Neurobehavioural links from infant screen time to anxiety

Summary

Background

Infant screen time is linked to many negative outcomes, including anxiety, but the underlying neural correlates and pathways remains understudied. We aimed to assess the directional association between infant screen time, development of brain network topology, decision-making behaviour and anxiety symptoms in adolescence.

Methods

Using data from the Growing Up in Singapore Towards healthy Outcomes (GUSTO) cohort study, we examined the effects of total daily screen time for infants on developmental outcomes using structural equation modelling. Specifically, we looked at the developmental slopes of network integration for the seven major brain cortical networks between ages 4.5, 6.0, and 7.5, decision-making behaviour assessed using the Cambridge Gambling Task (CGT) and anxiety symptoms assessed using the Multidimensional Anxiety Scale for Children, 2nd Edition (MASC).

This study included 168 children from the GUSTO cohort with data on infant screen time (ages 1–2), diffusion MRI (ages 4.5–7.5), data on decision-making performance (CGT at age 8.5), and anxiety symptoms (MASC at age 13). Brain network integration was derived from diffusion MRI and each participant's developmental slopes were modelled using latent growth models. Structural equation modelling assessed pathways linking early screen time to adolescent anxiety, mediated by brain network development and decision-making.

Findings

Higher infant screen time was associated with a steeper decline in visual-cognitive control network integration from ages 4.5–7.5 years (β = −1.03 (−1.61, −0.46)), which mediated increased CGT deliberation time at age 8.5. Deliberation time, in turn, was associated with greater anxiety symptoms at age 13. A full serial mediation pathway was significant, linking infant screen time to later anxiety via accelerated brain network maturation and decision-making behaviour (β = 0.033 (0.002, 0.160)).

Interpretation

Higher infant screen time is linked to accelerated topological maturation of the visual and cognitive control networks, leading to prolonged decision latency and increased adolescent anxiety. Sensory processing impairment may underlie this novel neurodevelopmental pathway, highlighting a potential target for early intervention.

Funding

This research was supported by the Singapore National Research Foundation, Singapore Ministry of Health’s National Medical Research Council (NMRC), Singapore, Institute for Human Development and Potential, Agency for Science, Technology and Research, Singapore, the Hope for Depression Research Foundation, USA, the Toxic Stress Network of the JPB Foundation, USA, and the Jacobs Foundation, Switzerland.

Research in context.

Evidence before this study

Global screentime exposure has been steadily increasing, even among young children and infants. This is a growing area of concern, being highlighted by the WHO, US Surgeon-General, as well as local advisory bodies. Screentime exposure has been associated with many negative outcomes, such as anxiety and poorer emotional regulation. However, the neural and behavioural correlates underlying these pathways are understudied.

Added value of this study

Using a longitudinal birth cohort, we demonstrate that higher levels of screentime exposure in infancy accelerates the rate of development of the visual and cognitive control networks in childhood. This accelerated maturation is further associated with slower decision-making and elevated anxiety problems in adolescence. Our results demonstrate a plausible pathway through which the negative impact of infant screentime is operationalised and show that infant screentime has severe and long-lasting consequences on the child's neurodevelopment and behavioural outcomes.

Implications of all the available evidence

Elevated infant screentime may affect the child's later neurodevelopment and exacerbate the risk of future mental health problems. While screens are becoming increasingly ubiquitous in our daily lives, care must be taken to minimise the child's exposure, especially during critical developmental windows.

Introduction

Screens are pervasive in our modern society, and current trends suggest that screen use will only continue to rise. Globally, the amount of time young children spend on screens has increased.¹ Infants generally engage in two to three hours of screen time daily,² far exceeding the recommendation by WHO. These high levels of screen time is an area of growing concern. The US Surgeon-General published an advisory in 2023³ which highlighted growing concerns about the mental health impact of screen time, supporting delaying children's exposure to social media and encouraging parents to set limits and monitor usage. Locally in Singapore, the Ministry of Health published a similar advisory in 2023, further updated in 2025,⁴ emphasising the importance of limiting screen time and evidence of negative outcomes associated with excessive screen time.

These high levels of screen time are concerning as the first two years of life are critical for brain development. Brain volume doubles in the first year and thereafter increases by an additional 15%–80% of its final volume in the second year.⁵ Higher-order brain networks have also been shown to develop dramatically in the first few years of life, with long-range synchronisation and increased network connectivity in the first year and increased segregation of networks in the first two years of life.⁶ These profound brain changes during infancy lay the critical foundation for future brain development, with enduring effects on future behaviours and cognitive capacities.

There is a growing body of evidence demonstrating the impact of screen use on brain development throughout childhood. For instance, Hutton et al. found that excessive screen usage during childhood has detrimental effects on white matter tracts that support language and reading abilities.⁷ In a separate study, high levels of screen time in childhood have been linked to decreased fractional anisotropy in the medial lemniscus tract.⁸ More recently, a study from our group revealed that infant screen time is significantly associated with the degree of integration between the cognitive control and emotion processing networks.⁹ However, most of these studies rely on cross-sectional neuroimaging data, which only allows us to capture a single snapshot of brain growth. Brain development across childhood is a progressive and dynamic process, and examining brain changes over time using longitudinal neuroimaging data is crucial.¹⁰

The majority of published work on the effect of childhood screen use on brain development focuses on changes at the macrostructural and microstructural levels.^7,8,11 These regional and tract-based changes, although important, are less useful when studying complex constructs such as decision-making behaviour and anxiety. These complex behavioural constructs rely on effective interaction between numerous brain networks, and the development of these brain networks is particularly sensitive to environmental exposures.¹² Hence, network topology metrics are ideal for investigating the potential neural correlates that link infant screen use to subsequent behavioural outcomes. Brain network topology is being increasingly recognised as a crucial determinant of complex brain function¹³ and has been linked to various neurodevelopmental and mental health conditions.14, 15, 16, 17

Decision-making constitutes a higher-order cognitive function that integrates sensory, cognitive, and affective processes, thereby serving as a theoretically relevant construct with the potential to mediate the association between screen time and subsequent anxiety symptoms. Converging evidence shows that children and adolescents with anxiety display atypical decision-making profiles,¹⁸ suggesting that maladaptive decision-making both contributes to and is exacerbated by anxiety symptoms.¹⁹ Behavioural studies and theoretical models have demonstrated that anxiety influences decision-making behaviour.20, 21, 22 Notably, neuroimaging studies have also revealed overlapping brain structures involved in decision-making and anxiety.^23,24 Importantly, decision-making abilities consolidate during mid-childhood, when our behavioural assessments were conducted, making it a developmentally appropriate pathway linking early screen exposure, neural network maturation, and later anxiety symptoms.

We leveraged data from the deeply phenotyped Growing Up in Singapore Towards healthy Outcomes (GUSTO) birth cohort to carry out an exploratory analysis examining the link among infant screen time, developmental trajectories of brain network topology, decision-making behaviour in mid-childhood, and anxiety symptoms in adolescence, using a structural equation modelling (SEM) approach (Fig. 1). Data on infant screen time were collected via parental questionnaires at ages 1 and 2 years. We utilised our unique longitudinal neuroimaging dataset obtained at ages 4.5, 6.0, and 7.5 years to model the developmental trajectories of structural network topology, employing latent growth models to derive an intercept and a slope variable for each network topology measure. The longitudinal neuroimaging data utilised in our study is highly notable as it allows us to distinguish between-subject and within-subject variability, in contrast to the more commonly used approach of estimating neurodevelopmental trajectory from cross-sectional imaging data, adopting age as a proxy for time.²⁵ This was followed by the assessment of decision-making behaviour at age 8.5 using the Cambridge Gambling Task (CGT) and the assessment of anxiety symptoms at age 13 using the Multidimensional Anxiety Scale for Children, 2nd Edition (MASC). To our knowledge, no previous study has investigated the association among infant screen use, brain development, and behavioural outcomes at multiple later timepoints. Our primary hypothesis is that infant screen time would predict alterations in the developmental trajectory of network integration between some brain regions. As a follow-up hypothesis, we postulate that these neural alterations would in turn relate to decision-making behaviour, which would subsequently predict anxiety symptoms in adolescence.

Fig. 1.

Study design and aims. Diffusion data at each timepoint was processed independently and used to generate the structural connectivity for a 114-region cortical parcellation. This structural connectivity matrix was clustered using a community Louvain algorithm to calculate the network integration metric for each pair of the seven overarching networks. Each individual network integration measure was combined longitudinally, and a slope and intercept estimate for each integration measure was obtained using latent variable modelling. Our study examined the longitudinal relation between infant screen time, derived brain network trajectory, decision-making behaviour and anxiety in adolescence.

Methods

Participants

Participants were recruited from the Growing Up in Singapore Towards healthy Outcomes (GUSTO) cohort,²⁶ an ongoing longitudinal birth cohort in Singapore that aims to investigate developmental influences on later life health outcomes. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines for cohort studies.²⁷

Women aged 18 years or older in their first trimester of pregnancy across all socioeconomic backgrounds were recruited from two main public hospitals in Singapore between June 2009 and December 2010 for the GUSTO cohort.²⁶ Mother and child were followed up during and after pregnancy, with mothers having signed informed consent and children assenting at 7 years old. Only children with gestational age at birth ≥34 weeks, birth weight ≥2000 g, and a 5-min Apgar score ≥8 were included to minimise effects of potential birth complications on brain development. A summary of cohort demographics is outlined in Table 1. This study included 168 GUSTO-enrolled children (54.2% boys) with both screen time measures and diffusion MRI data (see Supplementary Figure S1 for exclusion flowchart).

Table 1.

Summary of demographics for study cohort.

	n	Mean	S.D.
Sex	923
Male	483
Female	440
Environment latent Variable	450	0.00	0.78
Maternal Age at Birth	485	30.76	5.17
Household Income (SGD)	464
< 2000	80
2000–5999	253
> 6000	131
Highest Maternal Education	482
Primary	29
Secondary/Technical	188
GCE ‘A’ levels/University	265
Screen time (hrs per day)
Y1	388	1.64	1.52
Y2	848	2.53	2.15
Y3	861	2.84	2.21
Y4	642	1.52	1.33
Average (Y1–Y2)	905	2.17	2.01
Neuroimaging Data
Y4.5	350
Y6.0	416
Y7.5	497
Present for all 3 timepoints	179
Cambridge Gambling Task
Delay Aversion	439	0.21	0.35
Deliberation Time (s)	439	3.87	1.63
Overall Proportion Bet	439	0.51	0.17
Quality of Decision-making	439	0.79	0.18
Risk adjustment	439	0.32	1.01
Risk Taking	439	0.54	0.18
MASC Total Score	661	56.12	26.57

Ethics

All investigations were conducted according to the principles expressed in the Declaration of Helsinki and approved by the National Healthcare Group Domain Specific Review Board (D/2009/021 and B/2014/00411) and the SingHealth Centralised Institutional Review Board (D/2018/2767 and A/2019/2406). Written informed consent was obtained from all guardians on behalf of the children enrolled in this study.

Screen time measures

Data on infant screen time were collected via study-specific parental questionnaires at ages 1 and 2 years. Parents were asked to estimate the average screen time for a typical weekday and a typical weekend in the past week. For each timepoint, average daily screen time was calculated by taking a weighted mean of the weekday and weekend screen time estimates. Our primary exposure variable, infant screen time, was calculated by taking the mean of the screen time from the two timepoints.

Cambridge Gambling Task

The Cambridge Gambling Task (CGT),²⁸ a sub-test of the Cambridge Neuropsychological Test Automated Battery (CANTAB), was administered at age 8.5 years to assess decision-making behaviour. In this task, ten boxes are presented in varying ratios of red and blue. Participants are asked to decide whether the yellow token is hidden in a red or blue box and stake a proportion of their cumulative points on their choice being correct. The stake is added to their total points if the choice is correct or subtracted if it is incorrect. Participants are instructed to gain as many points as possible.

This task measures six different facets of decision-making: (1) quality of decision-making, (2) impulsivity/delay aversion, (3) risk-taking, (4) deliberation time, (5) risk adjustment, and (6) overall proportion bet. Quality of decision-making measures the percentage of trials where the participant chooses the correct colour (i.e. the colour with the majority of boxes). Delay aversion measures the difference in average betting ratios across ascending and descending conditions but only include trials where the participant chooses the correct colour. Risk-taking measures the average betting ratio across all trials where the participant chooses the correct colour. Deliberation time is the time between the start of the trial and the bet choice. Risk adjustment represents the tendency for participants to increase their bets when the odds are better. Lastly, overall proportion bet records the average betting ratio across all trials, regardless of the participant's accuracy. These six measures are the standard outcomes of the CGT²⁹ and examine different facets of decision-making. These measures have been widely used in studies utilising CGT.^30,31

As decision-making is a multidimensional construct with no prior evidence pointing to a single Cambridge Gambling Task (CGT) parameter as the primary mediator of the screen time-anxiety association, we included all six standard CGT parameters in our analyses. This comprehensive approach allowed us to capture the full spectrum of decision-making behaviour while permitting the data to identify the most relevant construct. To minimise the risk of false positives, all six parameters were tested within the same model, and our inferences were restricted to associations that remained robust under this framework.

Multidimensional Anxiety Scale for Children

The Multidimensional Anxiety Scale for Children, 2nd Edition (MASC) was administered at age 13 years. The MASC is a self-reported questionnaire consisting of 50 items, scored on a four-point Likert scale. The items are grouped into six subscales (separation anxiety, generalised anxiety disorder, obsessions and compulsions, harm avoidance, social anxiety, and physical symptoms), with a total score summing all six scales. In this study, we used the total score as a general measure of anxiety.

Sociodemographic data

Sociodemographic data were collected from parental questionnaires. A confirmatory factor analysis was performed with the lavaan package in R, version 4.4.1³² to construct a latent environment variable (ENV) from three observed variables: (1) household income, (2) maternal education, and (3) maternal age at recruitment. All three variables significantly contributed to the latent variable (all p < 0.001), with standardised estimates reported in previous studies.^9,33,34 Model fit indices indicate an excellent fit (Comparative Fit Index (CFI) = 1.00, Standardised Root Mean Square Residual (SRMR) = 0.00). This latent environment variable was included as a covariate of no interest in all analyses.

MRI acquisition

Participants at age 4.5 (mean: 4.58, standard deviation: 0.08) and 6.0 (mean: 6.04, standard deviation: 0.12) years underwent magnetic resonance imaging (MRI) of the brain utilising a 3-T scanner (Magnetom Skyra; Siemens, Germany) at KK Women's and Children's Hospital, Singapore. A third scan session was performed at age 7.5 (mean: 7.46, standard deviation: 0.14) years utilising a 3-T (Magnetom Prisma; Siemens, Germany) at Centre for Translational MR Research, National University of Singapore.

On both scanners, the exact same sequence parameters were used to acquire T1 MPRAGE (Magnetization Prepared - Rapid Gradient Echo) and diffusion-weighted images. 3D T1-weighted MPRAGE images were acquired with the following imaging parameters: repetition time = 2000 ms, echo time = 2.08 ms, field of view = 192 × 192 mm², matrix size = 192 × 192, number of slices = 160, slice thickness = 1 mm. Diffusion-weighted images were acquired using a single-shot echo planar imaging (EPI) sequence with the following imaging parameters: field of view = 192 × 192 mm², voxel size = 2 mm isotropic, repetition time = 8200 ms, echo time = 85 ms, flip angle = 90°, 30 non-collinear directions, b values = 0, 1000 s/mm², acceleration factor = 3.

MRI data processing

Preprocessing

The diffusion data was mainly processed using FMRIB's Software Library (FSL, v6.0).³⁵ For diffusion data acquired at 4.5 and 6.0 years, eddy correction was carried out using an older version (FSL v5.0.2.2). FMRIB's eddy tool³⁶ iteratively predicts and estimates the eddy-current distortions and performs a single resampling of the data including input from the susceptibility estimates. At 7.5 years, TOPUP correction was applied prior to eddy. Specifically, the b0 volumes from the diffusion dataset was extracted and concatenated with the reverse phase-encoded images to estimate the susceptibility induced distortions using the TOPUP tool in FSL.³⁷ Next, images were corrected for eddy currents which include susceptibility-by-motion interactions. Simultaneously, outlier slices are replaced with a Gaussian Process prediction with a set threshold of 4 standard deviations. The corrected data from all timepoints is then skull-stripped once more to provide a brain-tissue-only image which is further de-noised using a local PCA method, which reduces random noise from the acquisition process. This cleaned image is used for subsequent analysis. The eddy generated average absolute motion estimates over all volumes were generated using FMRIB's eddy_quad function and used to assess MRI motion. To maximise power in our analysis, no participants were excluded based on motion parameters. Instead, a sensitivity analysis was carried out on a subset of participants that passed motion criteria to demonstrate that the findings were independent of motion (Supplementary Table S1).

Probabilistic tractography

Next, within-voxel probability density functions of the principal diffusion direction were estimated using Markov Chain Monte Carlo sampling in FSL's BEDPOSTX tool.³⁸ A spatial probability density function was then estimated using FSL's PROBTRACKX tool³⁸ in which 5000 samples were taken for each input voxel with a 0.2 curvature threshold, 0.5 mm step length, and 2000 steps per sample. Cortical regions were parcellated into 114 individual ROIs, as described in Yeo et al.³⁹ These 114 ROIs can be grouped into seven major resting state networks based on their functional connectivity profiles (Supplementary Table S2). Pairwise structural connectivity was obtained from the total number of tractography streamlines connecting the two ROIs. This procedure generated a 114 × 114 structural connectivity matrix, which was used subsequently to generate measures of brain network topology. We also regressed out the effect of ROI volumes for each pairwise structural connectivity. This process was repeated independently for all three timepoints.

Computation of brain network topology

A tuned Louvain community detection algorithm⁴⁰ was used to cluster the ROIs into community clusters. Specifically, the Louvain community detection algorithm generated a group assignment for each ROI, which was then used to create an allegiance matrix for each of the 114 ROIs to one of the 7 major cortical networks (visual, somatomotor, salience, dorsal attention, cognitive control, default mode, and limbic networks).^9,41 This analysis was iterated 100 times to generate a stable average allegiance matrix for estimation of network integration. The integration coefficient is defined as the probability of a region being assigned to the same community as regions from another network. These network topology measures were normalised using the distribution from 10,000 iterations with randomly permuted structural connectivity matrices to account for differences in the number of regions for each network.

Data harmonisation

As the third timepoint data was collected at a different imaging site, longitudinal comBAT (v0.0.0.90)⁴² was used to harmonise the data across timepoints. Summary of regions with significant site effects pre-comBAT harmonisation are tabulated in Supplementary Table S3.

Latent growth models of brain trajectory

We used latent growth modelling⁴³ to model the change in network integration over time, generating an intercept and gradient (slope) variable for each of the 21 measures of network integration (Supplementary Table S4). A weight of 1 was assigned for the effect of the intercept and a weight of (age - 2) was assigned for the slope effect on the network measures at each timepoint. This is to position the intercept at 2 years, where we presume to be the start of divergent trajectories due to infant screen time. Latent growth modelling is a type of structural equation modelling (SEM) used in longitudinal studies to model and assess the growth trajectory of a variable over time.

Rather than focussing on specific brain regions or networks, we adopted a whole-brain approach, based on the premise that screen time may exert diffuse effects on overall brain development. Moreover, because our constructs of interest, namely decision-making and anxiety, are underpinned by multiple brain regions and networks, a broad-based approach was both necessary and conceptually justified.

Statistics

All statistical analysis were carried out in R (v4.4.1). All SEM and latent growth models were conducted using the lavaan package³² in R. All models were bootstrapped 10,000 times to estimated bias-corrected confidence intervals. Furthermore, beta coefficients were standardised to assist with interpretation of effect sizes. Spearman correlations between all primary variables are tabulated in Supplementary Table S5. Only participants with complete data were used for each step of the analysis (Supplementary Figure S1). Our derived latent environment variable (ENV) was included as a covariate in all analyses.

• (i)Infant screen time as a predictor of brain network development.

We included infant screen time in the latent growth model of each network integration measure to test for significant association with the slope variables. Bonferroni correction was applied to ensure overall rate of false positives remains at 5% despite multiple comparisons. Slope variable(s) that are significantly associated with infant screen time would be incorporated as candidate mediators in subsequent analysis.

• (ii)SEM analysis: Mediating role of brain network development.

We used a SEM approach to test the associations between infant screen time, slopes of brain network integration, and all six CGT outcomes. Specifically, we tested whether the CGT outcomes were significantly correlated with infant screen time either directly or indirectly via the slopes of brain network integration from (i).

• (iii)Correlation analysis between CGT outcomes and anxiety symptoms in adolescence.

Next, we examined if anxiety symptoms were associated with any of the CGT outcomes using Spearman correlation as not all the data was normally distributed. This was done with all six CGT outcomes and Bonferroni's correction for multiple comparisons was carried out.

• (iv)SEM analysis: Pathway from infant screen time to anxiety symptoms in adolescence.

Lastly, we carried out a serial mediation analysis, whereby significant slopes and CGT outcomes from previous analyses were used as mediators in the association between infant screen time and anxiety symptoms in adolescence.

Role of funders

The funders played no role in the study design, data collection, analysis, interpretation, or the writing of the manuscript. None of the authors have been paid to write this article by a pharmaceutical company or other agency.

Results

Infant screen time as a predictor of brain network development

Infant screen time was found to be significantly associated with the slope of the visual-cognitive control network integration (β = −1.03 (−1.61, −0.46), p < 0.0003, bootstrapped permutation testing) (Fig. 2). Specifically, higher infant screen time was associated with a steeper decrease in visual-cognitive control network integration between ages 4.5 and 7.5 years, suggesting an element of accelerated maturation. This result remained significant after Bonferroni's correction for multiple comparisons. No other network integration slopes were significantly associated with infant screen time (Supplementary Table S6). Notably, follow-up analysis showed that screen time at age 3–4 years was not significantly associated with all 21 network integration slopes (Supplementary Table S7). This highlights the critical impact of infant screen time on brain network development compared to screen time between ages 3 and 4.

Fig. 2.

Latent Variable Modelling of the trajectories of brain network integration. A) Only the slope of the visual-cognitive control network integration was significantly correlated to infant screen time. All other slopes and intercepts were not correlated with infant screen time (Supplementary Table S6). Bootstrapped permutation testing was carried out to test for significance. B) Scatterplot (n = 168) showing the association between the slope of the visual-cognitive control network integration and infant screen time. The solid line represents the line of best-fit and the shaded area indicates the 95% confidence interval. C) Trajectories of visual-cognitive control network integration across time for children with reduced (>1 S.D. below mean, blue, n = 32) and elevated (>1 S.D. above mean, red, n = 27) screen time. Individual lines represent individual subjects and the thick lines with error bars represent population average.

SEM analysis: mediating role of brain network development

We utilised a mediation analysis to examine whether the visual-cognitive control network integration slope significantly mediated the association between infant screen time and the six CGT outcome measures. Only deliberation time was significantly associated with the visual-cognitive control network integration slope (β = −0.27 (−0.71, −0.05), bootstrapped permutation testing). The visual-cognitive control network integration slope fully mediated the association between infant screen time and deliberation time (β = 0.20 (0.01, 0.91), bootstrapped permutation testing) (Fig. 3). Specifically, higher infant screen time was associated with a steeper decrease in visual-cognitive control network integration, which is subsequently associated with increased deliberation time. Full results are provided in Supplementary Table S8.

Fig. 3.

SEM analysis on the mediating role of the slope of the visual-cognitive control network integration on the association between infant screen time and the six CGT outcomes. Increased infant screen time is significantly associated with a sharper decrease of the visual-cognitive control network integration, which is in turn associated with increased deliberation time only. No other CGT outcomes were significant (Supplementary Table S8). Bootstrapped permutation testing was carried out to test for significance. SEM model fit parameters indicate an excellent fit (CFI = 1.00, RMSEA = 0.0002, SRMR = 0.037).

Correlation analysis between CGT outcomes and anxiety symptoms in adolescence

We tested the correlation between the 6 CGT outcomes and anxiety symptoms as indexed by the total score in MASC. Only deliberation time was significantly correlated with total score in MASC (r = 0.20, p = 0.006, Spearman's correlation). All other correlation values are tabulated in Supplementary Table S5.

SEM analysis: pathway from infant screen time to anxiety symptoms in adolescence

We constructed a serial mediation model to analyse the pathway between infant screen time and anxiety symptoms at age 13, illustrated in Fig. 4. The slope of the visual-cognitive control network integration and CGT deliberation time were used as candidate mediators. Only the full mediation pathway was significant (β = 0.03 (0.01, 0.13), bootstrapped permutation testing). The direct and both partial indirect pathways were not significant (Supplementary Table S9). These results show that higher infant screen time is linked to a steeper decrease in the slope of the visual-cognitive control network integration between 4.5 and 7.5 years, leading to increased deliberation time at age 8.5 and finally elevated anxiety scores on the MASC at age 13.

Fig. 4.

Full SEM analysis on the pathways between infant screen time and anxiety as indexed by the Multidimensional Anxiety Scale for Children, 2nd Edition (MASC), with visual-cognitive control network integration slope and Cambridge Gambling Task Deliberation Time as serial mediators. Full pathway estimates are presented in Supplementary Table S9. Bootstrapped permutation testing was carried out to test for significance. SEM model fit parameters indicate an excellent fit (CFI = 1.00, RMSEA = 0.0003, SRMR = 0.044).

Discussion

Infant screen time has been linked to a wide range of negative outcomes, including internalising disorders such as anxiety and depression,⁴⁴ but the underlying pathways and mechanisms of effect are unknown. The levels of infant screen use observed in our study cohort are concerning, both in their magnitude and in their clear divergence from WHO recommendations of no screen exposure before age 2. By ages 1 and 2 years, children were already exposed to more than one and two hours of daily screen time, respectively. This context underscores the public health significance of our study, as identifying the neurodevelopmental pathways that link early screen exposure to later socioemotional outcomes is critical in an environment of rapidly escalating screen use. Furthermore, these figures likely understate current exposures, as our data (collected between 2010 and 2014) precede recent evidence of further increases globally, particularly during the COVID-19 pandemic.^45,46 Thus, the already elevated levels of screen exposure we observed a decade ago are likely even higher today, making the developmental implications of our findings especially urgent.

In our study, we present a plausible neurodevelopmental pathway linking infant screen time with later decision-making behaviour and anxiety symptoms, operationalised and propagated through changes in brain network development across early- to mid-childhood. Our findings provide novel evidence for an association between infant screen time and the developmental trajectory of the visual-cognitive control network integration, whereby higher levels of screen time were associated with a steeper decrease in network integration between ages 4.5 and 7.5 years. During this period of development, an increase in within-network recruitment and a decrease in between-network integration are expected, reflecting the specialisation of brain regions to their respective functions.⁴⁷ Hence, a faster decrease in visual-cognitive control network integration may reflect accelerated maturation of these brain networks, possibly related to sensory overstimulation from high levels of infant screen time. Notably, we also found that the developmental trajectory of the visual-cognitive control network integration mediated the association between infant screen time and deliberation time in decision-making, and as hypothesised, these subsequently led to higher levels of anxiety symptoms in adolescence.

The association between screen time in early childhood and developmental brain changes is well-established,^7,9,11,48 with multiple studies reporting the involvement of brain regions critical for sensory processing and cognitive functioning. For example, higher screen time in early childhood was found to be associated with lower cortical thickness in primary visual areas and the lingual gyrus, a higher-order visual-association area. At the behavioural level, screen time in early childhood was reported to be associated with developmental delays in multiple cognitive domains49, 50, 51 and abnormal sensory processing.^52,53 It is undeniable that sensory processing and cognitive functioning are closely intertwined. Based on Mesulam's model of sensory processing organisation,⁵⁴ sensory input undergoes a decomposition process as an unimodal input converges on the trans-modal cortex. This process has been proposed to be associated with the development of higher-order cognitive functions.⁵⁵ This literature corroborates our study's result that infant screen time is associated with topological organisation of the visual and cognitive control networks, which may later contribute to a child's higher order cognitive functioning such as decision-making.

Specifically, in our study, the developmental trajectory of the visual-cognitive control network integration was found to mediate the association between infant screen time and deliberation time, assessed using CGT, a visually oriented neurocognitive task. Different individuals have different signal thresholds prior to making a decision,⁵⁶ which contributes to the inter-individual variability in decision latency. The threshold theory of decision-making attributes variations in deliberation time to the growth rate in frontal neuronal activation toward the fixed threshold.⁵⁷ As the achievement of signal threshold is highly reliant on the processing of sensory input, it is indeed not surprising that altered development in the topological organisation of the visual and cognitive control networks would contribute to inter-individual variations in decision latency. It is important to note that in our study, we did not observe a significant association between deliberation time and the quality of decision-making (Supplementary Table S5). This may be related to the fact that in the CGT, the increase in deliberation time will not result in an increase in information needed for better decision-making as the quantity of information presented is fixed.

Sensory processing impairment has also been linked to various forms of anxiety disorders58, 59, 60 and may constitute a fundamental component that integrates findings from our serial mediation model. High levels of infant screen usage may disrupt the development of visual and cognitive control networks, leading to sensory processing impairments that could impact both deliberation time and anxiety symptoms. Notably, there exists an overlap of brain regions implicated in decision-making behaviour and anxiety,^23,24 providing additional evidence of their association and potentially a common neural pathway; the topological maturation of brain networks implicated in sensory processing. However, it remains a possibility that the developmental impact of infant screen use may be indirect through the displacement of parent-child interaction, as parent-child interaction and sensory processing are also closely intertwined.⁶¹ In the ALSPAC cohort, screen use was found to be associated with anxiety, but this association is diminished when time spent alone is considered as a covariate.⁶² This finding underscores the influence of displaced social interaction as a confound when assessing the effects of screen use.

Several study limitations need to be considered when interpreting our findings. Firstly, brain network topology, decision-making behaviour, and anxiety symptoms were not assessed at every developmental stage. Although this prevents definitive conclusions about bidirectional effects, the temporal spacing of our measures, from infant screen exposure, to brain development in early childhood, to decision-making in mid-childhood, and anxiety in adolescence, provides a plausible developmental ordering. Future work should incorporate repeated measurements across development to better test potential bidirectional relationships. Furthermore, as this was an exploratory study exploring an observational cohort, the results should be interpreted as plausible pathways of effect of infant screentime and does not preclude alternative pathways or mechanisms of effect of infant screentime. Another limitation is the unidimensional emphasis on screen time without consideration for the content and context of viewing. Future study should seek to differentiate the impact of media content quality on brain development and cognitive outcomes. In our analysis, we only controlled for maternal age, maternal education level, and total household income. Other confounding factors, including diagnostic family history, parent-child interaction, and sleep quality, were not included as the measures were either not collected or incomplete. However, prior work has demonstrated that socioeconomic indicators such as parental education and income exert broad and powerful effects on children's neurodevelopment, often encompassing variance explained by more proximal factors. Lastly, only linear slopes of brain development trajectories were examined as data was only available at 3 different timepoints. Future studies should target to include more timepoints to test non-linear trajectories of brain development.

Higher levels of infant screen time are linked to accelerated topological maturation of the visual and cognitive control networks, possibly related to sensory overstimulation. This leads to prolonged decision latency and higher levels of anxiety in adolescence. Hence, public health strategies to reduce infant screen time may yield significant beneficial effects on children's cognitive and mental health outcomes. Notably, sensory processing impairment may underlie this neurodevelopmental pathway and could serve as a potential target for intervention. However, further mechanistic studies are warranted to substantiate this finding. It is imperative that the significance of these discoveries not be understated because they represent important knowledge advancement with implications for prevention and intervention.

Contributors

PH: data curation, formal analysis, investigation, methodology, writing-original draft. SYC: data curation, investigation, writing-review and editing. KXZ: investigation, writing-original draft. JC: data curation, formal analysis, investigation, methodology. AMAM: data curation, formal analysis, investigation, methodology. ECNL: investigation, project administration. SS: funding acquisition. HJZ: data curation. MVF: data curation, funding acquisition. YSC: funding acquisition. MJM: funding acquisition, supervision, writing-review and editing. APT: formal analysis, investigation, funding acquisition, supervision, writing-review and editing. PH and APT have accessed and verified the underlying data. All authors read and approved the final manuscript.

Data sharing statement

The GUSTO cohort data is part of a multi-site collaboration and hence, is not able to be shared openly. The data can be made available upon reasonable request to the GUSTO executive committee. All software and code developed in-house can be made available upon reasonable request to the corresponding author.

Declaration of interests

ECL received grant from MOH-001784-00: Promoting early relational health with PlayReadVIP to prevent socioeconomic disparities in child development (unrelated to the manuscript) and works in an unpaid role KidSTART Ltd (Charity status), International Community School. All other authors declare that they have no conflicts of interest.