Charting brain functional development from birth to 6 years of age

Abstract

Early childhood is crucial for brain functional development. Using advanced neuroimaging methods, characterizing functional connectivity has shed light on the developmental process in infants. However, insights into spatiotemporal functional maturation from birth to early childhood are substantially lacking. In this study, we aggregated 1,091 resting-state functional MRI scans of typically developing children from birth to 6 years of age, harmonized the cohort and imaging-state-related bias, and delineated developmental charts of functional connectivity within and between canonical brain networks. These charts revealed potential neurodevelopmental milestones and elucidated the complex development of brain functional integration, competition and transition processes. We further determined that individual deviations from normative growth charts are significantly associated with infant cognitive abilities. Specifically, connections involving the primary, default, control and attention networks were key predictors. Our findings elucidate early neurodevelopment and suggest that functional connectivity-derived brain charts may provide an effective tool to monitor normative functional development.

Main

Normative growth charts are widely used to monitor biometric parameters (for example, height, weight and head circumference) and serve as highly effective screening tools to identify individuals with atypical growth during early childhood¹. Deviation from the normative growth charts indicates potential abnormal development so that early interventions can be implemented to diminish adverse long-term sequelae¹. Head circumference, for instance, has been commonly used to monitor brain size and growth during the first years after birth². Head circumference values, however, do not provide insights into the underlying neural substrates leading to deviations from normal head circumferences. The human brain undergoes highly temporal dynamic maturation processes during early childhood, and the pace of maturation varies among those with different cognitive abilities³. Specifically, remarkable global and regional structural brain changes and maturation of functional networks occur^4,5,6. It has been proposed that these neural maturation processes are associated with behavioural developmental milestones between early infancy and childhood, including improved visual^7,8, sensorimotor⁹ and language ability⁹, as well as executive functions^3,10. Therefore, it is of vital importance to chart early brain development as a step toward understanding the emergence of cognitive ability.

A seminal study has recently reported lifespan growth charts of brain structure, such as brain volume, surface area and cortical thickness, revealing essential neuroanatomical milestones¹¹. Notably, altered individual centile scores, which were estimated using the established reference curves, were significantly associated with neuropsychiatric disorders. These important findings suggest that neuroanatomical charts might act as a standardized tool to monitor atypical neuroanatomical variations. However, the development of growth charts for brain functional characteristics remains less explored, which could yield complementary insights to those provided by neuroanatomical charts.

Resting-state functional MRI (rsfMRI), which assesses the temporal synchrony of blood oxygen level-dependent contrast signals among different brain regions, has been extensively used to characterize brain functional topologies¹². Unlike task-based fMRI, which requires participants to perform tasks or receive external inputs—often impractical for infants and young children—rsfMRI captures time-series images while participants are at rest^13,14. Therefore, it is not surprising that rsfMRI has been widely used as a highly effective and non-invasive tool for characterizing brain functional development during early infancy and childhood^{14,15,16,17,18,19}. To this end, numerous studies have recently used rsfMRI to shed new light on our understanding of early brain functional development, including the temporal sequence of maturation starting from primary sensory to higher-order functional networks²⁰, the presence of a small-world topology in infants²¹, distinct developmental stages during the first 2 years after birth²² and the emergence of neural flexibility in infants²³. Although brain structural developmental characteristics starting from birth and extending to toddlerhood have been widely reported^24,25,26,27, to the best of our knowledge, normative reference curves of brain functional development from birth into early childhood are still lacking. One major barrier to establishing reference standards for brain functional development lies in the experimental requirement of imaging typically developing children¹⁴. Children younger than 2 or 3 years of age are usually scanned during natural sleep to obtain high-quality rsfMRI with minimal motion artefacts^{23,28,29,30,31}. In contrast, children older than 3 years of age are often scanned awake while watching movies^14,28. Because functional connectivity (FC) profiles in adults undergoing rsfMRI scans differ depending on sleep or wakefulness^32,33,34,35, it is likely that similar differences are present in rsfMRI from children. As a result, the potential confounds resulting from different consciousness/imaging states in children at different ages need to be addressed before constructing functional charts spanning from infancy into early childhood.

To this end, we aimed to (1) discern the functional differences between sleep and awake states and develop approaches capable of harmonizing these differences, (2) establish functional developmental charts from birth to early childhood and (3) determine the potential associations between brain growth charts and cognition. Specifically, we first determined whether rsfMRI FC differs between sleep and awake states in children 3–6 years of age who underwent two paired imaging sessions during sleep and wakefulness, respectively. Subsequently, we trained and validated elastic net regression models to harmonize the FC differences observed between sleep and awake states. Next, we included five paediatric datasets for a total of 1,091 scans^{22,28,36,37,38} and constructed FC growth charts from birth to 6 years. Finally, in a subset of children who underwent concurrent cognitive assessments and MR imaging, we determined the associations between individual deviation scores from the normative growth charts and the Mullen Scales of Early Learning (MSEL) assessment³⁹.

Results

Functional differences between awake and asleep states

In this study, a whole-brain parcellation comprising 100 cortical regions (Schaefer surface atlas)⁴⁰ and 10 subcortical regions (AAL volumetric atlas)⁴¹ was used. The Schaefer brain atlas has been widely utilized in both paediatric and adult studies due to its detailed and functionally informed parcellation^42,43,44. To align with the number of brain regions typically analysed in neonatal studies^21,22, we employed the 100-ROI version of the Schaefer atlas. We observed that the FC matrices differed between awake and asleep states (Fig. 1a), with an overall higher FC in the awake state as compared with the asleep state ($P < 0.001$, two-sided, Supplementary Fig. 1 and Table 1). A matrix similarity (Pearson’s correlation) of $0.31\pm 0.018$ (${\rm{mean}}\pm {\rm{se}}$) and mean absolute difference (MAD) of $0.31\pm 0.006$ between paired FC matrices of each participant were obtained, implying notable differences on average between the FC matrices of the two states (Fig. 1b, grey bars). Statistical comparisons identified connections with significantly different FC strengths between the two states ($P < 0.05$, two-sided, FDR corrected, Fig. 1c and Supplementary Table 2). Of significantly altered connections, over 85% were internetwork connections and more than 80% exhibited higher FC during the awake state when compared with the asleep state. These internetwork connections were largely associated with the dorsal attention, ventral attention, default and subcortical networks. Finally, the most altered brain regions between the two states included calcarine, fusiform, middle and anterior cingulate, middle temporal, angular, temporal pole, putamen, caudate and thalamus (Fig. 1d). These regions largely support basic and higher-order visual function^45,46,47, attention^46,48,49, movement control⁵⁰ and consciousness⁵¹.

Fig. 1: Differences in brain FCs and the performance of harmonization between asleep and awake states.

a, FC matrices from a representative infant participant in both awake and asleep states, along with the predicted asleep state derived from the experimentally acquired scan during the awake state of the same participant. b, Averaged similarity index and mean absolute difference (MAD) between asleep and awake states, as well as harmonization performance using the ComBat algorithm, from nested 10-fold cross-validation using images from BCP-augmented data. Error bars represent standard errors. c, Statistical comparison of individual functional connections between asleep and awake states; significantly different connections depicted (linear mixed-effects model, $P < 0.05$, FDR corrected, two-sided). d, The number of significantly altered FCs was calculated for each region. The most affected regions, which displayed the highest number of significantly altered FCs, include fusiform, calcarine, anterior and middle cingulate cortex, middle temporal gyrus, angular gyrus, temporal pole, putamen, thalamus and caudate.

With the observed state-induced differences in FC patterns, particularly in brain regions associated with higher-order cognition, approaches capable of harmonizing these differences before constructing developmental FC charts are needed. To this end, we trained elastic net regression models to predict FCs during a sleep state using awake-state FCs as inputs. Due to the limited sample size of successfully scanned participants, we augmented the sample size using a graph-based data augmentation method (Supplementary Information 1 and Supplementary Fig. 2). Figure 1a presents a representative FC matrix for an infant during sleep, along with the predicted sleep FC matrix derived from the participant’s awake FC matrix. On average, an FC matrix similarity of 0.77 and a MAD of 0.18 were achieved between the actual and predicted sleep FC matrices (Fig. 1b, orange bars). Consistent findings were observed for harmonizing inter- and intranetwork FC strengths between the two states (Supplementary Fig. 3). As an alternative approach, we compared the performance of the ComBat harmonization algorithm^52,53 in harmonizing across states, which underperformed our elastic net regression method (Fig. 1b blue bars and Supplementary Fig. 3).

Brain FC charts during early childhood

With the ability to account for FC differences between the two states, we subsequently sought to characterize the developmental curves of eight functional networks (Fig. 2a) spanning 0–6 years of age. To construct the growth charts, we included a total of 1,091 scans from five paediatric datasets (Fig. 2b and Supplementary Table 3). While the ComBat approach exhibited lower performance in harmonizing two consciousness states, it has been demonstrated to be effective in mitigating potential biases arising from different imaging cohorts^52,53,54,55. We thus employed the ComBat method to account for discrepancies in imaging parameters across different datasets (Supplementary Table 4). Harmonization models from augmented samples were applied to awake scans to estimate their sleep FC matrices, which were then combined with all sleep scans.

Fig. 2: Brain growth charts of the functional connectivity within canonical functional networks during early childhood.

a, Cortical and subcortical parcellations. b, Distribution of MRI scans across five cohorts, separated by dashed horizontal lines. Each dot represents scans at the corresponding age on the x axis. c, Developmental patterns of FC of resting-state functional networks (solid lines) and 95% confidence intervals (dotted lines). Transition ages, determined as the first sign change point of the first derivative of the trajectory, are also labelled (dashed vertical lines).

The network-level developmental curves exhibited complex and functional domain-specific temporal patterns, including both increases and decreases in FC strengths (Fig. 2c). It has been suggested that increased intranetwork FC strength may indicate functional maturation processes in the normative paediatric brain⁵⁶, while decreased intranetwork FC strength may indicate functional specialization processes⁵⁶. Accordingly, as depicted in Fig. 2c, the visual network exhibited a rapid maturation process until 5 months, followed by functional specialization and then relatively stable FC after 48 months (blue curve). In contrast, the somatomotor network showed a rapid functional specialization process immediately after birth and became relatively stable after 18 months (red curve). A third developmental pattern was observed for the limbic, default and ventral attention networks, which showed an early and rapid functional maturation process, followed by temporally stable FCs (green curves). Nevertheless, the age reaching the peak/transition point differed among these three networks, with limbic at 10 months, followed by default at 16 months and ventral attention at 21 months. It is also worth highlighting the difference in developmental patterns for the two attention networks. Unlike the ventral attention network, which showed a rapid functional maturation process during the first 21 months after birth, the dorsal attention network (cyan curve) was relatively stable during early infancy. After 18 months, the dorsal attention network began to show a protracted increase in functional maturation. Finally, the control network displayed a continuous, protracted increase in functional maturation (pink curve), while the subcortical network exhibited a relatively stable and high FC throughout the first 6 years of age (orange curve).

Importantly, functional brain networks work collaboratively to execute complex cognitive processes⁵⁷. Therefore, we also evaluated internetwork FCs to characterize how functional networks interacted during early childhood. Accordingly, of the 28 possible network pairs, we observed 7, 9 and 12 pairs of always positive, always negative and functional transition patterns, respectively (Fig. 3).

Fig. 3: Brain growth charts of the functional connectivity between pairs of canonical functional networks during early childhood.

a, Network pairs with positive connections. b, Network pairs with negative connections. c, Network pairs whose connections transitioned between positive, negative and non-significant connections. Transition ages, as well as ages indicating statistically zero FC, are also labelled (dashed vertical lines). Dashed horizontal lines indicate zero FC for reference.

The network pairs with consistently positive FC, indicating network integration, were mostly linked with higher-order functional networks, such as the default mode, control and attention networks (Fig. 3a). While these observed network integrations associated with the higher-order networks could be understood as the foundations for performing complex cognitive functions, two important features of network interactions are apparent. First, although the dorsal and ventral attention networks exhibited a pattern of network integration, the FC between them was stable and relatively low, suggesting functional specialization between them. Second, while most of these internetwork interactions exhibited a stable or initial increase in FC strength, the interaction strength between the default and control networks steadily decreased with age, although it remained positive.

The network pairs with negative FC from birth, indicating competition between the networks, largely involved connections between basic and higher-order/emotional processing functional networks (Fig. 3b). The default network was a part of approximately half of these network pairs. The networks interacting with the visual network showed an initial increase in negative FC during early infancy, followed by a decrease in negative FC. In contrast, the networks interacting with the somatomotor network showed a decrease in negative FC during early infancy, followed by either stabilizing FC (for somatomotor-default and somatomotor-limbic pairs) or an increase in negative FC (for the somatomotor-control pair). Finally, we observed increasing negative FC between the dorsal attention and default networks during the first 57 months and decreasing negative FC between the dorsal attention and subcortical networks during the first 23 months.

Finally, interactions indicating temporal transitions were observed among 12 pairs of networks, involving both higher-order/emotional processing functional networks (Fig. 3c). These functional charts depicted the developmental alterations across functional integration, disassociation (statistically zero FC) and competition throughout this period, highlighting the process of forming a more efficient brain topology. Despite the complexity of the developmental curves during early infancy, five of these network pairs resulted in functional disassociation by 72 months (visual-dorsal attention, visual-limbic, dorsal attention-limbic, ventral attention-control and control-subcortical). Conversely, six of these network pairs resulted in functional competition by 72 months (visual-somatomotor, visual-control, somatomotor-dorsal attention, ventral attention-limbic, ventral attention-default and limbic-subcortical). Notably, our results suggested that the anticorrelation between the ventral attention and default networks observed in adults was not established until ~2 years of age. Last, we also observed a trend towards functional integration between the limbic and control networks, with a positive interaction emerging shortly after birth and strengthening until 26 months of age.

To further demonstrate the necessity and efficacy of our awake-to-sleep prediction models in constructing brain developmental trajectories, we conducted detailed comparisons of brain functional charts before and after harmonization (Supplementary Information 2 and Supplementary Fig. 5). Collectively, these findings provide valuable insights into the complex and dynamic nature of intra- and internetwork development during early childhood, underscoring the need to examine both intra- and internetwork interactions to gain a comprehensive understanding of early brain functional development.

Association between brain FC charts and cognitive abilities

Finally, we aimed to determine whether the developmental curves were associated with cognitive ability. We calculated maturation deviation scores for each scan and used multiple linear regression to assess their association with MSEL-measured cognitive outcomes. A total of 364 scans from 212 BCP participants (95 males, 117 females) aged 2.9–61.2 months, with concurrent MSEL behavioural evaluations, were included. Figure 4a shows the performance of the models.

Fig. 4: Prediction of Mullen Scales of Early Learning.

a, Model performance of the multiple linear regression between the maturational deviation scores and the MSEL scores ($n=364$, ${P}_{{\rm{ELC}}} < 0.001$, ${P}_{{\rm{GM}}}=0.448$, ${P}_{{\rm{FM}}}=0.001$, ${P}_{{\rm{EL}}} < 0.001$, ${P}_{{\rm{RL}}}=0.001$, ${P}_{{\rm{VR}}}=0.007$). Error bars represent 95% confidence intervals of regression coefficients, with the centre of error bar indicating the estimated regression coefficient. b, Prediction performance of the multiple linear regression models evaluated using 50% holdout testing for 100 runs (${P}_{{\rm{ELC}}} < 0.001$, ${P}_{{\rm{GM}}}=0.353$, ${P}_{{\rm{FM}}}=0.064$, ${P}_{{\rm{EL}}} < 0.001$, ${P}_{{\rm{RL}}}=0.007$, ${P}_{{\rm{VR}}}=0.002$). On each box, the red central line represents the median, and the upper and lower edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers represent 1.5× the interquartile range, with red crosses showing observations outside of this range. Statistical significance was assessed using permutation tests; ^**$P < 0.01$, ^***$P < 0.001$. ELC, early learning composite score; GM, gross motor; FM, fine motor; EL, expressive language; RL, receptive language; VR, visual reception.

The maturational deviation scores exhibited significant predictive power for the early learning composite score ($P < 0.001$, ${R}^{2}=0.237$) and for the subdomains of the MSEL assessment, including fine motor score ($P=0.001$, ${R}^{2}=0.179$), expressive language score ($P < 0.001$, ${R}^{2}=0.199$), receptive language score ($P=0.001$, ${R}^{2}=0.184$) and visual reception score ($P=0.007$, ${R}^{2}=0.164$). However, the model did not show significant predictive ability for the gross motor score ($P=0.448$, ${R}^{2}=0.108$). These findings remained consistent after FDR correction, with full statistical details provided in Supplementary Tables 5 and 6. Moreover, the regression coefficients for a subset of intra- and internetwork growth charts were statistically significant ($P < 0.05$, two-sided), with several connections involving basic and higher-order functional networks identified as significant contributors. Furthermore, internetwork connections were particularly important for predicting cognitive outcomes across multiple domains, highlighting the need to consider internetwork interactions in relation to cognitive abilities. Finally, while the full model did not predict the gross motor score, the connection between the somatomotor network and the control network was significant ($P=0.022$, two-sided).

To further assess the replicability and predictive ability of our models, we conducted a split-half validation analysis in which the participants were randomly divided into a training sample and a testing sample with a ratio of 1:1 that was repeated 100 times. Our results showed that the model derived from the training samples maintained good predictive power for not only the early learning composite score ($P < 0.001$, ${R}_{{\rm{mean}}}=0.26$), but also for the subdomains of the MSEL assessment, including expressive language ($P < 0.001$, ${R}_{{\rm{mean}}}=0.23$), receptive language ($P=0.007$, ${R}_{{\rm{mean}}}=0.18$) and visual reception ($P=0.002$, ${R}_{{\rm{mean}}}=0.21$) (Fig. 4b). These findings provide evidence of the replicability of cognitive prediction based on our intra- and internetwork developmental charts.

Discussion

The human brain undergoes dramatic development during early childhood. With advanced neuroimaging methods, characterizing trajectories of FC has provided a comprehensive understanding of the developmental processes of the infant’s brain. In this study, we aimed to characterize FC trajectories beyond infancy into childhood by harmonizing rsfMRI scans acquired when infants were asleep and when children were awake. To do so, we first confirmed differences in FC between awake and asleep conditions, and then developed approaches to harmonize FC between the two conditions. In this study, we utilized elastic net regression to harmonize differences induced by varying states of consciousness. In addition, we evaluated the results using other machine-learning algorithms, including Gaussian process regression and support vector regression. These methods yielded largely identical results, confirming the robustness of our harmonization approach across different computational techniques. Additional discussions on these findings are provided in Supplementary Information 2, 3 and 7. We then characterized brain development from 0 to 6 years by establishing normative growth curves for eight functional networks as well as interactions across pairs of these networks. These growth charts revealed distinct neurodevelopmental patterns and thus characterized the complex development of functional brain networks, including maturation and specialization for intranetwork charts as well as functional integration, competition and transition processes for internetwork charts during the first 6 years of age. Finally, we demonstrated that individual deviations from these growth charts were associated with cognitive ability determined using the MSEL assessment.

We observed an early increase in FC with age for most of the functional networks, except for the dorsal attention, somatomotor and subcortical networks. It is possible that these developmental patterns indicate the emergence of specific cognitive functions. For example, visual functions undergo rapid maturation during the first few months after birth⁸. An adult-like default network topology is observed around 1 year old^17,20. Maturation of bottom-up attentional processes, thought to be subserved by the ventral attention network, occurs earlier than maturation of top-down attentional processes (dorsal attention network)⁷. Thus, the observed temporal order of each of these networks reaching its peak FC could reflect specific developmental milestones related to different cognitive functions emerging. Notably, the distinct developmental patterns of the two attentional networks parallel the developmental sequence of bottom-up and top-down attention control. Finally, the protracted and continuous increase in intranetwork FC in the control network could reflect the protracted emergence of executive functions, which emerge during the first few years after birth and undergo notable strengthening throughout childhood and beyond⁵⁸.

It is worth highlighting the developmental trajectory of the visual network, which was the only network to show a marked decrease in FC after reaching its peak. The visual network of the employed functional atlas comprises both medial visual (V1) and lateral visual regions (V2 and V3). Therefore, the initial strengthening of FC within the visual network may indicate synergetic development of basic and higher-order visual functions before 5 months of age. As higher-order visual functions, such as facial/object recognition and visual attention, emerge around 5–7 months^59,60,61, the prolonged reduction in FC beginning ~5 months may reflect functional specialization of the different subsystems of the visual network. In addition, synaptic pruning, critical for functional specialization, emerges in the visual regions around 4 months of age⁶², which is consistent with our observed transition point from increased to decreased FC strength. On the other hand, the somatomotor network exhibited a rapid decrease in FC immediately after birth until 18 months of age. It has been well documented that development of the somatomotor network occurs prenatally^16,63, and its topology is adult-like at birth^20,56. Thus, the initial reduction in FC strength may represent a functional specialization process between sensory and motor functions that begins at birth.

The subcortical network displayed a unique developmental curve in our study, with relatively stable and high FC throughout the first 6 years of age. The subcortical network is associated with many essential basic and cognitive functions, including emotion processing, memory formation, reward processing and automatic functions (such as regulating heart rate and blood pressure), many of which are critical from the very beginning of life^64,65,66,67. For example, research has shown that newborns and infants exhibit the ability to respond to emotional stimuli, a response that is partially regulated by subcortical regions⁶⁸. The regulation of basic physiological functions, managed by subcortical areas, is also essential from birth⁶⁷. Therefore, the stable and high FC of the subcortical network from birth may reflect its early maturation and foundational roles in cognitive and physiological functioning in the first 6 years of age.

Unlike the intranetwork FCs, which were all positive, the growth charts of internetwork interactions exhibited both positive and negative values. While positive internetwork interactions typically indicate functional integration between networks, the potential biological significance of negative or a lack of (that is, FC = 0) internetwork FC is not as well established. One theory is that a competitive relationship exists when two networks exhibit a negative interaction pattern (anticorrelation)⁶⁹. One of the most representative examples of this is the relationship between task-positive networks and the default network^69,70. Notably, we did not employ global signal regression in our study, precluding the possibility of artificially introducing negative correlations in our findings⁷¹. Therefore, we interpreted negative internetwork interactions as network ‘competition’. Finally, we also observed internetwork interactions with statistically zero FC at certain age periods, suggesting that these two networks may be independent. It is important to note that these internetwork interactions largely reflect the average pattern between two networks, which may differ on a regional level. One may thus argue that the observed near-zero patterns could result from the presence of both strong positive and strong negative FC strengths between a pair of networks. However, the histogram plots of the interregional FC strength across all available scans show a zero-mean Gaussian instead of a bimodal distribution between networks (Supplementary Fig. 6), indicating that these near-zero patterns are probably the results of weak FCs. Consequently, we defined statistically zero FC between two networks as functional ‘disassociation’. Below, we will separately discuss the three identified internetwork interaction patterns. Nevertheless, we acknowledge that the underlying biological mechanisms of many of the observed internetwork interaction patterns in the developing brain remain largely unknown.

Integration

We observed positive FC among seven network pairs, most of which involved higher-order functional networks (for example, the control, default and attention networks). This may not be surprising given the continuing maturation of higher-order functional networks and the need to integrate across these networks to perform complex cognitive functions⁵⁷. Despite the observed temporal delay of the increase in FC within the dorsal attention network compared with that of the ventral attention network, their integration remained temporally stable, with a positive but weak FC. This stable level of integration may facilitate the coordination of attentional processes and ensure efficient transitions between attending to internal and external stimuli⁷². The interaction between the ventral attention network and the subcortical network exhibited a rapid increase in functional integration until ~3 years of age before stabilizing. This pattern of interaction may guide bottom-up attention for emotional stimuli^73,74,75. Finally, although adult studies have largely reported an anticorrelation between the default and control networks⁷⁶, in children aged 0–6 we observed a positive but decreasing internetwork FC. This suggests that the anticorrelations observed in older samples may not have yet emerged in early childhood.

Competition

The general consensus from adult studies is that a competitive relationship exists between the default network and task-positive networks⁶⁹. While we found positive (but decreasing) FC between the default and control networks in our sample, FC between the default network and several other networks was negative (dorsal attention, visual, somatomotor and subcortical). These findings support the possibility that some competitive relationships, particularly as related to sensorimotor and attention processes, already exist during the first 6 years of age. Despite consistently negative FC, the temporal patterns of the negative interactions with the default network differed among the four networks. Specifically, the visual network exhibited a weak competition with the default network at birth, followed by minor fluctuations. In contrast, a strong competition with the somatomotor network was observed at birth, but the strength diminished with age. Meanwhile, the competition between the default and dorsal attention networks increased during the first 57 months, then stabilized. This negative association, which suggests competitive interaction between goal-directed attentional processes and internally focused processes, is consistently observed in adult studies^69,76. Thus, our results confirm that this relationship starts during early infancy. Finally, we observed weak and relatively stable competition between the default and subcortical networks. This appears consistent with the reported negative relationships between the default network and brain structures in the subcortical network, such as the putamen and globus pallidus, that have been observed in adults⁷⁷.

It is worth noting the different interaction patterns between the two attention networks with the subcortical network. Unlike the ventral attention network, which maintained positive integration with the subcortical network, the dorsal attention network initially exhibited a competitive relationship with the subcortical network. The observed anticorrelation between them may reflect an immature interplay between top-down attentional control and stimulus detection/emotion regulation during early infancy. The gradually weakened anticorrelation during the first 23 months may potentially be indicative of functional dissociation between these two networks at a later age.

Transition

We observed FC trajectories that indicated a transition between positive, negative and statistically zero FC among 12 pairs of networks, reflecting the dynamic nature of functional brain development during early childhood. Here we focused on the polarity patterns at later ages and divided them into three main categories: positive (integration), negative (competition) and disassociated network pairs. Interestingly, only the connection between the limbic and control networks exhibited positive FC after an initial disassociation at birth. Considerable evidence suggests that functional coupling between the limbic and frontal control systems plays an important role in self-control⁷⁸ and emotion regulation strategies, an essential skill that can improve learning⁷⁹ and mitigate the increased risk of anxiety and stress-related disorders in adolescence^80,81. Our results provide additional insights into the developmental processes of forming emotion regulation during the first 6 years of age. In particular, the rapid integration of the limbic and control networks reflects the formation of emotion regulation during the first 2 years after birth, followed by temporally stable integration after 26 months old. This finding appears to be consistent with infants’ preferences to rely more on dyadic regulation than on self-soothing and attentional distraction, while toddlers rely on more independent, attention-seeking and escape behaviour during periods of maternal unavailability⁸².

Six internetwork interactions showed trajectories that resulted in competition (negative FC) at later ages, particularly between the ventral attention and default networks. Adult studies have reported an anticorrelation between the ventral attention and default networks⁶⁹. Our findings indicate that this competition is not established until ~2 years of age, contrasting our observation of the competitive relation between the dorsal attention and default network from birth. Nevertheless, both attention networks exhibited increased competition with the default network during early childhood. These findings support the idea that there is an emergence of competition between externally directed attention and internal cognition in infants, laying the neural foundation for effective task performance⁸³. Similar transitions were observed between the somatomotor and dorsal attention networks, as well as between the limbic and subcortical networks. The transition from early integration into functional competition may indicate a process of functional specialization of these networks, which is consistent with the theory of interactive specialization during functional brain development⁸⁴.

Finally, five internetwork interactions displayed functional disassociations at later ages, largely involving the visual and attention networks. Because brain networks do not work in isolation but rather collaboratively to accomplish complex cognitive demands, the observed dissociation may reflect the substantially diminished interactions between visual and attention networks during sleep. More studies are needed should approaches become available to image children younger than 3 years old in an awake condition to address this hypothesis.

Using the shallow model of maturation, in which an upshifted or downshifted data point is indicative of altered maturation⁸⁵, we found that individual deviations from normative growth charts were associated with MSEL scores during early childhood. These findings confirm that altered neural maturation underlies individual differences in cognitive abilities. Notably, a subset of connections showed statistically significant regression coefficients in the multiple linear regression analyses. These connections demonstrated spatial heterogeneity across different MSEL domains, suggesting that they may play distinct functional roles in cognitive performance. This highlights the complex relationship between brain connectivity and cognitive abilities, with specific connections contributing differently to various aspects of cognitive ability. It is worth noting that the connections involving the task-positive networks, which includes the attention networks and the control network, and task-negative network (that is, default network), emerged as significant contributors to behavioural prediction in our study. Indeed, the proper functioning of these networks provides the foundation for executive functions and cognitive processes^10,86. Abnormalities associated with the task-positive networks and the default network, as well as impaired cognitive/executive functions, have been widely observed in neurodevelopmental disorders, such as attention-deficit/hyperactivity disorder and autism^{87,88,89,90,91}. Therefore, our findings highlight the significance of task-positive networks and the default network in monitoring cognitive development. On the other hand, connections involving visual and somatomotor networks emerged as significant predictors in four of the six association analyses. The importance of visual functions in cognitive development is well established. For example, measures of infant visual performance, such as attention and fixation, have been shown to predict neurocognitive development^92,93,94. Similarly, a close relationship between motor development and cognitive development has been reported⁹⁵. In addition, recent imaging studies have underscored the significance of visual and somatomotor connections in predicting behavioural abilities^23,43. Our results thus support previous findings, concluding that visual and somatomotor network functions in early infancy are crucial for the development of cognition.

This study has several potential limitations that should be taken into consideration. First, ‘the limited sample size for the training data constrained the awake-to-sleep predictions’. To address this, data augmentation techniques were employed (Supplementary Information 1) to enhance the sample size, and the model’s generalizability was further assessed using two independent datasets (Supplementary Information 4 and Supplementary Fig. 7). Second, ‘the results were derived from participants during sleep, not wake’. It is important to acknowledge that the brain FC charts in this study were derived from participants during sleep due to existing imaging protocols^{14,28,29,30,31}. These protocols typically require scanning infants during natural sleep to ensure high-quality images, while older children are often scanned awake. In this study, we recruited children aged 3–6 years for paired natural sleep and awake scans, constructing brain FC charts using the awake-to-sleep harmonization model. Ideally, separate harmonization models would be developed with paired sleep and awake rsfMRI scans from children under 3 years to establish accurate wakeful-state brain FC charts. However, obtaining high-quality wakeful rsfMRI from unsedated infants and very young children is currently impractical and costly, as they cannot reliably remain still and follow instructions. Third, ‘employing a learning-based approach to estimate asleep FCs from awake scans has led to reduced variance in the predicted data’. This reduction is probably because BOLD signals are contaminated by noise, and machine-learning models typically ignore noise that does not contribute to prediction accuracy, resulting in more generalizable models with lower variance compared with the raw data. In addition, this study used only awake FCs as input features, which may limit the observed variance in the predicted data. Incorporating additional valuable features could potentially enhance the prediction accuracy.

Methods

Participants and imaging acquisition

In this study, a total of 501 participants (male/female: 238/263) ranging in age from birth to 6 years with 1,091 scans were obtained from five different cohorts (Fig. 2b). These cohorts included the Multi-visit Advanced Pediatric Brain Imaging Study (MAP)²², the UNC/UMN Baby Connectome Project (BCP)²⁸, the Pediatric Imaging, Neurocognition and Genetics (PING) study³⁶, the Pixar study³⁷ and the Calgary study³⁸. The experimental protocols were approved by the Internal Review Board of the University of North Carolina at Chapel Hill, as well as the local corresponding institutions. Written informed consent was obtained from the parents of all participants. Participants from our leading studies^22,28 were financially compensated, while the compensation information for the other studies was not disclosed publicly^36,37,38. In addition, we collected samples from existing imaging cohorts, focusing on individuals within the age range of 0–6 years. No statistical methods were used to pre-determine sample size, but our sample size, which combined participants from five imaging cohorts, is larger than those reported in previous publications^23,25,30.

In the MAP study, a total of 73 infants (male/female: 34/39) with 269 longitudinal scans were included. Participants were recruited from UNC-Chapel Hill newborn hospital. The participants were imaged from birth to 6 years of age using 3 T Siemens scanners. T1-weighted images were acquired with the following parameters: ${\rm{TR}}=1,820{\rm{ms}}$, ${\rm{TE}}=4.38{\rm{ms}}$ and ${\rm{resolution}}=1\times 1\times 1\,{{\rm{mm}}}^{3}$. RsfMRI images were acquired using the EPI sequence with the following parameters: ${\rm{TR}}=2000{\rm{ms}}$, ${\rm{TE}}=32{\rm{ms}}$, ${\rm{resolution}}=4\times 4\times 4\,{{\rm{mm}}}^{3}$ and 150 volumes. To ensure imaging quality, participants younger than 2 years of age were imaged during natural sleep, while participants older than 2 years of age were imaged while watching movies.

In the BCP study, a total of 297 infants (male/female:140/157) with 627 longitudinal scans were included. Participants were recruited from existing registries at UNC and UMN on the basis of state-wide birth records as well as from broader community resources (for example, community centres and targeted day-care centres) to ensure the sample approximates the racial/ethnic and socioeconomic diversity of the US census. The participants were imaged from birth to 6 years of age using 3 T Siemens scanners. T1-weighted images were acquired with the following parameters: ${\rm{TR}}=2,400{\rm{ms}}$, ${\rm{TE}}=2.24{\rm{ms}}$ and ${\rm{resolution}}=0.8\times 0.8\times 0.8\,{{\rm{mm}}}^{3}$. RsfMRI images were acquired using the EPI sequence with the following parameters: ${\rm{TR}}=800{\rm{ms}}$, ${\rm{TE}}=37{\rm{ms}}$, ${\rm{resolution}}=2\times 2\times 2\,{{\rm{mm}}}^{3}$ and 420 volumes. Participants younger than 3 years of age were imaged during natural sleep, while participants older than 3 years of age were imaged while awake watching movies. Furthermore, 21 BCP participants (3–6 years) underwent paired asleep and awake scans, allowing for the study of rsfMRI differences between the two imaging states (BCP supplement data).

Paediatric participants aged 3–6 years from the PING (26 participants: 13 males and 13 females), Pixar (47 participants: 20 males and 27 females) and Calgary (58 participants: 31 males and 27 females) studies were also included in this study. Specifically, participants of the PING study were recruited through local postings and outreach activities conducted in the greater metropolitan areas of Baltimore, Boston, Honolulu, Los Angeles, New Haven, New York, Sacramento and San Diego. Participants of the Pixar study were recruited from the local community in Massachusetts, USA, and participants of the Calgary study were recruited from the local community in Calgary, Canada. These participants were scanned in various awake states: uncontrolled (PING), watching movies (Pixar) and viewing passive video movies (Calgary), respectively. The demographic information and imaging protocols for these participants are summarized in Supplementary Tables 3 and 4.

Infant-dedicated surface-based MRI preprocessing

For each scan, brain structural images were processed using the infant Brain Extraction and Analysis Toolbox (iBEAT v.2.0, www.ibeat.cloud)^27,96,97. This toolbox segments the brain into grey matter, white matter and cerebrospinal fluid (CSF), and reconstructs the topologically correct and geometrically accurate infant cortical surfaces. Specifically, the reconstruction process involved first reconstructing the inner surface, which represents the interface between grey matter and white matter. This surface was then deformed to reconstruct the outer/pial surface, representing the interface between grey matter and CSF. With vertex-wise corresponding between inner and outer cortical surfaces, the middle cortical surface, which sits purely inside the grey matter and shares the same distances to the inner and outer cortical surface, was finally reconstructed.

RsfMRI data were preprocessed using FSL^98,99,100. The preprocessing steps included discarding the first 10 volumes, motion correction and bandpass filtering (0.01–0.08 Hz). After initial preprocessing, the rsfMRI images were aligned with the T1 images using boundary-based registration. The mean signals from white matter, CSF and 24 motion parameters were further removed using the linear regression model. Finally, the preprocessed rsfMRI images were mapped onto their reconstructed middle cortical surface and spatially smoothed with a small kernel (2 mm full width at half maximum) using Freesurfer and Connectome Workbench, as recommended previously¹⁰¹. Scans with excessive motion, defined as a mean frame-wise displacement (FD) greater than 0.5 mm, have been excluded from this study.

In this study, a whole-brain parcellation comprising 100 cortical regions (Schaefer 100 surface atlas)⁴⁰ and 10 subcortical regions (AAL volumetric atlas)⁴¹ was used. The brain surface atlas was deformed back to the participant’s surface space using spherical demons surface registration¹⁰², while the brain volumetric atlas was warped to the participant’s volumetric space using the advanced normalization tools (ANTs)¹⁰³. Additional results using the UNC infant-dedicated surface-based atlas¹⁰⁴ and the Schaefer 400 atlas⁴⁰ are provided in Supplementary Information 5 and Supplementary Fig. 8. For each scan, the FC matrix was derived by calculating Pearson’s correlation coefficient between the time series extracted from each pair of regions of interest. Furthermore, brain cortical regions were grouped into seven functional systems on the basis of the Yeo-7 networks parcellation¹⁰⁵, while subcortical regions were independently classified as a subcortical network (Fig. 2a). Intranetwork FC was defined as the average FC within each network, and internetwork FC was defined as the average FC between each pair of networks.

Two-stage harmonization of FC

In this study, we proposed a two-stage harmonization pipeline to minimize potential cohort-related differences and imaging-state-induced differences (Supplementary Fig. 9). To account for potential cohort-related differences in imaging parameters, the ComBat harmonization method was employed to mitigate multicohort effects on brain FC^52,53. During this harmonization process, we designated the BCP cohort as the reference site using the ‘ref.batch’ parameter in the R ‘neuroCombat’ package, available at https://github.com/Jfortin1/ComBatHarmonization?tab=readme-ov-file. We selected the BCP dataset as the reference site because it encompasses both asleep scans from ages 0–3 years and awake scans from ages 3–6 years. This setup allowed us to harmonize potential cohort effects from the MAP dataset’s asleep scans (0–2 years) and to correct potential cohort effects from the awake scans of the MAP (3–6 years), Calgary, PING and Pixar datasets. Specifically, FCs of asleep scans were harmonized from other datasets to align with the FCs of BCP asleep scans, covering the age range of 0–3 years. Similarly, the FCs of awake scans from the other datasets were harmonized to match the FCs of BCP awake scans, spanning the ages of 3–6 years. Details of the reference-based batch adjustment can be found in Supplementary Information 6. This harmonization approach allowed us to adjust for cohort-specific differences and ensure consistency in the assessment of brain FC across multiple cohorts.

A learning-based method was further established to harmonize state-related differences in rsfMRI data. Due to the limited number of paired asleep and awake scans, we employed a graph-based data augmentation method to effectively increase the sample size (Supplementary Information 1), and elastic net regression was employed to estimate the asleep FCs on the basis of the experimentally acquired awake FCs¹⁰⁶. Specifically, FCs of given awake scans were input into the model to predict each FC of the corresponding asleep scans. The elastic net equation¹⁰⁶ is written as

$${\hat{\beta }}_{0},\hat{\beta }={{\rm{argmin}}}_{{\beta }_{0},\beta }\left(\frac{1}{2N}\mathop{\sum }\limits_{i=1}^{N}{\left({y}_{i}-{\beta }_{0}-{x}_{i}^{T}\beta \right)}^{2}+\lambda {P}_{\alpha }\left(\beta \right)\right),$$

(1)

where

$${P}_{\alpha }\left(\beta \right)=\frac{(1-\alpha )}{2}{{||}\beta {||}}_{2}^{2}+\alpha {{||}\beta {||}}_{1}.$$

(2)

Here, ${y}_{i}$ is the response FC of asleep scan i, N is the number of observations and ${x}_{i}$ is a vector containing all FCs from awake scan i. The tuning parameters of the elastic net regression, α and λ, were determined by 10-fold cross-validation with the lowest mean square error over all possible α and λ values (α and λ ranging from 0.01 to 1 with an increment of 0.01). The final regression model can be defined as

$$Y=X\beta +{\beta }_{0}.$$

(3)

In this study, we used a nested 10-fold cross-validation approach, involving two levels of cross-validation to ensure robust model evaluation and effective hyperparameter tuning. The inner-loop 10-fold cross-validation was employed to tune the hyperparameters, while the outer-loop 10-fold cross-validation was used to assess the model’s performance. Additional results using different machine-learning algorithms with nonlinear kernels, such as Gaussian process regression and support vector regression, are provided in Supplementary Information 7 and Supplementary Figs. 10–12.

The performance of the prediction models was also evaluated using two independent datasets (Supplementary Information 4): one publicly available dataset consisting of 3 infants aged 1–3 years from Yale University⁴² and another comprising 12 adults from Taipei Medical University. The experimental protocols were approved by the Internal Review Board of the local corresponding institutions. Written informed consent was obtained from the parents of all infant participants and directly from the adult participants.

Statistical analysis

FC matrix similarity and mean absolute difference were used to evaluate the overall similarity and difference between paired scans (for example, sleep vs wake, sleep vs predicted sleep). Specifically, FC matrix similarity was determined as the Pearson’s correlation between paired matrices, while MAD was calculated as the mean of the absolute FC differences across all connections between paired matrices.

A linear mixed-effects model was conducted to compare statistical difference in FCs between paired awake and asleep scans, covarying for age and mean frame-wise displacement, and including a random intercept for each participant. Statistical significance was assessed using a two-sided t-test on the regression coefficients, testing the null hypothesis that the coefficient is equal to zero. To validate the assumptions of the statistical tests, normality was assessed with the Shapiro–Wilk test, and equality variances was evaluated using Levene’s test. $P < 0.05$ after FDR correction was considered as statistically significant.

A generalized additive mixed model (GAMM)¹⁰⁷ was used to delineate the normative growth charts of FC after harmonization, given its ability to apply objective curvature fitting. Considering the longitudinal design of the BCP, MAP and Calgary studies, each participant was modelled with a random intercept. The model, which included age and mean FD as covariates and participant ID as the random intercept, could be written as:

$${Y}_{i}(t)=f(t)+{\alpha }_{i}+{e}_{i}(t)+{{FD}}_{{it}}$$

(4)

where ${Y}_{i}(t)$ is the brain FC for the ${i}{{\rm{th}}}$ participant at time t, ${\alpha }_{i}$ is the random intercept effect for the ${i}{\mathrm{th}}$ participant, ${e}_{i}(t)$ is the random noise and ${{\rm{FD}}}_{{it}}$ is the mean FD for the ${i}{\mathrm{th}}$ participant at time t. Specifically, ${\alpha }_{i}$ and ${e}_{i}(t)$ are independent and identically distributed with a normal distribution $N(0,{\sigma }_{\alpha }^{2})$ and $N(0,{\sigma }_{e}^{2})$, respectively. With the observations of brain FC at different ages, an estimate $\hat{f}(t)$ of $f(t)$ was obtained by using cubic spline, where smoothing parameters were chosen by restricted maximum likelihood. The 95% confidence intervals of each fitting plot were also evaluated through the Bayesian posterior covariance matrix¹⁰⁷. The model curve fitting plots used the R packages mgcv and itsadug¹⁰⁸. Sex-stratified developmental patterns were also uncovered by including a fixed sex term and a non-parametric age–sex interaction term in the GAMM model (Supplementary Information 8 and Supplementary Fig. 13). Additional results using a different fitting method (that is, Gaussian process regression) are provided in Supplementary Information 9 and Supplementary Fig. 14.

Neurodevelopmental milestones can be derived from normative growth charts. We identified these milestones as the transition age of each chart by locating its first maximum or minimum points. Specifically, the transition age was computed as the first sign change point of the first derivative of the function $Y(t)$. The trajectories of the first derivative of all brain growth charts are shown in Supplementary Fig. 15.

The maturational deviation score (that is, residual) was also calculated for each scan in relation to each growth chart. These deviation scores reflect the differences in each participant’s FC relative to the expected expression of their biological age. Through multiple linear regression, the association of these deviation scores with behavioural performance was evaluated across all growth charts using the Mullen Scales of Early Learning³⁹. A total of 364 scans from 212 BCP participants (95 males, 117 females) aged 2.9–61.2 months, with concurrent MSEL behavioural evaluations, were included in this analysis. To account for potential motion effects, mean FD was also included as a covariate in the model. Specifically, within each regression model, two types of P value were obtained. Individual P values were obtained for each variable to assess their statistical significance within the regression model, and variables with P values less than 0.05 were considered statistically significant. In addition, an overall P value was determined for each regression model to evaluate its capacity to significantly predict the response variables. To minimize potential false discoveries in the prediction of Mullen scores, a false discovery rate (FDR) correction was employed to adjust for multiple comparisons across the five subdomain scores.

To further assess the predictive ability of our proposed method, the scans were divided into a training sample and a testing sample with a ratio of 1:1 and 100 repeats. The multiple linear regression model was fitted in the training samples and applied to the testing samples to evaluate independent prediction performance. To determine whether the model achieved better-than-chance accuracy, permutation tests were performed by shuffling behavioural measures across the scans and repeating the 50% holdout testing 10,000 times.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.