Determinants of Maternal Breast Milk Cortisol Increase: Examining Dispositional and Situational Factors

Background. Breast milk is a rich nutritional source, containing numerous proteins, carbohydrates, and hormones that impact long-term offspring development. Strikingly, predictors and correlates of breast milk composition remain largely unknown. Building on a previously discovered increase in breast milk cortisol concentration from 2-12 weeks postpartum, we investigated potential predictors of maternal breast milk cortisol in the first three months post-delivery by examining a suite of maternal dispositional (e.g., attachment, adverse childhood experiences or ACEs) and situational factors (e.g., partner support, self-efficacy). Methods. Data from 73 mothers were collected prenatally, at birth, and 2-, 6-and 12 weeks postpartum. The analyses, which sought to predict postnatal changes in breast milk cortisol, included a pool of theoretically-sound constructs (Table 1) and an exploratory data-driven approach. We fit models differing in complexity as preregistered: 1) Random Forest models, capable of modeling interactions and non-linear relationships, and 2) Bayesian linear models, allowing to model change over time while less prone to overfitting. Results. Overall, we found that no single variable had strong predictive value beyond the known predictors of cortisol, such as time since awakening and time of collection. However, results from both models suggest that ACEs carry information that warrants future investigations, pointing towards a negative relationship with cortisol concentration in breast milk, albeit with a minimal effect size. Conclusion. Using sophisticated models, we found that early life stress may play a role in physiological stress markers in breast milk in the first three months postpartum, with potential implications for offspring development.


Introduction
Breastfeeding is widely recognised to carry positive effects both for the infant and the mother (Labbok, 2001;Victora et al., 2016) and to hold significant societal and public health implications (Silfverdal, 2023), such as reducing early infant mortality and morbidity (WHO, 2013). Next to being a rich source of nutrition, breast milk contains numerous components such as proteins, carbohydrates, hormones, growth factors, and other bioactive molecules (Yi & Kim, 2021) that play a fundamental role in long-term cognitive and behavioural developmental outcomes (Hechler et al., 2018;Horta & Victora, 2013;Jain, 2022). The process by which milk components in a mother's milk can influence the development of her offspring is known as lactocrine programming (de Weerth et al., 2022). Notably, despite broad consensus on the importance of human milk for infants, not only as an essential nutrient but also for long-term cognitive developmental outcomes (Kumari et al., 2023), surprisingly little attention has been devoted to mapping the determinants J o u r n a l P r e -p r o o f of human milk composition and variations postpartum, a period of profound psychobiological changes in the mother (see Aparicio et al., 2020) and the infant. One such change that has been documented, yet left largely unexplored, is a significant increase in an essential hormone, namely cortisol concentrations in the breast milk in the first three months postpartum (Hechler et al., 2018). Accordingly, unveiling the predictors of this phenomenon is our empirical quest.
Cortisol, one of the most critical glucocorticoid hormones, is produced in the adrenal glands mostly in response to stress and is vital in lactation, as it stimulates milk production (Svennersten-Sjaunja & Olsson, 2005). Already during pregnancy, cortisol produced by the mother crosses the placenta and helps to regulate the development of the fetal HPA (hypothalamic-pituitary-adrenal) axis or stress system (Meyer et al., 2021); after birth, cortisol from the mother continues to be essential for infants as it helps to regulate the infant's cortisol levels and promote healthy growth and development (Hinde, 2013). Indeed, in infants, cortisol regulates various physiological processes and brain development (Russell & Lightman, 2019).
However, variations in maternal cortisol levels may impact the offspring's development. Indeed, there is some evidence that heightened maternal breast milk cortisol is linked to offspring fear behaviour in humans (e.g., temperament; Glynn et al., 2021;Grey et al., 2013;Nolvi et al., 2018) and primates (Hinde et al., 2015). Instead, others reported associations between higher maternal breast milk cortisol and better infant neurobehavioral functioning in the first two weeks of life cross-sectionally (Hart et al., 2004) and lower child BMI percentile throughout the first two years of age, respectively, longitudinally (Hahn-Holbrook et al., 2016). Altogether, this set of studies underscores the link between early infant exposure to breast milk cortisol and developmental outcomes, warranting attention to maternal breast milk cortisol concentrations in the early postpartum weeks. Thus, the finding showing increased breast milk cortisol concentrations in the first 12 weeks of an infant's life in a healthy community sample (Hechler et al., 2018) deserves empirical attention.
While increasing challenges in juggling new motherhood with other (pre-existing) tasks over the first months, such as maternal psychosocial stress, partly explain this increase in cortisol (Aparicio et al., 2020), some women may be more vulnerable to heightened cortisol concentrations throughout the first three months post-delivery than others. To identify what contributes to higher risk, maternal prenatal dispositional and postnatal situational factors need to be examined. Maternal prenatal factors (e.g., personal life history, history of caregiving, or dispositional traits) are recognised to have an enduring impact on maternal psychophysical well-being, particularly during the first few postpartum months, which is a vulnerable period for new mothers (Grekin & O'Hara, 2014). In particular, maternal exposure to adverse childhood experiences (ACEs; Mersky & Janczewski, 2018;including childhood trauma, Choi et al., 2017) and poor quality of parent-child attachment relationships (Flykt et al., 2010) both pose significant risks for maternal psychophysical health and the quality of subsequent mother-infant interaction (Brindle et al., J o u r n a l P r e -p r o o f 2022; Costa-Martins et al., 2016). However, other protective factors may mitigate the adverse effects of poor maternal history on stress physiology, such as resiliency, which is successfully coping with adversity and crises (Irwin et al., 2016), or empathy that may facilitate the mother's connection with the baby postdelivery and alleviate postnatal stress (Bak et al., 2021). The postpartum experience is influenced by various situational factors that emerge after delivery, including postpartum psychological symptoms (Ramakrishna et al., 2019), daily hassles related to work overload, financial difficulties or arguments with the partner (Page & Wilhelm, 2007), help with caregiving tasks, the initiation of childcare, separation from the baby, own caregiving sensitivity, and partner sensitivity towards the baby's needs, infant crying (i.e., colic; Howell et al., 2009), and feelings of parenting self-efficacy (Leahy-Warren & McCarthy, 2011).
Understanding factors that potentially influence breast milk cortisol concentrations can inform prevention programs designed to decrease maternal postnatal stress in the future. We sought to determine the potential determinants of maternal breast milk cortisol levels and increase in the first three months postpartum, exploring the effects of dispositional and situational factors. In a data-driven approach, we leveraged data from a healthy prospective sample, named the BINGO study (Dutch acronym for Biological Influences on Baby's Health and Development), investigating early predictors of child development from pregnancy until early childhood, in which the increase in cortisol concentrations across three time-points in the first 12 weeks postpartum was documented (Aparicio et al., 2020;Hechler et al., 2018). Due to limited knowledge regarding factors that influence cortisol trajectories in milk, we took an exploratory approach: no hypotheses were formulated, and two complementary algorithms were employed to identify relevant predictors of breast milk cortisol. We employed Random Forest and Bayesian models to test linear and nonlinear effects and interaction effects of the proposed array of convenience variables on absolute cortisol concentrations at age 2-, 6-and 12 weeks and the increase from 2 to 12 weeks of infant age.

Participants
This project is part of the longitudinal BINGO study (Dutch acronym for Biological Influences on Baby's Health and Development), investigating early predictors of child development from pregnancy until early childhood in mothers and their partners (Hechler et al., 2018). The study was approved by the Ethics Committee of the Faculty of Social Sciences, Radboud University, Nijmegen, The Netherlands (ECSW2014-1003-189). Initial exclusion criteria were an unhealthy, complicated pregnancy, insufficient mastery of the Dutch language, excessive alcohol use (i.e., alcohol dependency), and drug use. Participants signed up via the project's website or folders handed out in midwife practices, pregnancy courses, and baby stores in the region Nijmegen-Arnhem (Netherlands). Women who were interested in potential participation J o u r n a l P r e -p r o o f received additional study information by mail. Ninety-eight expectant women signed up for the study by providing written informed consent, and 88 mothers (M age = 31.62 in years; SD = 3.79) and 57 partners (M age = 32.93 in years; SD = 4.16) were eligible for participation. The exclusion criteria after birth were: gestational age at birth <37 weeks (gestational week 35, n = 1), 5-minute Apgar score <7, birth weight <2,500 g, congenital malformations (brain damage at birth, n = 1), complications during pregnancy after initial contact, maternal antibiotic use, no milk samples (i.e., not collected any milk sample, n = 3), and maternal illness, such as fever during the preceding week, or chronic diseases and pregnancy-related illnesses that required second line obstetric care. Five mothers stopped their participation in the study after birth. Four of the remaining 78 mothers did not start breastfeeding. One only had one milk sample with an unlikely high and meaningless cortisol concentration value, resulting in a final sample of 73 women (see Figure 1 for a flowchart of the sample).

Procedure
Mothers first visited the laboratory during the third trimester of pregnancy, M = 34.54 weeks, SD = 6.02 weeks, during the late afternoon (after 15:00) or in the early evening (M = 17:28, SD = 01:53). After giving consent, participants filled in several questionnaires concerning their demographic and psychosocial information, and then performed other lab-based tasks as a part of the larger project that are not relevant for the current study.
Postnatally, mothers were asked, and reminded the day before by email, to collect milk samples on the day after the infant reached the age of 2 weeks (M age = 14.63 days; SD = 1.60), 6 weeks (M age = 43.46 days; SD = 4.79), and 12 weeks (M age = 85.38 days; SD = 2.31). At ages 2 weeks and 12 weeks, mothers were sent additional questionnaire packages at home (see Table 1 for an overview of the instruments collected and the time of assessment).
When the infant was 6 weeks old, parents were visited at home. This infant age was chosen because infant crying increases from birth onwards and reaches a peak around 6 weeks of age, also known as the crying peak (Barr et al., 2006), and crying is known to trigger caregiving behaviour (Zeifman, 2001). All visits took place during the late afternoon (after 15:00) or in the evening (M = 17:40, SD = 01:59). During the home visit, parents were first asked to fill in some questionnaires and then perform a working memory task and handgrip dynamometer task, not relevant for the current study. Afterwards, parents were asked to undress, change the diaper, and redress their 6-week-old infant, interacting with their infant as they typically would. For ethical reasons, parent-infant interactions were only carried out when the infant was not overly distressed. The interaction was filmed as unobtrusively as possible by the experimenter and was 15 min long (in cases when the parent finished before the 15 min, he/she was asked to continue interacting with the infant until 15 min were completed). Changing an infant at this age constitutes a mild physical stressor that J o u r n a l P r e -p r o o f may elicit crying and fussing (Jansen et al., 2010). Infants are usually fussier and cry more in the evening than in the morning (Barr et al., 2006). This procedure was chosen to elicit caregiving behaviour by infant crying. When both parents participated, mothers and fathers interacted with their infants separately, for 15 minutes each. Mothers interacted with the infant first.

Collection of milk samples
Breast milk was collected by the mothers, after washing their hands, breasts, and nipples with water, in collection cups provided by the researchers (unpublished results of our lab have shown that using water to wash the breast prior to sampling yields the same results as using soap or mild antiseptics). Prior to feeding the infant, mothers collected 20 mL of the first milk in the morning [mean (SD) time at 08:36 (02:48) am] by hand expression (in cases where mothers needed to use a breast pump instead, parts that came in contact with the milk were boiled prior to collection). Mothers noted the date/time of sample collection, maternal illnesses, and medication intake the preceding week. After collection, samples were immediately stored in the mothers' freezers (-18 --20 ºC). After the last sample was taken (infant M age = 85.38 days; SD = 2.31), samples were collected in a portable freezer by a researcher to be stored at -80 ºC at Radboud University, Nijmegen, in the Netherlands. Aliquots of the samples were afterwards shipped by temperature-controlled shipment to the Utrecht University Medical Centre, the Netherlands. Cortisol was quantified by Liquid Chromatography-tandem Mass Spectrometry (LC-MS/MS), adding cortisol-D4 as an internal standard. The limit of quantification for cortisol was 1.0 nmol/L. About 4 ml was needed for cortisol analysis. Taking cortisol circadian variations into account and in line with previous research (Glynn et al., 2007;Grey et al., 2013;Nolvi et al., 2018), we included the collection time of breast milk sampling and the time interval between waking up and collection time of breast milk sampling in our statistical models.

Instruments
An overview of the instruments, the assessment time, the type of assessment, and whether the assessment concerns the mother, the father, or the infant is presented in Table 1. , 1991). Mothers were asked, "Which of the four descriptions describes best how you feel?" each to be answered on a 7-point Likert scale from 1=strongly disagree to 7= strongly agree. Accordingly, participants can be categorised into four dimensions: preoccupied, dismissing, unresolved, and secure. This instrument has been widely used to assess adult attachment, including mothers' and fathers' own attachment (Johansson et al., 2021). Here, we used one categorical variable containing the four classifications and each category on a continuous scale with higher scores indicating a better endorsement of the attachment dimensions listed above. The instrument is not meant to assess mothers' attachment to their infant, but their J o u r n a l P r e -p r o o f own attachment orientations in relationships, which may also affect their caregiving practices (Van IJzendoorn, 1995).

Maternal attachment was assessed with the Relationship Questionnaire (Bartholomew & Horowitz
Childhood trauma was assessed with the 25-item Childhood Trauma Questionnaire (CTQ; Bernstein et al., 2003) on a 5-point Likert scale (1=never true to 5=very often true). The instrument comprises 5 subscales: physical, sexual, and emotional abuse, and emotional and physical neglect. The CTQ total score was used here as a continuous variable, calculated as the mean of all 25 items, with higher scores indicating more exposure to trauma. Reliability analysis revealed a Cronbach α coefficient of .82.
Childhood adversity was assessed with the Adverse Childhood Experience (ACE; Mersky & Janczewski, 2018), which contains 10 categories of adversities (emotional, physical, and sexual abuse, emotional and physical neglect, witnessing domestic violence, growing up with mentally ill or substance abusing household members, loss of a parent, or having a household member incarcerated), with response possibilities yes=1 and no=0. The instrument was adapted for our sample by excluding community violence and war/collective violence items, as such events are extremely rare in The Netherlands. The higher the score, the more adversities (sum score) were experienced early in life.
Ego-resiliency was assessed with the Ego-Resiliency Scale (Block & Block et al., 1980). This is a 14-item one-dimensional questionnaire on a 4-point Likert scale (1=does not apply at all to 4=applies strongly), with higher scores suggesting more trait ego-resiliency. Internal consistency analysis yielded a Cronbach alpha of .75.
Empathy was assessed with the Interpersonal Reactivity Index (IRI, Davis, 1980) on four dimensions with seven items each on a five-point Likert scale (0-4): fantasy scale (FS), namely the tendency to identify strongly with fictitious characters, perspective taking (PT) or the tendency to adopt another's psychological perspective, empathic concern (EC), that is, the tendency to experience feelings of warmth, sympathy, and concern toward others, and personal distress (PD) or the tendency to have feelings of discomfort and concern when witnessing others' negative experiences. Scores for each subscale represent the mean of the seven items, with higher scores indicating a higher tendency. Reliability analysis revealed a Cronbach α coefficient of .80.

Self-Efficacy for Parenting
was assessed with the Short Form Self-Efficacy for Parenting Tasks Index-Toddler Scale (SEPTI-TS; Coleman & Karraker, 2003). This is a 26-item questionnaire on a 6-point Likert Scale (1=strongly disagree to 6=strongly agree). Questions concern seven domains: emotional availability, nurturance, valuing the child, empathetic responsiveness, protection from harm or injury, discipline and limit setting, play, teaching and instrumental care, and establishing structure and routines. In this study, we used the total mean efficacy score, with larger values corresponding to more self-efficacy for parenting.
Reliability analysis revealed a Cronbach α coefficient of .91.

J o u r n a l P r e -p r o o f
Anxiety was assessed with the Dutch translation of the State Anxiety Inventory (STAI- S Spielberger et al., 1983), which consists of 20 statements related to feelings of anxiety at the present moment. Participants scored on a 4-point scale from 1 ("not at all") to 4 ("very much") how they felt at that specific moment, and scores were summed up. Results indicated excellent internal consistency in this sample (Cronbach α = 0.91). Higher scores indicated more feelings of anxiety (sum scores).
Depression was measured with the Edinburgh Postnatal Depression Scale (EPDS; Cox et al., 1987). The EPDS consists of 10 items for which participants indicated whether they experienced depressive symptoms in the past seven days. Responses were scored on a 4-point Likert scale ranging from 0 to 3, with higher scores indicating more depressive symptoms (sum scores). Internal consistency was good in this sample (Cronbach α = 0.87).
Daily hassles were measured with the Everyday Problem Checklist, which assessed the occurrence and intensity of daily hassles (Vingerhoets et al., 1989). Participants indicated whether a daily hassle (49 items presented) had occurred in the past two months and how much it had bothered them on a 4-point Likert scale ranging from 1=I do not mind at all to 4=I do mind a lot. The mean intensity rating of daily hassles was computed by dividing the sum of total (negative) valence by the frequency of the events, with higher values indicating more experienced negativity. Internal consistency in this sample was good, with Cronbach's α of 0.84. Shared childcare activities, namely the chores for taking care of the baby, were assessed on a diary-like instrument at weeks 2, 6, and 12 on a scale from 1 to 5, with 1 indicating that it is always the father and 5 indicating that it is always the mother to perform each of the 18 chores assessed (e.g., feed the baby, change the baby). A middle score indicates an equal share of childcare activities, and a higher score indicates the mother is doing more infant care.
Parental sensitivity was assessed based on ratings of video-recorded observations using the 9-point Ainsworth's rating scale (1978). These scales have been extensively validated in various cultures and are well applicable for rating the quality of parental behaviour with very young infants in natural settings.
Sensitivity (versus insensitivity) refers to the extent to which the parent timely and adequately responds to the infant's needs and signals. Highly sensitive parents are aware of all, including subtle, signals from their infant, accurately interpret these signals, and react to them promptly and appropriately. In contrast, insensitive parents are often unaware of their infants' signals, either by ignorance or failure to perceive subtle communications, may not understand their infants' signals, may react inappropriately or late to these signals, or not react at all (Ainsworth et al., 1978). The rating scales range from 1=highly insensitive/interfering through 5=inconsistently sensitive/mildly interfering to 9=highly sensitive. Higher scores here suggest higher sensitivity. About 30% of the videos were rated twice for reliability, yielding an interrater agreement of .82.

J o u r n a l P r e -p r o o f
Childcare practices for the baby were assessed retrospectively at 12 months in a diary-like instrument.
Mothers indicated the baby's age when additional care was present (age of entry at childcare in months) and the type of childcare (centre-based versus non-centred-based and mixed). Non-centred-based childcare includes grandparents, host parents, other family members, and babysitters.
Infant crying was assessed at weeks 2, 6, and 12 with the Baby Day Diary (Barr et al., 1988) for three consecutive days at each time point. The measurement of fussing and crying in the Baby Day Diary is valid and produces data comparable to actual audio recordings (r = 0.90; Barr et al., 1988). Each day, mothers reported the following infant behaviours: fussing, crying, unsoothable crying, sleeping, feeding, and being awake without crying. This was indicated with lines/symbols assigned to each behaviour on time bars. Each 24-hr period was represented on a sheet of paper with four horizontal 6-hr time bars, subdivided into periods of five minutes (Barr et al., 1988). Mothers were asked to fill this diary every two to three hours retrospectively. The diary data were prepared for analysis by noting the number of times fussing, crying, and unsoothable crying occurred (i.e., frequency) and how long the behaviours lasted (i.e., bout length). No distinction was made between fussing, crying, and unsoothable crying in the analyses, and the behaviour is henceforth referred to as "crying." This was done separately for each of the three days, and then the mean daily frequency and bout length were calculated. The multiplication of the mean daily frequency and mean bout length rendered the mean total duration of crying. An index of the history of crying (mean duration in minutes per 24 hr per all three days of recording in the diary) was calculated with the area under the curve to the ground (AUC) with the three points of assessment.

Statistical Analysis
All analyses were conducted in R. The code is publically available (https://doi.org/10.5281/zenodo.6586667). Data preparation included the following steps. Cortisol was analysed on the log scale; otherwise, the assumption of normality of the residuals could not be met. Next, missing values were imputed using predictive mean matching (m = 50, k = 5) using the mice package (Buuren, 2019). Predictive mean matching has been shown to work well under conditions similar to this study's (Buuren, 2019). The missingness ranged from 1.4 % to 26% for all variables except for father sensitivity, which had 43.8% missingness (i.e., several mothers participated without their partners).
Therefore, we performed the entire analysis with and without this variable and ran a complete case analysis

Random Forest
We fitted Random Forest models using the R package ranger (Wright & Ziegler, 2017) and tuneRanger (Probst et al., 2019). The tuneRanger package optimises hyperparameters using out-of-bag error estimation.
After tuning the hyperparameters mtry, node size, and sample fraction, we obtained the 2 out-of-bag estimate as an accuracy measure. The out-of-bag error estimation allows utilising the total sample while estimating out-of-sample accuracy. This procedure was performed twice for each time point: once for a model that only contained the time-specific known covariates (collection time, time since awakening, postnatal week, and previous cortisol measurement) and once for a model that contained all variables. We repeated this step for each imputed dataset and obtained the median 2 . Next, we generated a distribution of 2 values under the null hypothesis by repeating the above steps 100x for each of the 50 imputed datasets after permuting the outcome variable. The resulting distribution was used to obtain the p-value. To identify predictors that improve predictive performance, we obtained permutation-based variable importance scores and calculated p-values based on the method described by Altmann et al. (2010).

Bayesian Linear Models
We fitted Bayesian Linear Models and performed projection predictive variable selection (Piironen & Vehtari, 2017) using the packages rstanarm (Goodrich et al., 2022) and projpred (Piironen et al., 2020)  we extracted the predictors that achieved the lowest root mean squared error, reported the frequency at which each predictor was selected, and summarised the posterior distributions of the corresponding coefficients.

Results
Descriptive analyses of the sample characteristics and the main variables in the study are provided in Table 2. Figure 2 illustrates individual trajectories of changes in cortisol concentrations in the breastmilk from week 2 to 12.

Predicting Breast Milk Cortisol per Time-Point: Random Forest Regression
For each time point, we fitted 2 models. One model only included the time-specific covariates (the base model), and one included all variables (the full model). The Random Forest algorithm could predict cortisol levels at 2 ( 2 = 0.056, = 0.047), 6 ( 2 = 0.074, = 0.036), and 12 ( 2 = 0.120, = 0.008) weeks using the base models. However, the full models performed worse when looking at the out-of-bag 2 accuracy score at 2 ( 2 = 0.049, = 0.062), 6 ( 2 = 0.024, = 0.15), and 12 ( 2 = 0.067, = 0.037) weeks. This outcome indicates overfitting to noisy variables in the full models and that no single variable has a strong predictive value for breast milk cortisol levels.
Nevertheless, we tried to identify which variables in the full models may have predictive value next to the variables in the base models by calculating permutation-based importance scores (Table 3, Figure   S1) is among the top 3 predictors for all models explaining the superior performance of the base models. Among the variables not in the baseline models, empathy (fantasy) and state anxiety at week 6 and early life adversity in the form of ACEs at week 6 are among the significant predictors indicating that these variables may help to predict breast milk cortisol levels at the corresponding time points. However, the very low accuracy of the full models indicates that these predictors have poor explanatory power.

Predicting Change in Breast Milk Cortisol over Time: Bayesian Linear Models
We performed feature selection once per imputed dataset using Projection Predictive Feature Selection with a Bayesian linear regression model that predicts change in cortisol over time. Table 4 summarizes the posterior distribution of the coefficients for each variable summarized over all imputed datasets. It furthermore includes the frequency of how often the projpred algorithm selected this variable. The higher J o u r n a l P r e -p r o o f the frequency, the more confident we can be that the predictor has predictive value for the outcome. If a predictor is only selected in a few imputed datasets, it is more likely to be a chance finding.
One of the highly correlated time point variables (either collection time or time since awakening) was always selected. Furthermore, cortisol measured at the previous time point was selected across most imputed datasets. As expected, the model suggests a negative relation between the cortisol concentrations and the time variables and a positive relation between adjacent cortisol concentrations assessment. The very small effect sizes, the overlap of the 95% credible interval with zero, and the moderate frequency of the time variables can be explained by their high correlation (r = .903, p < .001). In other words, adding another does not add much information about breast milk cortisol when knowing one of the time variables. The variable ACE was selected across 58% of datasets and was negatively related to breast milk cortisol.
However, the effect size was very small, and the 95% credible interval overlapped with zero, indicating that it is a weak predictor of the change in cortisol over time.

Discussion
Summarizing the results of two complementary analyses, we found that the relatively large number of studied predictors could not explain variance in the milk cortisol concentrations at each time point (2 weeks, 6 weeks, and 12 weeks) and in the increase in cortisol over the weeks. In other words, although we achieved significantly better prediction accuracy than would be expected if the predictors had no predictive value at all, the prediction accuracy is very low, and no single predictor has strong predictive value beyond the known predictors of milk cortisol, such as time since awakening and time of collection. Interestingly, however, maternal adverse childhood experiences (ACEs) were detected by both statistical approaches to be related to lower milk cortisol concentrations at 6 weeks and a larger decrease in cortisol over the weeks, albeit with a very small effect size. With even higher uncertainty, as only identified by a single modelling approach, empathy and postnatal anxiety were associated with higher milk cortisol at week 2. Below, two main critical messages of this study are discussed: 1) the relation between ACEs and breast milk cortisol, and 2) the lack of meaningful predictive power of many variables on human milk cortisol.
Maternal adverse childhood experiences negatively predicted the cortisol concentrations at week 6 and the cortisol increase from week 2 to week 12. These findings point towards a blunted cortisol production in mothers reporting more early adverse experiences. This is in line with recently documented findings of a meta-analysis aggregating associations between ACEs and different measures of stress reactivity (i.e., cardiovascular indices, salivary and plasma cortisol), except for breast milk measures, over 80 studies (Brindle et al., 2022). Indeed, to date, the evidence appears to indicate that exposure to repeated stressors likely results in chronic HPA dysregulation, particularly down-regulation (Kalmakis et al., 2015). However, the discussion on the factors and the mechanisms responsible for hypo-or hyper-regulation of the HPA axis J o u r n a l P r e -p r o o f remains ongoing (Brindle et al., 2022). Moving away from the somewhat simplistic hypo-and hyperregulation dichotomy, a sizeable p-curve meta-analysis investigating the distribution and the shape of pvalues reported in the literature found that HPA axis calibration follows an inverted u-shaped curve, with severe adverse events being associated with down-regulation and mild adversities to heightened cortisol production (i.e., salivary, plasma cortisol; Hosseini-Kamkar et al., 2021). This pattern was explained in the framework of the general adaptation syndrome model by Selye (1946), which posits that upon repeated/acute stressful experiences that result in hyper-activation, the stress system becomes depleted, resulting in more efficient energy expenditure, namely hypercortisolism. Given our sample's highly educated and low-risk nature, it is unsurprising that participants did not generally experience many adversities (see Table 2). Accordingly, it may have been reasonable to expect an increase in cortisol.
However, this remains to be further corroborated in larger samples, wherein a distinction in the severity level of adversity rather than simply the frequency of adverse events could provide additional insights (Cohodes et al., 2023;McLaughlin et al., 2011). Notably, the literature discussed above mostly concerns salivary cortisol and, in some instances, plasma cortisol, meaning that this discussion assumes that cortisol concentrations obtained from different bodily materials (saliva, milk, blood, hair) would behave similarly under the same circumstances. Whether this is the case is an essential empirical question for future work, warranting efforts towards greater methodological integration in physiological research to prevent the accumulation of scattered, too often mixed, findings.
Another striking gap in the psychoneuroendocrinology field is that we our results provide the first evidence of an association between early maternal adversity and lower milk cortisol in human mothers, these results need to be replicated in more extensive studies and studied in relation to infant development. While little is known about cortisol in breast milk, fewer studies examined links between (heightened) cortisol concentrations in breast milk and infant outcomes. These studies reported relations between higher milk cortisol and a more fearful temperament (Grey et al., 2012;Nolvi et al., 2018;Glynn et al., 2007), lower BMI percentile (Hahn-Holbrook et al., 2016), and better neurodevelopmental outcomes in the first two weeks of life (Hart et al., 2004). In sum, the results of the current and previous studies underscore that cortisol in mother milk may be associated with infant development. However, the causality of such relations and the potential effects of higher or lower milk cortisol concentrations on child development remains to be determined.
The second key message of the current paper is the lack of associations between a substantial number of maternal dispositional and situational factors and cortisol concentrations in breast milk (i.e. both absolute concentrations and increase over time). Including more variables than the base variables in the random forest's models produced results indicative of poor predictive power and/or noisy measurement of some of the variables. However, despite including a plethora of potentially relevant variables, many remain that were not included. The first three months postpartum are accompanied by life changes that demand high and immediate parental adaptation to new routines and, importantly, sleep cycles (Gay et al., 2004).
Indeed, changes in sleep patterns or lack thereof and fatigue, infant behaviour (crying, waking, feeding), relationship quality and other social changes, all of which we did not assess in this study, may explain cortisol variations in breast milk. For example, breastfeeding mothers may experience changes in their circadian rhythm due to the demands of caring for an infant. Additionally, the frequency and timing of feedings can affect the release of hormones, including cortisol, melatonin, and prolactin, affecting the mother's sleep-wake cycle and circadian rhythm (Brunton et al., 2008;Glynn et al., 2007). Evidence from plasma cortisol samples revealed that chronic circadian misalignment yields reduced cortisol concentrations (Wright et al., 2015). We could reasonably hypothesize that parents of a newborn also suffer from chronic circadian misalignment, especially more susceptible individuals like those with a history of adversity. Such factors should be investigated in relation to milk cortisol as well as with the goal of determining potential implications for the mother's health and child development. These studies may inform future supportive interventions for parents in the perinatal period.
Despite providing crucial initial evidence on how early maternal adverse experiences may transition into the next generation, our study also suffers from some limitations, for instance, a relatively small sample drawn from a low-risk community sample. A selection bias may have been introduced, as it often is the case for young parents willing to contribute to longitudinal studies. Next to additional aspects characterizing the early postpartum period that may affect cortisol concentrations in breast milk, day-today variability in cortisol, possibly partly due to measurement error, deserves attention in our discussion and future studies. Measurement errors could be due, amongst others, due to maternal self-reports, selfcollection of milk samples and initial home storage, and having gathered only one milk sample per day for and 50% high, we may also infer that it covers a heterogenous, albeit low-risk, sample. However, as can be seen in the descriptives table, mothers report a generally low incidence of adverse events and trauma and most shared caregiving tasks with their partners. This may explain why although ACEs seemed to carry explanatory value to milk cortisol variations, the effect size was small.

Conclusion
In conclusion, no single variable had a substantial predictive value for milk cortisol concentrations or their increase within the first 12 weeks postpartum in lactating mothers. However, maternal adverse childhood experiences (ACEs) were related to lower breast milk cortisol at week 6 and a smaller milk cortisol increase from week 2 to week 12, albeit with a small effect size. These results that warrant replication suggest that mothers with more ACEs may pass lower cortisol levels to their infants through breast milk. M-diary=diary filled in by the mother, C=breast milk collected by hand expression; AUC=area under the curve. M=mean, SD=standard deviation, %=percentage, Min=minimum, Max-maximum, w=week, P=pregnancy assessment; N*=73, unless otherwise indicated; the different sample size for each variable is due to a missed assessment point or round as a whole of the mother or the father. Note that although postnatal anxiety was only assessed at week 6, we included it in all models as a potential predictor because the instrument assesses symptoms in the past month.