Introduction
Executive function (EF) is an umbrella term for multiple neurocognitive functions that enable individuals to control their thoughts, actions, and behaviors, to achieve goals, and solve problems effectively (Diamond, Reference Diamond2013). Language skills (LS) are crucial for children to achieve their developmental potential, facilitating educational attainment and the acquisition of a diverse range of other skills, including EF (Cortés Pascual et al., Reference Cortés Pascual, Moyano Muñoz and Quílez Robres2019; Gilkerson et al., Reference Gilkerson, Richards, Warren, Oller, Russo and Vohr2018; Spiegel et al., Reference Spiegel, Goodrich, Morris, Osborne and Lonigan2021).
During the first years of life, risk and protective factors affect childhood development and play a pivotal role in impairing or fostering the emergence of EF and LS (Merz et al., Reference Merz, Wiltshire and Noble2019). Although the exact mechanisms are not fully understood, as well as which particular aspects of EF are relevant to specific LS, and vice versa, there is evidence for specific factors that have been tested individually. Parenting characteristics (Madigan et al., Reference Madigan, Prime, Graham, Rodrigues, Anderson, Khoury and Jenkins2019), infant temperament (Kucker et al., Reference Kucker, Zimmerman and Chmielewski2021; Ursache et al., Reference Ursache, Blair, Stifter and Voegtline2013), genetic profile (Hatoum et al., Reference Hatoum, Morrison, Mitchell, Lam, Benca-Bachman, Reineberg, Palmer, Evans, Keller and Friedman2022), nutrition (Sania et al., Reference Sania, Sudfeld, Danaei, Fink, McCoy, Zhu, Fawzi, Akman, Arifeen, Barros, Bellinger, Black, Bogale, Braun, van den Broek, Carrara, Duazo, Duggan, Fernald and Fawzi2019), and gut microbiota (A. L. Carlson et al., Reference Carlson, Xia, Azcarate-Peril, Goldman, Ahn, Styner, Thompson, Geng, Gilmore and Knickmeyer2018), mediated by specific patterns of DNA methylation (Peter et al., Reference Peter, Fischer, Kundakovic, Garg, Jakovcevski, Dincer, Amaral, Ginns, Galdzicka, Bryce, Ratner, Waber, Mokler, Medford, Champagne, Rosene, McGaughy, Sharp, Galler and Akbarian2016), may all play complex and interconnected roles (Berens et al., Reference Berens, Jensen and Nelson2017; Ioannidis et al., Reference Ioannidis, Askelund, Kievit and van Harmelen2020) impacting neural networks associated with EF and language skills (LS).
Identifying modifiable factors influencing these trajectories is essential for developing interventions to target such factors. For example, the promotion of parenting skills (Lucassen et al., Reference Lucassen, Kok, Bakermans-Kranenburg, Van Ijzendoorn, Jaddoe, Hofman, Verhulst, Lambregtse-Van den Berg and Tiemeier2015) and treatment of maternal mental health problems (Park et al., Reference Park, Brain, Grunau, Diamond and Oberlander2018) are potential targets for various potential interventions. Hence, to understand risk and protective factors for EF and LS development during infancy, it is critical to longitudinally assess the comprehensive exposure to such factors and their underlying biological processes operating at multiple levels (McEwen et al., Reference McEwen, Bowles, Gray, Hill, Hunter, Karatsoreos and Nasca2015).
Therefore, we aimed to develop an integrated predictive model for the emergence of EF and LS at three years of age using multimodal longitudinal data. To accomplish this, we designed and established a prospective cohort study named Germina, in which we are assessing mother-infant dyads across the first three years of life, collecting data from multiple domains: sociodemographic, clinical, behavioral, cognitive, developmental, brain development and maturation, genetic, epigenetic, and microbiome data. Our objective in the current article is to provide a detailed description of our study design and methods and provide preliminary findings of an integrated multimodal model to predict infant development measured by the Bayley-III at three months of age.
Methods
Design and setting
Germina is an ongoing prospective cohort study designed to assess mother-infant dyads from 3 to 36 months of age. Our study aims to assess families from diverse economic and ethnic backgrounds, as well as infants with low, moderate, and high risk profiles of developing EF deficits at 36 months of age. We recruited families with 3-month-old infants in the metropolitan area of São Paulo (Southeast of Brazil), which consists of 39 municipalities and includes more than 20 million people. It is a major economic hub, accounting for more than 10% of Brazil’s gross domestic product, and it is known for its diversified industry.
Ethical approvals were obtained from the Ethics Committee for the Analysis of Research Projects (Comissão de Ética para Análise de Projetos de Pesquisa, CAPPESq) and the National Council of Ethics in Research (Comissão Nacional de Ética em Pesquisa, CONEP; ref.: CAAE 49671221.2.0000.0068). In accordance with the Declaration of Helsinki, all mothers provided written informed consent before completing any study measure. Cases of developmental, health, and mental health issues are being referred to specialized health services.
Sampling and recruitment
We designed a non-probabilistic sampling process to enroll mother-infant dyads in our study. The goal was to recruit at least 500 participants in 14 months. Eligibility criteria to be enrolled in our cohort were: (a) maternal age between 20–45 years; (b) infant age between 3 months 0 days and three months and 29 days; (c) gestational age at birth of ≥37 weeks; (d) infant birth weight ≥2,000g; (e) no illegal substance or alcohol use during pregnancy; (f) no history of severe maternal mental disorders (e.g., psychosis, bipolar disorder); (g) no childbirth complications needing medical procedures (e.g, perinatal asphyxia, shoulder dystocia, excessive bleeding); (h) infant not previously diagnosed with genetic syndromes or auditory/visual deficiencies; (i) availability for in-person assessments in our labs in the city of São Paulo.
We used a diverse range of enrollment strategies: social media, traditional media (online newspapers, TV), advertisements, partnerships with various institutions, and word of mouth. In addition, we partnered with designers to develop a brand to communicate more effectively with interested families and to have a social media presence (www.projetogermina.com.br). To ensure the final sample consisted of diverse families from various backgrounds and risk profiles, we established partnerships with community leaders and institutions (e.g., NGOs, hospitals, etc.) located in different municipalities and neighborhoods of the metropolitan area of São Paulo.
Interested families received a URL to a webpage with information explaining the study objectives and procedures and a survey containing questions to verify eligibility criteria and infant risk profile. We defined low, moderate and high risk profiles for EF deficits at age 36 months using a data-driven approach applied to longitudinal data from the 2004 Pelotas Cohort, a population-based Brazilian birth cohort previously used to define criteria for child growth development. Considering the available data in the Pelotas cohort study, we selected characteristics with the most potential to negatively influence EF development. Based on results from a logistic regression (N = 3,452) conducted to predict attentional control at 11 months of age, we selected the following characteristics: total family income, maternal education, maternal skin color, recipient of cash-transfer government programs, tobacco smoking during pregnancy, single mother, and having two or more children. See Table S1 for additional details on the score composition. Then, we conducted a receiver operating curve analysis to validate the ability of our risk score to classify children with attentional control scores below the 10th percentile. Finally, based on the total risk score (0–35), we divided our Germina sample into three risk tertiles: tertile 1 = low risk (0–8); tertile 2 = moderate risk (9–15); tertile 3 = high risk (16–35).
To cover developmental processes monthly from 3 months to 36 months of age, we randomized mother-infant dyads at baseline to three schedules of assessments depicted in Figure 1. This approach is similar in intent to an accelerated longitudinal design. We balanced the randomization process to ensure that each Schedule had a similar number of participants in each EF risk profile (low, moderate, high). Figure 1 describes assessment schedules by time-point.
Timeline of assessment schedules of the Germina cohort by time-point.
Assessments
The Germina study includes extensive and comprehensive assessments across multiple domains throughout the first three years of life. Table 1 summarizes the data we are collecting at each time-point (see Table S2 for a complete list with references). After our staff confirmed eligibility criteria via phone call, the participants were invited to be assessed at T1. All assessments are composed of two parts: (1) an online interview and (2) an in-person lab assessment. The online interview focuses on verbally administering scales and questionnaires to the mother in Brazilian Portuguese. The in-person lab assessment focuses on administering observational measures, task-based assessments, and biological data collection with the mother-infant dyad.
Assessments are being conducted by a team of psychologists who underwent a 2-month full-time training on all measures used in our study. Experts on multiple domains relevant to the Germina cohort trained the psychologists. Before starting each time-point, psychologists undergo new training to revise previous instruments and to understand how to administer new instruments. Weekly supervision sessions discuss assessments and challenges encountered during data collection. Data is being entered online in the Research Electronic Data Capture (REDCap) system (Harris et al., Reference Harris, Taylor, Thielke, Payne, Gonzalez and Conde2009). Next, we describe the instruments being used for data collection across all time-points. T1 procedures are described in the past tense. Subsequent time-points are described in the present tense to emphasize it is an ongoing process. Additional details regarding assessment procedures, data collection, data and sample processing, and instruments (Table S1) can be found in the Supplementary materials.
Sociodemographic information and environment
We assess socioeconomic status (SES) using the Brazilian economic classification criteria (ABEP, 2022) and the MacArthur scale (Ferreira et al., Reference Ferreira, Giatti, Figueiredo, Mello and Barreto2018). Family food insecurity is measured by the Brazilian Food Insecurity Scale (Escala de Insegurança Familiar, EBIA) (Santos et al., Reference Santos, Lindemann, Motta, Mintem, Bender and Gigante2014). Intimate partner physical violence is assessed by a brief questionnaire based on the WorldSAFE study protocol (Miranda et al., Reference Miranda, Paula and Bordin2010). To assess infant abuse and neglect, we use the Conflict Tactics Scales: Parent-child Version (CTSPC) (Bonfim et al., Reference Bonfim, Santos, Menezes, Reichenheim and Barreto2011), which measures the frequency and severity of abuse and neglect towards the infant. The Affordances in the Home Environment for Motor Development (P. Caçola et al., Reference Caçola, Gabbard, Santos and Batistela2011; P. M. Caçola et al., Reference Caçola, Gabbard, Montebelo and Santos2015) scale is used to assess the quality of the home environment in terms of opportunities for motor and cognitive stimulation. The Chaos, Order, and Hubbub Scale (CHAOS) (Matheny et al., Reference Matheny, Wachs, Ludwig and Phillips1995) is used to assess the level of noise and disorder in the family home environment.
Maternal mental health and behaviors
We selected several measures to assess various maternal mental health domains. A modified version of the Alcohol, Smoking, and Substance Involvement Screening Test to measure use/abuse of substances and alcohol. The Edinburgh Postnatal Depression Scale (EPDS) to assess maternal depression. The Perceived Stress Scale (PSS) (Siqueira Reis et al., Reference Siqueira Reis, Ferreira Hino and Romélio Rodriguez Añez2010) to assess perceived general stress levels. The Generalized Anxiety Disorder Scale-7 (GAD-7) (Moreno et al., Reference Moreno, DeSousa, Souza, Manfro, Salum, Koller, Osório and Crippa2016) to assess anxiety symptoms. The Adult ADHD Self-Report Scale (Polanczyk et al., Reference Polanczyk, Laranjeira, Zaleski, Pinsky, Caetano and Rohde2010) to evaluate symptoms of attention-deficit/hyperactivity disorder (ADHD). The Single-item sleep quality scale (Snyder et al., Reference Snyder, Cai, DeMuro, Morrison and Ball2018) to measure sleep quality using a one-item scale in which participants rate their sleep quality from 0 (very poor) to 10 (excellent). The Multidimensional Scale of Perceived Social Support (Slavin et al., Reference Slavin, Creedy and Gamble2020; Zimet et al., Reference Zimet, Dahlem, Zimet and Farley1988) 3–item version to assess maternal social support. Lastly, the Adult Temperament Questionnaire to assess temperament across multiple dimensions (Evans & Rothbart, Reference Evans and Rothbart2007).
Mother-infant interaction
Mother-infant dyads are asked to participate in a semi-structured social interaction activity comprised of two segments: (1) mothers are asked to play and talk with the infants as they would at home for 5 minutes without objects/toys; (2) mothers are asked to choose among culturally-appropriate objects/toys to choose from and to use for play with the infants for 5 minutes. A protocol based on studies of predictability of maternal behavior (Davis et al., Reference Davis, Stout, Molet, Vegetabile, Glynn, Sandman, Heins, Stern and Baram2017) is used to code the following behaviors: (1) mother is holding the baby; (2) mother is touching the baby; (3) mother is touching or holding a toy or other object; (4) the mother points towards object to draw infant’s attention to it; (5) infant is looking at the mother, or the object the mother is holding/playing with; (6) maternal vocalization directed towards the infant. We also used the Coding interactive behavior (Feldman, Reference Feldman1998) to quantify the following constructs, divided in three domains: caregiver behaviors (overriding-intrusiveness, acknowledgement, vocal appropriateness, consistency of style, affectionate touch), infant behaviors (child gaze, joint attention, negative emotionally, withdrawal, vocalization, verbal output, initiation), and dyadic interaction (dyadic reciprocity, adaptation regulation).
Infant/child behavior and mental health
We chose a range of measures to evaluate different aspects of infant behavior and mental health. The Brief Infant Sleep Questionnaire (Del-Ponte et al., Reference Del-Ponte, Xavier, Bassani, Tovo-Rodrigues, Halal, Shionuma, Ulguim and Santos2020) is a scale designed to evaluate infants’ sleep patterns regarding sleep duration, quality, and regularity. Temperament, defined here as innate or biologically determined behavioral differences in how infants/children react and respond to their environment, is being measured by the Infant Behaviour Questionnaire-Revised (IBQ-R) (Klein et al., Reference Klein, Putnam and Linhares2009) when the infants are aged 3–16 months and the Early Childhood Behavior Questionnaire (ECBQ) (Putnam & Rothbart, Reference Putnam and Rothbart2006) when the infants aged 18–36 months. We use the Child Behavior Checklist for Ages 1 ½–5 to assess the children’s mental health at later time-points (Bordin et al., Reference Bordin, Rocha, Paula, Teixeira, Achenbach, Rescorla and Silvares2013).
Development
We adopted the Bayley scales of infant and toddler development Third Edition (Bayley-III) to assess developmental milestones (Madaschi et al., Reference Madaschi, Mecca, Macedo and Paula2016). The Bayley-III consists of a series of tasks and behavioral observations. The assessment comprises the following domains: cognitive, language, motor, and social-emotional development. The Bayley-III was previously translated and adapted to Brazilian Portuguese.
To assess an infant’s adaptive functioning in everyday life, we use the Brazilian Portuguese version of the Vineland Adaptive Behavior Scales — Third Edition (Vineland-3) (Sparrow et al., Reference Sparrow, Cicchetti and Saulnier2019), a semi-structured interview that measures caregiver-reported adaptive abilities across the domains of communication, daily living skills, socialization, and motor skills.
We also measure weight, length/height, and head circumference using standard pediatric procedures and deriving age-adjusted and sex-specific scores using WHO growth standards.
EF
EF is measured using a multi-method approach throughout development. First, parent-rated indices of self-regulation are measured at T1 to T5 using the Duration of Orienting, Soothability, and Regulation scales of the IBQ-R (T1–T3) (Klein et al., Reference Klein, Putnam and Linhares2009) and the Attentional Focusing, Attentional Shifting, Inhibitory Control, Soothability and Effortful Control scales of the ECBQ (T4–T5) (Putnam & Rothbart, Reference Putnam and Rothbart2006). Second, from 18 months onwards (T4–T5), children complete three behavioral tasks that index key EF processes during early development. These are Reverse Categorization (S. M. Carlson et al., Reference Carlson, Mandell and Williams2004), which measures conflict suppression and cognitive flexibility, Spin the Pots (Hughes & Ensor, Reference Hughes and Ensor2005), which measures working memory, and the Prohibition Task (Friedman et al., Reference Friedman, Miyake, Robinson and Hewitt2011), which measures inhibitory control. Then, at T5 we administer additional measures of EF: Picture Memory and Zoo Locations working memory subtests of the Wechsler Preschool and Primary Scale of Intelligence Fourth Edition (WPPSI-IV) (Raiford & Coalson, Reference Raiford and Coalson2014); Statue subtest of inhibitory control of the NEPSY-II neuropsychological test (Argollo, Reference Argollo, Malloy, Fuentes, Mattos and Abreu2010); Stroop Day-Night test of inhibitory control (Montgomery & Koeltzow, Reference Montgomery and Koeltzow2010); Dimensional Change Card Sort test of cognitive flexibility (Doebel & Zelazo, Reference Doebel and Zelazo2015); Behavior Rating Inventory of Executive Function Preschool Version (BRIEF-P) scale of general executive functioning (including inhibition, attentional shift, emotional control, working memory, planning) (Sherman & Brooks, Reference Sherman and Brooks2010).
Electroencephalography
Infant neurodevelopment is assessed with electroencephalography (EEG), a safe, noninvasive, real-time measure of brain activity recorded using scalp electrodes. At each assessment, infants complete short passive-viewing tasks while their EEG activity is recorded using a 128-channel sponge-based net. Passive-viewing tasks include a 2-minute baseline video of abstract shapes, a face processing task with images of faces and nonsocial stimuli, and (for T1 and T2 only) a visual evoked potential (VEP) task with checkerboard stimuli. EEG metrics are computed on a task-by-task basis and include power spectral density (Levin et al., Reference Levin, Méndez Leal, Gabard-Durnam and O’Leary2018), phase synchrony (debiased weighted phase lag index; dwPLI) (Vinck et al., Reference Vinck, Oostenveld, van Wingerden, Battaglia and Pennartz2011), and other estimates of oscillatory activity and functional connectivity during the baseline video (resting state), the N1 and P1 event-related potential components for the VEP task (Jensen et al., Reference Jensen, Kumar, Xie, Tofail, Haque, Petri and Nelson2019) and the N290 and Nc components for the face processing task (Halit et al., Reference Halit, de Haan and Johnson2003). EEG indices will be analyzed in relation to EF and language outcomes. Full details on EEG data recording and processing are provided in the Supplementary Materials.
Genotyping
Infants’ saliva samples were collected at T1 with the Oragene OG-575 DNA collection kit (DNA Genotek) and will be recollected at T3 if required. Genomic DNA is being extracted with prepIT.L2P (DNA Genotek) or QIAsymphony DNA mini kits (QIAGEN) and quantified with the Qubit DNA BR system (Thermo Fisher). DNA samples below the desirable concentration are spun in Genomic DNA Clean & Concentrator columns (Zymo Research). Whole-genome sequencing libraries are being prepared with 300-500 ng of DNA with the DNA PCR-free kit and sequenced at 10x coverage in the NovaSeq 6000 system (Illumina). Sequencing data is processed following GATK Best Practices for Germline Short Variant discovery (Van Der Auwera & O’Connor, Reference Van Der Auwera and O’Connor2020) using GATK (v4.0.9.0).
Copy number variations (CNV) were identified with CNVkit with a target average bin size of 10,000. For each sample, another ten samples with similar coverage are used as controls. Annotation is carried out with AnnotSV. We are filtering out CNVs present in more than one individual, and with AnnotSV ACMG_class = 1 or 2 (i.e. benign and likely benign respectively). CNV quality is determined by CNVkit metrics and visualization of CNVkit scatter plots. The remaining CNVs are manually curated for pathogenic status and checked with Nexus software (Bionano).
Polygenic scores (PGS) were generated from GWAS summary statistics of genetically correlated phenotypes in a multi-PGS approach to predict EF and LS (Krapohl et al., Reference Krapohl, Patel, Newhouse, Curtis, von Stumm, Dale, Zabaneh, Breen, O’Reilly and Plomin2018). PGS calculations were accomplished with PRSIce-2, which allows the selection of the best-fit PGS across GWAS p-value thresholds. Model covariates include sex, genetic principal component, and environmental factors.
Epigenetics
Saliva was collected at T1 and T4 using Oragene OG-575 DNA collection kit (DNA Genotek). Genomic DNA was extracted with prepIT.L2P (DNA Genotek) or QIAGEN and quantified with the Qubit DNA BR system (Thermo Fisher),. Genomic DNA was treated with sodium bisulfite using an EZ DNA Methylation Kit (Zymo Research, CA, USA), according with manufacturer’s instructions and submitted to methylation arrays using Infinium MethylationEPIC v1.0 BeadChip (850K), Illumina platform.
Epigenetic indices were obtained as biomarkers of early life physiologic stress exposure based on previous studies (McGill et al., Reference McGill, Pokhvisneva, Clappison, McEwen, Beijers, Tollenaar, Pham, Kee, Garg, de Mendonça Filho, Karnani, Silveira, Kobor, de Weerth, Meaney and O’Donnell2022). Glucocorticoid index exposure was obtained from 24 CpGs sites. We adopted an unweighted score defined as the sum of DNA methylation (standardized beta values) across previously found glucocorticoid sensitive CpGs to indicate glucocorticoid exposure. We also adopted an Inflammation-related epigenetic polygenic risk score (i-ePGS) (Barker et al., Reference Barker, Cecil, Walton, Houtepen, O’Connor, Danese, Jaffee, Jensen, Pariante, McArdle, Gaunt, Relton and Roberts2018) as a biomarker of low-grade inflammation in a developmentally sensitive framework.
Microbiome
Stool samples are collected at T1, T2, and T4 and stored in sterile tubes at −20°C followed by −80°C. In some cases, home collection with −20°C storage was allowed. DNA extraction from stool samples utilized the ZymoBIOMICS™ DNA Miniprep Kit with ZMCS mock community for quality control. For taxonomic and functional profiling, stool samples underwent metagenomic sequencing (Illumina NextSeq 2000, shotgun sequencing, paired-end 2x150bp).
The metagenomic data was analyzed by Bioinformatics tools using the bioBakery workflows for taxonomic and functional profiles. Key tools included KneadData (quality control), MetaPhlAn (taxonomic composition), and HUMAnN (pathway abundance and gene families). Alpha diversity (Shannon, Chao1, Simpson indices) and beta diversity (Bray-Curtis similarity) were calculated using Phyloseq.
Data analysis
Given that we have multimodal (multi-view) data (Lahat et al., Reference Lahat, Adali and Jutten2015), there are two main approaches to developing a predictive model: (1) join all modalities into a single one and use a method for single-view data, or (2) treat each modality independently with a multimodal data method. For our objective, we chose approach (2) since it allows us to evaluate the influence of each modality on the prediction. Thus, we used the data-driven sparse Partial Least Squares (ddsPLS) method to deal with multimodal data, and used the scores obtained for each data mode jointly with a regression model (e.g, linear, SVM, Random Forest, etc.) to build a regression model, and extended the multiple holdouts framework (Monteiro et al., Reference Monteiro, Rao, Shawe-Taylor and Mourão-Miranda2016) to assess the generalizability of the regression model. The multimodal framework allows us to test whether there is an association between the outcome and predictors. The framework comprises the data-driven sparse partial least squares (ddsPLS) (Lorenzo et al., Reference Lorenzo, Cloarec, Thiébaut and Saracco2022). We built the prediction model in the following way: (1) ddsPLS was used in the first step to obtain a score vector for each data mode, and (2) the score vectors were used to build a linear regression model to predict our outcome. To evaluate whether the model performs better than the null model, we permuted the outcome variable 10,000 times, built models with the same hyperparameters, and calculated p-values. This process is repeated for ten splits of the data from the training and holdout sets. Finally, we selected the models with p-values smaller than 0.05 after applying the Bonferroni correction. Figure S1 shows the multiple holdouts framework pipeline for single- and multi-modal data.
As proof of concept for our proposed multimodal model, we analyzed predictors of infant development measured by the Bayley-III at T1 (age 3 months). To do this, we used confirmatory factor analysis (CFA) to extract a latent factor variable of infant development encompassing Bayley-III raw scores (cognition, receptive communication, expressive communication, fine motor, and gross motor). We used modification indices to determine pairs of residual correlations that could improve the model. An adequate model fit was considered using the following fit indices and criteria: Tucker Lewis index values >0.90, Comparative fit index values >0.96, Root Mean Square Error of Approximation (RMSEA) values <0.05, and Standardized Root Mean Square Residual values <0.08.
Our multimodal model to predict infant development included 15 domains: socioeconomic and demographic, maternal mental health, home environment, infant behavior, anthropometry, genetics, epigenetics, EEG (power and dwPLI in the delta, theta, low alpha, high alpha, beta and gamma frequencies), and microbiome (alpha diversity, species, and functional pathways abundances). See Table S3 for details of each variable included in each of the 15 domains. Analyses were conducted in 557 mother-infant dyads. To deal with missing values we used the K nearest neighborhood imputation method using k = 3. The imputation method was used each time the data was partitioned in train and test sets. After applying the multiple holdout framework, we obtained four models with significant p-values, whose R2 values in the holdout set ranged from 0.09 to 0.23. We selected the model with the largest R2 in the holdout set. In the following sections, we present the findings corresponding to this model. Analyses were conducted using Python version 3.10.5 and R version 4.3.2.
Results
Recruitment and characteristics of the sample
From November 2021 to January 2023, 1189 mothers/caregivers were screened. Of those, 622 were further assessed to confirm eligibility. The most common reasons for exclusion at this stage were infant age ≥4 months, unavailability for assessments, and inability to be contacted by the research team. Finally, 557 mother-infant dyads were enrolled and assessed at T1. Mother-infant dyads were assigned to the three assessment schedules as follows: 186 to Schedule A, 186 to Schedule B, and 185 to Schedule C. See Figure 2 for the flow diagram of the Germina cohort.
Flow diagram of the Germina cohort.
Maternal mean age at baseline was 33.7 years (SD 5.3). The majority of the mothers were white (65.2%), had a college degree (78.6%), and were married/living with a partner (90.5%). Food insecurity was present in 21.4% of the sample, while 7.0% of the families were enrolled in a government welfare program (e.g., cash-transfer program Bolsa Família). Mothers, on average, presented low to moderate levels of depression (EPDS mean = 7.6), anxiety (GAD-7 mean = 6.2), and stress (PSS mean = 17.8). Regarding their offspring, 19.9% presented a moderate to high risk of developing EF deficits. All Bayley-III scores were slightly above the normative mean (see Table 2).
Infant development latent factor
We ran an initial CFA model composed of the 5 Bayley-III raw scores measured at T1. Except for RMSEA, all fit indices indicated a good fit according to our adopted criteria. Modification indices suggested the inclusion of the following residual correlations of pairs of variables: cognition and fine motor; receptive communication and expressive communication; expressive communication and fine motor; cognition and receptive communication; receptive communication and gross motor. Thus, we ran a second CFA model, including the modifications suggested to improve fit. This second and final model presented a good fit based on all fit indices. Model fit indices are shown in Table S4, a diagram of the final model showing factor loadings is depicted in Figure S2, and Figure S3 depicts the distribution of the latent factor score.
Data integration and predictive model
We applied the multiple holdout framework for multimodal data. We obtained a significant p-value<0.05 after Bonferroni correction in the omnibus test with an R2 = 0.28 in the training set and an R2 = 0.23 in the holdout set. This implies that the outcome variable (infant development measured by the Bayley-III at T1) was associated with the data modes included in the model.
The following data modes were significantly associated with infant development in the linear regression model: socioeconomic and demographic (B = 0.29, p < 0.001), epigenetics (B = 0.11, p = 0.005), EEG theta (B = 0.14, p = 0.002), EEG beta (B = 0.11, p = 0.009), and microbiome functional pathways (B = 0.08, p = 0.044) (see Figure 3A and table S5). Furthermore, in Figure 3B and Table S6, we present a measure of the importance of each mode. These values were obtained by running linear models using the training data removing each data mode score to verify the impact on the R2 in the test data. Thus, the importance measure is the percentage by which data mode affects the R2 in the test data. We observed that the data modes with the largest impact were socioeconomic and demographic, epigenetics, microbiome functional pathways, and EEG theta and beta.
Findings from data integration and predictive model. (a) Coefficient plot of regression model. (b) Relative importance of data modalities.
Discussion
We described the rationale, methods, and design of the Germina prospective cohort study. We outlined a comprehensive strategy to assess mother-infant dyads across five-time points during the first three years of life, detailing specific methods to collect multimodal data. In addition, we reported preliminary findings of an integrated model designed to predict infant development measured by the Bayley-III at three months of age using data from 15 domains. We successfully enrolled 557 mother-infant dyads and assessed them at T1 (3 months of age). To the best of our knowledge, this represents one of the largest single-site early childhood development cohorts with multimodal data. Furthermore, it is one of the largest samples from a low and middle-income country, where such studies are scarce. Our preliminary findings revealed five domains associated with early infant development: socioeconomic and demographic characteristics, epigenetics, EEG theta and beta activity, and microbiome functional pathways.
Socioeconomic characteristics are known to significantly impact infant development in the first years of life. Populational data from 66 countries showed that developmental delay is substantially more prevalent in countries in the poorest decile than in the wealthiest decile (C. G. Victora et al., Reference Victora, Hartwig, Vidaletti, Martorell, Osmond, Richter, Stein, Barros, Adair, Barros, Bhargava, Horta, Kroker-Lobos, Lee, Menezes, Murray, Norris, Sachdev, Stein and Black2022). Meta-analytic data have shown that family SES can negatively impact cognitive and language development in the first years of life (Letourneau et al., Reference Letourneau, Duffett-Leger, Levac, Watson and Young-Morris2013). Low SES is linked to a diverse range of environmental factors associated with deleterious effects on infant development. For instance, on average, low SES families engage less in interactions and expose infants to fewer words when compared to high SES families (Gilkerson et al., Reference Gilkerson, Richards, Warren, Montgomery, Greenwood, Kimbrough Oller, Hansen and Paul2017). Also, there is a dose-response relationship between maternal educational level and cognitive stimulation (Barros et al., Reference Barros, Matijasevich, Santos and Halpern2010), indicating that mothers with lower education provide fewer opportunities for cognitive stimulation for their infants. In addition, poverty-related stress, such as financial hardship, can lead to feelings of powerlessness and negative emotional states. Low-income families also often encounter threatening and uncontrollable events, and face destabilizing situations, therefore, having to cope with more demands than high-income families (Bradley & Corwyn, Reference Bradley and Corwyn2002).
Socioeconomic characteristics, such as some of the variables we measured in our study, are fortunately modifiable risk factors. Thus, well-designed interventions can alter these factors, potentially preventing adverse outcomes. Recently, evidence from a randomized controlled trial suggested that cash transfers to impoverished families significantly impacted language development in 1–3 years old children (Fernald & Hidrobo, Reference Fernald and Hidrobo2011). Interventions aimed at supporting impoverished families can also positively affect the home environment and infant development outcomes (Jeong et al., Reference Jeong, Franchett, Ramos de Oliveira, Rehmani and Yousafzai2021). Despite this, it must be noted some socioeconomic factors we analyzed are difficult to change, especially in a short amount of time (e.g., educational level), or not possible to change (e.g., age).
Our findings also showed that the epigenetics factors analyzed, glucocorticoid and inflammation exposure, were associated with infant development. Glucocorticoids are stress-related hormones with an intricate bidirectional relationship with the immune system; they also play a vital role in health, pregnancy outcomes, and fetal development (Moisiadis & Matthews, Reference Moisiadis and Matthews2014; Schepanski et al., Reference Schepanski, Buss, Hanganu-Opatz and Arck2018). However, glucocorticoid overexposure is associated with poor developmental outcomes (Tao et al., Reference Tao, Du, Chi, Zhu, Wang, Meng, Ling, Diao, Song, Jiang, Lv, Lu, Qin, Huang, Xu, Liu, Ding, Jiang and Ma2022). Previous findings revealed a connection between exposure to stress hormones during fetal or early postnatal development and the overall occurrence of neurological and psychiatric disorders (Mesquita et al., Reference Mesquita, Wegerich, Patchev, Oliveira, Leão, Sousa and Almeida2009). The fetal hypothalamic–pituitary–adrenal axis is especially susceptible to glucocorticoids long-term programing, and these effects can persist throughout life (Moisiadis & Matthews, Reference Moisiadis and Matthews2014). A previous study in Brazil found accelerated gestational epigenetic age in infants exposed to glucocorticoids and low-grade inflammation indices (Euclydes et al., Reference Euclydes, Gomes, Gouveia, Gastaldi, Feltrin, Camilo, Vieira, Felipe-Silva, Grisi, Fink, Brentani and Brentani2022).
The significant positive prediction of developmental outcomes on the Bayley-III by oscillatory power and phase connectivity in the theta and beta frequencies is consistent with several previous studies on associations between oscillatory activity and neurocognitive development in infants. For example, (Xie et al., Reference Xie, Jensen, Wade, Kumar, Westerlund, Kakon, Haque, Petri and Nelson2019) reported associations between cognitive development and functional connectivity in the theta and beta frequencies in infants growing up in poverty in Bangladesh, with growth faltering predicting lower functional connectivity in these frequencies, which in turn predicted poorer cognitive outcomes (Xie et al., Reference Xie, Jensen, Wade, Kumar, Westerlund, Kakon, Haque, Petri and Nelson2019). Power in the theta band has also been positively associated with cognitive development in infants growing up in non-adverse environments (Braithwaite et al., Reference Braithwaite, Jones, Johnson and Holmboe2020). The current findings are also consistent with the crucial role that theta and beta oscillations play in executive function and cognition in the adult cognitive neuroscience literature (Cavanagh & Frank, Reference Cavanagh and Frank2014; Cohen, Reference Cohen2014). It should be noted, however, that other studies have not found associations between theta or beta oscillatory activity and infant development (e.g., Shephard et al., Reference Shephard, Fatori, Mauro, de Medeiros Filho, Hoexter, Chiesa, Fracolli, Brentani, Ferraro, Nelson, Miguel and Polancyk2019). Thus, replicating the current findings in other samples would be crucial to determine the frequency-specificity of the findings.
Our integrated model revealed the importance of microbiome functional pathways in predicting infant development. The role of the microbiome in neurodevelopment has gained increasing attention as a result of epidemiological studies investigating typical and atypical neurodevelopment (Naspolini et al., Reference Naspolini, Schüroff, Figueiredo, Sbardellotto, Ferreira, Fatori, Polanczyk, Campos and Taddei2024), as well as neurobehavioral outcomes and their relation with microbiome functional pathways (Korteniemi et al., Reference Korteniemi, Karlsson and Aatsinki2023). Lately, shotgun metagenomic sequencing has allowed the description of microbial mechanisms affecting neurodevelopment, by identifying microbiome functional pathways. Laue et al., reported proline and vitamin B6 biosynthesis and catechol degradation as important microbiome functions linked to BASC-2 scores (Laue et al., Reference Laue, Karagas, Coker, Bellinger, Baker, Korrick and Madan2022). Understanding the gut microbiome-brain axis in early life is important since early interventions can have long-term consequences due to the dynamic and plastic characteristics of both the gut microbiome and the brain.
Several challenges in implementing our cohort should be noted. Our risk score for EF deficits at 36 months of age, derived from an analytical approach using data from the 2004 Pelotas Cohort, showed that only 20% of our sample presented moderate/high risk profiles. Our risk score was mainly composed of environmental characteristics (e.g., family income, education), so it denoted that most families enrolled were not from low SES settings. Indeed, the mean family income was 11.7 thousand (BRL), approximately nine times the Brazilian minimum wage in Brazil. Besides, 76.5% of the mothers reported having at least a college degree, while in Brazil it is estimated that only 19% of the general population has a college degree.
We believe one of the reasons for the low prevalence of moderate/high infant risk profiles in our study was related to our recruitment process. We sought to recruit families using an assortment of enrollment strategies. However, the most effective approach was the use of social media. This probably led to the limited reach of families with low SES. This could be due to the availability of smartphones or computers to access the internet, inadequate internet speed, low interest and difficulty of low SES families in participating in observational studies, and difficulties in urban mobility (e.g., travel time and expenses). Moreover, we had limited time to recruit the sample due to the need to have infants in the age range of relevant outcomes for our study (EF at 36 months) in the following four years within the funding period. Thus, extending the recruitment period beyond 14 months would have been problematic, jeopardizing the data collection of outcomes because of financial constraints.
In addition, our inclusion criteria excluded preterm, low-birthweight infants, and adolescent mothers, characteristics that are associated with socio-demographic factors related to SES, such as low total years of schooling, low income, and inadequate prenatal care (J. D. Victora et al., Reference Victora, Silveira, Tonial, Victora, Barros, Horta, Santos, Bassani, Garcia, Scheeren and Fiori2018). This may have biased our recruitment process towards enrollment of higher SES families. Moreover, mother-infant interaction variables were not included in our model due to this data not being available at the time of analysis, because of the time consuming nature of behavior coding procedures. This domain has a crucial influence in infant development during the first year of life (Rocha et al., Reference Rocha, Dos Santos Silva, Dos Santos and Dusing2020), thus it could have changed our findings.
Also, our model did not include aspects related to father involvement. We collected limited information regarding the father because there is limited availability of validated measures for father involvement in the Brazilian context. While mothers often play a central role, it is important to recognize that other caregivers, such as fathers and grandparents, also contribute significantly to development by providing cognitive stimulation. However, it must be noted that in Brazil, census data reveals that in households with children up to age 4, mothers constitute the primary caregivers in 84% of cases (IBGE, 2017).
Additionally, our analytical strategy precluded determining which specific variables within domains had a stronger influence on our outcome. ddsPLS prioritizes variable selection based on covariance related to the outcome variable, rather than individual predictive power. This focus on maximizing covariance can lead to the selection of variables with strong collective influence on the outcome, but weak individual predictive ability. Thus, isolating and ranking the predictive importance of individual variables becomes challenging.
In conclusion, studies such as Germina, which are capable of capturing the complexity of early child development by integrating different data modalities, have the potential to move the field forward and bring new insights into the emergence of EF and LS. We leveraged data from a cohort of mother-infant dyads, analyzing the impact of 15 data domains on infant development measured by the Bayley-III at three months of age. Our integrative model indicated that socioeconomic, demographic, epigenetics, EEG, and microbiome significantly predicted infant development and ultimately demonstrated the feasibility of implementing the planned models. The results can further delineate the predictors of EF and LS during the first three years of life.
Supplementary material
For supplementary material accompanying this paper visit https://doi.org/10.1017/S0954579425000069
Acknowledgments
We would like to thank all the families involved in our study for their ongoing support. We also would like to thank our team of assessors: Ana Luiza Polimeno, Camila Deak, Fabiana Barbosa, Julia Nicolella, Lais Maciel, Márcia Lima, Mariane Tinti, Natalia Ablen, Nathalia Moravski. We would like to thank the following institutions responsible for allowing us to set up laboratories for data collection: Instituto de Psiquiatria, Liga Solidária e Educandário Dom Duarte, and Centro de Pesquisa sobre o Genoma Humano e Células-Tronco. We acknowledge the support of the following institutions in our recruitment process: Hospital Israelita Albert Einstein, Hospital Municipal Santa Catarina, and Centro de Saúde Escola Samuel Barnsley Pessoa.
Funding statement
Our work is supported by the Wellcome Leap 1kD Program. G.C.G was also supported by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) and Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) (440245/2022-2); A.C.C. by Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP) (2023/04619-7); E.S. by FAPESP (20/05964-1); D.C.P.A. by CAPES (88,887.887517/2023-00); D.C.P.A. and A.C.M.S. by Fundação de Apoio ao Ensino, Pesquisa e Assistência do Hospital das Clínicas da Faculdade de Medicina de Ribeirão Preto da Universidade de São Paulo.
Competing interests
GVP has been in the past 3 years a member of advisory board of Shire/Takeda and Medice and a speaker for Shire/Takeda, Novo Nordisk, and Aché. He received travel expenses for continuing education support from Shire/Takeda and royalties from Editora Manole. The other authors declare no conflict of interest.
Data availability statements
The data necessary to reproduce the analyses presented here are not publicly accessible. The analytic code necessary to reproduce the analyses presented in this paper is available upon request. The materials necessary to attempt to replicate the findings presented here are not publicly accessible. The analyses presented here were not pre-registered.
