Over the past two decades, a steady increase in prescription medication use during pregnancy, specifically in the first trimester, has been observed (Werler et al., 2023). This rise is mainly attributed to the increasing maternal age at the first pregnancy and the higher prevalence of chronic conditions present in persons of reproductive age, such as depression and hyperglycemia (Hayes et al., 2020). Consequently, medication use during pregnancy has become inevitable in persons of childbearing age. Despite that, there is a notable lack of evidence-based information available for healthcare providers and pregnant persons or those wishing to become pregnant to assist them in their complex decision-making process, which is a growing public health concern.

Although progress has been made in advancing therapeutic research during pregnancy, pregnant individuals continue to be underrepresented in clinical trials (Yakerson, 2019; Wesley et al., 2021; Sewell et al., 2022). Electronic healthcare claims data have proven to be a feasible and reliable resource for assessing medication safety during pregnancy (Davis et al., 2024; Thurin et al., 2022; Huybrechts et al., 2019; European Network of Centres for Pharmacoepidemiology and Pharmacovigilance (ENCePP), 2023). These real-world data provide detailed information, namely, on medication exposure at various stages of pregnancy (i.e. trimesters and pre- and postnatal periods). Furthermore, with unique anonymized identifiers and the ability to link children to their mothers, these longitudinal population cohorts allow researchers to evaluate the impact of in utero medication exposure on both maternal and child health outcomes (Thurin et al., 2022; Huybrechts et al., 2019). These mother–child cohorts offer a unique opportunity to investigate a broad range of rare exposures and outcomes with precision due to their large sample sizes (Huybrechts et al., 2019; Su et al., 2023). Therefore, the use of secondary data becomes invaluable in studying existing and newly introduced medications for postmarketing surveillance to supplement current evidence for decision-making (Wesley et al., 2021; Huybrechts et al., 2019).

Particularly, claims and other large healthcare utilization databases have been used by several groups, and algorithms to identify pregnancies, gestational age, and linkage to infants have been developed and described before, with several studies conducted using the US claims data sources (Ailes et al., 2016; MacDonald et al., 2019; Weaver et al., 2023; Moll et al., 2021; Sumner et al., 2021; Ailes et al., 2023; Matcho et al., 2018). The Sentinel System in the US (Lyons et al., 2024) and the adoption of the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) across Europe to create a federated network (Thurin et al., 2022) have improved over time. Similar advances have been made in Canada (Bérard and Sheehy, 2014; Bérard et al., 2022). These initiatives have improved the quality of observational research in pregnancy, thus increasing the validity of study findings and reducing the knowledge gap for making informed decisions regarding medication use during pregnancy and outcomes in exposed children.

The Canadian Mother–Child Cohort (CAMCCO) initiative on drug Safety in pregnancy), based on the Quebec Pregnancy Cohort (QPC) model, was established to support the creation of harmonized provincial mother–child cohorts across Canada (Bérard and Sheehy, 2014; Bérard et al., 2022). However, US and Canadian insurance claims data differ mainly in the scope and structure, with US data being fragmented across private insurers and Canadian data being more centralized and population-based due to its publicly funded healthcare system. Leveraging Canadian and US claims data represents a great opportunity to address research questions on the safety and effectiveness of medication use during pregnancy. It is important to note that when referring to the US claims data, since the same data sources change providers and health and insurance policies change over time, definitions must be updated and codes have to be adapted in order to create new cohorts to be used either individually or as part of multi-country studies. Furthermore, demonstrating the reproducibility of previous work across different settings is crucial to ensure the reliability of cohorts used in individual or pooled analyses.

In this regard, our objective is to create the AMerican PREGNANcy Mother–Child CohorT (AM-PREGNANT) using the Merative™ MarketScan® research databases (hereafter referred to as MarketScan®). We adapted previously described algorithms and carefully updated ICD-9 and ICD-10 definitions. To demonstrate that cohort creation is feasible, characteristics are reliable, and the cohort can be used in multi-country studies, we present the baseline characteristics of pregnant persons and their children included in AM-PREGNANT, in addition to the prevalence of medication dispensing in pregnant individuals and their linked children. A descriptive comparison of summary statistics is also presented, referencing US estimates of maternal and gestational age distributions.

We identified 7,991,200 pregnancies from 6,079,647 persons aged 15–45 years between 1 January 2003 and 31 December 2021. The selection process of pregnancies included in AM-PREGNANT is shown in Figure 1. AM-PREGNANT included 4,767,208 pregnancies among 3,626,555 continuously enrolled persons, and the study period was between 1 January 2004 and 31 December 2020 (Figure 1). Deliveries represented 76.9% of pregnancy outcomes, followed by 17.3% of spontaneous abortions and 5.9% of induced/planned abortions (Figure 1).

The linked mother–child cohort was composed of 2,554,964 pregnancies linked to 2,578,990 children. Among them, 1,310,341 were continuously enrolled children linked to singleton pregnancies, while 24,663 children were linked to 13,412 multi-fetal pregnancies (Figure 1).

AM-PREGNANT represents a large and up-to-date mother–child cohort using the MarketScan® database built for contributing to perinatal pharmacoepidemiological research, including its use in multi-country studies. We demonstrated the feasibility of addressing research questions for both pregnant individuals and their children using a large sample size, with pregnancy outcomes, linked children outcomes, and medication dispensing data described over 17 years (2004–2020). Linked and non-linked pregnancies presented slight differences, which should be taken into account when defining research questions, applying methods, and interpreting the results in terms of external validity and its related concepts (i.e. target populations, generalizability, and transportability) (Webster-Clark et al., 2023). The same rationale is also needed when the study population is composed of linked children with and without continuous enrollment.

In AM-PREGNANT, maternal age at delivery was higher among linked pregnancies (32 years) than among non-linked pregnancies (28 years). This pattern aligns with the findings of Weaver et al. (2023), who reported the same age distribution for linked versus non-linked pregnancies. Similarly, Ailes et al. (2023) observed that non-linked pregnancies were more common among younger individuals (15–24 years at delivery), further supporting the association between younger maternal age and non-linkage. In this work, we also demonstrated that these pregnancies involved individuals with higher rates of alcohol and substance use, slightly poorer health—including a higher proportion of infections and depression—a lower prevalence of filled prescription medications, both overall and across time-windows of use. This information is valuable and can be used in future studies to address selection bias that could arise from different mechanisms of selection related to both the exposure and the outcome (Fox et al., 2021). Selection bias due to loss of follow-up is another limitation that could be addressed based on our cohort description. We defined a strict continuous enrollment period, allowing a maximum 30-day gap between enrollment periods for both the mother and the linked children. Ailes et al. (2023), in order to make the criterion less strict, showed that a 60-day gap would not have any impact on linked and non-linked pregnancies. We decided to maintain strict enrollment criteria for both mothers and children in order to minimize the risk of introducing selection bias. In AM-PREGNANT, in the linked-child cohort, children with a gap of enrollment greater than 30 days presented a lower prevalence of medication dispensing (45.6% vs. 62.9%). Therefore, increasing the sample size may also increase the risk of introducing both differential selection and measurement errors.

A common challenge in the creation of pregnancy cohorts using claims and secondary data is the assignment of the LMP (Thurin et al., 2022; Huybrechts et al., 2019). This information is needed to appropriately define the duration of pregnancy and is not recorded in the majority of these data sources. To address this, we used a previously described algorithm and developed a code definition based on the previous studies (MacDonald et al., 2019; Hornbrook et al., 2007; Margulis et al., 2013). Due to the inability to perform a validation study to assess the accuracy of our estimates, we compared AM-PREGNANT gestational ages with CDC WONDER national estimates. We found comparable results in addition to trends indicating the impact of adopting ICD-10 codes for defining gestational age after 2015. The slight variation in the average gestational age between AM-PREGNANT and CDC WONDER of less than a week was consistent with the observation by Ailes et al. (2023). The results presented in this study suggest that AM-PREGNANT closely mirrors the source population, strengthening the suitability of the cohort in the assessment of exposures and outcomes in perinatal pharmacoepidemiological research. Nevertheless, it will be important, depending on the research question, to conduct sensitivity analyses and acknowledge differences between our study sample (AM-PREGNANT) and the source population to account for potential sources of bias and how we might adjust the estimate for better inferences (Lesko et al., 2022). Our estimates of the average maternal data were also comparable to the US national estimates when using the CDC WONDER database (Centers for Disease Control and Prevention (CDC), 2023c). We acknowledge that the use of external data is not ideal, and discrepancies between national estimates and cohort study findings can arise due to various factors, including study design, population differences, and methodological variations. Moreover, healthcare databases may cover specific populations. Specifically, the percentage of pregnancies covered by MarketScan® may vary depending on the specific cohort or time-period under consideration (Packnet, 2023). Nevertheless, previous research showed that private insurance was the most common principal source of payment for deliveries in 2021 (51.7%), providing coverage for more than half of mothers giving birth in the US (Valenzuela and Osterman, 2023). Therefore, despite the need to cautiously interpret results and consider the broader context of the research question, similar distributions of selected variables reassure the feasibility of the algorithm used for building AM-PREGNANT.

For describing AM-PREGNANT, harmonized definitions aligned with CAMCCO were used, considering its potential inclusion in multi-country studies (Bérard and Sheehy, 2014; Bérard et al., 2022). Recommendations and methods for defining covariates are continuously evolving (Ehrenstein et al., 2024), and the methods adopted in this study reflect the best available knowledge to date. Several other characteristics can be defined using MarketScan®, and it is known that variability in health event occurrences when the same algorithms are applied across different data sources may occur (Ehrenstein et al., 2024). Therefore, the algorithms to operationalize variable definitions might change depending on the research question under study. Definitions used in this project are listed in Supplementary Material. When used in multi-country studies, methods for addressing the differences in distribution of key covariates will be adopted (Webster-Clark et al., 2023; Webster-Clark et al., 2024a; Shu et al., 2023; Webster-Clark et al., 2024b).

Overall, the prevalence of covariates we presented was similar to that reported in previous validated linkage studies (MacDonald et al., 2019; Moll et al., 2021; Sumner et al., 2021; Ailes et al., 2023; Matcho et al., 2018; Hornbrook et al., 2007; Andrade et al., 2017). In terms of prescription medication filled, we found that 77% of pregnancies filled prescriptions for at least one medication, which was within the expected results given that non-prescription medications are unlikely to be captured in an insured population (Centers for Disease Control and Prevention (CDC), 2023b; Mitchell et al., 2011). Studies exploring medication use during pregnancy usually exclude vitamins, vaccines, and other medications that could be acquired over the counter (Mitchell et al., 2011). We chose not to exclude these classes of medications since the aim of our study was to provide a comprehensive overview of medication dispensing in this cohort, for both pregnancies and linked children. The observed increase in prescription medication filled during pregnancy is probably related to the fact that vitamins and pregnancy-related medications were frequently prescribed early in pregnancy, potentially at a higher rate than the rate at which other medications were discontinued. We investigated the trend by excluding vitamins and their combinations, but an increase in use was still observed during pregnancy (Supplementary File S3; Table 2). Previous work had also excluded medications used exclusively topically, locally, or intravenously (Mansour et al., 2024), which could have altered the observed trend. The first trimester is a critical phase in terms of exposure to medication due to the organogenesis period. The patterns of medication dispensing in pregnancy observed in AM-PREGNANT provide the possibility of investigating monotherapy, polytherapy, and other risk factors that may be associated with various perinatal outcomes in both mothers and their children. It should be noted that medication dispensing was observed during the second and third trimesters for spontaneous and planned abortions, which were 13% and 19%, respectively. According to the CDC, the vast majority of abortions occur during the first trimester of a pregnancy (Diamant et al., 2024; Kortsmit et al., 2023). Specifically, in 2021, 93% of abortions occurred during the first trimester (at or before 13 weeks of gestation). An additional 6% occurred between 14 and 20 weeks of pregnancy, and approximately 1% were performed at 21 weeks or more of gestation (Diamant et al., 2024; Kortsmit et al., 2023). Therefore, prescription medication filling observed in the third trimester for the non-live birth outcomes represents prescriptions that overlap with the beginning of the time-window and may reflect the capture of abortions occurring in later gestational ages.

In addition to the prevalence of medication dispensing, we demonstrated the top 10 most frequently filled classes of medications during pregnancy. Anti-infectives were the most prevalent class filled during pregnancy, and this pattern is similar to that of previous pregnancy cohort studies in both Canada and US (Bérard and Sheehy, 2014; Bérard et al., 2022; Mansour et al., 2024). The most prevalent class of prescription medication filled was also anti-infectives for linked children. AM-PREGNANT, therefore, represents an opportunity to assess several outcomes in both mothers and children using these medications.

The overall prevalence of preterm birth in the linked-child cohort was slightly higher than the US national estimates (11.9% vs. 10.1%, 2018–2020) (Centers for Disease Control and Prevention (CDC), Division of Reproductive Health, 2022). This overestimation is expected, as the estimate refers to children who were linked to a pregnancy with available child information. When assessing preterm birth prevalence using a definition in which the denominator included all preterm births among pregnancies ending in a live birth, regardless of linkage to a child, a prevalence of 10.2% was found, which is similar to the US national estimates. We compared this estimate with other Canadian mother–child cohorts, and a higher prevalence was found compared to Canadian estimates (Supplementary File S3; Table 3). A greater proportion of extremely and very preterm birth was observed among children without continuous enrollment. In the US, there is an issue related to the healthcare system in which multiple insurers compete for employment-based insurance contracts and individual enrollees. As a result, turnover among commercial insurance enrollees reflects not only individuals switching insurers but also changes in employment and employers switching insurance providers (MacDonald et al., 2019; Fang et al., 2022). Consequently, dependent enrollees—such as children—are often missed and not properly followed over time, leading to discrepancies when comparing enrolled and non-enrolled populations. The differences observed between AM-PREGNANT and Canadian mother–child cohorts are, therefore, expected, as we are comparing an employed and insured population in the US with a publicly covered population in Canada. The weighting of the cost-benefit to relaxing continuous enrollment among linked children will need to be evaluated for future studies. Other characteristics do not suggest that a sicker population is being excluded, but assumptions would be needed in case these characteristics become outcomes of interest. When using claims US data, mortality data are not available; therefore, it is not known whether these children died or not, and perhaps death was one of the reason for not being continuously enrolled.

For LBW, estimates observed in AM-PREGNANT were half of the US national estimates, 4.7% vs. 8.2% in 2020 (Osterman et al., 2023). It is known that weight at birth is not available in claims databases, including MarketScan® data, where estimates rely exclusively on recorded diagnosis codes. Performance of these codes has been validated by Chomistek et al. (2023) in the Optum Research Database with PPVs close to 100%. Kasman et al. assessed LBW using ICD-9, ICD-10, and DRG codes limited to extreme immaturity and extremely low birth weight. In their study, a 4.7% prevalence of LBW was estimated (2007–2016) (Kasman et al., 2020). We adopted the Agency for Healthcare Improvement (AHRQ) definition, which comprises a broader range of ICD-9 and ICD-10 codes used for assessing pediatric quality indicators in the US healthcare system (Agency for Healthcare Research and Quality (AHRQ), 2023b), and searched for codes in both mothers’ and children’s files. We still have a possible underestimation of this measure. These results reinforce the need for improving algorithms related to infant growth when using administrative databases, as this measure is poorly captured.

Other linked-children characteristics were assessed in AM-PREGNANT, including medication dispensing. Follow-up time of children when using insurance data is limited by continuous enrollment (MacDonald et al., 2019). AM-PREGNANT average follow-up time for continuously enrolled children was 3.3 years, which is a great opportunity for studying acute conditions. When studying long-term outcomes, some limitations may arise.

When evaluating the prevalence of major congenital malformations adopting the harmonized definition used in previous Canadian studies (definition A), higher estimates were observed (13.1% of live births, Table 2) than the US estimates (approximately 4% of live births) (Centers for Disease Control and Prevention CDC, 2008). The estimate decreased to 4% when requiring at least two diagnosis codes recorded on two different days. Huybrechts et al. demonstrated that when major congenital malformations are defined based on a single diagnostic code, the prevalence of malformations exceeds 10%, approximately three times higher than that reported in the US population (approximately 4%) (Huybrechts et al., 2019; Centers for Disease Control and Prevention (CDC), 2008), which is the same as the results we obtained. Nevertheless, we decided to demonstrate the one-code definition in AM-PREGNANT, given that we aimed to use harmonized definitions with QPC and CAMCCO initiative (Bérard and Sheehy, 2014; Bérard et al., 2022). Nevertheless, in future studies using AM-PREGNANT, mainly when adopting a common data model in multi-country studies where major congenital malformations are the outcomes of interest, at least two diagnosis codes, or one diagnosis and related procedure, as previously described (Huybrechts et al., 2019), should be adopted when using the US cohort.

AM-PREGNANT has several strengths. We built upon previous algorithms and assembled the largest and most up-to-date mother–child cohort from a US representative sample of patients with employer-provided health insurance. Inclusion criteria were as follows: 1) continuous enrollment in the insurance plan for a given baseline period; 2) full coverage, including prescription benefits; 3) an appropriate enrollment type (e.g., fee-for-service or capitated plans, provided that they do not underreport encounter claims); and 3) linkage to infants for pregnancies ending in a live birth (Huybrechts et al., 2019). We, therefore, obtained baseline characteristics, including pregnancy outcomes, similar to the US national estimates, demonstrating the feasibility and reliability of this cohort.

Limitations of AM-PREGNANT are similar to those previously reported when constructing pregnancy-linked cohorts using claims data sources. LMP estimation continues to be one of the main concerns when using claims data. The timing of medication exposure during pregnancy is critical when evaluating adverse effects on infant development. Thus, erroneous estimates of LMP will misclassify exposure time. To reduce the risk of inaccurately estimating the LMP, we used previously established and validated codes (Hornbrook et al., 2007; Margulis et al., 2013; Margulis et al., 2023), and as a result, we obtained gestational age estimates similar to the national estimates. The improved specificity of ICD-10 codes enhanced gestational age ascertainment in the later years of the study; however, their limited availability earlier in the study period may have led to incomplete or inconsistent gestational age data. To assess the impact of this transition, we reviewed the distribution of outcomes by gestational age across both coding eras. In AM-PREGNANT, 81% of deliveries (live births, stillbirths, and mixed births) were captured using ICD-10 codes post-2015, compared to 76% using ICD-9 codes before the transition. The prevalence of spontaneous abortions remained consistent at approximately 17% in both periods, suggesting stable classification for this outcome. However, we observed a lower prevalence of elective abortions post-2015 (3%) compared to that in the pre-2015 period (7%), a trend also reported by Sa et al. (2020). This discrepancy may reflect underreporting in claims data, particularly given that a significant proportion of individuals seeking abortion care pay out-of-pocket due to the lack of insurance coverage or privacy concerns. These findings underscore the importance of considering coding system transitions when interpreting trends over time.

Another challenge is related to the identification of pregnancy outcomes other than live births. The adoption of a hierarchical algorithm, similar to those used by MacDonald et al. (2019), Ailes et al. (2016), and Ailes et al. (2023), ensured the identification of pregnancy outcomes similar to the US population for deliveries (i.e., live births, stillbirths, and mixed births) and spontaneous abortions (Benson et al., 2023). The identification of induced/planned abortions appeared to be underestimated; however, this has been an issue even when assessing surveillance data (Kortsmit et al., 2022). Administrative claims data offer various advantages for pharmacoepidemiological research, but bias due to the misclassification of exposure, outcome, and covariates is usually a concern. In terms of exposure to medication, despite acknowledging the difference between filling a prescription for a medication and taking the medication, especially given potential discontinuation once pregnancy is recognized, it is well-known that a strength of using administrative claims to evaluate medication effects is the comprehensive data on prescription fills, which are more reliable than self-reported use or physician-ordered prescriptions. In addition, lack of capture for over-the-counter medications may affect both maternal and child exposures, leading to misclassification in classes of medications used for treating common conditions of pregnancy, such as gastrointestinal diseases (e.g., antacids), fever and pain (e.g., ibuprofen, acetaminophen, and other anti-inflammatory medications), cold, flu, allergies (e.g., anti-histamines and decongestants), and other medications that can be purchased out-of-pocket by the user without the need for a prescription or those that are not covered by insurance (Funk and Landi, 2014). Another limitation is the lack of information on medications dispensed during hospital stays. Although pharmacy claims data may not always accurately reflect actual medication exposure, methods to address such misclassification have been described (Funk and Landi, 2014; Walraven, 2018) and are increasingly used (Brown et al., 2024; Petersen et al., 2021). However, because MarketScan® data do not capture inpatient medication dispensing, there is potential for exposure misclassification and confounding by indication, which cannot be fully addressed in the analysis. As such, research questions that rely specifically on in-hospital medication data should be avoided. Regarding generalizability, the MarketScan® population likely includes healthier pregnant individuals with private insurance coverage, which may lead to more favorable outcomes. Therefore, findings are generalizable to the privately insured population but may not be applicable to individuals covered by Medicaid or other public insurance programs. Johnson et al. (2023) demonstrated that 42.1% of all births in the US in 2019 were covered by Medicaid, indicating that nearly half of the population may experience higher rates of inadequate or delayed prenatal care—factors known to be associated with adverse infant outcomes. This is a methodological concern that will be taken into account when conducting multi-country studies using AM-PREGNANT. Although this cohort is representative of the US pregnant population, these differences in terms of disparities are an opportunity to triangulate results using data from different settings. The practice of strengthening causal inferences by integrating results from multiple approaches –each with different and largely unrelated sources of potential bias, was employed by Lawlor et al. (2016), including the use of different sources. Methods for pooling results considering the external validity when using distributed data networks were described already (Webster-Clark et al., 2023), and these approaches will be taken into account for future studies.

AM-PREGNANT represents a reliable and promising resource for addressing queries and research questions on the safety and effectiveness related to medication use during pregnancy. By aggregating harmonized data from multiple populations, namely, the US and Canada (with CAMCCO DATA), we can increase statistical power to detect associations, which would, in turn, allow us to answer novel research questions as they pertain to rare exposures and outcomes, as is generally the case in perinatal pharmacoepidemiology. By following successful initiatives such as the ConcePTION Common Data Model (Thurin et al., 2022), in addition to applying appropriate methods when using different data sources from different countries and contexts, we can triangulate results to strengthen the evidence for answering causal questions (Lawlor et al., 2016). AM-PREGNANT will provide data on privately insured subjects, allowing comparisons between Canada and the US, as well as across Canadian provinces—an aspect rarely explored in other perinatal epidemiology studies. Our cohort will add value to the research program by identifying and quantifying differences in prescribing practices and medication dispensing during pregnancy and childhood across Canada, its provinces, and the US, and it will help in identifying the risk profiles of prescription medications filled during pregnancy and in the pediatric populations.

It is important to note that the prevalence of chronic conditions such as depression and hyperglycemia varies significantly across regions and populations globally. As such, trends in prescription medication use may reflect not only true changes in disease burden but also differences in healthcare systems, diagnostic practices, and access to care. Future studies should explore regional and international variations to better understand the broader applicability of our findings in other contexts, such as the evaluation of medications and their outcomes in pregnancy and children living in developing countries.

We have assembled AM-PREGNANT, which represents an important resource for the assessment of prescription medication safety for both mothers and their children. The large numbers of individuals included for both mothers and the linked mother–child cohort, even when using a conservative continuous enrollment requirement, provide an excellent resource for assessing rare exposures and outcomes. Preterm birth rates, mean maternal age, and gestational ages were comparable to US population estimates, reassuring the validity of this cohort in terms of feasibility, reliability, and generalizability. AM-PREGNANT represents an additional data source to be incorporated when performing multi-country studies in the field of perinatal pharmacoepidemiology.