A technical framework for using electronic medical records to create and maintain a cross-sectional research cohort of pregnant people in the Gulf South Authors: Elizabeth Crull, Anna V Legrand, Daniele Farrisi, Kristie Bardell, Suzan L Carmichael, Megan Jacobs, Maeve E Wallace, Thomas W Carton, Elizabeth Nauman Published: 2026 ABSTRACT AbstractObjectiveTo develop a cohort identification algorithm that captures pregnancy episodes in electronic health record data and characterize demographic and clinical data for eligible patients.Materials and MethodsWe used data extracted from EHRs and conformed to the PCORnet Common Data Model between July 2016 and June 2023 for this study.ResultsOur phenotype identified 74 480 pregnant people associated with 95 503 pregnancies between July 2016 to June 2023. The cohort was 44.4% non-Hispanic White, 37.5% non-Hispanic Black, and 11.4% Hispanic. The median maternal age in the cohort was 29. Most pregnancy episodes (80.1%) resulted in live birth. The rate of deliveries with a severe maternal morbidity (SMM) was 179 per 10 000 deliveries, 51.1% of which were associated with a non-Hispanic Black person (p < .01). The preterm birth rate was 12.8%, and the percentage of deliveries with timely postpartum care was 73.9%. The rate of deliveries where the person had a diagnosis of gestational diabetes was 9.1%, and the rate for hypertensive disorders was 24.2%.DiscussionClinical information about a pregnancy episode can be challenging to visualize using EHR data alone, even as it is increasingly critical to be able to do so for public health and policy research. The methods we describe result in a database of pregnancy episodes that distills complex data into a useful format for pragmatic researchers.ConclusionWe established a replicable method of identifying pregnancy episodes using electronic health record data using coding ontologies common in clinical practice and research. We demonstrated proof of concept by running simple maternal health metrics on the cohort. Background and significance There is an urgent need to better understand troubling maternal health trends in the United States, especially in states with stark disparities along racial and socioeconomic lines and lack of access to the full spectrum of obstetric and gynecologic healthcare due to political shifts.1 A 2024 report released by the Centers for Disease Control and Prevention described a nationwide decrease in gestational age at birth, with preterm and early-term birth rates rising by 12% and 20%, respectively, between 2014 and 2022, especially among Black people, who experienced higher rates of births delivered at less than full term compared to White and Hispanic people.2 Recent research indicates that Black people have higher odds of developing preeclampsia during pregnancy and are disproportionately affected by gestational hypertension3,4 and that the rate of gestational diabetes is also rising.5 For example, a 2021 study showed an increase in gestational diabetes at first live birth went from 47.6 to 63.5 per 1000 live births between 2011 and 2019, especially among Asian Indian and Puerto Rican ethnicity.6 A qualitative study conducted in 2021 shows that Black and Hispanic women are more likely to report emotional distress and postpartum complications following a childbirth complication.7 This is borne out in quantitative studies that show that rates of severe maternal morbidity (SMM) is higher among Black and Hispanic women.8,9 SMMs are serious complications of delivery such as sepsis, shock, or organ failure that can be fatal if they are not quickly diagnosed and treated. Access to timely, high-quality, and standardized maternal health data is critical for facilitating rapid insights into drivers of poor health outcomes such that they can be identified, and racial and ethnic disparities can be narrowed. The Office of the Secretary Patient-Centered Outcomes Research Trust Fund identified three key challenges in leveraging maternal health data for clinical research and public health surveillance. These challenges include incomplete and inconsistent collection of race, ethnicity, and other self-reported demographic data, difficulties with obtaining and linking data, and the lack of focus on factors outside of traditional obstetric complication measures, such as social determinants of health, behavioral health and substance use disorder, and chronic health conditions.10 By virtue of the variability of care environments and possible outcomes, as well as the temporal factors inherent to pregnancy, it is an epidemiological challenge to capture the entire clinical snapshot of a specific pregnancy episode. It is especially challenging to link together pre-pregnancy, prenatal, delivery, and postpartum data in the absence of robust longitudinal surveillance systems.11,12 Pregnant people may receive care in multiple clinics or even different health systems during the time surrounding a single pregnancy, depending on their insurance status, which is subject to fluctuation based on financial circumstances or other life factors. The advent of electronic health record (EHR) data, especially that which have been transformed into a standard data model, has led to innovations in disease surveillance and population health research.13–16 A computable phenotype is an algorithm for identifying a cohort of patients meeting certain criteria within an EHR and considers data found in multiple views of a single patient such as diagnosis, encounter, procedure, and medication data.17,18 For the purposes of evaluating maternal health trends, computable phenotypes are a valuable tool to define a cohort of pregnant people and capture clinical data from before, during, and after their pregnancy, rather than just in the context of isolated events like postpartum readmission to a hospital or pregnancy-associated death. In 2023, the Patient Centered Clinical Research Network (PCORnet) Maternal Morbidity and Mortality Workgroup developed several recommendations to enhance the capacity for comparative effectiveness research involving pregnant people using the PCORnet Common Data model infrastructure. They identified a need to accurately identify cohorts of pregnant and postpartum individuals for inclusion in clinical studies of pregnancy conditions like hypertensive disorders, gestational diabetes, and depression.19 While access to a patient’s EHR at the point of care enables clinicians to navigate between aforementioned views, data extracts made available to researchers through HIPAA-compliant regulatory agreements inherently lack such end-user latitude. Read-only, static data exports can be large, contain non-intuitive clinical vocabularies, and require extensive processing before analyses can begin. The methods described in this paper serve as a response to this knowledge gap and a starting point for researchers working with EHR data extracts to run preliminary cohort identification queries in a preparatory-to-research capacity. Objective Our primary goal was to define a cohort of pregnant people and their associated pregnancy episodes using data transformed from electronic health records to a Common Data Model. Our secondary goal, using a single analytic extract, was to generate descriptive demographic and clinical distribution tables to demonstrate the utility of the database. Materials and methods Data source We used data that were conformed to the Patient Centered Clinical Research Network (PCORnet)’s Common Data Model (CDM). PCORnet is a national collaboration of clinical research networks that are charged with extracting EHR data from local, engaged health systems; transforming it in accordance with the PCORnet CDM technical specifications; and storing the data as static datasets ready for research. Data from participating systems are refreshed quarterly and undergo routine, rigorous quality audits.20,21 The Research Action for Health Network (REACHnet) clinical research network is a node of PCORnet that is comprised of eight major health systems across California, Louisiana, Mississippi, and Texas. Two REACHnet-member health systems comprising 14 hospitals with labor and delivery units in Louisiana and Mississippi contributed EHR data to this project. EHR data extracts are HIPAA-compliant, patient-level limited data sets. While they do not contain direct identifiers such as patient name, social security number, phone number, or residential address, they contain encounter dates and birth dates, which are considered protected health information. Thus, the data cannot be classified as fully de-identified. Iterative development and testing of the cohort identification algorithm and subsequent summary statistics were performed using Statistical Analysis Software (SAS) Version 9.4 between July 2022 and September 2023. Institutional Review Board approval was obtained on June 24, 2022. Pursuant to our goal of transparency and knowledge sharing, the code lists that serve as the backbone of the programs we developed are available in Supplements 1–3, available as supplementary material [JAMIA Open] online. The pregnancy database is maintained as a longitudinal resource and refreshed on a quarterly basis to incorporate newly accrued clinical data. Each refresh follows a standardized extract–transform–load (ETL) process that applies consistent inclusion criteria, pregnancy identification algorithms, and variable definitions across refresh cycles. The analyses presented in this manuscript were conducted using a single, static extract and were intended as a one-time descriptive exercise to characterize the demographic and clinical composition of the database and to demonstrate its suitability for future research. Subsequent quarterly refreshes update the underlying database but do not alter the results reported here. Eligibility criteria People were included in the cohort if they had a pregnancy episode (PE) with one of the following outcomes between July 1, 2016 and June 30, 2023: live birth, stillbirth, ectopic pregnancy, trophoblastic disease/hydatidiform mole, or spontaneous abortion (miscarriage). People had to be aged 12 to 55 at the time of the outcome for the PE to be included. We selected a broader age range than is typical to capture rare but documented pregnancies among younger adolescents and older individuals, and to align with commonly cited bounds for menarche and menopause. Outcomes were identified by querying for relevant diagnostic [International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM)] and procedure [Current Procedural Terminology (CPT); International Classification of Diseases, Tenth Revision, Procedure Coding System (ICD-10-PCS)] codes and assigning the encounter a pregnancy outcome and index date. Index date was defined as the date of the qualifying diagnosis or procedure. Our phenotype did not include PEs with an indication of elective abortion (or related complications) to prevent patient re-identification in demographic tabulations and because participating hospitals did not offer elective abortion procedures while the procedure was still legal in these states. Finally, we did not consider clinical episodes that indicated pregnancy but did not have a defined outcome, e.g., a positive pregnancy lab result not associated with a subsequent loss or delivery. The full list of inclusion codes is available in Supplement 1, available as supplementary material [JAMIA Open] online. Pregnancy interval spacing algorithm To minimize duplicate counting of a single PE, we employed a rules-based hierarchical model to prioritize placement of PEs on an individual patient timeline during the measurement period and to establish spacing between multiple episodes. For example, we wanted to ensure that a follow-up appointment related to an original PE was not counted as a new PE (e.g., a miscarriage that occurred on January 10, 2021 with a follow-up on January 30, 2021). Live births (LB) were pulled in first with temporal spacing established on either end of the delivery date. Then, stillbirths and pregnancies with abortive outcomes were allowed in. All PEs needed to follow the time-based rules (adapted from previously published work22,23 and set forth in in Supplement 4, available as supplementary material [JAMIA Open] online) to be included. For example, an LB immediately followed by another LB needed to be spaced apart by a minimum of 182 days (26 weeks). A spontaneous abortion (SA) followed by an LB needed to be spaced apart by 168 days (24 weeks). Gestational age calculation ICD-10 contains a branch of codes under “Persons encountering health services in circumstances related to reproduction” that are used to indicate the weeks of gestation on the maternal record, when known. These codes contain a prefix “Z3A” followed by a dot separator and an integer that corresponds to the number of weeks of gestation in most cases (there are some exceptions, which will be described), e.g., Z3A.38 means 38 weeks gestation of pregnancy. These codes do not provide granular detail on the number of days of gestation, thus there is no way to differentiate between 38w1d and 38w6d of gestation, for example. After establishing PEs that met inclusion criteria, all PEs were scanned for presence of a Z3A.* code on the index date. If present, the integer was extracted and assigned as the gestational age in weeks. If no Z3A code was present or if the Z3A code present was non-specific, a gestational age estimate was assigned to the pregnancy based on the pregnancy outcome. Z3A.01 and Z3A.49 codes were automatically assigned a gestational age of six and 43 weeks, respectively. Supplement 5, available as supplementary material [JAMIA Open] online contains the logic we used to calculate or impute gestational age for each pregnancy. PEs were then removed if they had an implausible gestational age for the assigned outcome, defined in Supplement 6, available as supplementary material [JAMIA Open] online. An estimated pregnancy start date was calculated for each PE by converting the assigned gestational age (weeks) to days and subtracting that number of days from the index date. A full list of gestational age codes is available in Supplement 2, available as supplementary material [JAMIA Open] online. The practice of applying Z3A codes to pregnancy encounters was common at both data-contributing health systems, with low missingness described later in the Results section. Cohort characterization and selected metrics for evaluation We described the final analytic cohort of patients and pregnancies in the cohort by providing the distribution of maternal race and ethnicity, gravidity, maternal age in years, gestational age in weeks, pregnancy outcome, insurance type, and the state of residence. Race and ethnicity values in the PCORnet Common Data Model are based on guidance from the U.S. Office of Management and Budget (OMB) and is compatible with the 2010 U.S. Census. Patients with missing race were retained for the purpose of transparency but, in an effort to facilitate interpretation of results, are not included in data visualizations comparing outcomes across race and ethnicity groups. We chose six common indicators of maternal health to program, execute, and compare with published rates. The Centers for Disease Control and Prevention (CDC) maintains a list of indicators of severe maternal morbidity (SMM). There are 21 indicators in total, 16 diagnoses and five procedures.24 Research shows that a flag for a blood products transfusion procedure has poor positive predictive value for other severe maternal morbidities (i.e., it does not always correlate to an emergency situation),25 so SMM can be presented two ways: with the procedure considered, and with the procedure not considered. We present both in this paper. We defined SMM as at least one diagnosis or procedure for an SMM within six weeks of delivery. Only pregnancies resulting in delivery were considered.24,25 We also looked at preterm birth, defined as delivery before 37 weeks of gestation.26 We flagged deliveries as being preterm if their gestational age was calculated or imputed as less than 37 weeks (with a live birth or stillbirth outcome) or if they had an ICD-10 diagnosis code indicating preterm labor and delivery within seven days of delivery date. We then chose two common diagnoses associated with pregnancy. Gestational diabetes (GDM) was defined as at least one record of an ICD-10 code with the prefix “O24” indicating diabetes mellitus in pregnancy, childbirth, and the puerperium. Hypertension during pregnancy was defined as at least one record after 20 weeks gestation of an ICD-10 code with the prefix “O1” indicating edema, proteinuria, and hypertensive disorders in pregnancy, childbirth, and the puerperium.27 To enhance capture of pregnancy-associated clinical events rather than pre-existing chronic conditions, ICD-10 codes that denote chronic diabetes (prefix “E11”) and chronic hypertension (prefix “I1”) were not considered. Finally, we selected a healthcare utilization metric: timely postpartum care, defined by the National Committee for Quality Assurance (NCQA) as an appropriately timed postpartum visit during the puerperium.28 Deliveries were flagged as having timely postpartum care if they had an outpatient encounter for the following: cervical cytology (i.e., pap smear), a postpartum visit, or a postpartum bundled service. The procedure or lab code needed to have been present between seven and 84 days of delivery. All six indicators were calculated as a rate, with the numerator being the number of events and the denominator being the number of deliveries meeting criteria during a defined period. Because they are not as common and for clarity, SMM and non-transfusion SMM events are presented as counts per 10,000 deliveries. The remaining metrics (preterm birth, hypertensive disorders of pregnancy, gestational diabetes, and postpartum care) are presented as a percentage. To enable accurate flagging of events, the timeframe for each metric was established according to how much follow-up period was required. The denominator for preterm birth, hypertensive disorders of pregnancy, and gestational diabetes was deliveries between July 1, 2016 and June 30, 2023 because no data follow-up period is needed to flag for these events. The denominator for SMM and non-transfusion SMM events was deliveries between July 1, 2016 and April 30, 2023 to allow for six weeks of data follow-up. The denominator for postpartum care was deliveries between July 1, 2016 and March 31, 2023 to allow for 12 weeks of data follow-up. A full list of clinical codes used to define these indicators can be found in Supplement 3, available as supplementary material [JAMIA Open] online. All rates were stratified by maternal age and race and ethnicity. Because clinical event identification was categorized on the binary as absence or presence of a clinical code, it was assumed that absence of the qualifying code meant that the event did not occur. In other words, there was no categorization into missingness. Results Our phenotype identified 151 282 unique encounters (i.e., unique combination of patient and admit date) with an eligible pregnancy inclusion code prior to de-duplication and application of maternal age, time-spacing, and gestational age plausibility filters. Figure 1 diagrams attrition of the cohort as the stepwise filters were applied. 55 214 records were dropped after the PE spacing algorithm (Supplement 4, available as supplementary material [JAMIA Open] online) was applied to consolidate duplicate qualifying events. 87 PEs were then dropped because the maternal age was <12 or >55 at index date. Pregnancy cohort definition and attrition, by outcome type. We scanned the remaining 95 981 PEs for a Z3A code to enable gestational age estimation (Figure 2). 81 321 PEs had a specific Z3A code available at index date (84.7%). Gestational age was imputed for the remaining 14 660 PEs based on outcome alone (N = 12 080), presence of a Z3A.49 code indicating greater than 42 weeks of gestation (N = 39), or presence of a Z3A.01 code, indicating less than 6 weeks of gestation (N = 2541). As a final step, our plausibility filter removed 478 PEs where the calculated gestational age was not biologically consistent with the pregnancy outcome. The final analytic cohort consisted of 74 480 pregnant people associated with 95,503 pregnancy episodes with an outcome between July 1, 2016 and June 30, 2023. Gestational age calculation flow diagram. Gestational age calculation flow. PE, pregnancy episode; Z3A, ICD10 code prefix associated with gestational age in weeks; LB, live birth; SB, stillbirth; SA, spontaneous abortion; ECT, ectopic pregnancy; TRO, trophoblastic disease; GA, gestational age. Summary statistics for the final analytic cohort (pregnant people and PEs) are provided in Table 1. The cohort was 44.4% non-Hispanic White, 37.5% non-Hispanic Black, 11.4% Hispanic, 2.4% Asian American or Pacific Islander, and less than one percent American Indian or Alaska Native. Race/ethnicity data were missing for four percent of the cohort. Most people in the cohort (77.9%) had no more than one pregnancy during the measurement period. The median gestational age for all pregnancies was 38 weeks (interquartile range: 35-39; range: 6-43). The median maternal age in the cohort was 29 (interquartile range: 24-33; range: 12-55). Pregnancy outcomes were 80.1% live birth, 17.2% spontaneous abortion, 2.0% ectopic pregnancy, less than one percent stillbirth, and less than one percent hydatidiform mole. The clinical encounters associated with the defining pregnancy outcome were mostly covered by Medicaid (56.2%), and 35.0% were covered by private insurance. Payer data were missing for 8.8% of the cohort. Table 1.Characteristics of patients and pregnancies identified by the pregnancy cohort identification algorithm.CharacteristicNo. (%)CharacteristicNo. (%) N (patients)74 480 N (pregnancy episodes)95 503Race/ethnicityMaternal age, years American Indian or Alaska Native257 (<1) Mean (SD)28.9 (6.0) Asian American, Native Hawaiian or Pacific Islander1768 (2.4) Median (IQR) [Range]29 (24-33) [12-55] Hispanic8485 (11.4)Gestational age, weeks Non-Hispanic Black27 908 (37.5) Mean (SD)32.6 (11.5) Non-Hispanic White33 084 (44.4) Median (IQR) [Range]38 (35-39) [6-43] Missing2978 (4.0)Pregnancy outcomeGravidity* Live birth76 491 (80.1) 158 042 (77.9) Stillbirth610 (<1) 212 872 (17.3) Spontaneous abortion16 412 (17.2) 3+3566 (4.8) Ectopic1917 (2.0) Trophoblastic disease73 (<1)Insurance type Medicaid/government53 667 (56.2) Private33 445 (35.0) Other or missing8391 (8.8)State Louisiana90 701 (95.1) Mississippi3614 (3.8) Other1110 (1.2) Missing78 (<1)*During measurement period (July 1, 2016—June 30, 2023). Characteristics of patients and pregnancies identified by the pregnancy cohort identification algorithm. During measurement period (July 1, 2016—June 30, 2023). We selected and programmed six metrics to evaluate the pregnancy cohort and provide an initial overview of clinical indicators of maternal health (Table 2). For the six-year period, the SMM rate was 179.3 per 10 000 deliveries and 169.7 per 10 000 deliveries when adjusted to exclude blood products transfusion. The most common SMMs were sepsis (37 per 10 000 deliveries), pulmonary edema/acute heart failure (34 per 10 000 deliveries), and acute renal failure (27 per 10 000 deliveries) (Figure 3). Rates of severe maternal morbidity among deliveries in the cohort, by clinical event. Table 2.Select maternal health indicators by maternal age and race/ethnicitySevere maternal morbidityNon-transfusion severe maternal morbidityaPreterm birthHypertensive disorders of pregnancybGestational diabetesPostpartum care Nc 73 95473 95477 10177 10177 10173 942 No. of flagged deliveries (%) 1326 (1.8)1255 (1.7)9853 (12.8)18 649 (24.2)7013 (9.1)54 646 (73.9)Maternal age, yearsd Mean (SD)29.5 (6.2)29.6 (6.2)29.2 (6.1)29.6 (6.0)31.3 (5.6)29.1 (5.7) Median (IQR) [Range]30 (25-34) [14-55]30 (25-34) [14-55]29 (25-34) [12-50]30 (25-34) [12-50]31 (27-35) [15-52]29 (25-33) [12-53]Race/ethnicity, n (%)e American Indian or Alaska Native<11 (<1)<11 (<1)28 (<1)72 (<1)29 (<1)215 (<1)Asian American or Pacific Islander32 (2.4)28 (2.2)199 (2.0)229 (1.2)367 (5.2)1492 (2.7) Hispanic111 (8.4)100 (8.0)979 (9.9)1393 (7.5)912 (13.0)6041 (11.1) Non-Hispanic Black677 (51.1)656 (52.3)4432 (45.0)8710 (46.7)2336 (33.3)19 056 (34.9) Non-Hispanic White458 (34.5)433 (34.5)3913 (39.7)7781 (41.7)3053 (43.5)26 213 (48.0) Missing44 (3.3)35 (2.8)302 (3.0)464 (2.5)316 (4.5)1629 (3.0)a“Non-transfusion” means that blood products transfusion is not considered when flagging SMM events.bHypertensive disorders of pregnancy include: pre-existing hypertension complicating pregnancy, with or without preclampsia; pregnancy-induced edema and proteinuria with and without hypertension; pre-eclampsia and eclampsia; and unspecified maternal hypertension.cDenominator varied by indicator, depending on measurement period definition.dAge calculated at date of delivery, not date of flagged event; cutoff age 12-55.eParenthetical percentages are proportion of all people in the numerator, i.e., column percentage. Select maternal health indicators by maternal age and race/ethnicity “Non-transfusion” means that blood products transfusion is not considered when flagging SMM events. Hypertensive disorders of pregnancy include: pre-existing hypertension complicating pregnancy, with or without preclampsia; pregnancy-induced edema and proteinuria with and without hypertension; pre-eclampsia and eclampsia; and unspecified maternal hypertension. Denominator varied by indicator, depending on measurement period definition. Age calculated at date of delivery, not date of flagged event; cutoff age 12-55. Parenthetical percentages are proportion of all people in the numerator, i.e., column percentage. We stratified results by race and ethnicity and determined that statistically significant disparities existed for six maternal health indicators (Tables 2 and 3, Figures 4 and 5). To calculate race and ethnicity-level rates, we divided the number of events among a given race category over the total number of people in that race category. Deliveries among non-Hispanic Black people had the highest rate of preterm birth (15.9% compared to 12.8% overall), hypertensive disorders (31.2% compared to 24.2% overall), and SMM (251 per 10 000 deliveries, any SMM event; 243 per 1000 deliveries, any SMM event except blood products transfusion). Similarly, 51.1% of all SMM events in the cohort were experienced by non-Hispanic Black people, despite comprising only 37.5% of the cohort. Non-Hispanic Black people also had the lowest rate of timely postpartum care (70.7% compared to 73.9% overall). The rate of any SMM event was lowest among Hispanic people (133 per 10 000 deliveries); the rate of any SMM event except transfusion was lowest among American Indian or Alaska Native people, though the subgroup-specific rate was masked due to low cell counts. Preterm birth rates were lowest among American Indian or Alaska Native people (10.2%). The rate of hypertensive disorders was lowest among Asian American or Pacific Islander people (12.3%), and the lowest rate of gestational diabetes was observed in non-Hispanic Black people (8.4%). Chi-square tests to measure the difference between observed and expected frequencies had p < .001 for all six indicators when tabulated by race category (Table 3). Racial disparities for select maternal health indicators—preterm birth, hypertensive disorders of pregnancy, gestational diabetes, and timely postpartum care (July 2016—June 2023). Racial disparities for select maternal health indicators—severe maternal morbidity (July 2016—June 2023). Table 3.Select maternal health indicators, stratified by race and ethnicity.DenominatorFlagged pregnancy countsabcSMM1aSMM2aPretermbHTNbGDMbPostpartumcAmerican Indian or Alaska Native (n)264275264<11*<11*287229215 %<4.2<4.2§10.226.210.581.4Asian American or Pacific Islander (n)17861859178632281992293671,492 %1.81.610.712.319.783.5Hispanic (n)8367868283651111009791,3939126,041 %1.31.211.31610.572.2Non-Hispanic Black (n)26 96627 94226 96467765644328710233619 056 %2.52.415.9168.470.7Non-Hispanic White (n)33 83835 49633 83045843339137781305326 213 %1.41.31121.98.677.5Missing (n)27332847273344353024643161,629 %1.61.310.616.311.159.6Total73 95477 10173 94213261255985318 649701354 646 %1.81.712.824.29.173.9 Chi-square (p-value)<.001<.001<.001<.001<.001<.001Acronyms: SMM1, Severe Maternal Morbidity; SMM2, Severe Maternal Morbidity, Non-Transfusion, HTN, hypertension; GDM, gestational diabetes mellitus. Note: There are three different denominators for these indicators, the timeframes of which are listed below. In summary, measurement periods needed to be adjusted according to the indicator of interest to allow for proper look-forward.aDenominator = Deliveries from July 1, 2016 to April 30, 2023.bDenominator = Deliveries from July 1, 2016 to June 30, 2023.cDenominator = Deliveries from July 1, 2016 to March 31, 2023.*Exact value suppressed to protect patient confidentiality.§Indicates the lowest rate compared to other subgroups; exact value suppressed to protect patient confidentiality. Select maternal health indicators, stratified by race and ethnicity. Acronyms: SMM1, Severe Maternal Morbidity; SMM2, Severe Maternal Morbidity, Non-Transfusion, HTN, hypertension; GDM, gestational diabetes mellitus. Note: There are three different denominators for these indicators, the timeframes of which are listed below. In summary, measurement periods needed to be adjusted according to the indicator of interest to allow for proper look-forward. Denominator = Deliveries from July 1, 2016 to April 30, 2023. Denominator = Deliveries from July 1, 2016 to June 30, 2023. Denominator = Deliveries from July 1, 2016 to March 31, 2023. Exact value suppressed to protect patient confidentiality. Indicates the lowest rate compared to other subgroups; exact value suppressed to protect patient confidentiality. Discussion We conducted extensive literature review and developed a customizable, outcomes-based cohort identification algorithm to capture and categorize pregnancy episodes occurring within 14 hospitals across two health systems over a six-year period. We also extracted relevant demographic and clinical data surrounding pregnancy episodes to enable longitudinal analyses at the person and pregnancy level. Our spacing algorithm removed duplicate records for the same PE and ensured that PEs followed pregnancy interval requirements (in weeks). Our gestational age imputation and subsequent plausibility filter ensured that we only retained PEs that were suitable for analytical research. Results from descriptive queries highlighted significant differences in maternal health outcomes by race and ethnicity, especially for non-Hispanic Black women who had the highest rates of severe maternal morbidity during delivery, preterm birth, and hypertensive disorders. They and Hispanic women were the least likely to have a postpartum care visit in our cohort, compared to other race and ethnicity groups. Our results align with recent literature documenting substantial racial and ethnic disparities exist in the quality and frequency of health care that pregnant patients receive during pregnancy, as well as differences in the clinical, social, and financial contexts in which patients enter and experience pregnancy.29,30 The onus is on public health practitioners, policy makers, politicians, population health researchers, and clinicians to identify social and clinical determinants of maternal health, provide recommendations to reduce racial and ethnic disparities accordingly, and continue to evaluate progress. Though they do not provide a comprehensive picture, clinical databases such as the one described in this paper are merely one tool that can be leveraged by researchers and policy makers to achieve these aims. To our knowledge, this is the first EHR-based pregnancy database in the Gulf South that was designed less for surveillance efforts on a specific or rare event of interest from an absolute perspective (e.g., pregnancy-related death) and more for utilization among health services researchers seeking to conduct observational research on pregnancy, childbirth, and the postpartum period. Our hope is that this database will serve as a tool to address the maternal health crisis10 more effectively. Existing epidemiological and programmatic efforts in this space are well resourced and focus on specific snapshots within a pregnancy, such as vaccination, postpartum depression screening, and maternal mortality. With this database, we intended to capture a more comprehensive and longitudinal view of pregnant people’s clinical experiences before, during, and after pregnancy. The EHR data that make up the database are refreshed quarterly, allowing for rapid and nearly real-time insights into risk factors and health outcomes in this population. We determined that the pregnancy database was an appropriate resource to calculate maternal health metrics that are based definitionally and exclusively on items found in maternal medical records. Indicators of severe maternal morbidity, the definition of which is established by the CDC, were easily calculable using the clinical data, specifically hospitalization data, we extracted from participating systems’ EHR systems. Similarly, the database lent itself well to user-friendly querying for clinical encounter data surrounding a pregnancy outcome for indicators of timely prenatal and postpartum care or specific diagnoses that exacerbate pregnancy risks like hypertension and gestational diabetes. This allowed us to generate summary statistics and produce simple–but actionable and near current–reports on a sample of pregnancies in the state of Louisiana. The customizable nature of the programs that comprised our workspace allowed us to be highly responsive to requests from external partners and other entities interested in leveraging these data for their own maternal health improvement work. Limitations This work was limited by many factors including lack of generalizability (i.e., geographic representativeness), data depth and availability issues inherent to EHR-based research (i.e., data quality), and the absence of neonatal linkage (i.e., data robustness). This work was also limited by resource constraints that prevented validation against the source data (i.e., EHR applications). Although the two health systems contributing data to this project comprised major birthing centers in metropolitan areas of Louisiana, the cohort was not representative of the state as a whole and skewed more urban and more racially diverse than the general population of the state. For example, because participating hospitals are in urban areas, patients in this cohort were less likely to experience living in a maternal health care desert, which meant that this work was not able to discern the effects of access to quality maternity care. Focus on ensuring geographical diversification, perhaps by identifying rural hospitals to contribute data, would be a beneficial next step. However, given resource constraints common in non-urban settings, it is not guaranteed that such hospitals would have the research and data infrastructure needed to participate in these types of initiatives. The phenotype is likely to perform well at sites that similarly conform EHR data to the PCORnet Common Data Model, which is contingent upon funding to carry out this work. Next, use of EHR data for research is, by definition, restricted to data that are approved for specific uses by population health researchers. The EHR data we leveraged for this project did not have unstructured clinician or patient-reported notes. It also did not have discrete, non-clinical social determinants of health (SDH) data points such as household income, reliable access to transportation, food security, weeks of parental leave, or family and community support. Since conclusion of these analyses, the PCORnet Distributed Research Network Operations Center (DRNOC) has prioritized linkage to SDH data and laid out data specifications for population of patient responses to screenings related to depression (PHQ-2/9), food insecurity (Hunger Vitals), health literacy, transportation, housing stability, and more. Further, it has charged PCORnet sites with geocoding patient addresses to enable more granular area-level analyses such as neighborhood indicators of health and wellbeing. Efforts in upcoming work should aim to rerun these analyses with these SDH data as covariates for tabular analyses or predictor variables for more sophisticated analyses on health equity. We were unable to link maternal and infant records during this phase of the project due to resource constraints. Future efforts should focus on data linkage between the maternal and child health records, such that dyads can be established, and more sophisticated metrics calculated. Such linkage would allow for extraction of neonatal health data points of interest like birthweight and perinatal complications for neonates. Subsequent endeavors should prioritize and properly resource the integration of neonatal data into pregnancy databases. Finally, because conducting chart review on a subset of patients identified by our algorithm was outside the scope of this project, we were unable to validate the algorithm. Our team only had access to transformed limited data sets, not source EHR data itself, complicating chart review. While the logic used to establish pregnancy intervals–such as enforcing minimum spacing, applying median-based imputation, and using clustering to de-duplicate records–is internally sound and consistent, its precision was not verified against a gold standard. Although we expect any resulting imprecision to be minimal and within an acceptable margin of error, derived estimates for pregnancy dates, gestational age, and pregnancy classification may nonetheless be inaccurate. For instance, fragmented records from a single pregnancy episode may appear to our methodology as separate episodes, and conversely, distinct pregnancies may inadvertently be merged. Similarly, in the absence of independent external validation sources—such as vital records, fetal death registries, or manual chart adjudication—misclassification present in the source EHR (e.g., stillbirth recorded as live birth) cannot be identified or corrected and would be propagated through the standardized transformation to the PCORnet Common Data Model. These potential sources of misclassification may influence derived estimates of maternal health outcomes described in this paper such as severe maternal morbidity and preterm birth. Thus, further resourcing and access to a gold standard data source are essential to assess the validity of this approach and evaluate how closely the constructed cohort resembled true pregnancy episodes as documented in source data. Conclusion By leveraging EHR data from participating health systems, we have been able to successfully identify pregnancy episodes and create a nearly real-time database with demographic and clinical data. This showcases the potential for leveraging EHR data to create a robust and timely resource for advancing maternal and child health research and identifying disparities in health outcomes. Supplementary Material