Infant attachment does not depend on neonatal amygdala and hippocampal structure and connectivity

Infant attachment is an antecedent of later socioemotional abilities, which can be adversely affected by preterm birth. The structural integrity of amygdalae and hippocampi may subserve attachment in infancy. We aimed to investigate associations between neonatal amygdalae and hippocampi structure and their whole-brain connections and attachment behaviours at nine months of age in a sample of infants enriched for preterm birth. In 133 neonates (median gestational age 32 weeks, range 22.14–42.14), we calculated measures of amygdala and hippocampal structure (volume, fractional anisotropy, mean diffusivity, neurite dispersion index, orientation dispersion index) and structural connectivity, and coded attachment behaviours (distress, fretfulness, attentiveness to caregiver) from responses to the Still-Face Paradigm at nine months. After multiple comparisons correction, there were no significant associations between neonatal amygdala or hippocampal structure and structural connectivity and attachment behaviours: standardised β values − 0.23 to 0.18, adjusted p-values > 0.40. Findings indicate that the neural basis of infant attachment in term and preterm infants is not contingent on the structure or connectivity of the amygdalae and hippocampi in the neonatal period, which implies that it is more widely distributed in early life and or that network specialisation takes place in the months after hospital discharge.


Introduction
In the first year of life, interactions between infants and their caregivers shape attachment behaviours (Ainsworth et al., 1972;Bowlby, 1982).Attachment can be conceived as a behavioural system that coordinates proximity and exploration to promote safety and independent learning.Infant behaviours within the attachment system (i.e.attachment behaviours) include proximity-seeking towards the caregiver and displaying distress, which helps draw the caregiver's attention.An infant's attachment style reflects qualitative differences in the manner in which the infant organises attachment behaviours towards a primary caregiver.Securely-attached infants use caregivers as a secure base to explore from and return to.Under a threat, infants with an insecure attachment style avoid their caregiver (avoidant), fail to respond independently (resistant) or exhibit contradictory behaviours i.e. attachment disorganisation (Ainsworth et al., 1981;Main and Solomon, 1990).
Infant secure attachment positively associates with more effective emotion regulation (Cooke et al., 2019), and higher social competence (Groh et al., 2014) in later life.However, infant socioemotional development can also be impacted by adverse early life experiences.For example, preterm infants have higher likelihood to develop socioemotional difficulties, although there are uncertainties about whether low gestational age per se disrupts attachment (Arpi and Ferrari, 2013;Dean et al., 2021;Johnson and Marlow, 2011;Korja et al., 2012).Since attachment relationships precede socioemotional competence, identifying brain networks that underpin attachment and the factors that influence those networks may help inform interventions to foster socioemotional development.
Early contextual factors, such as aspects of parenting behaviour, only partly explain infant attachment (Madigan et al., 2006;Van Ijzendoorn et al., 1999), so it remains a question whether intrinsic infant characteristics support attachment behaviours that meet the goal of felt security.Individual differences in attachment behaviours may stem from infants' perception and processing of threatening stimuli, parental support, and subsequent emotion regulation (Bernier and Meins, 2008).The limbic system develops earlier and faster than the cerebral cortex (Utsunomiya et al., 1999), and may be crucial for emotional processing in infancy, while more diverse functional specialisation including cognitive control of emotion becomes established as children develop (Casey et al., 2008;Steinberg, 2017).Authors have suggested that greater functional connectivity between the amygdala and the salience network may reflect a greater tendency to perceive stressors as emotionally salient in infants (Hu et al., 2024).This is supported by significant associations between stronger left amygdala-anterior insula functional connectivity in the neonatal period and higher maternal-reported fear expression at six months of age (Graham et al., 2016;Thomas et al., 2019), greater amygdala-cingulate connectivity and lower positive affect in response to novelty at four months of age (Filippi et al., 2021) and preliminary associations between amygdala-salience network functional connectivity at three months and poorer recovery following a socioemotional stressor at six months of age (Hu et al., 2024).The amygdala structure may also play a role in early emotional processing, as suggested by a small study showing positive correlations between total amygdala volumes and escape response to a frightening stimulus at 12 months of age in very preterm infants (Cismaru et al., 2016).The amygdala structure has also been associated with the perinatal stress environment, which could modulate early parent-infant interactions (Atkinson et al., 2000).Maternal hair cortisol concentration, a stable marker of maternal hypothalamic-pituitary-adrenal axis activity during pregnancy, was associated with variations in newborns' amygdala microstructure and connectivity (Stoye et al., 2020).The structure and function of the hippocampus appears to relate to caregiving and attachment during early childhood.Among infants with comparatively larger left hippocampi in the neonatal period, lower maternal sensitivity at six months predicted higher attachment disorganisation scores at 18 months (Rifkin-Graboi et al., 2019).A small study also reported concurrent associations between infant's bilateral hippocampal volumes and maternal sensitivity at six months of age (Rifkin-Graboi et al., 2015), which predicts attachment security (De Wolff and Van Ijzendoorn, 1997).Importantly, the integrity of these nuclei and their networks may also be affected by preterm birth, as evidenced by the differences in hippocampal structure (Ball et al., 2012) and amygdala functional connectivity at rest (Rogers et al., 2017) apparent between term and preterm neonates at term-equivalent age.
Several studies have investigated longitudinal associations between parent-infant attachment relationships and brain structure in childhood (Bernier et al., 2019;Hidalgo et al., 2019;Leblanc et al., 2017), but the extent to which differences in brain structures around the time of birth associate with attachment within the first two years of life (Rifkin-Graboi et al., 2019;Tharner et al., 2011) is relatively understudied.This is important to understand for designing and implementing strategies to promote socioemotional development at the optimal time.MRI volumetric data has been used to capture infants' brain structural changes, but volumetric changes reflect macrostructural physiological or pathological processes, including growth, injury and dysmaturation (Boardman and Counsell, 2020).Microstructural measures of brain nuclei and their connections enable further inference about neurite density and organisation, and may be useful for understanding how the development of brain networks supports attachment.
By combining neonatal multimodal brain MRI and infant behaviours coded from structured interactions with a caregiver, we aimed to investigate associations between neonatal amygdala and hippocampal structure or connectivity and attachment behaviours at nine months of age in a sample of infants enriched for preterm birth.Our study prioritised these metrics over functional connectivity because previous research studying child attachment primarily focused on brain structural differences (Bernier et al., 2019;Hidalgo et al., 2019;Rifkin-Graboi et al., 2019), but less attention has been given to the microstructural organisation of nuclei and related networks.
We hypothesised that structural variation in these brain nuclei and their whole-brain connections at term-equivalent age associates with individual differences in attachment behaviours derived from infants' responses to the Still-Face Paradigm at nine months of age: infants' distress, fretfulness, and attentiveness to the caregiver.

Participants
Participants were recruited to a prospective longitudinal cohort study, the Theirworld Edinburgh Birth Cohort (TEBC) (Boardman et al., 2020).Infants were born between October 2016 to September 2021 at the Royal Infirmary of Edinburgh, UK.Preterm infants transferred to the hospital ex utero for intensive care, and infants with major congenital malformation, chromosomal abnormality, congenital infection, cystic periventricular leukomalacia, haemorrhagic parenchymal infarction, and post-haemorrhagic ventricular dilatation were excluded.Informed written parental consent was obtained.Ethical approval was obtained from the National Research Ethics Service (16/SS/0154), South East Scotland Research Ethics Committee, and NHS Lothian Research and Development (2016/0255).
Collected information included infant sex (male/female), gestational age (GA) at birth, birthweight, self-reported maternal characteristics and the Scottish Index of Multiple Deprivation 2016 (SIMD) rank, generated from postcode information collected via parental questionnaire.SIMD rank (Scottish Government, 2016) is a multidimensional score generated by the Scottish government ranking localities' deprivation according to local income, employment, health, education, geographic access to services, crime and housing.

Data acquisition
MRI scans were performed at term-corrected gestation according to the study protocol (Boardman et al., 2020).Briefly, a Siemens MAG-NETOM Prisma 3 T MRI clinical scanner (Siemens Healthcare, Erlangen, Germany) and a 16-channel phased-array paediatric head coil were used to acquire: a 3D T2-weighted (T2w) sampling scheme with application-optimised contrasts using flip angle evolution structural scan (voxel size = 1 mm isotropic), and axial diffusion MRI (dMRI) data.dMRI was acquired in two separate acquisitions to reduce the time needed to re-acquire any data lost to motion artefact.The first acquisition consisted of 8 baseline volumes (b = 0 s/mm2 [b0]) and 64 volumes (b = 750 s/mm 2 ); the second consisted of 8 b0, 3 volumes with b = 200 s/mm 2 , 6 volumes with b = 500 s/mm 2 and 64 volumes with b = 2500 s/mm 2 .In addition, an acquisition of 3 b0 volumes with an inverse phase encoding direction was performed.All dMRI images were acquired using single-shot spin-echo echo planar imaging with 2-fold simultaneous multislice and 2-fold in-plane parallel imaging acceleration and 2 mm isotropic voxels; all three diffusion acquisitions had the same parameters (TR/TE 3400/78.0ms).Further details can be found elsewhere (Boardman et al., 2020).
Infants were fed, wrapped, and slept naturally.Flexible earplugs and neonatal earmuffs (MiniMuffs, Natus) were used for acoustic protection.Infants were monitored throughout, and scans were supervised by a doctor or nurse trained in neonatal resuscitation.Each acquisition was inspected contemporaneously for motion artefact and repeated if there had been movement while the baby was still sleeping; dMRI acquisitions were repeated if signal loss was seen in three or more volumes.

Data pre-processing
Structural images were reported by a radiologist with experience in neonatal MRI (AJQ).Visual inspection and quality control of raw structural and diffusion images were performed by experienced image analysts (MBC, KV).T2w images were processed using the developing Human Connectome Project (dHCP) minimal processing pipeline for neonatal data to obtain the bias field corrected T2w image, brain masks, tissue segmentation and label parcellation; tissue volumes were then calculated (Makropoulos et al., 2018).
dMRI processing was performed as previously described (Blesa et al., 2021;Stoye et al., 2020;Vaher et al., 2022).For each subject, the two dMRI acquisitions were concatenated and denoised using a Marchenko-Pastur-PCA-based algorithm (Veraart et al., 2016).The eddy current, head movement and EPI geometric distortions were corrected using outlier replacement and slice-to-volume registration (Andersson et al., 2017(Andersson et al., , 2016(Andersson et al., , 2003;;Andersson and Sotiropoulos, 2016;Smith et al., 2004).Bias field inhomogeneity correction was performed by calculating the bias field of the mean B0 volume and applying the correction to all the volumes (Tustison et al., 2010).Neurite orientation dispersion and density imaging (NODDI) and diffusion tensor imaging (DTI) maps were calculated in the dMRI processed images to obtain fractional anisotropy (FA), mean diffusivity (MD), neurite density index (NDI), and orientation dispersion index (ODI).The DTI model was fitted in each voxel using the weighted least-squares method DTIFIT as implemented in FSL using only the b = 750 s/mm 2 shell.NODDI metrics of amygdalae and hippocampi were calculated using all shells and the recommended values for neonatal grey matter of the parallel intrinsic diffusivity (1.25 µm 2 /ms) using the original NODDI MATLAB toolbox.NODDI metrics of amygdalae and hippocampi whole-brain connections were calculated following the same approach, using the recommended values for neonatal white matter of the parallel intrinsic diffusivity (1.45 µm 2 /ms) (Guerrero et al., 2019;Zhang et al., 2012).Registration between dMRI and T2w was calculated using boundary-based registration (Greve and Fischl, 2009).
Nonlinear diffeomorphic multimodal registration with T2w and tissue probability maps using ANTs symmetric normalisation (Kamino et al., 2018;Avants et al., 2008) was then performed between the subject and the dHCP extended volumetric atlas (Fitzgibbon et al., 2020).The transformation was combined with the diffusion-to-structural transformation to obtain the final diffusion-to-template alignment.

Structural connectivity of amygdalae and hippocampi
Seed-based tractography from source regions of interest (ROIs, left and right hippocampi, and left and right amygdalae) to the whole brain was used to generate a whole-brain connectivity mask for each unilateral nucleus across subjects.The reason to take this approach in lieu of anatomically-constraint tractography was twofold.First, most relevant networks for attachment in human infancy are uncertain.Second, neonatal atlases do not currently include relevant regions with wellknown connections to the hippocampus and amygdala, such as the hypothalamus (Nieuwenhuys et al., 2007), likely due to the resolution required to parcellate these areas in the neonatal brain.Seed-based tractography from source ROIs to the whole brain allowed a more unbiased inclusion of tracts between source ROIs and highly-connected areas of interest that may not be captured by current neonatal atlases, while reducing the number of multiple comparisons.
Seed-based tractography from source ROIs to the whole brain was performed using constrained spherical deconvolution (Smith et al., 2012;Tournier et al., 2019).The required 5-tissue type file, was generated by combining the tissue probability maps obtained from the dHCP pipeline with the subcortical structures derived from the parcellation process (Blesa et al., 2021).Multi-tissue response function was calculated, with a FA threshold of 0.10 (Dhollander et al., 2019(Dhollander et al., , 2016)).The average response functions were calculated.Then, the multi-tissue fibre orientation distribution (FOD) for white matter and cerebrospinal fluid were calculated (Jeurissen et al., 2014), and global intensity normalisation on the FODs images was performed.Finally, a tractogram was created, generating 500,000 streamlines from the seed ROI, setting the backtrack option while the rest of the parameters were used as by default.
To find a balance between filtering out noisy streamlines and keeping biologically-relevant connections from unilateral nuclei, purifibre (Aydogan, 2022) was applied to each participant to retain 50 % of the fibres (Fig. S1).Then these tracts were binarised, and the binarised tracts were propagated to the common template using the previously computed registration.Masks across subjects were averaged, and the final image was generated by keeping only voxels common to 50 % of the subjects and propagating it back to each subject.This was the final mask of the whole-brain structural connectome used to calculate the mean values of the different metrics.

Summary of MRI metrics
In all, MRI metrics included volume, FA, MD, NDI and ODI of left and right amygdalae, and left and right hippocampi; FA, MD, NDI and ODI of whole-brain connections of left and right amygdalae, and left and right hippocampi.

Infant behavioural data
Caregivers were invited to attend follow-up appointments with their infants at nine months of age (corrected age for preterm infants) at a University of Edinburgh site at the Royal Edinburgh Hospital.There were several reasons to assess infant behaviour at nine months of age.The study data collection timepoints agree with the National Institute for Health and Care Excellence (NICE) guidelines for developmental follow-up of preterm infants, which warrants monitoring development between three and five, and by 12 months of corrected age (Kallioinen et al., 2017).Moreover, infants reach the phase of "Clear-Cut Attachment" at 8-12 months of age and start displaying behaviours that are organised on a goal-corrected basis (Ainsworth et al., 2015).Furthermore, infants undergo a major shift in their understanding of others' intentionality at 8-10 months of age, period often referred to as the "nine-month revolution" (Tomasello, 2009).Researchers have suggested changing social cognitive capacities during the "nine-month revolution" may support, explain, or coincide with attachment behaviours and cognitive models of infants' attachment figures in the first year of life (Sherman et al., 2015).
The visit lasted for approximately 2.5 h and comprised an extensive assessment battery including questionnaires and behavioural procedures.The Still-Face Paradigm (SFP) always took place close to the end of the appointment to avoid upsetting the infant near the beginning of the session.Further details can be found elsewhere (Boardman et al., 2020).

The Still-Face Paradigm
After consent for video-recording was obtained, infants were secured in a high chair.Caregivers sat facing the infant approximately 50 cm away.Two Panasonic HC-W580 video cameras were set up behind the caregiver and the infant to record the infant and caregiver's faces and hands.The researcher was seated out of view, and verbally cued caregivers.The procedure included five two-minute episodes (Haley and Stansbury, 2003); modified from the original protocol (Tronick et al., 1978): baseline, still-face, reunion, still-face, and reunion.During the baseline episode, caregivers were instructed to interact naturally with the infant, without using toys.During still-face episodes, caregivers were instructed to express a neutral facial expression, remaining still and looking slightly above the infant's head, avoiding eye contact and interaction with their infant.During reunion episodes, caregivers were instructed to interact normally with the infant again, without toys.Caregivers were always given the option to terminate the paradigm if infants exhibited severe distress.

Data coding
Williams and Turner's coding scheme (Williams and Turner, 2020) was selected to study attachment from the SFP at nine months of age.Williams and Turner's coding scheme includes the Happy-Distressed (HD), Not fretful-Fretful (NFF) and Attentive-Avoidant (AA) Global Rating Scales (Murray et al., 1996) to analyse infant behaviours during the reunion episode (Abbott, 2016;Williams and Turner, 2020).The coding scheme was applied as per the coding manual.To study attachment behaviours, the HD, NFF and AA raw scores were calculated per infant.These scores represented percentages of time infants engaged in HD, NFF or AA behaviours during the reunion episode; ranging from 0 (longer time displaying distress, fretful or avoidant behaviours according to the scale) to 1 (longer time displaying happy, not fretful or attentive behaviours to the caregiver according to the scale).Appendix S1 elaborates on the coding scheme, including the attachment style categorisation system that was not used in this work.
Importantly, scores derived from infant responses to the SFP mirror traditional attachment behaviours described by Bowlby (Bowlby, 1982) and Ainsworth (Abbott, 2016;Ainsworth et al., 2015) in the Strange Situation.Specifically, avoidance of proximity and contact in the Strange Situation is comparable to the AA scale, although the latter only involves avoidance of eye contact (other avoidant behaviours cannot be considered since infants' behavioural responses are constrained by sitting in a highchair in the SFP).Amount of crying in the Strange Situation is captured by the HD scale, which details the proportion of time spent in a distressed state during each episode.This is a dual ended scale and also includes markers of happiness or positive affect.Resistant behaviours in the Strange Situation are captured by the NFF scale, which considers fretful behaviours infants display when resisting or protesting such as angry vocalising, screaming, or arching the back away from the caregiver.Contact maintaining in the SFP is indirectly captured by the AA scale, along with positive affect.In other words, the infant is able to maintain contact with the caregiver by being attentive, engaged and positive towards the caregiver.Proximity seeking in the Strange Situation has no direct parallels in codes from the SFP since infants are unable to move towards or from the caregiver in the SFP (Abbott, 2016).
The SFP in the TEBC was designed to explore a broad range of infant behaviours, including infant stress responses which are more pronounced following repeated still-face episodes (Provenzi et al., 2016).Thus, a double-still face design was selected.In this study, only the first reunion episode of the five-episode SFP was coded because this provided a better equivalent to previous studies, where infants only experienced one stress episode (the still-face episode) before the reunion (Abbott, 2016;Williams and Turner, 2020).Participants were excluded based on SFP procedural violations initiated by others than the caregiver-infant dyad, and coded reunion time lower than 30 s (see Appendix S2 for further details).Videocoding was conducted using EUDICO Linguistic Annotator software (Wittenburg et al., 2006).
Reliability analyses were carried out for a larger sample of the study cohort (n = 167 infants, not all of whom contributed MRI data).Two types of reliability scores were calculated.For attachment dimensions (raw scores on the HD, NFF and AA scales), the intraclass correlation coefficient (ICC) was calculated.The measurement from a single rater was planned to be the basis of statistical analyses, so the ICC single-rater type was used (Koo and Li, 2016).Following Williams and Turner's coding scheme, two coders would need to agree on the absolute values of the HD, NFF, and AA raw scores to assign the same attachment style to a subject (Williams and Turner, 2020).Thus, a two-way random-effects agreement ICC was calculated for each scale.For attachment categories (attachment styles), the percentage of agreement (tolerance = 0) and Cohen's Kappa were calculated.For inter-rater reliability analyses, the second coder coded a randomly-chosen 10 % of all available videos, raw scores and attachment styles were extracted and compared.For intra-rater reliability analyses, the main coder coded a randomly-chosen 10 % of all available videos again, raw scores and attachment styles were extracted and compared with the first time videos were coded.

Statistical analyses
Data analyses and plots were generated using R, version 4.3.2(R Core Team, 2023).Relevant R packages are listed in Appendix S3.
To study associations between infant neuroanatomy and attachment behaviours, general linear regression was used with each scaled attachment behaviour (z-transformed HD, NFF, and AA raw scores) as outcome, each scaled adjusted brain structural measure as predictor, and covariates as additive predictors.The direction of these associations was not predicted as no previous study explored these nuclei microstructural characteristics in relation to infant attachment.To control for brain size and growth, derived MRI measures were adjusted for GA at scan prior to regression analyses by fitting a linear model of each ztransformed brain structural measure on GA at scan and retaining the residuals.Brain structural measures were not adjusted for total brain tissue volume because this variable showed high collinearity with age at scan (r Pearson = 0.713), and we expected age at scan was better suited to adjust dMRI metrics than total brain tissue volume.
Possible confounding variables were explored from screened and/or included covariates in studies investigating bidirectional associations between neuroanatomy and observational measures of attachment in infancy or childhood (Hidalgo et al., 2019;Leblanc et al., 2017;Rifkin-Graboi et al., 2015;Tharner et al., 2011).To be controlled for, identified variables that may modify the relationship between infant neuroanatomy and attachment behaviours had to meet either the definition of a confounder or a confound blocker (Wysocki et al., 2022).As confounders, we considered SIMD rank, infant sex, infant GA at birth, maternal age, maternal education, maternal antenatal smoking, maternal antenatal alcohol consumption, maternal antenatal depression, and maternal antenatal anxiety (Appendix S4).Infant birthweight also met the definition of confounder, but we did not include it as a potential covariate because of high collinearity with GA at birth (rho-Spearman = 0.862).To investigate correlations between attachment behaviours or brain measures and continuous variables, we used Pearson or Spearman correlation coefficients as a nonparametric alternative.To compare attachment behaviours or brain measures between groups defined by categorical nominal variables (e.g., male and female infants), we used two-sample t-tests or Mann-Whitney-Wilcoxon test as a nonparametric alternative.Tests to identify covariates were exploratory, so reported p-values were not corrected for multiple comparisons (Armstrong, 2014).Variables that were significantly associated with at least one attachment behaviour (HD, NFF or AA raw score) and at least one brain measure at p-value < 0.05 were retained and controlled for.In analyses of amygdalae and hippocampi structure, we included maternal antenatal anxiety as a covariate (Table S1).In analyses of amygdalae and hippocampi structural connectivity, no covariates were included (Table S2).All reported β coefficients were standardised.
For regression analyses, the linearity of the association, homogeneity of variance, and normality of the model residuals were checked by inspecting the model diagnostic plots.Multicollinearity of predictor variables was calculated with the variance inflation factor (O'brien, 2007).Most regression models met the model assumptions.NFF raw score data was skewed, and regression models for the NFF raw scores showed issues with linearity of the association, normality of residuals, and homogeneity of variance.A logit transformation of NFF raw scores improved conformation to the assumptions of linearity of associations, and normality of model residuals as indicated by the diagnostic plots; homogeneity of variance was then confirmed by Goldfeld-Quandt tests.Results from analyses using NFF logit-transformed data are reported in the main text, results from analyses using untransformed data are available in the Supplementary materials.
All reported p-values were adjusted for false-discovery rate using the Benjamini-Hochberg (BH) procedure within each family of experiments.Families of experiments were identified based on several assumptions.First, only the interaction of the left hippocampal volume and maternal sensitivity were found to be associated with infant attachment in a previous study (Rifkin-Graboi et al., 2015), so we expected lateralisation in the associations of amygdalae and hippocampi microstructure with infant attachment.Second, brain measures of nuclei structure mostly captured grey matter characteristics, whereas brain measures of whole-brain connections of nuclei captured white matter characteristics.Considering these different tissue configurations, grey matter and white matter metrics were expected to have different interpretations (Boardman and Counsell, 2020).Third, regression models for nuclei structure versus structural connectivity included different covariates.Fourth, all attachment behaviours (HD, NFF and AA raw scores) were derived from a single assessment (the SFP).Thus, eight families of experiments were identified: one for each unilateral nucleus' structure, and for each unilateral nucleus' whole-brain structural connectivity.

Sensitivity analysis
Although this variable did not show confounding effects, we explored whether results applied to the whole range of GA at birth by adjusting MRI measures for GA at scan and GA at birth prior to regression analyses.A linear model of each z-transformed brain structural measure was fitted to GA at scan and GA at birth and residuals were retained.General linear regression models were performed with each scaled attachment behaviour as outcome (including logit-transformed NFF raw scores), each scaled adjusted brain structural measure as predictor, and covariates as additive predictors (maternal antenatal anxiety for analyses of amygdalae and hippocampi structure, none for analyses of amygdalae and hippocampi structural connectivity).

Post-hoc power analyses
The included sample provided 90 % power to detect a small effect size of 0.11 in analyses of amygdalae and hippocampi structure, and 90 % power to detect a very small effect size of 0.08 in analyses of amygdalae and hippocampi structural connectivity (R packages are listed in Appendix S3).

Baseline characteristics
One hundred and thirty-three infants had available neonatal MRI and SFP data (119 mother-infant dyads; 13 father-infant dyads, one grandmother-infant dyad, Fig. S2).Median (range) GA at birth was 32 +0 (22 +1 -42 +1 ), birthweight was 2000 (370-4560) g, and SIMD rank was 4972 (137-6967).Demographic characteristics of the group are displayed in Table 1.Compared to infants who had neonatal MRI but did not attend a 9-month appointment and/or did not contribute SFP data (N = 159), the study group (N = 133) was slightly less premature (GA at birth median difference 10 days), had mothers of higher age (mean difference 1.3 years) and education, and came from less deprived areas on average (Table S3).

Segmentation and whole-brain connectivity masks of amygdalae and hippocampi
Amygdalae and hippocampi segmentations from a participant at term equivalent age viewed in native space are shown in Fig. 1A.The whole-brain connectivity mask of amygdalae mostly captured tracts between amygdala and the septal nuclei, lentiform nucleus, thalamus, hypothalamus, hippocampus, temporal and occipital lobe (Fig. 1B).The whole-brain connectivity mask for the left hippocampus mostly captured tracts between hippocampus and the lentiform nucleus, hypothalamus, mesencephalon, insula and temporal lobe (Fig. 1C).

Amygdalae and hippocampi macrostructure at birth and attachment behaviours at nine months of age
Correlations of attachment behaviours are reported in Fig. S3.There were no significant associations between volumes of amygdalae and hippocampi and attachment behaviours (Fig. 2A, Table 2, S4).

Amygdalae and hippocampi microstructure at birth and attachment behaviours at nine months of age
There were nominally statistically significant associations between

Amygdalae and hippocampi structural connectivity at birth and attachment behaviours at nine months of age
There were nominally statistically significant associations between left amygdala whole-brain connectivity NDI and attentiveness to the caregiver, which did not survive correction for multiple comparisons (β Fig. 1.Segmentations and connections of interest in 50 % filtered tracts from left amygdala and hippocampus.(A) Amygdalae (red) and hippocampi (blue) segmentations from a participant at term-equivalent age, viewed in native space (left hemisphere is shown in sagittal/lateral sections).(B, C) Tractograms after filtering 50 % of streamlines from a participant at term-equivalent age, viewed in native space.For visualisation purposes, afferent and efferent regions from each unilateral nucleus were inferred from previous anatomical descriptions of amygdalae and hippocampi structural connectivity (Nieuwenhuys et al., 2007).  2

, S4).
There were no other statistically significant associations between MRI metrics capturing the microstructure of amygdala or hippocampal whole-brain connections and attachment behaviours (Fig. 2B, Table 2, S4).

Sensitivity analysis
Results were consistent when controlling for GA at birth.Nominally significant associations between left amygdala MD and attentiveness to the caregiver did not survive correction for multiple comparisons (β (SE) = − 0.20 (0.10), raw p-value = 0.044, p-value adjusted = 0.400).Nominally significant associations between left amygdala whole-brain connectivity NDI and attentiveness to the caregiver did not survive correction for multiple comparisons (β (SE) = − 0.23 (0.11), raw p-value = 0.038, p-value adjusted = 0.461).There were no other significant associations between MRI metrics and attachment behaviours (Table S5).

Reliability analyses for infant behavioural data
For inter-rater reliability analyses, ICC was 0.92 for the HD raw scores, 0.86 for the NFF raw scores and 0.81 for the AA raw scores.The percentage of agreement was 76.50 % and Cohen's kappa was 0.65 for attachment styles.For intra-rater reliability analyses, ICC was 0.82 for the HD raw scores, 0.85 for the NFF raw scores and 0.96 for the AA raw scores.The percentage of agreement was 82.40 % and Cohen's kappa was 0.69 for attachment styles.

Discussion
This study did not find statistically significant associations between the structure and structural connectivity of amygdalae and hippocampi around birth and attachment behaviours at nine months of age in a sample of infants born between 22 and 42 weeks of gestation and their primary caregivers.Therefore, the data do not support the hypothesis that structural variation in these brain nuclei and their whole-brain connections at term-equivalent age associates with individual differences in attachment behaviours derived from infants' responses to the Still-Face Paradigm at nine months of age.Importantly, attachment behaviours included in this work contribute to, but are different from, the status of the infant-caregiver attachment relationship.It is possible that the structure of amygdalae and hippocampi is not a strong determinant of infant attachment in isolation, but may instead interact with environmental factors, such as maternal sensitivity, to predict attachment behaviour.This is supported by the reported significant interaction between neonatal left hippocampal volume and maternal sensitivity in predicting levels of infant disorganised behaviour, albeit in a smaller non-clinical sample (Rifkin-Graboi et al., 2019) compared with this study.
The results agree with the reported lack of associations between volumes of left and right amygdalae, or left and right hippocampi within  the first two weeks of life and infant attachment at 18 months of age (Rifkin-Graboi et al., 2019).Here, we show that the finding applies to a sample of term and preterm infants, with increased likelihood of atypical brain development and socioemotional outcomes.Furthermore, we assessed microstructure and connectivity, in addition to macrostructure, evaluated attachment at a critical period of infant development, were powered to detect even small effects, and used a statistical approach that minimised the risk of error.
To our knowledge, no study has explored associations between structural variation of amygdala or hippocampal networks and attachment in infancy or childhood.Greater neonatal functional connectivity between amygdala and anterior insula connectivity predicted higher levels of maternal-reported fear at six months of age and less increase in maternal-reported fear expression over the first two years of life (Graham et al., 2016;Thomas et al., 2019).Infant fear and emotion regulation play an essential role in attachment, as the presence of caregivers helps reduce fear and distress in the secure caregiver-infant relationship (Ainsworth et al., 1972;Hofer, 1994).However, maternal sensitivity did not modify the reported associations (Thomas et al., 2019) so amygdala connectivity may specifically associate with fear expression as an aspect of temperament.Thus, the absence of significant findings between amygdala connectivity and attachment behaviours in this study may align with the predominantly nil associations between temperament and attachment (Groh et al., 2017).Moreover, preliminary analyses showed maternal sensitivity positively associated with left hippocampal functional connectivity to regions important to social communication (left fusiform, left superior temporal cortex), and negatively associated with left hippocampal functional connectivity to regions important to memory (left entorhinal cortex) at six months of age (Rifkin-Graboi et al., 2015).It is possible that functional, rather than structural, characteristics of hippocampal networks play a role in the caregiver-infant relationship.A recent study indeed demonstrates that coupling between structural and functional connectivity is weaker in higher-level cognitive regions such as temporal areas compared to sensory regions (Preti and Van De Ville, 2019).
Finally, other brain nuclei or networks may more closely associate with infant attachment behaviours.For instance, the role of general cognitive ability in attachment has recently been highlighted (Del Giudice and Haltigan, 2023).It is thus possible that neural correlates of general cognitive ability, such as parieto-frontal networks (Basten et al., 2015;Hu et al., 2022;Langeslag et al., 2013), are more relevant for the development of attachment relationships than amygdala or hippocampal networks.
The strengths of this study include the use of a data-driven approach to study whole-brain connectivity that captured connections of the amygdala and hippocampus with structures in the midbrain, diencephalon and basal telencephalic structures; not frequently parcellated in current neonatal MRI atlases.We considered NODDI parameters to assess microstructure because ODI and NDI are functionally tractable (Kamiya et al., 2020).The potential role of the amygdala and hippocampus was assessed at term-equivalent age so possible impacts of caregiving after discharge from hospital on these structures were not expected.Based on prior literature, we explored a broad range of variables that may confound the association between neonatal brain measures and infant attachment and controlled for possible confounders.The relationships were described in a sample of infants enriched for preterm birth, and appear to apply across the whole GA range (Table S5).This suggests that the structural integrity of amygdalae and hippocampi and their whole-brain networks does not associate with later attachment behaviours in preterm and term infants.This work also has limitations.The SFP was not attempted or was stopped due to infant fussiness in 13 cases, and these infants were more frequently preterm in our sample (n = 10, Fig. S2).Moreover, we did not use the Strange Situation (Ainsworth et al., 2015), the most commonly used assessment of infant attachment.In addition, we did not consider caregiving behaviours such as maternal sensitivity, which predicts attachment security (De Wolff and Van Ijzendoorn, 1997) and could interact with brain neuroanatomy in predicting infant attachment (Rifkin-Graboi et al., 2019).
There are several research implications of this work.First, the structural neural substrates of attachment may include networks remote from the amygdala and hippocampi, and/or that these develop months after birth.These questions could be tested in future studies using longitudinal imaging throughout infancy, leveraging strategies to maximise participant retention (Teixeira et al., 2021).Second, if intrinsic infant factors around birth, such as neuroanatomy, play a negligible role in supporting attachment behaviours, interventions could focus on contextual factors of the infant postnatal environment, such as aspects of parenting behaviour, to cultivate infant attachment behaviours that positively associate with security and socioemotional development.Caregiving practices and infant social networks can vary significantly across cultures (Schmidt et al., 2021), so future studies could explore whether results of this work are generalisable to other populations, as well as other samples.

Conclusions
Structural variation in the amygdalae and hippocampi and their whole-brain connections at term-equivalent age do not contribute substantially to individual differences in attachment behaviours at nine months of age.
All maternal characteristics were self-reported measures.For missing data, percentages do not add up to 100 %.AA = Attentive-Avoidant scale; HD = Happy-Distressed scale; HNC = Higher National Certificate; HND = Higher National Diploma; GCSE: General Certificate of Secondary Education; NC = National Certificate, NFF = Not fretful-Fretful scale; SD = Standard Deviation; SIMD = Scottish Index of Multiple Deprivation 2016.a Missing data: ethnicity data was not available for n = 1 participants.b Missing data: maternal education data was not available for n = 3 participants.c Missing data: maternal smoking during pregnancy data was not available for n = 3 participants.d Missing data: maternal alcohol consumption during pregnancy data was not available for n = 2 participants.left amygdala MD and attentiveness to the caregiver, which did not survive correction for multiple comparisons (β (SE) = − 0.20 (0.10), raw p-value = 0.042, p-value adjusted = 0.406; Fig. 2A, Table 2

Fig. 2 .
Fig. 2. Associations between amygdalae and hippocampi structure and structural connectivity at birth and attachment behaviours in infancy.NFF raw scores were logit-transformed.Colour indicates effect size (standardised β coefficient of regression models).Only corrected p-values are reported, see Table 2 for further details.AA = Attentive-Avoidant scale; FA = fractional anisotropy; HD = Happy-Distressed scale; MD = mean diffusivity; NDI = neurite density index; NFF = Not fretful-Fretful scale; ODI = orientation dispersion index.

Table 1
Characteristics of the study sample.

Table 2
Associations between amygdalae and hippocampi structure and structural connectivity, and attachment behaviours.