The Development of Hippocampal‐Cortical Functional Connectivity in Infants and Toddlers

ABSTRACT

Infancy is a critical period for the development of the memory system, yet the functional neural changes that occur during this time remain poorly understood. In adults, long‐term memory relies on hippocampal‐neocortical coupling, which differs along the hippocampal long‐axis. In this study, we investigated resting‐state hippocampal‐neocortical functional connectivity along the long‐axis in 212 infants across the first two postnatal years. We found large increases in hippocampal connectivity with canonical adult memory regions across the first 6 months of age, accompanied by early functional differentiation along the hippocampal long‐axis. However, anterior and posterior hippocampal connections continued to fine‐tune with age with medial temporal and medial parietal memory‐related cortical regions, but also with areas associated with social cognition, salience, and attention in adults. These systems are known to strongly modulate memory formation and retrieval in mature brains. These findings trace the early maturation of hippocampal‐cortical coupling along the long‐axis, which may play an important role in evolving long‐term memory capacity with development.

This study traces the early maturation of hippocampal‐cortical coupling along the hippocampal long‐axis across the first 2 years of life, revealing early increases with canonical memory regions followed by prolonged fine‐tuning with memory, salience, and attention networks. These developing interactions may support evolving long‐term memory capacity in early development.

1. Introduction

Infancy represents a critical period in the development of the memory system. Although infants can encode and retrieve memories for some period of time, these memories are forgotten at an accelerated rate (Madsen and Kim 2016). Indeed, most adults are unable to recall experiences from before the age of 3. One contributing factor may be the immaturity of the neurobiological systems that support memory during this early period (Alberini and Travaglia 2017; Bauer 2006). In particular, the development of the hippocampus—a brain region essential for memory—is quite protracted in humans, with robust hippocampal growth occurring throughout the first postnatal year and continuing into childhood (Gilmore et al. 2012; Uematsu et al. 2012). However, structural development alone does not capture the full picture of maturity of the memory system. In healthy adults, episodic memories do not reside exclusively in the hippocampus, rather, the hippocampus is thought to serve as an index to mnemonic representations distributed across the neocortex (Gilmore, Quach, Kalinowski, Gotts, et al. 2021; Teyler and Rudy 2007). Cortical regions that are canonically involved in episodic memory in adults include midline medial temporal lobe, medial parietal, and medial prefrontal cortices alongside angular gyrus, superior and middle frontal gyri, as well as lateral and inferior temporal cortex (Audrain et al. 2022). Consequently, hippocampal‐neocortical coupling is conceived as critical for the consolidation and retrieval of episodic memories. Despite this, relatively little is known about the functional development of the hippocampus in infancy, particularly regarding the emergence and maturation of hippocampal‐neocortical coupling.

In recent years, it has become clear that the topological organization of functional coupling with the neocortex differs along the long‐axis of the hippocampus in adults. Specifically, the anterior hippocampus is relatively more functionally and structurally connected to ventral medial prefrontal cortex (PFC), and anterior temporal areas, while the posterior hippocampus is more connected to dorsal medial PFC, posterior ventral temporal cortex, occipital cortex, and much of the precuneus (Barnett et al. 2019, 2021; Frank et al. 2019; Grady 2020; Vogel et al. 2020). Differences also exist intrinsically along the long‐axis: the posterior hippocampus has smaller population receptive field sizes than the anterior hippocampus (Leferink et al. 2024), as well as faster signal dynamics (Bouffard et al. 2023; Brunec et al. 2018). These distinctions are thought to support qualitatively different features of memory that together engender sophisticated episodic remembering (Poppenk et al. 2013). In particular, it has been proposed that the anterior hippocampus and its connections construct coarse or gisty features of memory, while the posterior hippocampus and its interactions with posterior and sensory regions is important for elaborating on the rich details necessary for episodic memory (Audrain et al. 2022; Audrain and McAndrews 2022; Gilmore, Quach, Kalinowski, González‐Araya, et al. 2021; Robin and Moscovitch 2017). However, we know very little about the maturity of hippocampal‐neocortical functional connections along the long‐axis in the earliest years of life. Until recently, very little data were available with which to investigate functional neural changes from birth through childhood.

In the present study, we investigate in detail the development of hippocampal functional coupling with the neocortex along the hippocampal long‐axis across the first two postnatal years in a large cohort of 212 human infants and toddlers. We asked (1) At what point in development does communication between the hippocampus and canonical memory‐related neocortex in adults increase? (2) Do anteroposterior distinctions in hippocampal connectivity exist at birth, and does this differentiation increase or decrease with development? And finally, (3) Does the developmental trajectory of connectivity with the neocortex differ for anterior and posterior segments of the hippocampus?

2. Materials and Methods

2.1. Experimental Design

2.1.1. Participants

Neuroimaging data was acquired as part of the Baby Connectome Project (Howell et al. 2019). A total of 319 participants spanning 694 sessions were available for analysis. A small proportion of participants were scanned multiple times across sessions with age. In order to adhere to assumptions regarding statistical independence of data in our models we included data from only one session per participant, choosing the session with the highest whole‐brain temporal signal‐to‐noise ratio (tSNR) of the resting‐state data. Fifty‐four participants were excluded due to poor data quality quantified by excessive motion (AFNI's @1dDiffmag values > 0.3 mm/TR) or low whole‐brain tSNR (< 30). An additional 10 participants were excluded due to failed co‐registration during data preprocessing that could not be remedied without extensive manual alteration. As there were few participants older than 25 months (n = 38 spanning 26–69 months of age), we decided to focus our analyses on participants less than 26 months of age. This yielded 212 infants and toddlers between the ages of < 1 and 25 months (112 female, 110 male), which we divided into four age bins spanning 6–7 months in age: 0–6 months (n = 50), 7–12 months (n = 59), 13–18 months (n = 58), and 19–25 months (n = 45). The decision to bin participants into discrete categorical age ranges was made to allow the exploration of the relationship between age and connectivity without making assumptions about the shape of the association (i.e., linear or nonlinear, as required when modeling continuous data). The age bin cutoffs were constructed to allow a comparable number of participants in each age bin, with roughly equal age ranges. A histogram of the specific ages of participants in each age bin can be viewed in Figure S1. We additionally present continuous analyses and visualizations where appropriate, to ensure that age binning does not obscure underlying continuous effects.

2.1.2. Procedure and Data Acquisition

Detailed information regarding the imaging protocol is described by Howell et al. (2019). All imaging was acquired on 3T Siemens Prisma MRI scanners at the Center for Magnetic Resonance Research (CMRR) at the University of Minnesota and the Biomedical Research Imaging Center (BRIC) at the University of North Carolina at Chapel Hill, using a Siemens 32 channel head coil. We here utilize the T1w (MPRAGE, 0.8 mm isotropic voxels, TE = 2.24, TR = 2400/1060), T2w (3D variable flip angle turbo spin‐echo sequence, 0.8 mm isotropic voxels, TE = 564, TR = 3200), and two resting‐state T2*‐weighted functional scans (gradient‐echo EPI acquisition: Multiband 8, TR = 800 ms, TE = 37 ms, flip angle = 52°, FOV = 208 mm, 104 × 91 matrix, 2 mm isotropic voxels) with anterior‐to‐posterior and posterior‐to‐anterior encoding (420 TRs, 5 min 36 s for each encoding; total EPI duration: 840 TRs, 11 min 12 s) per participant. All participants were asleep during scanning.

2.2. Data Processing and Analyses

2.2.1. Preprocessing

Data processing was conducted with Freesurfer (RRID: SCR_001847), infant Freesurfer (Zöllei et al. 2020), and AFNI (RRID: SCR_005927). The T1w anatomical data were processed using either Freesurfer (for participants > 10 months of age) or infant Freesurfer (for participants ≤ 10 months of age) for tissue segmentation. Different versions of Freesurfer and registration protocols were required for the different age ranges, as the gray matter white matter contrast is poor in infants less than a year old, often resulting in failed convergence for adult Freesurfer and registration to common space. In practice, we found that Freesurfer tended to fail for participants ≤ 10 months of age, which led to the use of infant Freesurfer for this age group. Importantly, anatomical segmentations from the two different versions of Freesurfer were used only for (1) creating CSF and WM masks for nuisance regression (described below), and (2) creating preliminary hippocampal masks, which were visually inspected and manually corrected where necessary (see details in Section 2.2.2). Gray matter, white matter, and CSF volumes were linearly aligned to the EPI data using AFNI's 3dAllineate, and all analyses were performed in functional volume space rather than surface space. Therefore, minor differences in CSF or WM masks between Freesurfer versions are expected to have negligible impact on the functional results reported here. Indeed, comparing the overlap between CSF and WM masks between infants processed with each version of Freesurfer revealed high overlap (Figure S2C). In infants > 10 months, warping the T1w data to the EPI data produced a good co‐registration. For infants ≤ 10 months, the T1w scan was first aligned to the T2w scan (which had better gray matter/white matter contrast). The T2w scan was then co‐registered with the EPI data, and the warp was applied to the T1w data to align it with the EPI data. The registration was visually inspected for each participant to identify failures of gross misalignment, which were excluded from the analysis (see Section 2.1.1). We observed good spatial agreement between the two alignment strategies (Figure S2A,B).

The functional data were despiked (3dDespike), adjusted for slice‐time acquisition (3dTshift), warped to the midpoint of the P‐A and A‐P encodings and volume registered using AFNI's unWarpEPI, and frame‐to‐frame motion was estimated. The data were then smoothed with a 4 mm FWHM kernel (3dmerge) and rescaled to % signal change. We then applied a modified form of the ANATICOR (Jo et al. 2010) denoising approach with a nuisance regression (3dTfitter) containing: 6 motion parameters, local average of WM signal (averaged within a 15 mm radius sphere), average ventricle signal, as well as the top 3 principal components of the time series from all white matter and ventricle voxels (modified aCompCor, Behzadi et al. 2007; Stoddard et al. 2016). All regressors were detrended with a fourth‐order polynomial before denoising (3dDetrend), and the same detrending was applied during nuisance regression to the voxel time series. The residual time series were concatenated across the A‐P and P‐A resting‐state runs. Subject‐specific anatomical and functional data were first registered to an age‐specific MNI template, and subsequently, to the median age T1w MNI template (11–14 months) using nonlinear warps (3dQwarp).

2.2.2. Region of Interest Definition

The hippocampus was isolated from the Freesurfer or infant Freesurfer subcortical segmentations for each participant and manually corrected where necessary. We note here that Freesurfer generally performed well at identifying the boundaries of the hippocampus, but Infant Freesurfer sometimes failed for the youngest infants, likely due to poor gray matter/white matter contrast. Importantly, the hippocampus maintains consistent anatomical boundaries across ages. To address this, we visually inspected the boundaries of all 212 sets of hippocampi and manually corrected them where necessary to ensure anatomically accurate masks that did not extend into adjacent cortex or CSF across ages. The hippocampus was then manually segmented into anterior and posterior segments at the uncal notch (Poppenk et al. 2013) to create anterior and posterior masks for each hemisphere for every participant. The uncal notch was also identifiable in even the youngest infants. The resulting masks were co‐registered to the group analysis space (median age template) using the warps created during preprocessing.

2.2.3. Functional Connectivity Computation

We conducted a seed‐to‐voxel functional connectivity analysis using the left and right anterior and posterior hippocampal masks as seeds. Specifically, the average time‐course across voxels within each hippocampal region was calculated using AFNI's 3dmaskave. Next, the Pearson's correlation between each hippocampal ROI time‐course and the time‐course of each voxel in the brain was calculated using 3dTcorr1D. The resulting correlation maps (one for each ROI, reflecting whole‐brain functional connectivity) were Fisher‐z transformed to normalize the data for group analyses.

2.2.4. Group‐Level Contrasts

We ran a hierarchical mixed‐effects model using AFNI's 3dMVM to test for group differences in connectivity across the age bins in infants and toddlers aged 0–25 months. Specifically, connectivity was predicted as a function of long‐axis (anterior/posterior), age bin (0–6 months/7–12 months/13–18 months/19–25 months), and hemisphere (left/right). We also included whole brain mean tSNR and mean motion (estimated with AFNI's 3dDiffmag) as regressors of no interest to adjust effects for variability in tSNR and motion across participants. tSNR is useful for inclusion in these analyses as it is a composite of all sources of nuisance variability that impact the BOLD signal and functional connectivity, including changes in signal strength with development alongside unmodeled effects of motion and respiration. We were particularly interested in the interaction between long‐axis compartments and age bin. To directly explore the degree of differential anterior–posterior connectivity at each age, we additionally specified a priori pairwise contrasts comparing anterior > posterior hippocampal connectivity within each age bin. We also contrasted anterior and posterior hippocampal connectivity separately across successive age bins (7–12 months > 0–6 months, 13–18 months > 7–12 months, 19–25 months > 13–18 months) to better understand the main effect of age bin on hippocampal‐cortical connectivity. Finally, we specified linear contrasts within each age bin to test connectivity as a function of continuous age separately for anterior, posterior, and anterior > posterior difference maps, to determine whether our age bins were reasonably grouping together participants with similar connectivity. Voxels surviving a threshold FDR‐corrected to q < 0.01 corrected across all voxels in the brain were considered statistically significant. We chose this method of correction because it is robust to inflated false positives due to multiple comparisons. To further mitigate false positives, we additionally conservatively implemented a cluster‐forming threshold to remove very small clusters of less than 10 contiguous voxels.

2.2.5. Hierarchical Clustering Analysis

Having identified several neocortical clusters of voxels where connectivity was changing along the long‐axis as a function of age (quantified by a significant long‐axis × age bin interaction), we next sought to describe the patterns driving the interaction for each region identified. To facilitate interpretation, we clustered together neocortical regional clusters evincing a similar direction of interaction to form what we term here, “superclusters.” As we did not know the number of superclusters a priori, we used a hierarchical clustering approach. We extracted the average connectivity across participants and hemispheres between the anterior and posterior hippocampi and each neocortical cluster where there was a significant interaction for each age bin. We subtracted posterior from anterior hippocampal connectivity to get a measure of the difference in connectivity along the long‐axis for each neocortical cluster for each age bin. We calculated the difference in connectivity along the long‐axis because we were interested specifically in the interaction between long‐axis and age bin and did not want the clustering solution to be driven by the main effects of long‐axis (as in the case of including anterior and posterior connectivity values separately). The result was a 44 neocortical clusters × 4 age bins matrix of anterior–posterior differential connectivity. We then transformed this matrix into a distance matrix by quantifying the Euclidean distance between each of these observations. Finally, we clustered the distance matrix into superclusters using Ward's method.

The agglomerative clustering process starts with one cluster for each observation (in this case, the connectivity values between the hippocampus and each neocortical cluster) and combines observations with the smallest distance between them to form new clusters (i.e., superclusters). Ward's method sums the squared Euclidean distances between each member of a cluster and the cluster centroid, joining clusters which increase the sum of within‐cluster square errors the least (Ward 1963). A new distance is calculated between new clusters, and the process is repeated until there is only one cluster with all observations in it. We used the agglomeration schedule to determine the optimal number of superclusters, choosing the point in the agglomerative process where the distance between conglomerated clusters jumps, indicating the combination of distant clusters that are different enough that they likely shouldn't be combined.

2.2.6. Cluster Validation

We took several steps to validate the hierarchical clustering solution (see Supplementary Methods 1 in Supporting Information for details). Briefly, we first visually verified that the superclusters were distinct by plotting the first two dimensions of a principal components analysis. Second, we plotted differential anterior–posterior hippocampal connectivity over age bins for each neocortical cluster comprising the superclusters, to visually check that neocortical clusters were being grouped together sensibly based on the interaction between long‐axis and age bin. Third, we calculated silhouette statistics to assess overall supercluster quality, and to quantify the degree to which neocortical clusters within each supercluster fit the supercluster they were assigned to versus other superclusters. In order to perform statistics to compare the patterns of each supercluster, it was important to verify that the comparisons would not be biased by subject‐level noise as a result of performing the selection and comparisons on the same data (i.e., double‐dipping; e.g., Kriegeskorte et al. 2008). To evaluate this, the hierarchical clustering was repeated while leaving each subject out once (212 times). We then tabulated which clusters were grouped together into superclusters across all 212 iterations, which by definition does not depend on subject‐level noise. We then constructed “bias‐free” estimates of the supercluster patterns using a weighted average of only those pairs of clusters that were grouped together 100% of the time. The bias‐free, cross‐validated patterns could then be compared to the original patterns based on the entire group of subjects. We note here that the two ROIs that comprise the parahippocampal supercluster were only clustered together 87% of the time in our leave‐one‐out validation approach, precluding our ability to compute connectivity estimates free of subject‐level noise. We report the statistics on the original connectivity estimates for this supercluster throughout for completeness.

2.3. Statistical Analyses

Statistical analyses of bias‐free estimates of connectivity for each supercluster were computed in R studio version 2023.12.1.402 (Posit team 2024) (http://www.posit.co/). We used hierarchical mixed‐effects models to test connectivity as a function of supercluster, long‐axis, and age bin. We included a random intercept for each participant using the nlme package in R (version 3.1.158; https://CRAN.R‐project.org/package=nlme) (Pinheiro and Bates 2000). Random slopes for predictors were included when they improved model fit, as determined by reductions in Akaike information criterion (AIC) values. In the event that residuals were heteroskedastic, we specified the variance structure to allow the variance to vary across levels of the heteroskedastic fixed effects, as identified using Levene's test for unequal variance as part of the car package in R (Fox and Weisberg 2019) (version 3.1.0; https://cran.r‐project.org/web/packages/car/index.html). Denominator degrees of freedom and p‐values were estimated using the containment method as implemented in nlme. Post hoc comparisons were interrogated using pairwise t‐tests on the estimated marginal means from the omnibus model, using the emmeans package (version 1.7.5, https://cran.r‐project.org/web/packages/emmeans/index.html), and were FDR corrected to control for multiple comparisons.

3. Results

The developmental trajectory of hippocampal‐cortical functional connectivity in infants is largely unknown. In order to investigate this, we employed a seed‐to‐voxel analysis to measure functional connectivity of the anterior and posterior hippocampus to the whole brain, both separately and in contrast to one another. We tested for differences in connectivity in infants and toddlers (0–25 months old) in the BabyConnetome dataset (Howell et al. 2019) as a function of long‐axis (anterior/posterior), age (bins of 0–6 months/7–12 months/13–18 months/19–25 months), and hemisphere (left/right), controlling for fMRI signal quality and motion. To examine the effects of age and long‐axis with development in more detail, we also contrasted anterior and posterior hippocampal connectivity within and across successive age bins. Throughout this manuscript, we refer to cortical regions according to their associated cognitive function in adults for ease of description (e.g., “memory‐related regions”). We acknowledge here that (1) the regions identified do not necessarily encompass all regions involved in the cognitive functions described, at least in adults, and (2) The function of these cortical regions in infants has yet to be definitively demonstrated.

3.1. Hippocampal Coupling With Memory‐Related Cortex Increases Across the First 6 Months of Age

We first conducted separate analyses of functional connectivity between the anterior and posterior portions of the hippocampus with the cortex. Between 0–6 and 7–12 months of age, we observed large increases in both anterior and posterior hippocampal connectivity with what are memory‐related regions in adults, including the medial PFC, medial parietal, angular gyri, posterior inferior temporal cortex, anterior temporal cortex, and superior and middle frontal gyri (Figures 1A and 2A). For these same ages, decreases in connectivity between the anterior hippocampus and postcentral sulcus were also observed (Figure 1A). There were fewer changes in hippocampal connectivity between the 7–12 and 13–18 month age bins, although coupling between the anterior hippocampus and medial PFC continued to increase as connectivity with occipital and left posterior middle and inferior temporal gyri decreased. There were no reliable changes in connectivity between the 13–18 and 19–25 month age bins for either the anterior or posterior hippocampus.

FIGURE 1.

Effects of age on anterior hippocampal‐cortical connectivity across the first 25 postnatal months. Surface reconstruction of cortical regions showing age‐related changes in anterior hippocampal‐cortical connectivity. (A) Pairwise contrasts across successive age bins, showing greater connectivity with age (warm colors) as well as reduced connectivity with age (cool colors). (B) Cortical regions showing changes in anterior hippocampal‐cortical connectivity as a function of continuous age within each age bin. Warm colors denote greater connectivity with increasing age, and cool colors denote decreased connectivity with increasing age. Data were surface projected to the adult MNI template using Connectome Workbench software for visualization purposes (Marcus et al. 2011). k = 10, q < 0.01. L: Left hemisphere; mo: Months; R: Right hemisphere.

FIGURE 2.

Effects of age on posterior hippocampal‐cortical connectivity across the first 25 postnatal months. Surface reconstruction of cortical regions showing age‐related changes in posterior hippocampal‐cortical connectivity. (A) Pairwise contrasts across successive age bins, showing greater connectivity with age (warm colors), as well as reduced connectivity with age (cool colors). (B) Cortical regions showing changes in posterior hippocampal‐cortical connectivity as a function of continuous age within each age bin. Warm colors denote greater connectivity with increasing age, and cool colors denote decreased connectivity with increasing age. Data were surface projected to the adult MNI template using Connectome Workbench software for visualization purposes (Marcus et al. 2011). k = 10, q < 0.01. L: Left hemisphere; mo: Months; R: Right hemisphere.

We ran additional linear models examining hippocampal‐cortical connectivity as a continuous function of age within each age bin. Across the first 6 months, we observed large increases in hippocampal‐cortical connectivity to canonical cortical memory regions in adults for both anterior and posterior hippocampus (Figures 1B and 2B). There were no reliable differences in connectivity as a function of age within the older age bins. Together, these results indicate that the most extensive changes in hippocampal‐cortical connectivity occurred across the first 6 months and were primarily characterized by increases in coupling with what are memory‐related cortical regions in adults. However, it is not the case that there were no changes in hippocampal connectivity beyond 6 months. Exploratory linear models predicting hippocampal‐cortical connectivity as a continuous function of age from 7 to 25 months revealed increases in connectivity between the anterior hippocampus and motor cortex and medial occipital cortices, and between the posterior hippocampus and medial prefrontal, anterior insula, and medial occipital cortices with age. For both segments of the hippocampus, we observed decreases in connectivity with much of the ventral and lateral temporal cortex, medial and lateral parietal cortices, posterior insula, and cingulate sulcus within the frontal lobes with age (Figure S3). Thus, hippocampal‐cortical connectivity continues to develop beyond 6 months, albeit more gradually.

3.2. Functional Connectivity Along the Long‐Axis of the Hippocampus Is Differentiated in Even the Youngest Infants

Next, we tested for differences in connectivity between anterior and posterior hippocampal segments with age (Figure 3). We observed differentiation of connectivity along the long‐axis in all age bins, including the youngest infants aged 0–6 months. Specifically, we found that the anterior hippocampus was more functionally connected to medial PFC and the anterior temporal lobes, while the posterior hippocampus was more broadly connected to medial parietal cortex, cingulate cortex, posterior lateral and inferior temporal cortex, motor cortex, occipital cortex, and the middle frontal gyrus. When we ran linear models examining the difference in long‐axis connectivity as a function of age within each age bin, we observed no reliable changes within bins, suggesting the observed anterior–posterior distinctions were generally consistent across ages within each age bin.

FIGURE 3.

Difference in anteroposterior hippocampal connectivity with the neocortex by age bin across the first 25 postnatal months. Surface reconstruction of cortical regions with significantly greater connectivity with the anterior compared to the posterior hippocampus (warm colors), and regions with greater connectivity with the posterior relative to the anterior hippocampus (cool colors). Data were surface projected to the adult MNI template using Connectome Workbench software for visualization purposes (Marcus et al. 2011). k = 10, q < 0.01. L: Left hemisphere; mo: Months; R: Right hemisphere.

3.3. Maturation of Functional Connectivity Along the Hippocampal Long‐Axis Reveals Canonical Cortical Networks Associated With Cognition in Adults

While there were evident similarities in anteroposterior connectivity across the age bins, there were also some notable changes with development. We next sought to identify differences in long‐axis connectivity as a function of age. To answer this question, we turned to the interaction between long‐axis and age bin. We found 44 neocortical clusters where the interaction between long‐axis and age bin was significant (cluster size k = 10 voxels, p = 0.001, q < 0.01, Figure S4A), spanning the dorsal medial PFC, lateral PFC, premotor cortex, anterior insula, posterior parietal cortex, superior parietal lobule, medial temporal lobe cortex, inferior temporal cortex, superior temporal gyrus, and occipital cortex. Thus, while some degree of anteroposterior differentiation was present in even the youngest infants, there were numerous cortical areas where connectivity was still developing with the hippocampus along its long‐axis. We also found some neocortical clusters wherein hemisphere interacted with our variables of interest (long‐axis and age bin), which we report in Figures S5–S7 for the interested reader.

We next sought to quantify exactly how connectivity along the long‐axis was changing with development. As the neocortical clusters elicited by a significant interaction between long‐axis and age bin could reflect numerous different patterns of connectivity, we took a hierarchical clustering approach to sort these clusters into superclusters sharing similar patterns of connectivity with the hippocampus across age bins. Specifically, we subtracted each neocortical cluster's connectivity with the posterior from the anterior hippocampus to get a measure of differential connectivity along the long‐axis for each age bin and grouped neocortical clusters with similar profiles of hippocampal connectivity across age bins together. Our aim with this analysis was to facilitate interpretation by grouping the 44 neocortical clusters together according to the direction of the significant interaction.

Using this method, we identified six superclusters consisting of neocortical clusters that shared similar connectivity profiles with the hippocampus across age bins (for further details, see Figure S4B,C). We took several steps to validate the clustering solution, as detailed in Supplementary Methods 1 in Supporting Information. To statistically test connectivity differences across the long‐axis for each supercluster, we constructed estimates of supercluster patterns free of subject‐level bias using a leave‐one‐out validation approach (see Supplementary Methods 1 in Supporting Information). We used these bias‐free connectivity estimates for all reported statistical testing unless stated otherwise. A linear mixed‐effects model predicting connectivity as a function of long‐axis, age bin, and supercluster indicated a significant three‐way interaction between the predictors (F(12,1872) = 10.92, p < 0.0001), confirming that the identified superclusters describe reliable differences in connectivity. We present the spatial localization of each supercluster on the brain in Figure 4A, as well as the connectivity profiles across age bins for each supercluster averaged across the neocortical clusters included. We additionally plot the individual subject‐level data alongside locally smoothed (LOESS) regression lines to visualize how connectivity continuously changes within and across age bins (see Figures S8–S13, Tables S1–S3 for related results). Three of the identified superclusters consisted of core neocortical regions canonically associated with episodic memory in adults (Figure 4B,C), and three superclusters comprised canonical networks typically associated with other types of cognition (Figure 4D). Post hoc two‐way models revealed significant interactions between long‐axis and age bin for each supercluster (medial parietal: F(3,208) = 22.95, p < 0.0001; entorhinal: F(3,208) = 12.21, p < 0.0001; parahippocampal: F(3,208) = 29.31, p < 0.0001; cingulo‐opercular: F(3, 208) = 41.72, p < 0.0001; dorsal frontoparietal: F(3,208) = 19.29, p < 0.0001; mPFC‐STS: F(3,208) = 24.48, p < 0.0001). We describe pairwise comparisons below.

FIGURE 4.

Superclusters of neocortical regions sharing similar patterns of changing connectivity along the hippocampal long‐axis with age across the first 25 postnatal months. Hierarchical clustering was used to group neocortical clusters (n = 44) elicited by a significant interaction between hippocampal long‐axis and age bin on connectivity (k = 10, q < 0.01) into superclusters (k = 6) sharing similar patterns of connectivity. (A) Brain images show the clustering solution, with each color denoting a different supercluster. Line graphs detail connectivity changes of the anterior and posterior hippocampus with age for superclusters canonically associated with memory in adults (B, C) and with those classically associated with other cognitive processes (D). Graphs in the first (left) column show average connectivity of each supercluster with the anterior and posterior hippocampus for each age bin. Graphs in the second and third columns show the individual subject data for the same contrasts, alongside a locally smoothed (LOESS) regression line. Graphs in the fourth (right) column show the LOESS regression trendlines for the anterior and posterior hippocampus (from the second and third columns) together to facilitate comparison with age bin data (first column). Error bars reflect standard error across participants, and gray ribbons reflect 95% confidence intervals. Vertical gray lines denote age bin cutoffs. Asterisks (*) denote statistically significant differences between anterior and posterior connectivity within age bin (p < 0.05 FDR corrected). Data were surface projected using Connectome Workbench software (Marcus et al. 2011) to the adult MNI surface for visualization. L: left hemisphere; mo: months; R: right hemisphere. Plots in B and D reflect data from a weighted average after a leave‐one‐out validation analysis of regions clustered together 100% of the time across 212 iterations (see Supplementary Methods 1 in Supporting Information for details). For related results, see Figures S8–S13 and Tables S1–S3.

3.3.1. Development of Hippocampal Connectivity With Cortical Regions Canonically Associated With Memory in Adults

Three of the identified superclusters were comprised of regions canonically associated with episodic and spatial memory in adults, and included a medial parietal supercluster, as well as entorhinal and parahippocampal superclusters (Figure 4B,C). Across these superclusters, we generally observed a decrease in anteroposterior hippocampal differentiation with age. The medial parietal supercluster connectivity profile was characterized by greater posterior–anterior connectivity at 0–6 months (t(1872) = −5.17, p < 0.0001) which disappeared in subsequent age bins (7–12 months: t(1872) = −0.98, p = 0.33; 13–18 months: t(1872) = −0.25, p = 0.8; 19–25 months: t(1872) = −1.55, p = 0.12), due to an increase in connectivity with the anterior hippocampus across the first year (0–6 months vs. 7–12 months: t(208) = −5.82, p < 0.0001; 7–12 months vs. 13–18 months: t(208) = 1.02, p = 0.31; 13–18 months vs. 19–25 months: t(208) = 1.33, p = 0.22). There were no differences in connectivity with the posterior hippocampus across successive age bins (0–6 months vs. 7–12 months: t(208) = −1.81, p = 0.2; 7–12 months vs. 13–18 months: t(208) = 1.66, p = 0.2; 13–18 months vs. 19–25 months: t(208) = 0.24, p = 0.87).

The entorhinal supercluster consisted of only the left and right entorhinal cortex, and the parahippocampal supercluster consisted of the left and right parahippocampal cortex. These superclusters had opposing patterns of connectivity with development. At 0–6 months, the entorhinal supercluster showed a large difference in hippocampal connectivity with greater anterior than posterior connectivity (t(1872) = 10.82, p < 0.0001) which attenuated with age. This attenuation can be attributed to connectivity with the posterior hippocampus quickly increasing across the first 12 months (0–6 months vs. 7–12 months: t(208) = −3.49, p = 0.003; 7–12 months vs. 13–18 months: t(208) = 0.82, p = 0.49; 13–18 months vs. 19–25 months: t(208) = 2.61, p = 0.015), and connectivity with the anterior hippocampus gradually decreasing by 19–25 months (0–6 months vs. 7–12 months: t(208) = 0.01, p = 0.99; 7–12 months vs. 13–18 months: t(208) = 1.49, p = 0.21; 13–18 months vs. 19–25 months: t(208) = 3.06, p = 0.005). Despite this attenuation, connectivity with the anterior hippocampus remained stronger than with the posterior for all age bins (7–12 months: t(1872) = 7.79, p < 0.0001; 13–18 months: t(1872) = 7.42, p < 0.0001; 19–25 months: t(1872) = 5.44, p < 0.0001).

The parahippocampal supercluster showed the opposite pattern with a large difference in connectivity driven by greater posterior connectivity than anterior at 0–6 months (t(208) = −16, p < 0.0001) which diminished with age. This diminishing effect was partially driven by anterior connectivity quickly increasing across the first 12 months (0–6 months vs. 7–12 months: t(208) = −4.20, p = 0.0001; 7–12 months vs. 13–18 months: t(208) = −0.18, p = 0.86; 13–18 months vs. 19–25 months: t(208) = 2.70, p = 0.015), while posterior connectivity gradually attenuated by 19–25 months (0–6 months vs. 7–12 months: t(208) = 1.72, p = 0.1; 7–12 months vs. 13–18 months: t(208) = 0.32, p = 0.75; 13–18 months vs. 19–25 months: t(208) = 3.08, p = 0.005). Posterior hippocampal connectivity remained higher than anterior to this cluster across all age bins, despite the attenuation with age (7–12 months: t(208) = −9.54, p < 0.0001; 13–18 months: t(208) = −9.68, p < 0.0001; 19–25 months: t(208) = −8.01, p < 0.0001). We note here that the parahippocampal supercluster only achieved 87% consistency across iterations, precluding our ability to compute connectivity estimates free of subject‐level noise. We report the statistics on the uncorrected connectivity estimates for this supercluster for completeness.

To summarize, for all three memory‐related superclusters we observed evidence of differentiated connectivity with the hippocampal long‐axis very early in development, which diminished as connectivity with the weaker segment of the long‐axis strengthened across the first year. In medial temporal areas, connectivity gradually decreased with the predominant segment of the long‐axis across the first 2 years.

3.3.2. Development of Hippocampal Connectivity With Cortical Networks Associated With Attention, Salience, and Social Processing in Adults

Interestingly, the remaining three superclusters revealed networks typically associated with other types of cognition beyond episodic memory in adults (Figure 4D). One supercluster primarily consisted of the dorsal anterior cingulate and anterior insula, reminiscent of the cingulo‐opercular (or salience) network in adults. The connectivity profile of this supercluster was characterized by greater anterior–posterior connectivity at 0–6 months (t(1872) = 6.60, p < 0.0001) which gradually flipped with development such that by 19–25 months there was greater connectivity of this supercluster to the posterior hippocampus than the anterior (7–12 months: t(1872) = 3.75, p = 0.0002; 13–18 months: t(1872) = 1.28, p = 0.2; 19–25 months: t(1872) = −4.86, p < 0.0001). This pattern was driven by an increase in connectivity of this cluster with the posterior hippocampus across the first year (0–6 months vs. 7–12 months: t(208) = −2.55, p = 0.023; 7–12 months vs. 13–18 months: t(208) = −1.39, p = 0.25; 13–18 months vs. 19–25 months: t(208) = 0.41, p = 0.69), as well as a decrease in connectivity with the anterior hippocampus between the 13–18 month and 19–25 month age bins (0–6 months vs. 7–12 months: t(208) = 0.41, p = 0.89; 7–12 months vs. 13–18 months: t(208) = −0.33, p = 0.89; 13–18 months vs. 19–25 months: t(208) = 3.27, p = 0.005). Connectivity between the cingulo‐opercular network and hippocampus was therefore gradually inverting with age, with a bias of stronger connectivity with the anterior hippocampus shifting to the posterior hippocampus with maturation.

We also found a supercluster reminiscent of the dorsal frontal–parietal (or dorsal attention) network in adults, consisting of the superior parietal lobule, lateral PFC, and inferior temporal gyrus. There was no reliable difference in anterior compared to posterior hippocampal connectivity with this supercluster at 0–6 months (t(1872) = −2.32, p = 0.02), but there was greater connectivity with the posterior hippocampus than anterior from 7 to 12 months onwards (7–12 months: t(1872) = −8.21, p < 0.0001; 13–18 months: t(1872) = −5.62, p < 0.0001; 19–25 months: t(1872) = −6.65, p < 0.0001), due to an increase in posterior hippocampal connectivity across the first year (0–6 months vs. 7–12 months: t(208) = −3.19, p = 0.01; 7–12 months vs. 13–18 months: t(208) = 1.92, p = 0.11; 13–18 months vs. 19–25 months: t(208) = 0.18, p = 0.86). Connectivity of the anterior hippocampus with this supercluster did not change across successive age bins (0–6 months vs. 7–12 months: t(208) = 1.5, p = 0.29; 7–12 months vs. 13–18 months: t(208) = 0.44, p = 0.79; 13–18 months vs. 19–25 months: t(208) = −0.49, p = 0.79). Therefore, while there was no differentiation in anteroposterior hippocampal connectivity with the dorsal frontal parietal network in the first 6 months of life, by 1 year of age connectivity with the posterior hippocampus strengthened and remained stronger than that with the anterior hippocampus thereafter.

Finally, we identified a supercluster of regions consisting of superior temporal gyrus, vmPFC, and the occipital pole, which are primarily regions that activate during social and schematic processing in adults (Adolphs 2009; Gotts et al. 2012). This vmPFC‐superior temporal supercluster was characterized by the lack of an anterior–posterior difference in hippocampal connectivity at 0–6 months (t(1872) = −1.54, p = 0.124). Anterior connectivity became greater than posterior at 7–12 (t(1872) = 5.83, p < 0.0001) and 13–18 months (t(1872) = 4.18, p < 0.0001). Anterior–posterior connectivity was not reliably different by 19–25 months (t(1872) = 1.78, p = 0.077). This pattern was driven by an inverted U connectivity profile of the anterior hippocampus across successive age bins, where connectivity peaked between 7–18 months (0–6 months vs. 7–12 months: t(208) = −5.55, p < 0.0001; 7–12 months vs. 13–18 months: t(208) = 0.19, p = 0.85; 13–18 months vs. 19–25 months: t(208) = 2.42, p = 0.02). Posterior hippocampal connectivity did not reliably change across successive age bins (0–6 months vs. 7–12 months: t(208) = −0.55, p = 0.84; 7–12 months vs. 13–18 months: t(208) = 0.45, p = 0.84; 13–18 months vs. 19–25 months: t(208) = 0.56, p = 0.84). It was therefore anterior hippocampal connectivity with these social regions that was changing with development across the first 2 years of age.

4. Discussion

In this study, we examined the development of hippocampal connectivity along the long‐axis in the first 2 years of postnatal life. We were surprised to observe the most extensive increases in connectivity between the hippocampus and memory‐related cortical regions in adults occurred across the first 6 months. We also found evidence that an appreciable degree of anterior–posterior differentiation in connectivity was present even in the youngest infants (0–6 months), indicating that major anterior–posterior divisions in hippocampal connectivity are present very early on. Specifically, the anterior hippocampus was more strongly connected to ventral medial PFC and anterior temporal regions, while the posterior hippocampus was more connected to much of the posterior medial cortex, as well as the superior parietal lobule, lateral middle prefrontal gyri, and posterior temporal‐occipital cortex.

This general division is similar to the topological pattern observed in adults (Adnan et al. 2016; Barnett et al. 2019; Frank et al. 2019; Grady 2020; Robinson et al. 2016; Vogel et al. 2020), and was present across all age bins examined. This finding differs from conclusions drawn by a recent study that did not observe evidence of differentiation along the long‐axis in infants aged 9–11 months (Howell et al. 2020). However, the methods employed by their analysis differed from ours. The authors investigated average connectivity of the hippocampus with functional networks that were defined a priori based on past parcellations in adults. The present work and work from others suggest that hippocampal‐cortical connectivity is still developing over the first two postnatal years, becoming more topologically similar to adults (Liu et al. 2021). It is possible that averaging connectivity across networks defined with adult parcellations was simply not sensitive enough to reveal differences in anterior–posterior connectivity, compared to our seed‐voxel analysis which was not topologically constrained. Notably, in our work, the differential anterior–posterior connectivity maps of participants aged 19–25 months (the oldest participants here examined; Figure 3) did not look exactly like those of adults, highlighting this issue. In adults the anterior hippocampus is more connected to motor cortex, superior and inferior lateral frontal gyri, and the angular gyrus (Barnett et al. 2019; Frank et al. 2019; Robinson et al. 2016; Vogel et al. 2020), while by age 19–25 months these regions were either dominated by posterior hippocampal connectivity or equal connectivity between the two axes. Presumably, differential anterior > posterior connectivity will continue to fine‐tune through later ages not examined here, by either anterior connections increasing or posterior decreasing. Within the first 25 months, both processes were occurring depending on the cortical region examined, and so we did not find strong evidence that one segment along the long‐axis of the hippocampus was more mature than the other within the first 2 years.

Thus, we observed that much of memory system development, as indexed by topological patterns of connectivity, occurs quite early on. While primary sensory networks are topologically adult‐like at birth, networks associated with higher‐order cognition continue to mature post‐gestation. A primitive medial frontoparietal network (also known as the default mode network) becomes more similar to adults across the first post‐gestational year, followed shortly after by refinement of cingulo‐opercular and lateral frontoparietal networks (Gao, Alcauter, Elton, et al. 2015). The medial frontoparietal and dorsal frontoparietal networks don't become anticorrelated (as in adults) until approximately 1 year of age (Gao et al. 2013), which is when the hippocampus starts synchronizing with the medial frontoparietal network at rest (Gao et al. 2009, 2013). Hippocampal activity related to encoding also emerges around 1 year of age (Yates et al. 2025). We provide evidence that rudimentary functional architecture in the form of hippocampal synchronization with core memory areas exists by 6 months of age.

In addition to large, early increases in hippocampal connectivity with canonical memory regions in adults, we also described protracted and maturing hippocampal‐cortical interactions along the long‐axis with age. Specifically, we detected many neocortical clusters wherein hippocampal connectivity along the long‐axis was changing with development, as quantified by a significant interaction between hippocampal long‐axis and age bin. When we clustered these neocortical regions together based on similarity of the direction of the interaction, we observed different patterns of developing connectivity along the hippocampal long‐axis with areas of the neocortex typically associated with episodic memory and spatial processing in adults, namely entorhinal, parahippocampal, and medial parietal regions (Coutureau and Di Scala 2009; Ritchey and Cooper 2020; Rugg and Vilberg 2013). Specifically, we found evidence of stronger anterior connectivity to entorhinal cortex, and greater posterior hippocampal connectivity to parahippocampal cortex at 0–6 months. This strong differential connectivity along the hippocampal long‐axis with these medial temporal cortical regions attenuated by 7–12 months as connectivity increased with the weaker axis (i.e., connectivity between anterior hippocampus and parahippocampal gyrus, and between posterior hippocampus and entorhinal cortex increased). In the medial parietal supercluster, there was stronger posterior connectivity than anterior connectivity at 0–6 months that disappeared by 7–12 months as connectivity of the anterior hippocampus with these regions increased. These findings are generally consistent with work that shows that connectivity is greatly increasing within adult‐like functional networks within the first year of life (Gao, Alcauter, Smith, et al. 2015; Gao et al. 2013), and between the whole hippocampus and neocortex (Liu et al. 2021). We found that connectivity of anterior–posterior regions of the hippocampus with these canonical episodic memory areas in adults are quite functionally segregated in the first few months of life. With development, segregation lessens somewhat as the weaker segment along the long‐axis of the hippocampus becomes more integrated into the network.

We also identified protracted hippocampal interactions with functional networks commonly associated with cognition beyond episodic memory in adults (Uddin et al. 2019). Connectivity of the hippocampus with the cingulo‐opercular network (or the salience network, often involved in tasks requiring sustained cognitive control and vigilance in adults; Uddin 2017) changed linearly with age. At 0–6 months the anterior hippocampus was more connected to this network than the posterior hippocampus, but this pattern gradually changed over the course of development such that by 19–25 months the posterior hippocampus was more connected than the anterior, similar to the pattern described in adults (Barnett et al. 2019; Frank et al. 2019; Vogel et al. 2020). The dorsal frontoparietal network (or the dorsal attention network, as it tends to be active during goal‐directed tasks involving top‐down attention in adults; Corbetta and Shulman 2002) did not show a difference in connectivity along the long‐axis at 0–6 months, but showed greater connectivity with the posterior hippocampus by 1 year of age, effectively becoming more segregated and adult‐like (Frank et al. 2019; Howell et al. 2020; Vogel et al. 2020). Finally, there was a collection of areas that included the ventral medial PFC and superior temporal sulcus, which are regions that are involved in social and schematic processing in adults (Adolphs 2009; Gotts et al. 2012). We found that connectivity of the anterior hippocampus was changing with these regions in an inverted U pattern, producing strongest anterior > posterior difference at 7–12 months.

The fact that cingulo‐opercular and dorsal frontoparietal regions clustered into identifiable cognitive networks (as defined in adults) based on their developing connectivity along the long‐axis of the hippocampus is particularly striking because in adults, these networks are usually defined using independent component analysis or cortical seed‐to‐voxel connectivity, and typically do not include the hippocampus. The fact that the hippocampus does not couple strongly with these networks at rest in adults does not imply that hippocampal interaction with these networks is not important for cognitive functioning. Indeed, the hippocampus flexibly interacts with neocortical regions during mnemonic functioning depending on memory content and task demands (Gilboa and Moscovitch 2021; Sheldon and Levine 2018). It is well‐documented that attention and salience play important roles in shaping hippocampal memories, and further, our memories shape what we attend to (Aly and Turk‐Browne 2017; Lisman and Grace 2005). In infants and toddlers, behavioral work shows that the development of attention and recognition memory are closely associated (Richards et al. 2010; Rose et al. 2003). We speculate that these emerging functional interactions may be important for the formation of long‐lasting memories, and are worth investigating as key contributors to the maturation of long‐term memory capacity. Indeed, long‐term memory undergoes significant maturation during the first 2 years following gestation. By around 9 months of age, the capacity for ordered recall begins to emerge, and both the quantity and reliability of remembered information continue to improve throughout this early developmental period (Bauer 2002). Future work incorporating behavioral memory measures and direct manipulations in rodents could help clarify the link between hippocampal connectivity and memory, especially as it pertains to regions not canonically associated with episodic memory.

Finally, it is worth considering how the developmental trajectory of hippocampal connectivity continues beyond the first 2 years. Past work in older children and adolescents has shown mixed patterns of hippocampal connectivity along the long‐axis with age. For example, one study reported region‐specific shifts in 8–25 year olds, with some cortical areas exhibiting greater anterior connectivity in childhood that transitioned to greater posterior connectivity by adulthood, while other regions showed the opposite pattern (Tang et al. 2020). Another study in 8–21 year olds described a general shift from posterior to anterior hippocampal connectivity with age (Xie et al. 2024). Others have found that in early childhood (4–6 year olds), the hippocampus begins to integrate with memory‐related regions and segregate from regions associated with other functions, consistent with emerging specialization (Riggins et al. 2016). This body of research has generally been interpreted to reflect the development of functional specialization within the hippocampus, including along its long‐axis.

Although it is challenging to make definitive claims about specialization without accompanying behavioral data, to the extent that greater differentiation of anterior–posterior connectivity profiles indexes specialization in the present work, our results may suggest early integration within memory‐related networks (or between memory subnetworks) and a gradual emergence of specialization across networks supporting distinct cognitive functions. It's also possible that connectivity patterns in infancy reflect the foundational architecture required for later specialization of memory subnetworks. Theories of episodic memory in adults posit that the anterior hippocampus supports construction of core memory features, while the posterior hippocampus elaborates episodic details (Audrain et al. 2022; Gilmore, Quach, Kalinowski, González‐Araya, et al. 2021; Robin and Moscovitch 2017). From this perspective, both segments and their interactions must access a broad spectrum of cortical information to support episodic remembering. This need for cross‐axis integration may be particularly pronounced early in development, as cortical regions themselves become increasingly specialized. Finally, it is plausible that specialization along the hippocampal long‐axis follows a nonlinear trajectory, similar to white matter maturation (Paus et al. 2001), involving phases of connectivity expansion and pruning based on experience. Thus, the observed decrease in anterior/posterior differentiation in connectivity patterns with cortical memory regions in infancy across the 2 years examined here may represent an early stage of network formation that precedes later refinement and segregation. Future work spanning infancy through childhood is needed to test this hypothesis.

This work is not without limitations. Examining functional imaging data in infants and toddlers poses many challenges. As data in this experiment were collected at two different sites, we wanted to make sure that site was not driving observed effects. Out of the 212 infants and toddlers in this sample, 189 had site information documented at the time of the scan. None of the predictors in our models significantly interacted with site, even at a liberal threshold (k = 1, p < 0.05 uncorrected). As infants age there are changes in fMRI tSNR due to changes in motion‐related artifact as well as the developing hemodynamic response (Arichi et al. 2012). Indeed, motion, signal variability, as well as mean connectivity across the neocortex was increasing across the first 6 months in our sample (Figures S14–S16). An analysis of motion‐related differences across age bins can be found in Supplementary Methods 2 in Supporting Information. The tSNR and mean motion regressors included in our analyses should control for changes in signal with age. Nonetheless, it's possible that the lack of observed hippocampal‐cortical connectivity changes across later age bins (Figures 1 and 2) were somewhat masked by lower signal quality. Similarly, it's possible that the large increase in hippocampal‐cortical connectivity across the first 6 months is in part driven by increase in BOLD signal engendered by neurodevelopment. Should that be the case, infant hippocampal connectivity with memory‐relevant regions in adults could be established even earlier than we have here described. Regardless, it's clear that by 6 months post‐gestation, a rudimentary topological pattern of hippocampal connectivity with adult‐like memory regions is largely established. The fact that we observed different profiles of cortical connectivity with the long‐axis of the hippocampus across later ages (e.g., some cortical regions showed increases in connectivity with development, others decreases or inverted U shaped patterns; Figure 4) increases our confidence that fine‐tuning of hippocampal‐cortical connectivity with development are generally not driven by linear changes in BOLD signal, motion, respiration, or other artifacts (all of which are indexed by tSNR).

While our approach allowed nonlinear changes in connectivity across age bins to emerge, we did not explore finer nonlinear relationships within each age bin at a voxel‐wise level, which remains a direction for future research. Similarly, while we described relatively gradual changes in connectivity with age, it's possible that step‐wise developmental bursts occur within individuals, which would be uncaptured here after averaging across variable participants. Indeed, behavioral measures of memory indicate appreciable variability in memory ability at any given timepoint within the first 2 years (Bauer 2002). While the current Baby Connectome release has few participants with multiple sessions aligned across ages required to examine the developmental trajectory of connectivity within participants (rather than cross sectionally, as presented here), this possibility may be tested with future releases.

It is also worth noting that participants in this study were asleep during data collection, which may affect network coupling to some extent. Recent work has found that sleeping infant functional networks are more similar to sleeping adult functional networks than those of awake adults, although this finding is somewhat dependent on the network examined (Mitra et al. 2017; Yates et al. 2023). The dataset does not contain information on sleep quality, awakenings, or sleep architecture, all of which could feasibly impact connectivity. Indeed, there is evidence of decreases in rapid eye movement (REM) sleep and increases in non‐REM sleep during the first two postnatal years (Lenehan et al. 2023). Future work could examine how connectivity changes are related to sleep‐related changes.

Finally, only limited demographic and psychosocial information was available for participants, despite these factors playing an important role in early brain development. Although data on race (White/Asian/Black/More than one) and ethnicity (Hispanic/Non‐Hispanic) were provided, the Baby Connectome dataset lacks diversity, with approximately 80% of participants identified as White and ~90% as non‐Hispanic. This limits the generalizability of our findings, underscoring the need for future research initiatives to prioritize greater demographic diversity in participant recruitment.

In summary, we found evidence of some degree of differentiation of anterior–posterior hippocampal connectivity in even the youngest infants examined (0–6 months), alongside large increases in connectivity with cortical regions associated with adult memory function. Nonetheless, hippocampal‐cortical connectivity along the long‐axis continued to fine‐tune over the first 2 years of life. This extended period of development was not limited to classic memory‐related regions; rather, cortical networks canonically associated with attention and salience in adults also showed emerging connectivity along the anteroposterior axis of the hippocampus. Maturation of hippocampal‐cortical coupling may play an important role in the emergence of long‐term memory capacity with development, and future research is needed to clarify how these connectivity changes relate to evolving memory abilities.

Author Contributions

Conceptualization: Sam Audrain and Alex Martin. Methodology: Sam Audrain and Stephen J. Gotts. Formal analysis: Sam Audrain, Shawn C. Milleville, Stephen J. Gotts, Jenna M. Wilson, and Jude Baffoe‐Bonnie. Resources: Alex Martin. Data curation: Shawn C. Milleville. Writing – original draft: Sam Audrain. Writing – review and editing: Sam Audrain, Stephen J. Gotts, Alex Martin, Jenna M. Wilson, and Jude Baffoe‐Bonnie. Visualization: Sam Audrain. Supervision: Alex Martin.

Funding

This work was supported by the National Institutes of Health (ZIA MH‐002930).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Age distribution of infants and toddlers in each age bin.

Figure S2: Assessing pipeline differences in infants < 10 months compared to those > 10 months in age. A different registration procedure was used for participants < 10 months in age than those above 10 months due to poor gray matter/white matter contrast in young infants (see Section 2 for details). To check that the two groups were in good alignment, (A) we present averages of subject connectivity maps just prior to the transition (6–9 months, N = 35) and just after (11–14 months). There is good spatial agreement and no obvious shifts. (B) We calculated the average connectivity map for each aligned subject in the EPI data, which has visible anatomical structure. We then ran a spatial correlation between each subject's mean connectivity image with that of the group average (across the 212 participants). We reasoned that registration changes in the data at 10 months due to changes in scaling or shifting for example, would result in a shift in the correlation data at that point in time. However, the data is smooth from ~6 to 25 months. Note that there is a change in the correlation across the first 6 months where we have already described signal quality is changing (Figures S14–S16). Thus, the EPI data using the different alignment strategies are in good alignment with one another, and we do not expect any impact on the analyses due to this factor. (C) As we used different versions of Freesurfer in younger and older participants to create CSF and WM masks, we examined the overlap between them. Overlap was high, suggesting the different versions are unlikely to appreciably affect the denoising process.

Figure S3: Effects of age on anterior and posterior hippocampal‐cortical connectivity across 7–25 postnatal months. Surface reconstruction of cortical regions showing age‐related changes in anterior (top row) and posterior (bottom row) hippocampal‐cortical connectivity. Warm colors denote greater connectivity with increasing age, and cool colors denote decreasing connectivity with increasing age. Data were surface projected to the adult MNI template using Connectome Workbench software for visualization purposes (Marcus et al. 2011). k = 10, q < 0.01.

Figure S4: Hierarchical clustering quality. Hierarchical clustering was used to cluster neocortical clusters elicited by a significant interaction between hippocampal long‐axis and age bin on connectivity (44 cortical areas). (A) A map of the 44 neocortical clusters identified by the long‐axis by age bin interaction before hierarchical clustering. (B) Six superclusters were identified based on the point in the agglomeration schedule where there was a jump in distance between combined superclusters. (C) Dendrogram of the six‐supercluster solution. (D) Principal components analysis shows the distribution of the 44 neocortical clusters separated into 6 superclusters on a 2D plot. Colors denote the superclusters, large circles represent the center of each supercluster, and small circles with numbers represent the component neocortical clusters. (E) Silhouette statistics for the clustering solution. (F) Consistency of the hierarchical clustering solution across 212 iterations using leave‐one‐out validation. The matrix on the left is unthresholded, and the one on the right is thresholded at 100% consistency. Numbers along the x‐axis of (C) and (E) and within the plot in (D) represent neocortical clusters that comprise each supercluster, the descriptive labels of which can be found in Table S1.

Figure S5: Significant voxel‐wise interaction between hemisphere, hippocampal long‐axis, and age bin on connectivity. (A) Significant voxels surviving the three‐way interaction between hemisphere (right/left), hippocampal long‐axis (anterior/posterior) and age bin (0–6 months/7–12 months/13–18 months/19–25 months) on connectivity at q < 0.01, k = 10. (B) Descriptive information for each identified region. (C) Plots depicting the interaction for each region. Asterisks (*) indicate that the region identified overlapped with the hippocampal masks that were used as seeds in this analysis. Ant: anterior; L: left; mo: months; post: posterior; R: right.

Figure S6: Significant voxel‐wise interaction between hemisphere and hippocampal long‐axis on connectivity. (A) Regions of significant two‐way interaction between hemisphere (left/right) and hippocampal long‐axis (anterior/posterior) on connectivity at q < 0.01, k = 10. (B) Descriptive information for each identified region. (C) Plots depicting the interaction for each region. Asterisks (*) indicate that the region identified overlapped with the hippocampal masks that were used as seeds in this analysis. Ant: anterior; L: left; mo: months; post: posterior; R: right.

Figure S7: Significant voxel‐wise interaction between hemisphere and age bin on connectivity. (A) Regions of significant two‐way interaction between hemisphere (right/left) and age bin (0–6 months/7–12 months/13–18 months/19–25 months) on connectivity at q < 0.01, k = 10. (B) Descriptive information for each identified region. (C) Plots depicting the interaction for each region. Asterisks (*) indicate that the region identified overlapped with the hippocampal masks that were used as seeds in this analysis. Ant: anterior; L: left; mo: months; post: posterior; R: right.

Figure S8: Entorhinal supercluster. (A) Entorhinal supercluster plotted on the surface of the brain. (B) Descriptive information regarding the neocortical clusters included in the supercluster. Coordinates refer to peak activation of each neocortical cluster using the LPI convention in MNI space of the median age template (11–14 months). These neocortical clusters were identified based on the interaction between hippocampal long‐axis and age bin on connectivity at k = 10 voxels, q < 0.01. (C) Average difference in anterior–posterior connectivity across the neocortical clusters included in the supercluster. (D) Difference in anterior–posterior connectivity for each neocortical cluster included in the supercluster. L: left; mo: months.

Figure S9: Parahippocampal supercluster. (A) Parahippocampal supercluster plotted on the surface of the brain. (B) Descriptive information regarding the neocortical clusters included in the supercluster. Coordinates refer to peak activation of each neocortical cluster using the LPI convention in MNI space of the median age template (11–14 months). These neocortical clusters were identified based on the interaction between hippocampal long‐axis and age bin on connectivity at k = 10 voxels, q < 0.01. (C) Average difference in anterior–posterior connectivity across the neocortical clusters included in the supercluster. (D) Difference in anterior–posterior connectivity for each neocortical cluster included in the supercluster. L: left; mo: months.

Figure S10: Medial parietal supercluster. (A) The media parietal supercluster plotted on the surface of the brain. (B) Descriptive information regarding the neocortical clusters included in the supercluster. Coordinates refer to peak activation of each neocortical cluster using the LPI convention in MNI space of the median age template (11–14 months). These neocortical clusters were identified based on the interaction between hippocampal long‐axis and age bin on connectivity at k = 10 voxels, q < 0.01. (C) Average difference in anterior–posterior connectivity across the neocortical clusters included in the supercluster. (D) Difference in anterior–posterior connectivity for each neocortical cluster included in the supercluster. L: left; mo: months.

Figure S11: Cingulo‐opercular supercluster. (A) The cingulo‐opercular supercluster plotted on the surface of the brain. (B) Average difference in anterior–posterior connectivity across the neocortical clusters included in the supercluster. (C) Descriptive information regarding the neocortical clusters included in the supercluster. Coordinates refer to peak activation of the neocortical clusters using the LPI convention in MNI space of the median age template (11–14 months). These regions were identified based on the interaction between hippocampal long‐axis and age bin on connectivity at k = 10 voxels, q < 0.01. (D) Difference in anterior–posterior connectivity for each neocortical cluster included in the supercluster. L: left; mo: months.

Figure S12: Dorsal frontal parietal supercluster. (A) The dorsal frontal parietal supercluster plotted on the surface of the brain. (B) Average difference in anterior–posterior connectivity across the neocortical clusters included in the supercluster. (C) Descriptive information regarding the neocortical clusters included in the supercluster. Coordinates refer to peak activation of the neocortical clusters using the LPI convention in MNI space of the median age template (11–14 months). These regions were identified based on the interaction between hippocampal long‐axis and age bin on connectivity at k = 10 voxels, q < 0.01. (D) Difference in anterior–posterior connectivity for each neocortical cluster included in the supercluster. L: left; mo: months.

Figure S13: vmPFC‐STS supercluster. (A) The vmPFC‐STS supercluster plotted on the surface of the brain. (B) Average difference in anterior–posterior connectivity across the neocortical clusters included in the supercluster. (C) Descriptive information regarding the neocortical clusters included in the supercluster. Coordinates refer to peak activation of the neocortical clusters using the LPI convention in MNI space of the median age template (11–14 months). These regions were identified based on the interaction between hippocampal long‐axis and age bin on connectivity at k = 10 voxels, q < 0.01. (D) Difference in anterior–posterior connectivity for each neocortical cluster included in the supercluster. L: left; mo: months.

Figure S14: Changes in mean connectivity across the neocortex with age. We extracted mean whole‐brain connectivity for each participant across 200 cortical parcels (excluding the hippocampus and subcortical regions) defined with the Schaefer atlas (Schaefer et al. 2018) projected to infant space. As motion can affect connectivity we plot the results with (right) and without (left) controlling for mean motion. Black dots reflect mean connectivity for each participant. The black line is a loess regression trend with 95% confidence intervals. The vertical gray bars denote age bin cutoffs.

Figure S15: Changes in average standard deviation of connectivity across the neocortex with age. We extracted the average standard deviation of whole‐brain connectivity for each participant across 200 cortical parcels (excluding the hippocampus and subcortical regions) defined with the Schaefer atlas (Schaefer et al. 2018) projected to infant space. As motion can affect signal variability we plot the results with (right) and without (left) controlling for mean motion. Black dots reflect average standard deviation of connectivity for each participant. The black line is a loess regression trend with 95% confidence intervals. The vertical gray bars denote age bin cutoffs.

Figure S16: Changes in motion with age. We quantified motion within each participant using AFNI's 1dDiffmag. Black dots reflect diffmag for each participant. The black line is a loess regression trend with 95% confidence intervals. The vertical gray bars denote age bin cutoffs.

Table S1: Peak regions reflecting the voxel‐wise interaction between hippocampal long‐axis and age bin.

Table S2: Pairwise contrasts between anterior and posterior hippocampal connectivity within each age bin and supercluster for weighted connectivity as determined by leave‐one‐out validation.

Table S3: Pairwise contrasts of connectivity between age bins for anterior and posterior hippocampi with each supercluster for weighted connectivity as determined by leave‐one‐out validation.

Data Availability Statement

This work utilizes data collected as part of the UNC/UMN Baby Connectome Project Consortium. Data used in the preparation of this manuscript were obtained from the NIMH Data Archive (NDA). NDA is a collaborative informatics system created by the National Institutes of Health to provide a national resource to support and accelerate research in mental health. Dataset identifier: https://doi.org/10.15154/w07p‐s888. This manuscript reflects the views of the authors and may not reflect the opinions or views of the NIH or of the Submitters submitting original data to NDA. Code used for all statistical analyses can be found here: https://github.com/saudrain/paper_BabyHippos.