Caffeine exposure in utero is associated with structural brain alterations and deleterious neurocognitive outcomes in 9-10 year old children

Abstract

Caffeine, a very widely used and potent neuromodulator, easily crosses the placental barrier, but relatively little is known about the long-term impact of gestational caffeine exposure (GCE) on neurodevelopment. Here, we leverage magnetic resonance imaging (MRI) data, collected from a very large sample of 9157 children, aged 9–10 years, as part of the Adolescent Brain and Cognitive Development^sm (ABCD ®) study, to investigate brain structural outcomes at 27 major fiber tracts as a function of GCE. Significant relationships between GCE and fractional anisotropy (FA) measures in the inferior fronto-occipito fasciculus and corticospinal tract of the left hemisphere (IFOF-LH; CST-LH) were detected via mixed effects binomial regression. We further investigated the interaction between these fiber tracts, GCE, cognitive measures (working memory, task efficiency), and psychopathology measures (externalization, internalization, somatization, and neurodevelopment). GCE was associated with poorer outcomes on all measures of psychopathology but had negligible effect on cognitive measures. Higher FA values in both fiber tracts were associated with decreased neurodevelopmental problems and improved performance on both cognitive tasks. We also identified a decreased association between FA in the CST-LH and task efficiency in the GCE group. These findings suggest that GCE can lead to future neurodevelopmental complications and that this occurs, in part, through alteration of the microstructure of critical fiber tracts such as the IFOF-LH and CST-LH. These data suggest that current guidelines regarding limiting caffeine intake during pregnancy may require some recalibration.

1. Introduction

Worldwide, caffeine is one of the most broadly used neuromodulators, acting on numerous neurotransmitter systems ^1–3. Its immediate use has been implicated in acute alterations to cognitive domains such as working memory ^4,5, reaction time ⁶, mood ^47,8 and attention ⁷. Alterations in attention-related alpha-band neuro-oscillatory activity ^9,10, cerebral blood flow ^11,12, functional connectivity ¹³, and global brain entropy ¹⁴ have also been associated with caffeine. The vast majority of caffeine research has been focused on understanding the acute effects of its consumption on ongoing neurocognitive processes. However, caffeine has demonstrated effects on cognitive aging ^15,16 and brain metabolism ¹⁷, with long-term consumption demonstrating differences in performance on the mini mental state exam ^18,19, word recall ²⁰, and long-term memory ²¹. Despite some negative outcomes related to mood, the majority of these studies have identified improved cognitive processing in relation to caffeine consumption.

Most of the work concerning the effects of perinatal caffeine exposure on neurodevelopment has been related to prematurely born infants. Caffeine and other methylxanthines are used to support the poorly developed respiratory system of premature infants, primarily through their actions as adenosine receptor antagonists ²². Adenosine inhibits a cascade of metabolic events resulting in bronchodilation, enhanced diaphragmatic contractility, and stimulation of the respiratory center in the brainstem ²³. Although the brainstem is the intended therapeutic site of action within the central nervous system, adenosine receptors are found extensively throughout the brain ²⁴, raising the possibility that acute caffeine exposure might affect neurodevelopmental processes. Indeed, differential outcomes in brain microstructure and cognition have been identified in those treated with methylxanthines ^25,26. For example, adenosine interferes with γ-aminobutyric acid (GABA) receptors ²⁷, which plays a role in early cortical development ²⁸. Although these findings strongly suggest that perinatal caffeine exposure improve neurodevelopment, they may be confounded preexisting conditions (i.e. those born prematurely) and only address exposure after many crucial neurodevelopmental events have occurred throughout fetal development ^29–31. The neurodevelopmental implications of caffeine exposure on a healthy developing fetus are largely unaddressed by these findings.

Fetal exposure to caffeine throughout gestation is regulated by passive diffusion through the placenta ³². Both the fetus and placenta lack the necessary enzyme for metabolizing caffeine (cytochrome P450 1A2; ³³). Clearance of caffeine from the fetus is completely dependent on the mother’s caffeine metabolic capacity, which is decreased throughout pregnancy ³⁴. Therefore, even typical levels of caffeine consumption result in increased maternal blood caffeine concentrations during pregnancy ³⁵, potentially placing the fetus at increased risk of toxicity. However, maternal consumption of caffeine during gestation has been inconsistently associated with adverse events that relate to neurodevelopment such as premature birth, low birth weight, or congenital defects ³⁶. Early motor and cognitive development have generally been shown to be unaffected by caffeine exposure ³⁷. Previous investigations have, however, found an association between in utero caffeine exposure and measures of inattention and hyperactivity ³⁸ and poor social outcomes ³⁹, although when socioeconomic measures are accounted for, this relationship is typically mediated or eliminated ⁴⁰.

Although current support in the literature for the association between GCE and human fetal brain development is limited, animal models of GCE consistently demonstrate altered brain development. Mice exposed to caffeine in utero demonstrate alterations in GABAergic neuronal networks in the visual cortex ⁴¹, hippocampus ⁴², and delayed differentiation of glutamatergic neurons ^43,44. These physiological changes are likely related to altered performance on cognitive task (e.g., object recognition, memory) in mice exposed to caffeine throughout gestation ⁴⁵. Clearly GCE alters the development of individual neurons and circuits early in the developing mouse brain. However, these alterations are demonstrated in multiple networks throughout the brain and it remains to be seen which systems in early neurodevelopment are predominantly influenced by GCE.

Despite caffeine’s known neuromodulating effects postnatally and elevated blood caffeine levels throughout pregnancy, there is a lack of consistent evidence demonstrating its effects on neurodevelopment into adolescence. Previous studies attempted to explore the relationship between fetal GCE and neurodevelopment into adolescence purely through social and behavioral measures ^38–40. In contrast, the present study identifies plausible biomarkers of neurocognitive development and their association between gestational caffeine exposure (GCE), cognition, and psychopathology. More specifically, we test the following hypotheses: 1) whether regional brain microstructure is altered as a result of GCE and 2) if neurocognitive changes associated with GCE are due to alterations in regional brain microstructure. We tested these hypotheses in a sample of 9157 children by: 1) identifying brain regions whose changes in fractional anisotropy (FA) were predictive of GCE and 2) modeling the combined effects of GCE and microstructural changes in highly predictive regions with neurocognitive outcomes.

2. Materials and methods

2.1. Subjects

The sample used was obtained from the Adolescent Brain and Cognitive Development^sm (ABCD®) study, a longitudinal study designed to comprehensively track neurodevelopmental outcomes in a demographically representative sample of almost 12,000 children from 22 study sites across the United States; the University of Rochester is one of the ABCD study sites. Initial enrollment began when these children were aged 9–10 years, with the goal of following this cohort for a full decade. Recruitment and study design considered age, sex, and socioeconomic factors to better reflect US sociodemographic levels ⁴⁶. Details concerning demographic factors are described elsewhere ⁴⁷. Data were acquired through the National Institute of Mental Health Data Archive (Data Release 2.0.1, http://dx.doi.org/10.15154/1504041). The ABCD study protocol excludes individuals that would not be able to complete all of the administered tasks. This excluded individuals that lacked fluency in English, uncorrectable sensory deficits (e.g., legal blindness), diagnosis of a psychotic disorder, moderate to severe autism spectrum disorder, intellectual disability, or neurological issues (e.g., brain tumor, previous head injury with loss of consciousness > 30 minutes), or significant perinatal medical issues (e.g., gestational age < 28 weeks, birth-weight <1.2 kg, complications resulting in >1 month of hospitalization following birth). Participants also had to be able to complete an MRI scan.

Participants and their families were recruited through school- and community- settings across the 22 ABCD centers, following locally and centrally approved In/stitutional Review Board procedures. Most ABCD research sites rely on a central Institutional Review Board (cIRB) at the University of California, San Diego, which is the case for the University of Rochester ABCD site represented in the current report. The UCSD cIRB approved all research protocols reported herein. Written informed consent was obtained from parents, and written assent obtained from children. All data are strictly de-identified to protect participant anonymity.

A subset of the initial ABCD sample was used for the present study based upon availability and quality of key measures. These included caffeine intake throughout gestation, smoking throughout gestation, diffusion weighted imaging data (DWI). Individuals whose DWI data did not pass initial quality control were also excluded. Caffeine exposure throughout gestation was determined using parental reports of caffeine ingestion throughout pregnancy. The original response scale consisted of “At least once a day”, “Less than once a day but more than once a week”, “Less than once a week”, and “Never”. Subjects were placed in a GCE group if parents reported caffeine consumption at least weekly and all other subjects were placed in a minimal-gestational caffeine exposure (mGCE) group. This dichotomization was chosen because most evidence supports neurodevelopmental alterations given regular exposure to caffeine ^41–45 and the exposure to caffeine in those reporting less than once a week would be highly heterogenous and irregular. A summary of demographic measurements across both groups after exclusion criteria is included in Table 1.

Table 1.

Demographics for GCE and mGCE groups.

	GCE N (%)	mGCE N (%)	χ2 (P(>χ2))
N = 9157	4135 (78.8)	5022 (21.2)
Sex			1.21 (0.272)
Female	2004 (0.48)	2376 (0.47)
Male	2131 (0.52)	2646 (0.53)
Age (years)			4.63 (0.031)
>= 9 & < 10	2063 (0.50)	2619 (0.52)
>= 10 & < 11	2072 (0.50)	2403 (0.48)
Education			26.16 (2.93e-5)
< HS Diploma	196 (0.05)	317 (0.06)
HS Diploma/GED	392 (0.09)	491 (0.10)
Some College	1290 (0.31)	1385 (0.28)
Bachelor	1156 (0.28)	1533 (0.31)
Post graduate degree	1101 (0.27)	1296 (0.26)
Income			6.60 (0.037)
< 50 K	778 (0.19)	1045 (0.21)
>= 50 K & < 100 K	1132 (0.27)	1294 (0.26)
>= 100 K	2225 (0.54)	2683 (0.53)
Race/Ethnicity			160.70 (1.29e-34)
Asian	50 (0.01)	116 (0.02)
Black	393 (0.10)	744 (0.15)
Hispanic	687 (0.17)	1140 (0.23)
White	2559 (0.62)	2540 (0.51)
Smoking			178.70 (9.33e-41)
None	3834 (0.93)	4939 (0.98)
Yes	301 (0.07)	83 (0.2)
Premature (weeks)			13.30 (0.065)
0	3311 (0.80)	4090 (0.99)
>=1 & <2	80 (0.02)	131 (0.03)
>=3 & <4	351 (0.08)	375 (0.09)
>=5 & <6	211 (0.05)	235 (0.06)
>=7 & <8	116 (0.03)	127 (0.01)
>=9 & <10	35 (0.01)	36 (0.01)
>=11 & <12	11 (0.00)	7 (0.00)
12	11 (0.00)	21 (0.01)

Group differences were computed using a χ² test.

Note that “<” indicates those less than and not including the provided number (e.g., “>= 10 & < 11” means those 10 years of age and up to but not including those 11 years of age)

2.2. Cognitive and psychopathology measures

We identified two cognitive measures from the NIH Toolbox, the Toolbox Pattern Comparison Processing Speed Test and the Toolbox List Sorting Working Memory Test. Higher scores on each measure indicates superior performance. These tests were used as proxies of task efficiency (e.g., accurate responses with a reaction time incorporated) and working memory, respectively ^48–50. Task efficiency and working memory were chosen because effects of caffeine consumption on them have been consistently reported in prior work ^4–6. In so doing we were able to better represent the much larger body of neurocognitive and caffeine related research that does not specifically address neurodevelopment.

The child behavior checklist (CBCL) consists of eight syndrome scales, two composite scales, and six DSM-oriented scales ⁵¹. Previous work in this population derived four higher order measures of the CBCL through factor analysis (externalizing, internalizing, neurodevelopment, and somatization) ⁵². Higher scores on each of these measures indicated greater problematic behaviors within each domain. We utilized these four factors to address the variety of social and behavioral alterations previously associated with GCE ^38–40.

Visualization for these measures is provided in Figure 1.

Figure 1. Neurocognitive Measures Correlation Plot.

Lower left portion shows scatter plots and fitted line for each pair of neurocognitive measures. Along the diagonal is a histogram for each measure. The upper right portion displays the numeric correlation between each pair of measures.

2.3. Imaging measures

Images were acquired across 22 participating sites on Siemens, Phillips, or GE scanner 3 Tesla scanners. Acquisition protocols, inter-site reliability, and post-processing is described in detail elsewhere ^53,54. The present study focused on mean DWI derived measurements of major fiber tract pathways. Tracings of fiber tract pathways were acquired through semi-automated segmentations of white matter tracts, using AtlasTrack ⁵⁵. The original probabilistic map provided by AtlasTrack included 23 fiber tracts. We chose to include additional fiber tracts recently made available (e.g., tracts from the striatal inferior frontal) ⁵⁴, testing a total of 27 regions of interest (ROI). ROIs for each hemisphere were evaluated separately because hemispheric lateralization plays a role in cognitive functions such as language and visual processing ⁵⁶ and has been implicated in developmental disorders such as autism spectrum disorder ^57,58 and attention deficit disorder ⁵⁹. We specifically focused on FA values because it is a widely used measurement for characterizing fiber tracts ⁶⁰ and because adenosine plays a role in axon development ⁶¹.

2.4. Demographic measures

In addition to those demographic variables previously identified as critical in large multi-site studies and cognition (age, sex, ethnicity, education) ^62–64, we included prematurity of birth and whether the mother smoked during pregnancy due to their previously established association with gestational caffeine exposure ^25,35. Children exposed to smoking throughout gestation have demonstrated altered brain activity on inhibition tasks ⁶⁵ and fine motor skills ⁶⁶. Premature birth is associated with alterations in white matter and increased problematic behaviors ⁶⁷.

2.5. Statistical Analysis

Figure 2. depicts the causal diagrams motivating our analysis. The first two diagrams are based on relationships between GCE, demographics, brain development and cognition, which are already well established. The third diagram represents GCE as a potential modulator of the relationship between brain development and cognition, which allows composition of the final diagram and a plausible causal pathway from GCE to cognition. We first focused on identifying those fiber tracts having a significant relationship with GCE.

Figure 2. Causal diagrams depicting the relationship between demographics, brain development, GCE, and cognition.

1. There is a causal relationship between demographics and cognition, demographics and brain development, and brain development and cognition.

2. Demographics has a causal relationship with GCE and cognition.

3. GCE moderates the relationship between brain development and cognition

4a. Combining diagrams 1–3 results in three causal paths:

1 – Pathway A (black): various demographic measures that influence cognition

2 – Pathway D -> E -> C (blue): the main effect of GCE on caffeine

3 – Pathway B -> C (orange): the main effect of brain development on caffeine through brain a pathway from GCE to cognition. Cognition and GCE are conditionally independent, conditioning on brain development and demographics.

4b. Alternate representation including the interacting effect between GCE and brain development on cognition

Identification of major tracts affected by GCE was accomplished by performing a binomial linear mixed effects model where the response was GCE or mGCE grouping and the fixed effect was a centered and scaled mean FA value. This same model was repeated for each major fiber tract with alpha set to 0.05 and corrected for multiple comparison via false discovery rate ⁶⁸. Fibers found to have a significant relationship with GCE were parameterized in an analysis of covariance (ANCOVA) with GCE for each neurocognitive outcome. Significance of each effect was determined by comparing to a comparative model with the effect removed via a likelihood ratio test (LRT).

Analyses were performed using the Julia software MixedModels package ⁶⁹. The relationship between demographics and cognition was modeled using random effects measures. We used a recommended method of iterative model reduction to avoid over paramterization ⁷⁰. Each iteration removed the variable with the lowest contribution to the model variance selecting the final model based on the lowest Akaike Information Criterion (AIC) ⁷¹. This process was performed for each neurocognitive measure to find the optimal null model for comparison, resulting in the following outcome measure, variable pairs:

• Externalizing: Collection site, Sex, Race/Ethnicity, Education, Income, Smoking

• Internalizing: Collection site, Race/Ethnicity, Education, Income, Smoking

• Neurodevelopment: Collection site, Sex, Race/Ethnicity, Education, Income, Smoking

• Somatization : Collection site, Race/Ethnicity, Education, Income, Smoking

• Task efficiency: Collection site, Sex, Race/Ethnicity, Education, Income, Age

• Working memory: Collection site, Sex, Premature birth, Race/Ethnicity, Education, Income, Age

3. Results

Linear mixed effect models identified decreased mean FA in the left corticospinal tract (CST-LH) and the inferior fronto-occipital fasciculus of the left hemisphere (IFOF-LH) to be predictive of GCE (CST-LH: Z = −2.0359, p = 0.0436; IFOF-LH: Z = −2.967, p = 0.003). No other tracts were significant predictors of GCE (see Table 2). In both fiber tracts lower mean FA was associated with a greater probability of being in the GCE group and a higher mean FA was associated with being in the mGCE group.

Table 2.

Bivariate mixed effects model of GCE on mean FA values at each ROI

ROI	Effect	Std. Error	Z score	P(>|z|)	P(>χ2)
Fornix-Right	−0.012	0.007	−1.622	0.105	0.108
Fornix-Left	−0.0131	0.0072	−1.8249	0.068	0.0706
Cingulate Cingulum-Right	−0.005	0.006	−0.897	0.370	0.372
Cingulate Cingulum-Left	−0.0102	0.0064	−1.578	0.115	0.1166
Hippocampal Cingulum-Right	−0.007	0.007	−1.026	0.305	0.312
Hippocampal Cingulum-Left	−0.0012	0.0072	−0.1625	0.871	0.8726
Corticospinal Tract-Right	−0.007	0.007	−1.09	0.276	0.279
Corticospinal Tract-Left	−0.0136	0.0067	−2.0359	0.042	0.0436
Anterior Thalamic Radiations-Right	−0.005	0.007	−0.658	0.510	0.517
Anterior Thalamic Radiations-Left	−0.0022	0.0074	−0.3037	0.761	0.7646
Uncinate-Right	−0.014	0.008	−1.772	0.076	0.083
Uncinate-Left	−0.0135	0.0076	−1.7694	0.077	0.0838
Inferior longitudinal fasciculus-Right	−0.007	0.007	−1.062	0.288	0.292
Inferior longitudinal fasciculus-Left	−0.0064	0.0066	−0.9738	0.330	0.3337
Inferior fronto-orbital tract-Right	−0.009	0.007	−1.28	0.200	0.205
Inferior fronto-orbital tract-Left	−0.021	0.0071	−2.9669	0.003	0.0034
Forceps Major	−0.008	0.006	−1.269	0.204	0.206
Forceps Minor	−0.0073	0.0077	−0.9412	0.347	0.3514
Corpus Callosum	−0.014	0.007	−1.838	0.066	0.068
Superior Longitudinal Fasciculus-Right	−0.0042	0.0066	−0.6393	0.523	0.5257
Superior Longitudinal Fasciculus-Left	−0.005	0.006	−0.82	0.412	0.414
Superior Corticostriate-Right	−0.0076	0.0066	−1.1496	0.250	0.2536
Superior Corticostriate-Left	−0.007	0.006	−1.129	0.259	0.261
Striatal Inferior Frontal Cortex-Right	−0.0063	0.0071	−0.8793	0.379	0.3818
Striatal Inferior Frontal Cortex-Left	−0.013	0.007	−1.834	0.067	0.069
Inferior Frontal to Superior Frontal Cortex-Right	−0.0062	0.0069	−0.8994	0.368	0.3726
Inferior Frontal to Superior Frontal Cortex-Left	−0.006	0.007	−0.89	0.373	0.376

P(>|z|) – significance of ROI term

P(>Chisq) – significance of likelihood ratio test between model with ROI term compared to null model

p-values <0.05 in bold

FA in the CST-LH was positively correlated with cognitive measures and negatively correlated with psychopathology measures. This relationship was significant when comparing a model with the FA term to null models via LRT (working memory, p < 0.001; task efficiency, p < 0.001; externalizing, p = 0.0075; internalizing, p = 0.0176; somatization p = 0.0176; neurodevelopment, p < 0.001; see Figure 3 and Table 3 column B). Similarly, FA in the IFOF-LH was positively correlated with cognitive measures and negatively correlated with psychopathology measures. However, LRTs found this relationship was only significant in working memory, task efficiency, and neurodevelopment (working memory, p < 0.001; task efficiency, p < 0.001; externalizing, p = 0.1441; internalizing, p = 0.0830; somatization p = 0.2215; neurodevelopment, p < 0.001; see Figure 3 and Table 3 column B). Those in the GCE group had higher scores across all psychopathology measures than those in the mGCE group (externalizing, p = 2.3452e-7; internalizing, p = 3.2935e-8; somatization, p = 0.0063; neurodevelopment, p = 5.7835e-6). This relationship was not found between GCE and cognitive measures (working memory, p = 0.2393; task efficiency, p = 0.3119). The only significant interaction detected was between FA in the CST-LH and GCE when modeling task efficiency (p = 0.0245), where the relationship between task efficiency and the CST-LH was weaker in the GCE group.

Figure 3.

Along the y-axis are values for each neurocognitive measure. Along the x-axis is the scaled and centered mean FA of the CST-LH and IFOF-LH. Working memory, Task efficiency: increases in FA correlated to increased scores on both measures, with the exception of the relationship between task efficiency and FA in the CST-LH for those in the GCE group with with much smaller correlation. Externalizing, Internalizing, Somatization, Neurodevelopment: Those in the GCE group consistently had worse outcomes than those in the mGCE group. Note that the apparent positive relationship between FA and measures of internalizing and somatization are absent when accounting for collection site. This difference is reflected in effect measures derived from our final model and presented in Table 3.

Table 3.

Likelihood ratio tests

	CST-LH				IFOF-LH
	Effect	Std. Error	χ2	P(>χ2)	Effect	Std. Error	χ2	P(>χ2)
Working Memory
GCE	−0.0116	0.0101	1.38457	0.2393	−0.0107	0.0101	1.38457	0.2393
FA	0.05236	0.0121	22.0926	6.2400e-4	0.0704	0.0170	30.7372	9.6555e-5
Interaction	−0.01605	0.0100	3.8766	0.1233	−0.0149087	0.0100	3.3413	0.1881
Task Efficiency
GCE	−0.0089	0.0106	0.952696	0.3120	−0.0078	0.0106	0.952696	0.3120
FA	0.0689	0.0137	26.5171	2.3404e-4	0.0922	0.0172	38.3139	7.5649e-5
Interaction	−0.0233	0.0104	5.0619	0.0245	−0.0131	0.0104	2.6349	0.1045
Externalizing
GCE	0.0578	0.0106	30.4893	2.3452e-7	0.0576	0.0106	30.4893	2.3452e-7
FA	−0.0367	−0.0135	8.3095	0.00746	−0.0336	0.0141	6.66678	0.14414
Interaction	0.0094	0.0103	1.0154	0.3136	0.0100	0.0104	0.9825	0.3216
Internalizing
GCE	0.0646	0.0106	37.7547	3.2935e-8	0.0646	0.0106	37.7547	3.2935e-8
FA	−0.0332	0.0136	3.75291	0.0176	−0.0238	0.0142	30.7812	0.08295
Interaction	0.0159	0.0103	2.5076	0.1133	0.0148	0.0104	2.0213	0.1551
Somatization
GCE	0.0339	0.0106	10.4682	0.0063	0.0340	0.0106	10.4682	0.0063
FA	−0.0337	0.0135	6.83694	0.0176	−0.0170	0.0142	1.87568	0.2215
Interaction	0.0126	0.0104	1.5726	0.2098	0.0139	0.0104	1.7420	0.1869
Neurodevelopment
GCE	0.0504	0.0105	23.8137	5.7835e-6	.0495	0.0105	23.8137	5.7835e-6
FA	−0.0616	0.0136	21.8763	1.4758e-5	−0.0774	0.0143	30.7812	3.8778e-5
Interaction	0.0106	0.0103	1.2547	0.2627	0.0138	0.0103	1.9393	0.1637

Each model was compared to the null model (predicted the cognitive measure using the optimal set of potentially confounding variables) using a likelihood ratio test. All models being compared included the same potentially confounding variables included in the null model. The following models were compared:

- FA - Null: main effect of mean FA at a given tract

- GCE - Null: main effect of GCE

- Interaction -

(p-values <0.05 are in bold)

4. Discussion

This work constitutes the largest multi-site study to date of neurocognitive outcomes in adolescents as a function of GCE. This is also the first population study to investigate the potential influence of GCE on white matter microstructure that may be facilitating neurocognitive outcomes in adolescence, and the first analysis of the ABCD study focused specifically on GCE. We identified decreased mean FA values in the IFOF-LH and CST-LH as significant predictors of GCE. We subsequently identified significant relationships between six neurocognitive outcomes as a function of GCE or mean FA in the CST-LH and IFOF-LH. We also identified a significant interaction between the CST-LH and GCE on task efficiency scores, implicating a causal relationship between GCE and task efficiency scores taken nearly a decade after gestational exposure.

4.1. Cognitive measures

We found that working memory improved with higher FA values in the IFOF-LH. Previous investigation revealed alterations in alpha wave activity in relation to measures of task efficiency ⁹. Given the association of attention with alpha-band oscillations in occipito-temporal region ⁷², the IFOF-LH may represent a component of the attention network that is highly susceptible to caffeine. Additionally, fMRI-based analyses have identified portions of the frontal lobe to be susceptible to caffeine’s influences ^4,12. Although this relationship was also found in the CST-LH, it has historically been unrelated to measures of working memory ^73–75. Additionally, differential measures from the CST depending on subanatomical segmentation ⁷⁶ and complications due to fanning fibers in association regions near the CST further lead us to believe our finding concerning the CST-LH and working memory is representative of a more complicated global relationship than an isolated effect.

Measures of task efficiency also demonstrated a strong correlation with mean FA in the CST-LH and IFOF-LH. Although GCE did not appear to affect mean scores of task efficiency, the relationship between the CST-LH and task efficiency was nearly absent in the GCE group (see Figure 3), in contrast to a strong positive correlation between FA in the CST-LH and task efficiency in the mGCE group. This relationship would be consistent with the causal diagram proposed in Figure 2. where the effects of GCE can be detected through GCE moderating brain development in the CST-LH. Given that the measure we are using for task efficiency was derived from timed responses, our measure may be acting as a proxy for reaction time. This relationship is not without precedence, as previous literature has used this same measure from the NIH Toolbox as a component to create a composite score of reaction time ⁴⁸. Furthermore, this relationship would be consistent with previous literature associating the CST with reaction time ⁷⁷.

Previous literature has demonstrated that chronic caffeine consumption increases certain aspects of cognitive function in adulthood ⁷⁸. A recent investigation of the ABCD cohort indicated that chronic caffeine ingestion in adolescents decreased performance on measures of general intelligence, including those studied here ⁷⁹. The current investigation more compliments this body of literature by accounting for fetal exposure to caffeine. However, we did not find the addition of GCE alone was associated with lower scores on cognitive measures. A notable difference between our study design and the aforementioned is our unique inclusion of brain imaging biomarkers. Given our identification of an interaction between GCE and mean FA of the CST-LH when modeling task efficiency, differences between these studies may be related to differences in neurophysiological responses to caffeine throughout human lifespan.

4.2. Psychopathology measures

A notable advantage of the current study is the use of previous work that created higher order dimensional measurements of psychopathology derived from the CBCL⁵², as opposed to the summation of a small number of items composing subscales or discrete diagnostic measures as in previous studies ³⁸. Although this makes it more difficult to identify distinct outcomes associated with GCE, it likely facilitated greater association with individuals that exhibit related pathology. For example, previous studies struggled to identify strong associations between GCE and hyperactivity^38,80. The neurodevelopment measure used in the present study represented measures related to inattention, hyperactivity, and day-dreaming. Given increasing evidence that mental disorders have a complex hierarchical and dimensional relationship to one another ⁸¹, it is likely that our measure of neurodevelopment captured more individuals with behaviors related to hyperactivity than a single diagnostic or subscale would.

All measures of psychopathology were elevated in the GCE group and along with a relatively strong relationship between neurodevelopment and FA in both tracts. However, the relationship between psychopathology measures and both fiber tracts was unaffected by GCE. In the absence of a biological measure whose relationship with psychopathology is modulated by GCE, it is not clear what the relationship between GCE and elevated psychopathology is. This relationship could be entirely due to conditioning on demographic variables that act as a fork between GCE and neurocognitive outcomes, creating a spurious correlation (see Figure 2 diagram 2). It is also possible that this relationship is is due to GCE modulating neurodevelopment through some other region or mechanism unrelated to FA.

4.3. Brain function and structure

To our knowledge the only investigations identifying the affects of GCE on specific regions of the brain throughout neurodevelopment have been in animal models. The focus of these investigations has been predominantly on cortical structures, therefore our investigation of fiber tracts could not directly reproduce them. However, FA measures are often used as a proxy for fiber tract integrity where higher values are assumed to correlate with improved white matter health. This interpretation is a heuristic and may fail to capture other meaningful differences in white matter microarchitecture (e.g., crossing fibers) ⁸². For example, neurons that have been demyelinated due to mild head injuries may still have an oligodendrocyts capable of remyelination. These neurons may produce redundant, loosely bound layers of myelin along with relatively normal diffusion measures ⁸³. Therefore, common heuristics used to interpret FA values may fail to identify fibers that are compromised microstructurally. This trend would be consistent with our finding between GCE and the CST-LH. That is, GCE may cause damage that is mild enough such that by, the time adolescents were scanned almost a decade later, remyelination had already occurred, resulting in microstructural differences that were indistinguishable through measures of FA but nonetheless affected task efficiency.

However, our identification of the IFOF-LH may support previous literature asserting a change in cortical networks of the visual cortex as a result of GCE ⁴¹ with the caveat that the effects of GCE are not sufficient to induce changes to phenotype. It is also possible that FA is an Given previous studies identifying

4.4. Measures of caffeine exposure

There are multiple guidelines for how much caffeine can safely be consumed throughout pregnancy, but most sources agree that no more than 300 mg caffeine/day should be consumed ^84,85. Most mothers who do consume caffeine prior to pregnancy consume 129.9 mg/day and report decreased caffeine consumption throughout pregnancy ⁸⁶. However, caffeine is still commonly consumed in some form throughout pregnancy ⁸⁶. The current study relied on individual reports of how frequently individuals consumed caffeine 10 years after pregnancy, potentially introducing response bias and failure to report.

Additionally, reports did not indicate the amount of caffeine ingested or many factors that may have altered caffeine metabolism at the time of pregnancy (e.g., body weight, medications, diet, etc.) ⁸⁷. Therefore, the exact dose of caffeine exposure throughout gestation is highly uncertain given the measures used in the current investigation. This differs from many previous studies investigating the effects of caffeine on biological measures of brain development, in that most have used more precise measures of dosage GCE and investigated their affects over a much shorter period of time. Despite this imprecision and 10 years of other potentially confounding events, we still identified differential neurocognitive outcomes as a result of GCE.

4.5. Limitations

The current study focused on neurocognitive measures that did not isolate a particular pathology or cognitive domain. For example, the Pattern Comparison Processing Speed Test was used as an approximate measure of task efficiency and has been shown to correspond to several other cognitive domains ⁶³. Although this approach allowed us to concisely explore a large spectrum of potential neurocognitive interactions, it complicates interpretation and direct application of the findings. Similarly, we chose to focus on a single biomarker (mean FA) instead of exploring all available measures. Given previous findings that demonstrate adenosine receptors in limbic structures effect neurodevelopment ^42,43 and significant findings between all measures of psychopathology and GCE in the present study, ussing a more sensitive biomarker in future studies may be able to identify a causal pathway between GCE and psychopathology.

Our analysis utilized mixed linear models which have proven to be a robust approach to preventing Type 1 error and increasing generalizability of finding ^88,89. This approach also allowed us to account for misleading effects due variables such as collection site without overfitting (see Figure 3). However, this approach partitions variance across mixed effects and there are multiple approaches for calculating the standardized effects for mixed models. No one approach to calculating standardized effect sizes for mixed linear models is agreed upon and the interpretation is not necessarily clear ⁹⁰. Although we attempted to adhere to guidelines for reporting unstandardized effect sizes ⁹¹, it remains difficult to quantify the effects of individual model covariates. Therefore, we reported significance of individual terms based on comparison to models that isolated the contribution GCE and regional FA both separately and collectively.

Another potential concern with our analysis is the choice to iteratively refine the random effects parameters, meaning not all eight potentially confounding variables were included in every model. However, we followed an established and systematic method for determining the final variables included in the model ⁷⁰. For example, models of task efficiency did not include smoking as a random effect, but this was only done after finding that smoking accounted for less than 0.1% of the total variance among random effects and ensuring that models excluding smoking had a lower AIC.

It is important to acknowledge that interpretation of our findings is dependent on what is currently known concerning GCE (as summarized in the causal diagrams presented in Figure 2). Newly identified causal pathways linking GCE and cognition will provide opportunities in the future to refine our analyses.

5. Conclusion

With an unprecedented sample size, we present evidence that GCE alters the developmental trajectory of white matter and neurocognition into adolescence. We demonstrate that the IFOF-LH and CST-LH strongly correlate with measures related to task efficiency and working memory, finding GCE to consistently associate with higher measures of psychopathology (externalizing behaviors, internalizing behaviors, somatization, and neurodevelopmental problems). We also found that FA measures in the CST-LH tract had a weakly negative correlation with task efficiency, opposed to findings in the mGCE group. As a whole, these findings demonstrate that previously inconsistent social and behavioral findings associated with GCE may be more evident when investigated at subclinical levels and in association with relevant biomarkers. Therefore, further investigation should be pursued to validate the effects of GCE on neurocognitive outcomes. Our unique findings between mean FA values in the IFOF-LH and measures of neurodevelopmental problems suggest a greater need to characterize the quality of the relationship between white matter microstructure and neural health.

Highlights.

• Caffeine is a known neuromodulator that is commonly consumed throughout pregnancy.

• 4,135 mothers reported consuming caffeine more than once a week throughout gestation in a sample of 9,157.

• Decreased fractional anisotropy in the left IFOF and CST fiber tracts were associated with gestational caffeine exposure.

• Behavioral measures associated with externalizing, internalizing, somatization, and neurodevelopmental disorders were all elevated in those with gestational caffeine exposure

Abbreviations:

GCE

gestational caffeine exposure

mGCE

minimal gestational caffeine exposure