Background

Pharmacoepidemiology, pharmacovigilance and the reproductive years

After a medicinal product has been marketed, patient safety depends on accurate population surveillance, pharmacovigilanceglossary—detecting, assessing, and preventing adverse effects, and pharmacoepidemiologyglossary – describing the use and effects of drugs in large numbers of people (Table 1 Glossary has definitions). Pregnant and breastfeeding individuals and their infants should not be excluded from the protection afforded by pharmacovigilance [1]. Neither continuation nor discontinuation of medicines is without risk, but harm can be minimised by effective pharmacovigilance. This depends on comprehensive characterisation of drug-related benefits and harms, and any imbalance. For people of childbearing age, this should include information across the full reproductive life cycle: fertility rates; pregnancy loss; terminations, for all reasons; congenital anomalies; preterm birth; growth centiles; complications of pregnancy; complications of childbirth and the puerperium (e.g. haemorrhage); neonatal complications (pulmonary hypertension, hypoglycaemia, discontinuation syndromes); breastfeeding rates at different ages; infant and childhood outcomes, including cognitive functioning, neurodevelopmental disorders, education performance, long-term conditions, survival and reproductive success [2].

Table 1 Glossary: definitions of terms in this paper
Full size table

This paper aims to present the arguments for breastfeeding to be included alongside medicines use and neurodevelopmental outcomes in population databases and studies investigating the benefits and harms of medicines during pregnancy, the puerperium and postnatal period. We review:

  • 1) the current situation.

  • 2) how breastfeeding might be accommodated in pharmacoepidemiological models exploring the impact of medicines on breastfeeding as an outcome and on infants exposed to medicines in utero and via breastmilk.

  • 3) the challenges in obtaining comprehensive data.

Breastfeeding and medicines: locating the data and why we need them

To research the impact of medicine exposure during and after pregnancy on infant development, we shall need population databases linking data on medicine exposure plus breastfeeding plus infant development. Across Europe, few population databases hold data on all three together, and there is little uniformity in outcomes, definitions, methods and timing of assessments, as illustrated in Table 2.

Table 2 European Population-based data sources with data on breastfeeding plus medicines use during pregnancy plus neurodevelopment
Full size table

It took 30 years for the dose–response associations between in utero exposure to valproic acid derivatives and altered neurodevelopment trajectories [26], and congenital anomalies [27] to be incorporated into regulatory measures to curtail prescribing during pregnancy [28]. Valproate prescribing to women of childbearing age is declining [29], but, in England, it was prescribed to 247 pregnant women between April 2018 and September 2021 [30]. Subsequently, pharmacoepidemiologists have considered the effects of exposure to other medicines during pregnancy, particularly, antidepressants [31], and opioids [32]. However, although suboptimal breastfeeding is one of the main threats to global health [33], it appears to be a ‘blind spot’ in healthcare databases and pharmacovigilance.

It is rarely possible to obtain the population-wide picture of short- and long-term transgenerational outcomes from clinical trials, case series, spontaneous reports, and cohort studies, due to (largely unavoidable) selection biasglossary [21, 34]. Prospective observational birth cohort studies collect information on infant feeding, but they represent a self-selected sample of the population. For example, cohorts in Norway and Denmark recruited 41% and ~ 30% of the eligible populations [35, 36], whilst other cohorts lack linked prescription [37] or child development data [38]. Pharmacokinetic studies usually involve small numbers of participants, and, like animal studies, may not predict developmental or clinical outcomes [39]. Manufacturers’ medicine-related pregnancy registries capture insufficient data on pregnancy, infant follow-up, and breastfeeding: median (interquartile range) enrolment is reported as 36 (5–258) pregnancies and 12 (2–119) infants [40]. Patient safety researchers are, therefore, examining population databases; however, data quality and validity are not always completely evaluated [41, 42], and most have no data on breastfeeding and neurodevelopment [43, 44].

Breastfeeding

Breastfeeding requires an optimal biopsychosocial milieu [45]. Socioeconomic status (SES) glossary, cultural norms, availability of infant formula, and maternal intention are important, but the impact of prescribed medicines on breastfeeding should not be overlooked [46]. Exposure to some prescription medicines in pregnancy, labour and postpartum, may reduce breastfeeding initiation or continuation [24, 46, 47]. The complex physiology of lactation is vulnerable to disruption, particularly by medicines that affect serotoninergic pathways (including antidepressants) [48, 49], antagonise prolactin (amphetamines, oestrogens, ergotamine derivatives, aripiprazole, promethazine, possibly diuretics, injected corticosteroids), or reduce oxytocin release (alcohol, opioids, possibly sympathomimetics, anticholinergics, antidepressants) [50,51,52,53]. However, we do not know all the reasons underlying the lower breastfeeding rates amongst those using prescription medicines. People may be reluctant to breastfeed if the impact of the medicine on the infant is unknown [54]: doubts, hesitancies, and anxieties may compound physiological difficulties.

Neurodevelopment

Medicine exposure through breastfeeding is an important consideration in analyses of neurodevelopmental outcomes (neurodevelopmental disorders, cognitive performance, educational performance) [55,56,57,58], particularly if prolonged and exclusive [59, 60]. The effect of breastfeeding is demonstrated in many observational studies and a cluster randomised controlled trial of assistance with breastfeeding [61]. Similarly, a meta-analysis of observational studies reports that children with autism are less likely to have been breastfed (OR = 0.61, 95% CI 0.45, 0.83) [57].

Causation is not easily established: infants may be exposed to medicines in utero and/ or via breastmilk, and observational studies cannot discount the possibility that difficulties with breastfeeding, and early discontinuation, are due to neonatal irritability associated with early signs of neurodevelopmental problems [62]. Also, some prescribed medicines (antidepressants, valproic acid derivatives) and high dose alcohol may simultaneously predispose to neonatal irritability [63], and disruption of breastfeeding physiology [48, 64, 65]. Accordingly:

  • Breastfeeding warrants consideration as a health outcome measure, indicating a healthy mother-infant dyad.

  • Those concerned with medicines’ safety in pregnancy and breastfeeding should account for infant feeding when modelling both long- and short-term outcomes.

  • Inter-dependence between medicines and breastfeeding warrants scrutiny, alongside the full range of putative aetiologies of adverse outcomes. Currently, the optimum strategy is uncertain.

How should breastfeeding be accommodated in pharmacoepidemiology?

Without information on breastfeeding, it will be impossible to separate the effect of exposure to medicines in utero and/or during breastfeeding from the effect of ‘not breastfeeding’ for some medicines. Although prescribed medicines and breastfeeding may affect infant development in different ways, information on breastfeeding is needed to understand and minimise adverse outcomes in childhood. With definitions (Table 1 Glossary), examples and explanation of implications, we explore breastfeeding in relation to confounding, mediating, moderating, and colliding [9], and offer diagrammatic illustrations as tentative directed acyclic graphs (DAGs) for discussion [66, 67]. Breastfeeding may be:

  1. a)

    A confounder, when exploring whether exposure both during and after pregnancy affects the infant, assuming medicines reach the infant in utero and then via breastmilk.

  2. b)

    A mediator, when exploring how in utero exposure affects neurodevelopment, when medicines may affect initiation or duration of breastfeeding.

  3. c)

    A moderator, by countering any adverse effects of medicines on neurodevelopment.

  4. d)

    Vulnerable to colliding and volunteer bias when exploring the effects of medicines on breastfeeding success.

  5. e)

    Incorporated into a range of statistical models.

Confounding

Definition

The concept of ‘confounding’glossary, based on differences (or non-comparability) between exposed and non-exposed subjects, distinct from ‘selection bias’ glossary, has developed in the last half-century [68]. The definition of confounderglossary most widely adopted is: “a factor associated with both the exposure and the outcome, and not part of the causal pathway from exposure to outcome” [8], blurring effects [7, 69, 70] (Table 1).

All observational studies are vulnerable to confounding [67], including those considering the impact of maternal medicines on infant development. Some confounders are well known e.g., socioeconomic status (SES) glossary; others are known, but accurate information is almost impossible to obtain in fieldwork and routine care e.g., doses of recreational drugs consumed. However, there remain ‘unknown unknowns’ or ‘lurking or latent variablesglossary’ [71], such as local environmental pollution and cultural norms. Without randomisation, researchers can only speculate as to the effects of these unmeasured or unmeasurable confounders [13] on breastfeeding and childhood outcomes, based on knowledge of the data and their own milltir sgwar (communities, where they have lived for generations).

Example

Breastfeeding would be considered a confounder in the analysis of the effect of medicines used both during and after pregnancy on neurodevelopmental outcomes when prescribed medicines do not greatly affect breastfeeding physiology and reach the infant in utero and via breastmilk. Most medicines enter breastmilk in small quantities [39]; therefore, breastfeeding affects both:

  1. a)

    neurodevelopment directly [57, 61], probably via biological processes and

  2. b)

    infants’ total medicine exposure during both prenatal and postnatal periods of brain development (Fig. 1).

Fig. 1
figure 1

How does total medicine exposure affect infants’ neurodevelopment? Breastfeeding as a confounder.

In this example, breastfeeding affects both neurodevelopmental outcomes and total exposure to medicines (via the placenta + breastmilk). This relates to postnatal exposure, following prenatal exposure

Full size image

Implications: infants’ total exposure, including via breastmilk

Identifying associations between infant outcomes and medicine exposure via breastmilk is complicated by variations in medicine transfer from breastmilk to infant, depending on dose, timing of administration and breastfeeding, and supplementary formula feeding. Concentrations of medicines in infant blood samples, as proportions of maternal blood concentrations, vary widely, for example: 28.9% (0.6–90.3%), 17.2% (12.4–22.0%), 21.4% (17.9–24.9%), and 44.2% (35.2–125.3%) (median and full range) for lamotrigine, topiramate, valproic acid, and zonasamide respectively [72]. Similarly, appreciable but variable and unpredictable concentrations of citalopram, sertraline, venlafaxine and metabolites pass into breastmilk [73]. This variation in infant exposure may relate to infants’ ability to metabolise and eliminate medicines. Elimination is compromised in premature or sick infants [39] or if maternal or infant metaboliser or transporter status is unusual [74], suggesting a need for close infant monitoring [75, 76]. Although prenatal antidepressant exposure is reported to adversely affect cognitive development [77,78,79], there are few data on long-term outcomes of exposure via breastmilk [65]. Short-term effects of exposure to psychotropic medicines via breastmilk include sedation, irritability, restlessness, diarrhoea and suboptimal weight gain [80, 81], but how these relate to long-term neurodevelopment is unknown. Current practice recommendations rest on case series, indicating that opioids, clozapine, amisulpride, combinations of central nervous system depressants, amiodarone, oral retinoids, radio-iodine, topical and systemic free iodine, and chemotherapy pass into breastmilk, risking (at least) short-term transgenerational (mother to child) adverse drug reactions (ADRs) glossary following breastmilk exposure [39, 82]. Including breastfeeding in analyses of infant outcomes informs families as to whether the benefits of breastfeeding outweigh the risks of increasing infants’ total dose. Analysis of IQ in children followed to age 6 suggests that there are no disbenefits of breastfeeding for children of mothers using sodium valproate (n = 35), but more data are needed regarding breastfeeding when prescribed phenytoin (n = 36) [83], particularly for rare ADRs, such as methaemoglobinaemia and combination regimens [65]. Larger studies are needed to resolve the dilemma “should those prescribed medicines breastfeed?” (see moderating, below).

Mediating

Definition

Mediation occurs when the exposure is associated with an intermediate variable (the mediator), which is then associated with the outcome [84]. Mediatorsglossary, unlike confounders, lie on the causal pathway between exposure and outcome, and describe how, or even why, an association occurs [85, 86]. Breastfeeding is a mediator when:

  • a) medicines reduce breastfeeding, and then.

  • b) reduction or absence of breastfeeding affects neurodevelopment and health (Fig. 2).

Fig. 2
figure 2

How do medicines in pregnancy affect infants’ neurodevelopment? Breastfeeding as a mediator.

Breastfeeding as a mediator relates largely to prenatal exposure

Full size image

Mediator effects can be tested by exploring the relationship between exposure and outcome with and without the mediator [13], which is particularly important for breastfeeding after in utero exposure [87].

example

SSRI antidepressant exposure in late pregnancy is associated with reduced breastfeeding rates [88] at birth [89], discharge [90], two [91], and 6–8 weeks [47], and, in some studies, with delayed neurodevelopment, including motor control [92], social behaviours [77], and autistic spectrum disorders [78, 79]. Neurodevelopmental delay following prenatal and perinatal antidepressant use may be due, in part, to the medicines’ effect on breastfeeding, i.e., delay may be mediated (or caused) by reduced breastfeeding caused by medicines.

Implications: infant development

Antidepressants may disrupt the physiology of lactation by delaying alveolar secretary activation from 69 to 86 hours, due to serotonin-dependent changes in tight (inter-cellular) junctions [48] and disruption of local production of serotonin [49, 64]. In addition to direct effects on lactogenesis, SSRI exposure in trimester 3 affects monoamine metabolism and serotonin availability in infants, associated with a dose–response increase in restlessness, tremor, and incoordination [93]. These symptoms, and any neonatal withdrawal symptoms of irritability, may impede latching, making breastfeeding painful and difficult, promoting discontinuation. The same disturbances that increase difficulties with breastfeeding may underlie delays in fine motor development at three years [92] or autistic-like behaviours [77]. Adult insomnia [94] and sleep disorders associated with SSRIs [95] and their effects on mother-infant bonding [96] and breastfeeding may compound any direct effects on lactogenesis.

If absence of breastfeeding contributes to any suboptimal neurodevelopmental outcomes associated with antidepressant exposure, breastfeeding partly mediates the association. Any mediator effects of breastfeeding on development can only be explored where a database collects data on children’s neurodevelopment and medicines and breastfeeding.

Moderating

Definition

Moderatorsglossary affect the strength or direction of the relation between exposure and outcome [13]. Moderators explain when, and under which circumstances, associations occur, and are sometimes used to identify subgroups at risk, e.g., age bands or co-morbidities, where exposure and outcome may be more closely linked than in the full population [84, 97]. Breastfeeding is a moderator if it affects:

  • a) neurodevelopment directly via changes in neuronal architecture, and

  • b) the extent of the impact of prenatal and postnatal medicine use on neurodevelopment.

In regressionglossary analyses, moderating effects are usually tested with interaction variables: if these explain a statistically significant amount of model varianceglossary, moderator (or modification) effects are likely, i.e. associations identified depend on the value of the moderator [84], in this case, breastfeeding.

Example

If breastfeeding is a moderator, the impact of in utero exposure to medicines will depend on whether the infant is breastfed. In cohorts of infants exposed to antiepileptics (AEDs) (valproate, carbamazepine and lamotrigine) in utero, the prevalence of neurodevelopmental difficulties is lower in breastfed than formula-fed infants, despite the additional postnatal exposure [83, 98].

Implications: mitigating in utero exposure

This suggests that breastfeeding might mitigate harm emanating from AEDs or other medicines, and exposures should be explored separately in breastfed and formula-fed infants (Fig. 3). This would inform families regarding the benefits of breastfeeding while using medicines.

Fig. 3
figure 3

Does breastfeeding affect developmental outcomes for infants exposed to maternal medicines? Breastfeeding as a moderator.

Breastfeeding may affect developmental outcomes if infants have been exposed to AEDs prenatally. This relates to prenatal exposure, followed by postnatal use of medicines

Full size image

Genetic variations might be considered as moderators, defining sub-groups at risk of exposure via breastmilk [72]. Transgenerational adverse outcomes may be confined to genotypes vulnerable to changes in serotoninergic and corticosteroid substrates [99]. Also, CNS depression or sedation in breastfed infants mainly occurs in infants with low activity in blood–brain barrier efflux transporters (P-gp) [74]. Allelic variations in transporter proteins [100], and maternal or infant single nucleotide polymorphisms [101] are rarely recorded at population level, but may define subgroups at risk of ADRs, and should be considered as moderators in a priori subgroup analyses.

Colliding and Volunteer Bias: the case for whole population databases

Colliderglossary and volunteer biasglossary are examples of selection biasglossary, defined as systematic differences between participants and non-participants (Table 1).

Definition

Collider bias is the distorted (induced) association between two or more variables that both affect the likelihood of an individual being included in the dataset (sampled) [102]. A collider is a variable influenced by other variables: for example, when an exposure or risk factorglossary (such as medicines use) and an outcome (such as breastfeeding) both affect the likelihood of being sampled, they “collide”. Similarly, both being a ‘healthcare worker’ (exposure) and having a ‘severe COVID-19 infection’ (outcome) increase the chances of being tested for COVID, and thereby joining the dataset being analysed [102].

Volunteer samples may not represent the less affluent, smokers [21], or people with obesity [5]. Selective or volunteer recruitment and any deficit in representativeness risks collider bias [102]. This occurs when both exposure and outcome (or an antecedent of the outcome) influence recruitment or retention by their relation to volunteering, which then defines the sample [103,104,105]. The resulting collider bias can distort their relationships [102].

Example

Associations between variables may be vulnerable to collider bias [66, 106] if:

  1. a)

    Breastfeeding and medicine use both affect the selection of study participants, and

  2. b)

    the study sample over-represents these characteristics, (Fig. 4).

Fig. 4
figure 4

Does volunteer recruitment affect investigation of breastfeeding? Illustration of Collider Bias.

In this example, the outcome is breastfeeding. Its relationship with maternal medicines is influenced by the composition of the sample of women studied. In this example, both breastfeeding and using prescription medicines affected recruitment to the study

Full size image

Implications: breastfeeding as a study outcome

When exploring the impact of medicines on initiation or duration of breastfeeding, if recruitment were to favour participants who a) were not using medicines, and b) breastfed, these characteristics would be over-represented. This over-representation would distort the sample and generate associations between breastfeeding and ‘no medicines’ that may not appear in the wider (non-volunteer) population [102]. An analogy might be drawn between volunteer bias and the “streetlight effect”: looking for lost items under a lamp-post, because that is the only place where anything can conveniently be seen [107]. Any associations found under the light may be due to colliding and co-existence in the illuminated patch: whilst they are valid in the sample examined, they may not be true in the wider population [102].

Capturing the whole population of a country or region removes volunteer bias, and hence collider bias, because neither exposure nor outcome nor covariatesglossary drive study inclusion. Attempts to account for colliding in the statistical analysis involve untestable assumptions, and it is better to avoid this problem by capturing data on the unselected regional or national population [102]. However, many databases are vulnerable to ‘live birth bias’, as they fail to record miscarriages [108], and infants dying within their first few days may not be linked to population databases [109], which means they do not capture all pregnancies, risking collider bias.

Approaches to analysis

Regressionglossary models can identify associations, for example, between prescription medicines and exclusive formula feeding, but there are more complex questions, such as the inter-relationships between breastfeeding, maternal medicines, and infant development. Multilevel modellingglossary extends regression analysis to account for clustering of individuals, for example by hospital, primary care provider, region, or country. This allows for the possibility that, in any one cluster, exposures may differ, for example, each primary care provider may have an individualised prescribing pattern. Further analytic techniques move beyond regression models to explore causation:

  • Marginal Structural Modelsglossary were designed to accommodate:

    • ◦ time-dependent exposure (e.g., one-off medicine administration, changes in treatment or drug absorption throughout pregnancy) and

    • ◦ time-varying covariates, such as breastfeeding, that may be both confounders (prenatal plus postnatal exposure) and mediators (prenatal exposure) (above) [110]

Where prescription regimens change frequently, analysis depends on modelling assumptions [111]. This approach has, for example, been used to explore associations between breastfeeding, SES and adult health [112], and between breastfeeding, infant deaths and self-reported ethnic group [113].

  • Structural Equation Modelling.glossary offers a framework for theory-driven hypotheses to be tested (or falsified) in a single cohesive model, using large datasets [14, 114, 115]. For example, this approach has been used to explore: the effects of maternal and infant characteristics on breastfeeding techniques and breastfeeding, [116] associations between attention-deficit-hyperactivity disorder (ADHD), obesity and breastfeeding [117], and predictors of infants’ neurodevelopment, including breastfeeding [115]

These’ causal models’ are used where trials would be considered unethical (pregnancy, breastfeeding) or impractical (rare outcomes), and are scarce [118]. However, models cannot accommodate the possibility that omitted variablesglossary may bias associations of interest [112]. In contrast, large trials account for known and unknown confounding variables by randomisation, albeit within the recruited population. Return on investment in analytic techniques will be insufficient without comprehensive data collection.

Getting the full picture: challenges of comprehensive characterisation

Depending on how neurodevelopment is investigated, breastfeeding may be a confounder (when medicines are taken during pregnancy and whilst breastfeeding) or a mediator (when in utero exposure is considered) or a moderator (when defining ‘at risk’ subgroups) or a casualty of volunteer bias (in recruited cohorts). Many investigators consider breastfeeding sufficiently important to be an outcome itself [38], but it may be a casualty of volunteer bias in recruited cohorts [102].

Effective pharmacoepidemiology and pharmacovigilance need not only an understanding of causal pathways and unselected whole-population databases, but also comprehensive characterisation of the full range of variables affecting childhood outcomes [2, 119]. In addition to data on pregnancy dates, outcomes, and exposures to medicines and disease [2], regulators [1] recommend including data on: maternal age; obstetric and medical history; disease status and management; prescription of known teratogenic or foetotoxic medicines; folic acid and multivitamin use; smoking; alcohol intake; illicit drug use (with duration); lifestyle factors (exercise and nutrition); body mass index; and full family history of conditions possibly related to adverse perinatal and neurodevelopmental outcomes [2]. Breastfeeding is included as ‘follow up’ information [1]. Other parametersglossary may be important in determining perinatal and childhood outcomes, for example: SES; parental educational outcomes; infections or inflammation in pregnancy; healthcare contacts (including, but not limited to, antenatal monitoring); vaccinations; rurality [120]; distance from environmental pollutants [121,122,123]; and genetic/ epigenetic influences [101].

Environmental pollutants, including lead, mercury, and dioxins, pass into breastmilk [124] and may increase the risk of sub-optimal growth [125], allergy [126] and neurodevelopmental delay [79, 127]; however, breastfeeding mitigates the impact of prenatal exposure [128]. Organochlorides may reduce lactation [129]. As with medicines, long-term effects of exposure via breastmilk are under-investigated, and databases contain little information.

Inter-relationships between these myriad variables determining childhood outcomes complicate evaluation of co-exposures, and may converge on SESglossary. SES is associated with: breastfeeding status [24, 38, 46]; environmental pollution [130]; health, perinatal, developmental and educational outcomes [131]; morbidity [132]; depression [133]; smoking; substance misuse; and prescription medicines [25, 47], including antidepressant prescribing [24, 38, 46, 134]. Combining these diverse variables into a propensity score risks overlooking individual modifiable risk factors and targets for change, such as prescribing practices or breastfeeding support. Subsuming the impact of ‘not breastfeeding’ under SES allows it to be ‘drowned out’, obscured, and lost to pharmacovigilance.

Deprivationglossary [24, 38, 46], depression pre-pregnancy [47, 135] and antidepressants [47, 89,90,91] all lower breastfeeding rates [88, 136]. Both depression and antidepressants stimulate the hypothalamic–pituitary–adrenal axis, transfer of cortisol to the foetus, and epigenetic changes [137, 138] and their biological effects on neurodevelopment are difficult to disentangle [77]. Exactly how breastfeeding lies on the causal chain between deprivation (low SES) and poor school performance likely varies between individuals. Any impact of prescribed medicines on breastfeeding is of crucial importance, and any disruption of breastfeeding may have far-reaching consequences, as illustrated in Fig. 5.

Fig. 5
figure 5

How is deprivation linked to school performance? A putative causal chain.

School performance is affected by too many inter-related factors to be depicted in a single illustration. This figure illustrates just one scenario: we have prioritised clarity over complexity [103]

Full size image

The impact of prescribed medicines on reproductive health, childbirth and breastfeeding is not confined to transgenerational ADRs: other adverse effects, such as maternal weight gain, may affect breastfeeding directly or indirectly. For example, weight gain is associated with some antipsychotics, AEDs, antidepressants, and lithium; in turn, obesity complicates monitoring of pregnancy, glycaemic control, and childbirth [139, 140], and increases the risk of preterm birth, congenital anomalies and reduced breastfeeding [141,142,143].

How did we get here? Controlling the databases

The Cumberlege Report states that pharmacovigilance systems failed and are failing pregnant individuals prescribed valproic acid derivatives, due to inability to monitor adverse outcomes [144]( p.4). Congenital anomalies following in utero valproate exposure were reported in 1982 [145] and 1985 [27], but were regarded as isolated cases or attributed to co-prescribing. After accumulation of twenty years’ data from large databases demonstrated an association between in utero sodium valproate exposure and lower IQ, [26] UK authorities issued unequivocal prescribing instructions, in 2018 [28]. A similar argument might be made to address the absence of breastfeeding data. In 1994 and 2001, the American Academy of Paediatrics [146] recommended codeine for short term cough suppression whilst breastfeeding [147]. Codeine had been noted to cause apnoea [148] or sedation [149] in breastfed infants a decade earlier, but was not contra-indicated whilst breastfeeding, until the death of a breastfed infant from codeine exposure was reported in 2005 [150]. This case is controversial [151], but other case series are reported [65], and several opioids are probably harmful via breastmilk [23, 81]. Only one large database study is available [152], and assessment of codeine exposure is complicated by its availability (in low doses) without prescription. Until databases include breastfeeding, risks remain that either some transgenerational ADRs will escape detection or decisions on medicines approvals will be based on case series, with inherent risks of mis-interpretation [151].

To monitor adverse effects, maternal prescription records should be linked to all childhood outcomes and modifiable risk factors [144], and analyses defined explicitly a priori [153]. The usefulness of databases depends on comprehensive coverage, and the nature and detail of their data, including ‘women’s problems’ of miscarriage, pregnancy termination, breastfeeding (extent and duration), and all infections (including those sexually transmitted). Database studies, unlike large clinical trials, cannot rely on randomisation to account for unrecorded variables. Omission or redaction of data constrains examination of potential associations and renders hypotheses unfalsifiable [154] (p.44). Hence, ‘facts’ uncovered by research are dependent on and limited by the processes of inquiry and the zeitgeist of investigators and those establishing and controlling the databases [155, 156]. Any discrimination in data collection may lead to omission of key variables, ‘blind spots’, and obfuscation of transgenerational ADRs. We can only speculate as to reasons for the apparently selective exclusions of crucial variables [144]. Without full transparency, these decisions may be heard as echoes of patriarchal power and its ability to control resources [157], with multi-generational consequences [144].

Limitations of this paper

To our knowledge, this is the first paper to address the complex questions surrounding incorporation of infant feeding data into healthcare databases used to report patient safety. Other research approaches, such as case series, recruited cohorts and randomised trials, and analytical methods are detailed in guidelines [119], and are outside the scope of this discussion; however, concerns over external generalisation remain [21]. For simplicity, our tentative directed acyclic graphs do not offer comprehensive characterisation of the full range of variables affecting childhood outcomes and breastfeeding, listed above [1, 120, 122, 124, 129].

Conclusions

Families need to know whether prescribed medicines will make breastfeeding more difficult and if breastfeeding will leave infants vulnerable to ADRs from maternal medicines or confer benefits, as in the wider population [55,56,57,58,59,60,61]. Current data are inconclusive [44, 158]. Existing regression analyses may offer sufficient evidence to target low-risk interventions to those in most need: for example, records of antidepressant prescriptions in pregnancy should trigger additional breastfeeding support [24, 88, 89]. However, answering complex questions on transgenerational ADRs and how to avoid, monitor and mitigate them, will involve juxtaposition of high-quality linked data on medicines, childhood outcomes and modifiable risk factors, including breastfeeding, in whole-population databases. Comprehensive characterisation and robust analyses of drug-related benefits and harms necessitate information across the life cycle, from miscarriage to fertility of the next generations: breastfeeding should not be a ‘blind spot’.