Study population

The COPSAC2010 cohort includes 700 mother–child pairs with extensive phenotyping from 15 clinical visits and exposure assessments beginning in pregnancy [30]. The COpenhagen Prospective Study on Neuro-PSYCHiatric Development study (COPSYCH) was nested within COPSAC2010 at age 10 years [31].

Dietary exposures

Our main exposures included dietary pattern metabolite scores, assessed from global blood metabolomics profiles taken at 24 weeks gestation (HD4 platform, Metabolon (NC, USA)), derived from sparse partial least squares (sPLS) modelling, and breastfeeding duration (length of breastfeeding in days). Dietary pattern metabolite scores were modelled on dietary patterns during pregnancy, as identified from a validated 360-item semi-quantitative FFQ completed by mothers at recruitment (24 weeks gestation), with a 1-month recall period [32,33]. Principal component analysis (PCA) identified dietary patterns based on calculated estimates of energy, macronutrients, and micronutrients derived from pregnancy FFQs. The first principal component reflected a “Varied dietary pattern” (44.3% of variance), characterised by positive associations with a wide range of FFQ-derived food groups, including whole grains, fish, eggs, and nuts. The second principal component reflected a “Western dietary pattern” (10.7% of variance), characterised by positive associations with FFQ-derived food groups such as animal fats, refined grains, and high-energy drinks, and negative associations with fruits, fish, and vegetables (S1 Fig) [30]. Nutrient intake data and principal component loadings are detailed S1 Table. These dietary patterns were used to derive the metabolite scores in the sPLS models. Breastfeeding duration, measured as the length of breastfeeding in days, was also considered as a dietary exposure. Further details regarding the blood metabolomics analysis, derivation of dietary pattern metabolite scores via sPLS modelling, and data normalisation procedures for breastfeeding duration can be found in the Supplementary Methods. In subanalysis, dietary pattern metabolite scores were modelled for other maternal and child blood metabolome time points (maternal 1-week postpartum and children at 6 months, 18 months, and 6 years), trained on the pregnancy time point. Metabolite scores and breastfeeding length were internally z-scored at each time point, thus reflecting a change of 1 standard deviation (SD) within the population. When assessing the relationships between specific dietary exposures and neurodevelopmental outcomes, exposures were evaluated independently.

Cognitive ability

Cognition was assessed at 2.5 years [34] as reflected in the cognitive composite score (CCS) derived from the 3rd edition of the Danish [35] Bayley Scales of Infant and Toddler Development (Bayley-III) [36]. CCS was standardised with age-corrected means of 100 and a SD of 15, ranging from 50 to 150. For cognition at 10 years, the Full Scale Intelligence Quotient (FSIQ) was estimated, based on selected tests from the fourth edition of the Wechsler Intelligence Scale for Children-fourth edition (WISC-IV) [37], including Matrices, Vocabulary, Digit Span, Letter-Number Sequencing, Coding, and Symbol Search. Estimated FSIQ likewise utilised age-corrected means of 100 and a standard deviation of 15, ranging from 40 to 160. Previous studies have reported moderate longitudinal correlations between CCS and FSIQ [38].

Head circumference growth

Head circumference was measured using a tape measure at each of the 14 scheduled clinical visits, with the largest diameter serving as the measurement parameter (1 week–10 years). Foetal head circumference measurements were added from the week 20 prenatal screening programme using the Danish Føto-database [39].

In our main analysis, calibration of child head circumference was performed to account for potential confounding of related growth metrics. This was achieved by regressing head circumference against age (gestational age-adjusted), sex, length (0–18 months) or height (2–10 years), weight, and waist circumference at each visit. The residuals from this regression model were then used to create a calibrated measure of head circumference for each visit, and subsequently z-scored. Foetal head circumference was likewise calibrated for gestational age, sex, femur length, and abdominal circumference at measurement. Covariates with extreme outliers (those exceeding ± 6 SDs) and deviations exceeding ± 2 SD from an individual’s global growth trajectory were removed prior to calibration. Missing calibration covariates were imputed from linear mixed models on individual child growth trajectories. Head circumference measurements ± 3 SD were visually inspected to ensure data was in keeping with repeated growth measurements. In sensitivity analysis, head circumference measurements that lay outside of 2 SD of an individual’s global growth trajectory were removed (1.0%) and separately, head circumference was adjusted only for child age and sex.

We modelled head circumference longitudinally using latent class growth trajectories and linear mixed models. We considered latent class models with 2–5 latent classes. The optimal fit was a 3-class linear growth model, selected based on the smallest Akaike information criterion, Bayesian information criterion, and biological interpretation. The “lcmm” R-package was used [40]. Linear mixed models were applied with variable slopes and intercepts, the slopes and intercepts were treated as random effects, allowing for modelling of individual head circumference growth and head circumference in early life, respectively. Accordingly, these random effects were extracted, the “lme4” R-package was used for modelling [41] and both slopes and intercepts were scaled (estimates are interpreted as per SD change).

Polygenic risk scores

Maternal and child genotypes in the COPSAC2010 cohort were analysed using the Illumina Infinium HumanOmniExpressExome BeadChip. To calculate a polygenic risk score (PRS) for intelligence, a genome-wide association study (GWAS) from the UK Biobank was utilised (n = 391,124) measuring fluid intelligence [42]. For child head circumference PRS scores, we utilised a GWAS derived from child head circumference measurements (n = 29,192) [21]. As COPSAC2010 participated in this consortium study, we contacted the authors who re-ran the GWAS with COPSAC2010 excluded. The PRS-CS method, which integrates millions of SNPs across the genome for enhanced precision, was applied. This method is well-validated in European populations and ensures that the PRS captures genome-wide genetic predispositions rather than being limited to specific loci. The strong genetic basis of intelligence and head circumference, combined with the robust GWAS methodologies, supports the relevance of these PRS to our outcomes. For head circumference, parameters were manually tuned during analysis using a grid search approach to optimise the global scaling factor. PRS were z-scored to mean 0 and SD 1 for interpretation.

Information on covariates

Potential confounders were identified with a directed acyclic graph and existing literature [13,43–46]. In multivariable analysis, we included the following covariates: pre-pregnancy maternal body mass index (BMI), child sex, birthweight, gestational age, smoking during pregnancy, antibiotic use during pregnancy, pre-eclampsia during pregnancy, household income at birth (low (<€50,000), medium (€50,000–€110,000), high (>€110,000)), maternal education level at birth (numerically coded as: 1 (elementary and college education), 2 (medium education), 3 (university education)) and maternal age. All missing covariate data were imputed for analysis, based on other available covariate data.

To control for genetic confounding, we included maternal and child intelligence and head circumference genetic scores in further multivariable analysis. Likewise in sensitivity analysis, we further adjusted for any child neurodevelopmental disorders at 10 years, a child Western dietary pattern at 10 years [29], parental head circumference and outcomes from two prenatal randomised controlled trials: n3-long-chain polyunsaturated fatty acids supplementation and high-dose vitamin D during pregnancy [30].

Analysis plan

This study did not have a prospective analysis plan established prior to its commencement. Hypothesis-driven analyses were planned in mid-2023, based on existing literature and the study’s specific objectives. No data-driven changes to the analysis were made, and no modifications were implemented in response to peer review comments.

Statistical analysis

Linear regression models were used to examine the association between dietary variables and cognitive scores, as well as the association between head circumference and cognitive scores. Linear mixed models were used to assess the associations between dietary variables and head circumference growth, incorporating interaction terms with time and random effects to account for individual trajectories over multiple visits. In subanalysis, we employed nested linear and linear mixed models, comparing them using likelihood ratio tests (using the anova() function in R, specifying a Chi-square test) to assess the statistical significance of model differences. To handle the issue of missing metabolite scores (14.7% missing), essential for balanced anova comparison modelling, we imputed missing data using the missMDA package [47]. Causal mediation analysis was conducted with multivariable models using the mediation package in R (v4.5.0), employing 1,000 iterations. We removed the second twin child in twin pairs (n = 5) from our analysis, since the exposures of interest were not independent. The statistical analyses were designed in 2022 based on hypotheses derived from existing literature and the comprehensive data in the COPSAC2010 cohort. All data analyses were performed with the statistical software R version 4.1.1. Other R packages utilised in this analysis include tidyverse (v1.3.1), dplyr (v1.0.10), broom (v0.7.12), psych (v2.1.9), stats (v4.4.1), scatterplot3d (v.3.41), ggpubr (v0.4.0), lubridate (v1.8.0) and tableone (v0.13.0).

Inclusion and ethics

The study was approved by The Local Ethics Committee (H-B-2008-093), and the Danish Data Protection Agency (2015-41-3696). Written informed consent was obtained from participants at cohort recruitment, prior to any study-related procedures.

Baseline characteristics and dietary exposures

Dietary exposures included dietary pattern metabolite scores derived from maternal blood metabolome data (available for 684 mothers) and breastfeeding duration data (available for 687 mother–child pairs). Outcomes assessed were cognition, evaluated in 627 (90.2%) children at 2.5 years and 586 children (84.3%) at 10 years, and head circumference growth, based on 9,501 measurements collected from 20 weeks gestation to 10 years (8.5% data missing). Each child in the main analysis contributed a median of 15 head circumference measurements (interquartile range 13–15). There were no significant differences in baseline characteristics between participants and non-participants at the 10-year visit (S2 Table) nor between male and female children (S3 Table).

A Western dietary pattern metabolite score in pregnancy was characterised with increased intakes of animal fats, refined grains, and high-energy drinks, and a lower score was associated with increased intakes of fish, breakfast cereals, fruits, and vegetables (Fig 1A). A Varied dietary pattern metabolite score was characterised with increased intakes of coffee and breakfast cereals, and lower intakes of snacks and energy drinks (Fig 1B). The mean (SD) breastfeeding duration was 247 days (sd = 165) (Fig 1F).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Associations of the Western and Varied dietary patterns metabolite scores during pregnancy with food frequency derived food groups.

This figure presents the associations between the Western and Varied dietary patterns metabolite scores during pregnancy and food frequency-derived food groups. Panels A and B depict these associations for the Western and Varied dietary patterns, respectively. Panel C presents a scatter plot with a regression line showing the strong association between Western and Varied dietary patterns, with ellipses for high and low breastfeeding duration. Panels D, E, and F display the distribution of our data in density plots for the Western dietary pattern metabolite score in pregnancy, Varied dietary pattern metabolite score in pregnancy, and breastfeeding duration, respectively.

https://doi.org/10.1371/journal.pmed.1004454.g001

The Western dietary pattern was associated with numerous covariates (Table 1). Furthermore, a Western dietary pattern was negatively associated with a varied dietary pattern (p < 0.001) and breastfeeding duration (p < 0.001), reflecting the interconnected nature of these dietary exposures in our observational data (Figs 1C and S2). Detailed correlations can be visualised in a comprehensive heatmap (S3 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Covariates and dietary exposures stratified by high and low Western dietary pattern metabolite scores during pregnancy. This table presents the model covariates and dietary exposures, stratified by high and low (median split) Western dietary pattern metabolite scores during pregnancy (n = 684). It provides a comprehensive overview of the cohorts’ characteristics, including differences in other dietary exposures and covariates stratified by high and low Western dietary pattern groups.

https://doi.org/10.1371/journal.pmed.1004454.t001

Dietary patterns in pregnancy are associated with child cognition at 2.5 and 10 years

The mean CCS, assessed at 2.5 years, was 104.8 (sd = 9.7). Scores were higher amongst girls than boys (105.9 (10.3) versus 103.7 (9.1) (t test, p = 0.004)). FSIQ population mean, assessed at 10 years, was 102.7 (12.1). Girls had higher FSIQ cognitive scores compared to boys (104.5 (11.3) versus 101.1 (12.6), t test, p < 0.001).

In univariate analysis, the Western dietary pattern metabolite score in pregnancy (per 1 SD change) was negatively associated with CCS (β −1.43 [−2.18, −0.67], p < 0.001) and FSIQ at 10 years (β −2.45 [−3.42, −1.47], p < 0.001). In multivariable analysis, these results were consistent for CCS (β −1.24 [−2.16,–0.32], p = 0.008), whereas FSIQ no longer reached statistical significance (β −0.96 [−2.07,0.15], p = 0.09) (Tables 2 and S4 for WISC-IV composite scores). Findings were comparable after further adjusting for genetic confounding.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Associations between dietary pattern metabolite scores, duration of breastfeeding and cognition outcomes. This table presents the results of linear regression analyses assessing the associations between dietary exposures and cognition outcomes; specifically, the Bayleys-III cognitive composite score at 2.5 years and WISC: Full Scale Intelligence Quotient at 10 years. Estimates are interpreted as the effect of a 1 standard deviation increase of a Western, or Varied dietary pattern metabolite score, or logged breastfeeding duration. The table provides both unadjusted and adjusted associations, with the latter controlling for potential confounders including pre-pregnancy maternal body mass index, child sex, birth weight, gestational age, smoking during pregnancy, antibiotic use during pregnancy, pre-eclampsia, household income at birth, maternal education level at birth, and maternal age at birth. Further adjustments include maternal and child polygenic risk score (PRS) for intelligence, and child head circumference PRS.

https://doi.org/10.1371/journal.pmed.1004454.t002

There was no univariate association between the Varied dietary pattern metabolite score during pregnancy and CCS (β 0.42 [−0.35,1.20], p = 0.280), but a significant association with FSIQ (β 2.39 [1.40,3.37], p < 0.001). This association remained significant after multivariable adjustment (β 1.29 [0.27,2.30], p = 0.014), and after further adjusting for genetic confounding. Breastfeeding duration (per SD change) was not univariately associated with CCS (β 0.08 [−0.74,0.89], p = 0.85), but was associated with FSIQ at age 10 years (β 1.47 [0.45,2.50], p = 0.005). This finding became insignificant after multivariable adjustment (β 0.43 [−0.59,1.44], p = 0.412).

Sensitivity analyses adjusting for prenatal nutrient supplementation randomised controlled trials (RCT)s, child Western dietary patterns, parental head circumference, and any neurodevelopmental diagnosis at 10 years did not change the inference of our findings (S5 Table). There were no significant differences found in relation to child sex, as determined through statistical interaction (S6 Table).

In subanalysis, we assessed the temporality of the significant associations between the Western dietary pattern and CCS scores, as well as the Varied dietary pattern and FSIQ scores. To do this, we calculated dietary pattern metabolite scores at four additional postnatal time points: for mothers at 1-week postpartum and for children at 6 months, 18 months (for CCS), and additionally at 6 years (for FSIQ). For the Western dietary pattern, despite collinearity between scores (S7 Table), adding postnatal metabolite scores improved CCS model prediction (r2 increase from 0.035 to 0.043, p = 0.038). However, FSIQ prediction for the Varied dietary pattern was not improved when considering postnatal metabolite scores (p = 0.68).

Head circumference growth in early life is positively associated with child cognition at 10 years

Head circumference cross-sectionally was not univariately associated with CCS at 2.5 years (β −0.33 [−1.14,0.48], p = 0.420), but was significantly associated with FSIQ at 10 years (β 1.21 [0.25,2.16], p = 0.013).

For the latent class growth trajectory modelling, our analysis focussed on two classes; ‘Increasing’ (26%) and ‘Reference’ (74%) (Fig 2). In univariate analysis, children in the ‘Increasing’ head circumference latent class had 4.87 IQ points higher score at 10 years in FSIQ compared with children in the ‘Reference’ group (β 4.87 [2.62,7.12], p < 0.001). This association remained significant in the multivariable model (β 3.40 [1.21,5.60], p = 0.002) and when further adjusting for genetic confounding (Table 3, see S8 Table for WISC composite scores).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Latent class trajectories of calibrated head circumference z-scores.

This figure presents the two latent class trajectories analysed in the main analysis, namely “Reference” (74%, n = 511) and “Increasing” (26%, n = 177). The trajectories show the calibrated head circumference z-scores at 15 time points from the foetal ultrasound scan (20 weeks gestation) to the 10-year clinical visit. Head circumference measurements were adjusted for child length/height, weight, waist size, sex, and gestation-adjusted age at each clinical visit.

https://doi.org/10.1371/journal.pmed.1004454.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Associations between cognitive outcomes and head circumference growth trajectories. This table presents the results of linear regression analyses assessing the associations between cognitive outcomes and head circumference growth, as determined by latent class trajectory modelling and linear mixed modelling. The latent class trajectory model estimates compare the “Increasing” (26%) class to the “Reference” (74%) class. Estimates for linear mixed model slope and intercept are interpreted as the effect of 1 standard deviation within our population. The table provides both unadjusted and adjusted associations, with the latter controlling for potential confounders such as pre-pregnancy maternal body mass index, child sex, birth weight, gestational age, smoking during pregnancy, antibiotic use during pregnancy, pre-eclampsia, household income at birth, maternal education level at birth and maternal age at birth. Further adjustments include maternal and child polygenic risk score (PRS) for intelligence, and child head circumference PRS.

https://doi.org/10.1371/journal.pmed.1004454.t003

In linear mixed models, the slope, representing head circumference growth (per SD change), was similarly positively associated with FSIQ (β 2.03 [1.08,2.97], p < 0.001). This association remained significant in multivariable modelling (β 1.27 [0.34,2.20], p = 0.008) and when further adjusting for genetic confounding (Table 3). Model intercepts were also predictive of FSIQ in univariate (β 1.93 [0.94,2.93], p < 0.001) and multivariable models (β 1.29 [0.33,2.26], p = 0.009), and after genetic confounding adjustment (Table 3).

As a sensitivity analysis, we assessed the potential moderation of child genetics scores for head circumference and intelligence, on the association between head circumference and FSIQ at 10 years. In interaction analysis, we found no significant moderating effect of child head circumference PRS (p-interaction = 0.442), however, there was a significant and positive moderation of the child intelligence and head circumference on latent classes (p-interaction = 0.034), suggesting that for children who have a higher PRS for intelligence, and adhere to the ‘Increasing’ latent class, there is a stronger association with FSIQ (S4A Fig). Moderating effects were stronger in the WISC perceptual reasoning index (p-interaction = 0.020) (S4B Fig).

Despite strong associations with FSIQ at age 10 years, we found no association between head circumference growth measures and CCS scores at 2.5 years (Table 3).

A pregnancy Western dietary pattern is negatively associated and breastfeeding duration positively associated with head circumference growth

In univariate analysis, significant associations were observed between calibrated head circumference at age 10 years and both Western (β −0.19 [−0.2,−0.11], p < 0.0001) and Varied (β 0.17 [0.09,0.25], p < 0.0001) dietary patterns during pregnancy, as well as with breastfeeding duration (β 0.09 [0.00,0.17], p = 0.040). In multivariable longitudinal models, including all 15 time points, we observed that our three dietary variables of interest were all strongly associated with head growth, as indicated by significant interaction terms (p-interaction < 0.0001) (S5 Fig). After mutually adjusting for these variables, only the Western dietary pattern (p-interaction = 0.020) and breastfeeding duration (p-interaction = 0.0001) remained significant (S5 Fig). Therefore, we opted to exclude the Varied dietary pattern from subsequent modelling.

In multivariable linear mixed modelling, both the pregnancy Western dietary pattern (p-interaction < 0.0001) and breastfeeding duration (p-interaction < 0.0001) independently contributed to head circumference growth over time, with distinct associations (Fig 3 and S9 Table). The pregnancy Western dietary pattern showed a negative association, indicating smaller child head circumference over time, while breastfeeding length exhibited a positive association, suggesting larger head circumference with longer duration of breastfeeding. Comparable results were observed when using uncalibrated head circumference (p-interaction < 0.0001) and with head circumference data outliers (1.0%) removed (p-interaction < 0.001). Child sex did not significantly moderate the associations of the Western dietary pattern (p = 0.14) and breastfeeding length (p = 0.51) (S6 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Predicted head circumference z-scores from Western dietary pattern metabolite score during pregnancy and breastfeeding duration.

This figure presents the predicted head circumference z-scores (with 95% CL) of a Western dietary pattern metabolite score during pregnancy (p < 0.0001) and breastfeeding duration (p < 0.0001) on longitudinal measures of head circumference in a multivariable linear mixed model (± 1 SD of the exposure). Head circumference measures are adjusted at each visit for child length/height, weight, waist circumference, sex, and age. Panels A and C show the predicted head circumference estimates with 95% confidence limits for the Western dietary pattern metabolite score during pregnancy and breastfeeding duration, respectively. Panels B and D show the interaction term effects at each clinical visit for the Western dietary pattern metabolite score during pregnancy and breastfeeding duration, respectively.

https://doi.org/10.1371/journal.pmed.1004454.g003

In subanalysis, adjustments for maternal (p = 0.003) and paternal head circumference (p < 0.0001) significantly contributed to head circumference growth modelling. Yet, these further adjustments did not alter the estimates between the Western dietary pattern, breastfeeding duration, and head circumference growth, indicating no confounding from parental traits linked to the inheritable nature of these phenotypes.

To explore the temporal associations of a Western diet on child head growth, we included maternal metabolome-derived Western dietary scores from postnatal stages (1-week postnatal), and children at 6 months, 18 months, and 6 years into our models. Adding these four additional metabolite score time points did not improve our modelling (anova p = 0.63). A similar finding was observed for the Varied dietary pattern (anova p = 0.58). However postnatal Western dietary pattern metabolite scores showed independent associations with head circumference growth at 1 week postpartum (p-interaction < 0.001), 6 months (p-interaction = 0.011), and 18 months (p-interaction = 0.005), but not at 6 years (p-interaction = 0.904) (S7 Fig). Varied dietary pattern metabolite scores in children did not show independent associations with head circumference growth (p-interaction > 0.16) (S7 Fig).

To assess the dynamic genetic contributions to head circumference growth over time, we included child and maternal PRSs for intelligence, and child head circumference as interaction terms with clinical visits, to capture growth trajectories (S8 Fig). Child intelligence PRS (anova p = 0.014) and child head circumference PRS (anova p = 0.006) independently improved head circumference growth modelling, compared to models without these genetic interaction terms, whereas maternal intelligence PRS (anova p = 0.51) showed no improvement. In models mutually adjusted for child intelligence and head circumference PRS, the child intelligence PRS displayed a significant main effect (p = 0.002), implying a direct influence on head circumference, while the child head circumference PRS did not reach statistical significance (p = 0.071). Interaction effects revealed a differential pattern: the intelligence PRS interaction was not significant (p-interaction = 0.204), while the head circumference PRS interaction was significant (p-interaction = 0.038). Notably, the Western dietary pattern and breastfeeding duration interactions persisted (p-interaction ≤ 0.0001), emphasising their independent associations with head circumference, distinct from intelligence or head circumference genetic influences.

Mediating role of head circumference growth in the association between a varied dietary pattern in pregnancy and child cognition at age 10

In sensitivity analysis, we explored whether head circumference growth may mediate the significant association between the Varied dietary pattern and child cognition at age 10 using causal mediation analysis. We found that head circumference growth significantly mediated this association in multivariable modelling. The Average Causal Mediation Effect (ACME) was 0.182 (p = 0.022), accounting for 13.5% of the total effect (Proportion Mediated, p = 0.034), indicating a substantial mediation effect.

Governance

We are aware of and comply with recognised codes of good research practice, including the Danish Code of Conduct for Research Integrity. We comply with national and international rules on the safety and rights of patients and healthy subjects, including Good Clinical Practice as defined in the EU’s Directive on Good Clinical Practice, the International Conference on Harmonisation’s good clinical practice guidelines and the Helsinki Declaration. We follow national and international legislation on General Data Protection Regulation (GDPR), the Danish Act on Processing of Personal Data and the practice of the Danish Data Inspectorate.