	44 (59%)	43 (52%)	χ21 (n = 158) = 0.75, p = .39
Maternal Income, <$20,000	44 (56%)	52 (60%)	χ21 (n = 143) = 0.12, p = .73
Maternal Marital Status, Single	46 (62%)	45 (52%)	χ21 (n = 155) = 0.70, p = .40
GA at Birth, Weeks	39.01 (1.60)	37.85 (3.22)	t163 = −2.97, p = .002a
Preterm, <36 Weeks	4 (5%)	14 (16%)
Fetal Sex, Female	34 (43%)	38 (44%)	χ21 (n = 165) = 0.02, p = .88
rsfMRI Characteristics
Number of Low-Motion Volumes	169.28 (54.15)	158.31 (53.50)	t119 = −1.07, p = .29
Mean Translation, mm	0.23 (0.10)	0.25 (0.11)	t163 = 1.17, p = .24
Mean Rotation, mm	0.38 (0.17)	0.47 (0.20)	t163 = 3.15, p = .002a
3-Year Child Follow-up Measures
CBCL/1½-5 Aggression	8.76 (6.43)	8.13 (5.76)	t107 = −0.47, p = .64
CBCL/1½-5 Hyperactivity	2.47 (2.12)	2.53 (2.00)	t115 = 0.14, p = .89
CBCL/1½-5 Internalizing	7.48 (7.81)	7.36 (6.75)	t118 = −0.08, p = .94
Bayley Cognitive Composite	91.03 (7.13)	88.40 (10.74)	t123 = −1.49, p=0.14
Bayley Motor Composite	99.27 (10.40)	93.87 (14.64)	t120 = −2.38, p = .02a
Bayley Language Composite	93.04 (8.24)	89.91 (14.01)	t122 = −1.39, p = .17
Maternal Emotion Regulation Measures
Prenatal Internalizing Symptoms Composite	−0.10 (0.88)	−0.14 (0.95)	t162 = −0.26, p = .80
3-Year Internalizing Symptoms Composite	0.02 (0.85)	−0.26 (0.56)	t118 = −2.11, p = .04a
Quick to Anger	17 (24%)	25 (34%)	χ21 (n = 145) = 1.71, p = .19

Dyads included in our final analyses had an older GA at the scan and at birth, less motion during the fetal functional MRI scan, higher scores on the Bayley Motor Composite, and fewer maternal anxiety/depression symptoms at the 3-year follow-up visit compared with excluded dyads. There were no other sociodemographic differences or differences in rsfMRI data characteristics between dyads who were or were not included in our final analyses. 41 dyads were excluded because they were missing 3-year follow-up data, so excluded sample estimates for 3-year outcomes are based on the subgroup who completed the visit but were excluded from analyses for other reasons as described in Methods and Materials (n = 39).

CBCL/1½-5, Child Behavior Checklist for Ages 1½-5; GA, gestational age; GED, General Educational Development (test); HS, high school; MRI, magnetic resonance imaging; rsfMRI, resting-state functional MRI.

^ap < .05

Measures

Fetal fMRI

Fetal fMRI was obtained using a 3T Siemens Verio 70 cm open-bore system with a 550-g abdominal 4-channel Siemens Flex Coil (Siemens Corp.). For each participant, 360 axial frames (12 minutes) of blood oxygen level–dependent echo-planar imaging data were collected with the following scan sequence parameters: repetition time = 2000 ms; echo time = 30 ms; flip angle: 80°; slice gap: none; voxel size: 3.4 mm × 3.4 mm × 4 mm; matrix size: 96 × 96 × 25. This sequence was repeated when time permitted. On average, 169 low-motion rsfMRI frames were included in analyses (range across fetuses, 100–344).

Maternal Emotion Dysregulation

Mothers completed the Center for Epidemiological Studies–Depression Scale (41) and the Spielberg State-Trait Anxiety Inventory–Trait subscale (42) at the fetal scan and again at a 3-year follow-up visit to assess current depression and anxiety symptoms, respectively. Both measures demonstrated acceptable internal consistency in the present sample (Center for Epidemiological Studies–Depression Scale α = 0.91; Spielberg State-Trait Anxiety Inventory–Trait subscale α = 0.71). Due to correlation between these measures (fetal scan: r = 0.75, p < .001; 3-year visit: r = 0.46, p < .001), we standardized and averaged the Center for Epidemiological Studies–Depression Scale and Spielberg State-Trait Anxiety Inventory–Trait subscale total scores to create a single composite variable representing maternal internalizing symptoms during pregnancy and a second composite representing maternal internalizing symptoms at the 3-year visit. The 3-year internalizing composite was included as a covariate in all analyses examining maternal-reported toddler aggression. Additionally, we leveraged a single yes/no question from the Wayne Indirect Drug Use Screener scale (43) as a measure of maternal quickness to anger: “I get mad easily and feel a need to off some steam.” Mothers completed the Wayne Indirect Drug Use Screener at the fetal scan. In our sample, 17 (24%) mothers reported being quick to anger. All 3 maternal dysregulation variables were explored as potential modifiers of brain-behavior associations in children.

Toddler Aggression

Mothers completed the Child Behavior Checklist for Ages 1½-5 (CBCL/1½-5) (44) when their child was 3 years old. The CBCL/1½-5 aggressive behavior subscale was the primary outcome in the present study, which showed good internal consistency (Cronbach’s α = 0.92). Aggressive problems reported by parents on the CBCL/1½-5 have demonstrated construct validity with aggressive behaviors assessed in the laboratory (45), and this subscale demonstrates acceptable measurement equivalence in minoritized samples in the United States (46, 47, 48). Additional analyses used the attention subscale (α = 0.79) and internalizing subscale (α = 0.84) to determine specificity of findings (see Figure 1C and Table 1 for descriptive information). Other CBCL/1½-5 syndrome scales showed questionable or poor internal consistency in the present sample (αs < 0.68) and so were not examined in subsequent analyses.

Toddler Neurodevelopment

Children were administered the Bayley Scales of Infant and Toddler Development, Third Edition (49) during the 3-year visit by a trained examiner. Bayley composite scores were used to identify children with developmental delays in the present sample for descriptive purposes (Table S1). Given that language delay can co-occur with aggressive acts in children (32,33), the Bayley language composite was also included as a covariate in sensitivity analyses.

fMRI Preprocessing

Preprocessing was performed using a combination of FSL (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) and Statistical Parametric Mapping (SPM12) software (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/). First, low-motion segments were manually selected using the FSL image viewer. Next, Brainsuite (http://brainsuite.org/) was used to manually draw three-dimensional masks around single reference images, and these masks were applied to all other volumes within the corresponding low-motion segment (50). All manually drawn masks were quality assured by a second reviewer. A subset of images was independently masked by separate reviewers, and we calculated the percent overlap between resulting images (i.e., the dice coefficient). These dice coefficients revealed exceptional overlap in fetal brain masks created by separate reviewers (mean = 0.993, range = 0.986–0.997). Subsequent preprocessing included brain extraction, reorientation, within segment motion correction, normalization to a 32-week fetal brain template (51), concatenation of volumes across low-motion segments, realignment to correct for potential normalization misalignments between segments, reapplication of the fetal brain mask, independent component analysis (ICA)–based denoising with FSL Melodic toolbox, and spatial smoothing with a 4-mm full width at half maximum Gaussian kernel. Stringent ICA denoising has been shown to successfully remove motion artifacts from fMRI data without needing to censor high motion spikes and provides superior removal of motion-related signal contamination compared with traditional spike censoring and regression of motion parameters (52). For more information about the ICA denoising used in our fetal data, see Ji et al. (39).

Data Analysis

Identification of Resting-State Networks

A data-driven approach was used to define fetal frontal and limbic networks. Specifically, ICA decomposed the whole-brain data for all fetuses into 35 spatially independent components, each of which exhibited a unique time course profile. ICA was performed using the GIFT Functional MRI Toolbox, version 3.0b (https://trendscenter.org/software%20/gift/), with the Infomax algorithm as the optimization principle (53). Using the minimum description length approach, 35 components were derived based on image quality (54). Reliability and stability of the algorithm were ensured by repeating the component estimation 20 times using Icasso (55). Subject-specific spatial maps and time courses were obtained using the back-reconstruction approach (group-level ICA) (56) and converted to z scores. Three components were discarded because they represented signal from cerebrospinal fluid or other sources of noise, and the remaining group-level components were manually organized to be spatially consistent with established resting-state networks in neonates (57) and fetuses (58) as shown in Figure S3. The present analyses focus on a medial prefrontal network (consisting of 2 ICA-derived components) and a subcortical limbic network (consisting of 1 ICA-derived component). All ICA-derived networks are publicly available for download (https://www.brainnexus.com/projects-2/fetal-resting-state-ica-templates).

Calculation of Within- and Between-Network Functional Connectivity

We extracted the mean blood oxygen level–dependent time series from a 4-mm sphere surrounding the peak voxel within each component of interest (Figure 2, Table 2) for every fetus (n = 79). Peaks were selected on the aggregate component map and applied to each normalized fetal brain to account for variability and signal spread in individual component maps. There were 4 signal intensity peaks across the mPFC and limbic networks. Specifically, in the subcortical limbic network, 2 peaks were isolated in each of the left and right hemispheres, and a single peak was identified in each of the left and right mPFC components. Pearson correlations were calculated between all 4 extracted average time series to create an unthresholded resting-state functional connectivity (FC) matrix for each fetus, including both negative and positive associations. Within-network FC for each fetus was defined as the correlation coefficient between the 2 homologous regions of interest (ROIs) within a given network. Between-network FC was calculated by averaging the correlations between each ROI within the medial prefrontal network with the 2 ROIs in the subcortical limbic network. ROI-to-ROI analyses were Bonferroni corrected (p < .017).

Figure 2.

Fetal frontal and limbic networks based on group-level independent component analysis. The group-level independent component analysis decomposed all resting-state functional magnetic resonance imaging data for all fetuses into a set of spatially independent components, each of which exhibited a unique time course. Two networks consisting of independent component analysis–derived components were explored in the subsequent analyses: (A) a subcortical limbic network consisting of the bilateral amygdala and (B) a medial prefrontal network comprised of the left and right medial prefrontal cortex (mPFC). The peak of each component is marked with black edges.

Table 2.

Location of ICA-Derived Network Peaks

	x	y	z
Limbic Network FC
R Limbic	3	9	−17
L Limbic	−6	9	−17
mPFC Network FC
R mPFC	15	36	−2
L mPFC	−15	36	−2

A 4-mm sphere was drawn around the above coordinates to define our ICA-derived limbic and mPFC seeds. Coordinates reflect the central voxel of each region of interest, normalized to the Serag et al. fetal atlas (51).

FC, functional connectivity; ICA, independent component analysis; L, left; mPFC, medial prefrontal cortex; R, right.

Child Behavior Analyses

Descriptives and measures of variability were examined for all variables. The CBCL/1½-5 aggression subscale showed acceptable distribution normality (aggression skew = 0.53) (Figure 1B, C). Hierarchical linear regression was used to examine associations between fetal FC and behavioral outcomes at 3 years of age, with relevant covariates entered into the first step and fetal within-/between-network connectivity entered in the second step. Because of their theoretical relevance to early brain development, we included the following covariates in the first step of all analyses: number of rsfMRI frames included in the analysis, fetal motion parameters, fetal GA at scan, fetal sex, GA at birth, maternal education, and partner status at the fetal MRI scan. All analyses also controlled for a composite maternal anxiety and depression variable at the 3-year visit. Linear regression assumptions were assessed in several ways. Unstandardized residuals were visually examined using histograms to determine normality, and residuals were plotted against predicted values to ensure homoscedasticity. Finally, Cook’s D was used to identify potential outliers (i.e., values >1), as it considers both leverage and discrepancy. Moderation analyses were implemented with the PROCESS Macro and controlled for the same covariates included in our primary analyses (59). We also conducted exploratory seed connectivity analyses to examine whether variation in child aggression was related to altered frontal and limbic connections to other neural regions. These analyses are described in the Supplement.

Results

Toddler Aggression

Ninety-one percent (n = 72) of mothers reported that their toddler engaged in at least one aggression-related behavior. The most frequent problem behaviors reported were not being able to wait, being stubborn, and wanting attention (Figure 1B). Extreme acts of physical violence, such as hurting animals or attacking people, were rarely endorsed. Overall aggression scores in the sample are displayed in Figure 1C.

Fetal Frontolimbic Networks

The current study centered on 2 networks: a subcortical limbic network and a medial prefrontal network (Figure 2). The average strength of within- and between-network FC across fetuses is visually displayed in Figure 3, and fetal FC descriptive statistics are presented in Table S2.

Figure 3.

Network connectivity patterns vary across fetuses. On average, fetuses showed positive functional correlations (functional connectivity) between frontal and limbic regions. mPFC, medial prefrontal cortex.

Fetal Frontolimbic Connectivity and Toddler Aggression

Greater scores on the aggression scale at 3 years of age were associated with lower limbic–medial prefrontal FC before birth (β = −0.30, R² = 0.43, ΔR² = 0.09, p = .003, 95% CI b [−11.91, −2.49]) (Figure 4). This association was specific to limbic–medial prefrontal FC; neither within-network medial prefrontal FC (β = −0.16, R² = 0.36, ΔR² = 0.02, p = .14, 95% CI b [−5.63, 0.84]) nor within-network limbic FC (β = 0.01, R² = 0.34, ΔR² < 0.01, p = .90, 95% CI b [−4.33, 4.91]) in fetuses was associated with child aggression at 3 years of age.

Figure 4.

Fetal frontolimbic connectivity is associated with child aggression at 3 years of age. We conducted hierarchical linear regressions examining the association between fetal frontolimbic functional connectivity (FC) with child aggression at 3 years of age. All analyses controlled for number of resting-state functional magnetic resonance imaging frames included in the analysis, fetal motion parameters, fetal gestational age at scan, fetal sex, birth weight, gestational age at birth, maternal education, maternal age, partner status, and maternal anxiety and depressive symptoms at 3 years postpartum. (A) FC between the subcortical limbic network and medial prefrontal network before birth prospectively explained 9% of variance in maternal-reported child aggression at 3 years of age above and beyond covariates. Neither (B) within-network connectivity of the medial prefrontal cortex (mPFC) nor (C) within-network connectivity of the subcortical limbic network among fetuses was associated with child aggression after covariate control.

Next, we conducted follow-up analyses to determine whether effects were lateralized. Lower FC between the left subcortical limbic ROI and left mPFC was not associated with greater child aggression at 3 years of age (β = −0.18, R² = 0.37, ΔR² = 0.03, p = .10, 95% CI b [−6.50, 0.57]). Lower right ipsilateral FC between the subcortical limbic region and mPFC continued to be significantly related to greater child aggression (β = −0.23, R² = 0.39, ΔR² = 0.05, p = .03, 95% CI b [−7.21, −0.33]). Lower contralateral limbic-mPFC FC was also associated with greater child aggression (left mPFC to right amygdala: β = −0.24, R² = 0.39, ΔR² = 0.05, p = .03, 95% CI b [−7.51, −0.47]; right mPFC to left amygdala: β = −0.22, R² = 0.38, ΔR² = 0.05, p = .04, 95% CI b [−6.49, −0.15]). No lateralized associations survived Bonferroni correction for 4 comparisons (p < .01).

Specificity of Aggression Findings

We next examined whether fetal bilateral limbic-mPFC FC was uniquely associated with subsequent aggressive symptoms or was instead related to a range of child symptomatology. Toddler language abilities were not associated with fetal limbic-mPFC FC (β = 0.01, R² = 0.11, ΔR² < 0.01, p = .94, 95%CI b [−7.02, 7.58]), and frontolimbic FC continued to relate to child aggression even after adding toddler language ability to the model (β = −0.28, R² = 0.43, ΔR² = 0.07, p = .009, 95% CI b [−11.51, −1.74]). Toddler cognitive and motor development also did not explain our brain-aggression findings (see Supplement). Fetal limbic-mPFC FC was not associated with toddler inattention/hyperactivity symptoms after controlling for the same covariates included in our primary analyses (β = −0.11, R² = 0.29, ΔR² = 0.01, p = .31, 95% CI b [−2.59, 0.84]). However, lower fetal limbic-mPFC FC was associated with greater toddler internalizing symptoms after covariate control (β = −0.27, R² = 0.29, ΔR² = 0.07, p = .02, 95% CI b [−0.77, −0.08]).

Associations Between Maternal Emotion Dysregulation and Child Aggression

We also explored whether maternal emotion dysregulation correlated with child aggression and whether it interacted with fetal neural phenotype to predict child aggression. Maternal internalizing symptoms during pregnancy were not associated with toddler aggression at age 3 years (β = −0.18, 95% CI b [−3.05, 0.47], p = .15) and did not moderate the association between fetal limbic-mPFC FC and toddler aggression (b = −4.36, 95% CI [−9.57, 0.84], p = .10). Mothers with greater internalizing symptoms at the 3-year follow-up reported that their toddler engaged in more aggressive behaviors (β = 0.45, 95% CI b [1.81, 5.36], p < .001), but maternal symptoms at 3 years did not moderate the association between fetal limbic-mPFC FC and child aggression (b = −0.24, 95% CI [−6.16, 5.68], p = .93). Finally, there were no differences in aggression between children whose mothers endorsed being quick to anger versus mothers who denied quickness to anger (β = 0.05, 95% CI b [−3.00, 4.38], p = .71). Maternal quickness to anger did not moderate the association between fetal limbic-mPFC FC and toddler aggression (b = −3.24, 95% CI [−17.24, 10.75], p = .64). These moderation analyses should be considered preliminary given our smaller sample size. None of our maternal emotion dysregulation measures were associated with fetal frontolimbic FC (ps > .69).

Discussion

In a prospective study of 79 mother-child dyads, we found that lower intrinsic functional coupling between medial prefrontal and limbic regions before birth was associated with greater maternal report of aggressive behavior when children reached 3 years of age. This association was specific to between-network coactivation, as neither within-network connectivity of the mPFC nor within-network connectivity of the limbic network was associated with subsequent child aggression. Our results are consistent with extant fMRI studies showing links between aggressive behavior and altered frontolimbic circuitry in childhood (60), adolescence (61,62), and adulthood (63) and extend these findings to demonstrate prospective associations with frontolimbic connections measured before the onset of symptoms and prior to birth.

A critical finding from the present study was that individual differences in frontolimbic circuitry before birth, specifically weaker coupling of the amygdala and mPFC, precede and relate to greater subsequent aggression in toddlerhood. Rather than directly causing more mature emotion and behavior regulation abilities, which is typically reflected in negative functional coupling of limbic and frontal regions at older ages (22), enhanced excitatory inputs are believed to drive the positive coactivation of limbic and mPFC regions beginning in gestation, which is postulated to entrain this circuitry and stimulate development of inhibitory connections from the PFC to limbic areas postnatally (64, 65, 66, 67). Longitudinal research has found that frontolimbic activation in response to negative stimuli at 7 to 12 years of age prospectively predicts stronger intrinsic connectivity within this same circuitry on average 2 years later; this association was strongest when stimulus-elicited activity was measured earlier in development, a time when neural plasticity is more pronounced (66).

An alternative mechanistic explanation of our results may relate to the postnatal caregiving environment. It is possible that fetuses who show stronger positive frontolimbic connectivity in utero exhibit higher levels of behavioral and emotional arousal after birth, which elicits caregiving behaviors that externally regulate the infant (68). The repeated external regulation provided by caregivers early in life may scaffold the postnatal development of inhibitory connections from the mPFC to subcortical limbic regions, ultimately resulting in more effective behavioral control, less intense emotional reactions in early childhood, and fewer aggressive behaviors (69,70). This explanation is consistent with our finding that fetal frontolimbic circuitry also correlated with internalizing symptoms. It is possible that entrainment of frontolimbic circuitry in utero holds broader implications for child emotional reactivity and that heightened reactivity simply manifests most strongly as aggression during this early stage of development.

Heterogeneity in the development of frontal and limbic connections in utero is likely explained by a combination of genetic and environmental factors. MRI conducted with twin neonates suggests that while global measures of brain volume and structural variability in motor and visual regions are highly heritable, variation in the development of frontal regions tends to be more strongly influenced by environmental rather than genetic factors (71). Behavioral genetics studies in 7- to 9-year-old twins also estimate that unique and shared environmental factors are more strongly associated with interindividual variance in amygdala-prefrontal connectivity than shared genetics (72). Although examining neural development in utero accounts for the influence of postnatal environmental inputs, preconception and prenatal exposures likely shape developing frontolimbic circuitry. For example, maternal childhood maltreatment (19), negative affect and stress during pregnancy (31), and maternal inflammation during pregnancy (25) all are correlated with the intrinsic functional coupling of limbic and prefrontal regions in infants. Altered frontolimbic connections in utero may therefore reflect even earlier environmental exposures that have potent programming effects.

Our analyses focused solely on a single frontal network and a subcortical limbic network in the fetal brain, but these are not the only neural regions implicated in the unfolding of aggressive behaviors. Leading models propose that aggression emerges from a combination of circuitry alterations that subserve salience detection (e.g., amygdala, cingulate, and insula), reward processing (e.g., ventral striatum), and cognitive control (e.g., medial and lateral PFC and inferior parietal lobe) (12,14). Indeed, exploratory seed-based connectivity analyses in the present study (see Supplement) revealed that altered limbic and mPFC connections to the cingulate, inferior parietal, and cerebellum also correlated with increased aggression in toddlers. Frontolimbic connections are therefore part of a more complex network that may support individual differences in aggressive behaviors.

Of note, we examined normative variation in aggressive behaviors within a sample of typically developing children, and we lacked information about whether any children exhibited clinically significant levels of aggression that require intervention. Engagement of a typically developing sample rather than a clinical sample is of high interest given that aggressive behaviors in preschool-age children are a common response to frustration and an important aspect of development (73). Nonetheless, there is value in future research examining neural precursors of clinically significant levels of aggression. In addition, the families in this study were predominantly Black mother-child dyads from low- to middle-income households. We view this as a strength of the current study, given that racial and ethnic minority groups are markedly underrepresented in developmental (74) and neuroimaging research (75). Nevertheless, these analyses should be replicated in a population-representative sample to ascertain the generalizability of these results across diverse cultural, racial, and sociodemographic groups.

In vivo fetal fMRI is a powerful, but relatively new tool for parsing neural mechanisms that contribute to individual differences in behavior and clinical risk. The present results must therefore be interpreted with caution and in the context of analytic limitations. First, fetal fMRI data are contaminated by large fetal movements, resulting in significant data loss; it is common to discard 35% to 56% of rsfMRI frames due to fetal motion (36,76). In the present study, some fetuses had as few as 100 frames of rsfMRI data. Due to concerns that FC estimates are less reliable when examining small amounts of data (77), we repeated our analyses only in fetuses who had relatively greater amounts of data, and our results held (see Supplement). Nonetheless, our findings are based on a relatively small amount of rsfMRI data per fetus, so it will be imperative to assess whether these findings replicate in studies that are able to collect longer fetal rsfMRI scans. We also used a more liberal motion threshold than what is typically applied in adult imaging to maximize data retention. For this reason, we used a stringent ICA denoising strategy that was effective at minimizing motion-related artifacts in the present data (39). Emerging advanced motion correction approaches, such as slice-to-volume registration, may further improve our ability to reduce noise and enhance signal in fetal fMRI (78). Outside of motion-related challenges, fetal fMRI preprocessing presently requires a number of manual steps (e.g., selection of low-motion segments based on visual inspection, manual brain masking) that hold potential to introduce unmeasured bias. Several research groups are making directed efforts to automate and quantify the impact of different preprocessing steps (39,40,76), which will continue to enhance the rigor of fetal fMRI.

Fetuses were scanned during a fairly wide age range (i.e., between 26 and 39 weeks of gestation). Although histological, diffusion tensor imaging, and fMRI studies in the fetal brain suggest that limbic to forebrain connections are in place by 26 weeks (28), the brain undergoes significant growth and refinement of connections across this period of development (79). To facilitate group comparisons, we elected to normalize all fetal brains to a single template representing the average age of our sample. However, we cannot be certain that all spatially aligned areas are functionally identical across fetuses in this wide age range. ICA was used to define our ROIs based on underlying fMRI signal, which helps to ameliorate, but does not entirely address, this concern. Related to our use of ICA, our measure of within-network connectivity measures only cross-hemisphere connectivity of homologous regions and does not measure within-hemisphere connections within a given network. Additional research is needed to determine whether within-network associations with child aggression emerge when alternative methods are used to define fetal networks (80).

Aggression is a complex behavioral phenotype that is shaped by both environmental and biological factors. In the present study, we found that individual differences in fetal frontolimbic circuitry is one such factor that prospectively correlates with the emergence of aggressive behaviors. We postulate that stronger coactivation of frontolimbic regions during gestation may entrain this circuitry and stimulate the postnatal development of inhibitory connections from the PFC to limbic structures, which subsequently leads to greater emotional regulation in toddlerhood. Aggression was examined in the present study given its early developmental onset, ability to be measured reliably at young ages and in diverse samples, and public health significance. Yet we saw similar associations between fetal neurocircuitry and toddler internalizing symptoms. Stronger positive frontolimbic coupling before birth may thus be a neural phenotype that is associated with emotion dysregulation more broadly or even with a general psychopathology factor (81). In utero neural phenotypes are predisposing rather than prescriptive and are only the beginning of complex trajectories that subserve child health and wellness. Nonetheless, isolating neural beginnings that relate to the manifestation of human behavior is a necessary step in our efforts to understand the biological etiology of psychiatric illness.