SC COVID-19 Cohort-Pregnancy Database

With support from the National Institute of Allergy and Infectious Disease (R01A127203-4S1), our team has established a statewide S3C database for COVID-19 research since 06/2020 by integrating various state-level data sources including: (1) the COVID-19 testing data from the SC Department of Health and Environmental Control (SC DHEC), (2) hospital encounter data for inpatient hospitalisation, outpatient surgery, home health and emergency departments; (3) health utilisation data from large public and private health insurance plans (eg, Medicaid, State Health Plan, BlueCross BlueShield of SC); (4) EHR data from health systems (Prisma and MUSC) and (5) programme data from the SC Department of Mental Health. The database is updated every 6 months.

In this study, as part of the National Institutes of Health (NIH)’s ‘Implementing a Maternal health and PRegnancy Outcomes Vision for Everyone’ initiative supported by the Office of Director, NIH (3R01A127203-5S2), our team will create a population based S3C pregnant women cohort (S3C-P), which includes all women who gave birth between 2018 and 2021 in SC (>2 00 000 births, 57.2% white, 31.1% black, 4.6% Hispanic) and will add vital record data (birth and death certificates) to complement existing linkages from the parent S3C cohort. The identification of pregnancy status and COVID-19 infection will be cross-verified using EHR, claims data, laboratory reports and ongoing SC DHEC medical chart reviews among >4387 pregnant women with confirmed COVID-19 infections in SC as of December 2021. The SC Office of Revenue and Fiscal Affairs will collate databases and provide our team with a deidentified linked database system.35

National COVID-19 Cohort Collaborative

The N3C is a novel data consortium that integrates EHR and medical claims data from 92 healthcare systems and institutes across 50 states. N3C enables data sharing, computable phenotypes and collaborative data mining by harmonising EHR data of diverse standards using Observational Medical Outcomes Partnership Clinical Data Model. N3C was created to study potential risk factors and protective factors of COVID-19 and its long-term health consequences.34 As of 24 December 2021, N3C has aggregated 9.4 million patients (3.3 million COVID-19 patients) with their EHR dating back from January 2018, including 1.9 million women with COVID-19 (>61 k pregnant women) and 3.5 million women without COVID-19 (>209 k pregnant women). The participants in N3C represent diverse populations in the USA (eg, geographic, socioeconomic, racial/ethnic). Building on a secured cloud environment, N3C provides data harmonisation, privacy-preserving data linkage and high-performance data analytics. Our team has already gained access to the restricted N3C database including ZIP codes of patients and health systems and dates of services.

Nationwide social context databases

The 2015–2019 ACS and the 2018–2020 AHA Annual Survey will be used to calculate county-level residential segregation measures, racial discrimination in SES and ZIP-level accessibility to hospital-based obstetric units.

Time-varying local COVID-19 infection and social-distancing policy data

To better understand local pandemic settings, we will also add the Centers for Disease Control and Prevention (CDC) COVID-19 Case Surveillance restricted datasets for nationwide cases confirmed since 11 March 2020, and CDC’s state-level social distancing policies (eg, emergency declaration, stay at home order) in early pandemic and telehealth services expansion data and each corresponding date of enaction.36

SC Pregnancy Risk Assessment Monitoring System

SC Pregnancy Risk Assessment Monitoring System (PRAMS) as a part of national PRAMS is an ongoing survey of SC mothers who have recently given birth.37 These mothers are sampled from state birth certificates. After statistical weighting, PRAMS data are representative of all mothers who gave birth in SC. SC PRAMS added 11 COVID-19-related questions for mothers who delivered in August 2020 and after in their survey. SC PRAMS routinely collects detailed psychosocial and behavioural risk factors for each participant, which are not available in S3C and N3C. Residential ZIP codes will be used to add in social contextual variables and other ZIP-level or county-level characteristics. The unweighted sample size for SC 2018–2021 PRAMS will be at least 2000.

Outcome measures

The main outcomes of interest will be SMMM.38 We will adapt from a previously validated algorithm by using the International Classification of Diseases, 10th Revision, Clinical Modification diagnosis and procedure codes to identify women with one or more of the 21 SMM indicators developed by the CDC and updated by the Alliance for Innovation on Maternal Health programme at the time of childbirth. Maternal mortality will be identified using statewide death certificate data from the childbirth date to up to 1-year post partum. A composite variable of SMMM will be created to reflect SMM or maternal mortality incidence. We will also study MMM composite, which includes mortality and morbidities related to hypertensive disorders of pregnancy, postpartum haemorrhage and infections/sepsis that happen during pregnancy through 6 weeks post partum.39 Other outcomes to be studied include: (1) adverse maternal outcomes including intensive care unit (ICU) admission, invasive ventilation, receipt of extracorporeal membrane oxygenation, etc; (2) prolonged length of stay,40 41 and (3) hypertension, pulmonary embolism, type 2 diabetes, cardiovascular diseases (eg, heart attack, myocardia infraction, thrombus, stroke) diagnosed within 1 year after delivery.

COVID-19 status and severity

Eligible COVID-19 cases are those with a positive test for SARS-CoV-2 since 11 March 2020, during pregnancy. Data on symptom status (symptomatic, asymptomatic, unknown) are available, while severity will be defined using the WHO’s Clinical Progression Scale.42

Social context measures

The five dimensions of county-level residential segregation, including evenness, exposure, concentration, centralisation and clustering,43 44 will be determined for each race/ethnic group (eg, black, Hispanic) using the ACS Census tract data.45 46 Each index will be calculated across census tracts within residential counties. Higher values indicate higher levels of segregation. We will create the group indicator for segregated vs less segregated counties using the cutoffs for each dimension index.46 Additional hypersegregation index—segregations scores at >0.6 on at least four aforementioned dimensions—will be created to reflect the highest levels of segregation.46 Within-county racial/ethnic discrimination in SES will be calculated using the ACS county data,45 including black-white and Hispanic-white ratios of poverty, unemployment and home ownership rates.47–51 These measures will be linked to databases via maternal residence counties.

Healthcare institution

Using the AHA annual survey data on hospital location,52 we will identify loss of hospital-based obstetric units using our published validated algorithm.53–55 An indicator for whether a hospital or a hospital’s obstetric service was closed for each year will be created. In turn, women’s access to hospital-based obstetric care within 30 mile distance for years 2018–2020 using the ArcGIS fastest route network will be determined: (1) had access to; (2) no access to and (3) experienced the loss of all hospital-based obstetric units.

County-level COVID-19 infections and social distancing policies

The CDC’s COVID-19 Case Surveillance will be used to compute monthly cumulative rates of in-county residents that had been confirmed COVID-19 positive, hospitalised, admitted to an ICU, and with mechanical ventilation/intubation as a result of COVID-19 disease. Number of months elapsed since a county had each of the following policy orders will be calculated from a delivery date: emergency declaration, closures of bars, restaurants and/or other non-essential business, stay at home order and telehealth services expansion.

Impacts of the COVID-19 pandemic on racial/ethnic disparities in SMMM

We will examine the overall impacts of the COVID-19 pandemic on SMMM using the data from the S3C-P and N3C. We hypothesise that: (1) compared with prepandemic periods, SMMM has increased during the pandemic and racial/ethnic disparities have widened during the pandemic and (2) compared with pregnant women without COVID-19 infection, women with COVID-19 infection experienced higher proportions of SMMM, and racial/ethnic disparities in SMMM have amplified among COVID-19 infected women.

First, we will examine the distributions for all measures and clean the database (eg, outliers, data entry errors etc) using appropriate statistical techniques. Second, we will conduct preliminary analyses and examine descriptive statistics for outcome measures. Unadjusted and adjusted associations of SMMM with key variables and covariates will be assessed using appropriate statistical procedures (eg, tests of proportions, χ² tests, analysis of variance).

Women who gave birth between 1 January 2018 and 10 March 2020 will be categorised as before pandemic, while women, who gave birth between 11 March 2020 and 31 December 2021, will be considered as during the pandemic. The prepandemic versus pandemic impact on SMMM will be modelled via logistic regression. First, the crude model with the pandemic indicator only will address SMMM change before and after the pandemic. Second, to investigate whether racial/ethnic disparities in SMMM have widened during the pandemic, the crude model will be further adjusted with race/ethnicity, interactions between race/ethnicity and pandemic indicator and months elapsed since 11 March 2020, on delivery dates. Then additional variables will be added, including individual-level characteristics (eg, age, SES proxy (ie, Medicaid/uninsured status), parity, marital status, underlying health conditions). Variable selection and goodness of model fit will be evaluated using the AIC, BIC and likelihood ratio test.

We will also conduct analyses in women who delivered during the COVID-19 pandemic by comparing pregnant women with and without COVID-19 infection. We will first create the COVID-19 infection indicator and then will perform the similar analysis as those for the prepandemic versus pandemic impact. We will further adjust for county-level COVID-19 infections per capita and social distancing policies at the appropriate time points using logistic regressions with random effects accounting for correlations among counties.

Social contexts, racial/ethnic disparities in SMMM and distinct mediating pathways

We will study these issues using different databases and methods. First, we will examine the association between social contexts and changes in racial/ethnic disparities in SMMM before and during the pandemic. We hypothesise that racial/ethnic disparities SMMM are potentially disproportionately widened in communities with higher racial/ethnic economic disparities (measured by black-white ratios of economic disadvantages, Hispanic-white ratios of economic disadvantages) and in higher vs less segregated black or Hispanic counties (measured by residential segregation). We will conduct a parallel analysis between social contexts and SMMM using the data from S3C-P and N3C (figure 2). Similarly, we will examine the contributing roles of social contexts to racial/ethnic disparities in SMMM in the overall sample (prepandemic vs pandemic) and between COVID-19 positive vs COVID-19 negative women. The multilevel variables that will be investigated include individual-level characteristics, community-level characteristics (ZIP code accessibility of hospital obstetric units), and county-level characteristic (social contexts: residential segregation and racial discrimination on SES; COVID-19 infection per capita and social distancing policies). For exploratory analysis, the OR of SMMM among black and white and the social context level at the county level prepandemic versus pandemic will be visualised via the spatial temporal map using GIS. The summary statistics in SMMM with respect to race/ethnicity will be calculated according to individual-level and county-level characteristic.

We will model SMMM via multilevel hierarchical logistic regression. Women who reside at the same community-level or county-level will be accounted for via random effects, which will be further modelled in the regression model with a multivariate normal distribution to account for the correlations among community or county. We plan to use an incremental modelling strategy: (1) crude model (race/ethnicity and social context factors); (2) adjusting for individual level factors and (3) additional adjustment of additional community-level and county-levels characteristics. For N3C data, we will further adjust for state or Census region. To further examine whether social contexts moderate racial/ethnic disparities in SMMM and whether these disparities vary between pre-pandemic and pandemic, we will include two-way and three-way interaction terms in the model (eg, pandemic period*social context*race). To examine the added impact of COVID-19 infections on SMMM, we will repeat the model by restricting it to all women who gave birth during pandemic period (delivered after 11 March 2020) and including maternal COVID-19 severity status in the model. Models will be compared using the AIC and BIC criteria. In the modelling procedure, the outliers, missingness, multicollinearity and non-linear will be addressed accordingly, and sensitivity analysis will be conducted comparing models with or without the treatment of outliers or missing data. The magnitude and direction racial/ethnic disparities will be assessed through OR and its 95% CI.

Second, we will use a convergent parallel design to evaluate the underlying pathways between social contexts and racial/ethnic disparities in MMM via pandemic stressors, maternity care (prenatal and postpartum) and mental health condition. We hypothesise that social contexts might hinder maternity care and worsen mental health conditions among black and Hispanic women during the pandemic and exacerbate racial/ethnic disparities in MMM.

The quantitative analysis will be conducted using SC PRAMS data, which have unique data elements that are not available in S3C-P and N3C. The 2018–2021 SC PRAMS data will provide a more refined understanding of COVID-19 stressors, psychosocial stress and healthcare utilisation through the questionnaire, including pandemic stressors (financial, job loss, childcare, etc), individual mitigation practices, changes in prenatal and postpartum care, psychosocial stress, barriers to health services, intimate partner violence, prenatal and postpartum care utilisation, smoking, alcohol use, gestational weight gain and mental health. SC PRAMS asks respondents to assess how often they experienced depressive symptoms after delivery. Descriptive statistics will be used to examine pandemic-related changes in MMM and psychosocial and behavioural changes between prepandemic (delivered before 11 March 2020) and pandemic periods (11 March 2020 and after). The weighted hierarchical regression model will be applied to examine the association between social context and MMM. As previously mentioned, community and county levels will also be modelled in the regression. Different from the prior analyses above: (1) individual characteristics will mainly come from PRAMS or birth certificates; (2) the weight based on complex survey design will be modelled and (3) individual reports of healthcare utilisation and mental health condition included.

We will conduct in-depth interviews among 40 postpartum women of colour (~20 African American; ~20 Hispanic) stratified by COVID-19 infection status and 10 maternal care providers (MCPs) who serve pregnant and postpartum women in Black and Latino communities. The inclusion criteria of postpartum women include: (1) ≥18 years old; (2) either African American or Hispanic; (3) have given birth in 2021 and (4) living in SC. We will purposely recruit postpartum women and MCPs through local OBGYN clinics and community health organisations that serve a larger proportion of low-income black and Hispanic women. We will train female interviewers of the same race as the interviewees to obtain trust from the postpartum women participants, and interviews in Spanish will be conducted as needed. Guided by the conceptual framework, the main topics of the postpartum women interviews will include: (1) their perceptions toward their healthcare providers and institutions for perinatal care; (2) experience with prenatal and postpartum care; (3) stressors in the COVID-19 pandemic; (4) challenges in healthcare seeking (eg, appointments, clinic visits), especially from structural factors (racism and discrimination) and (5) their needs/recommendations for future healthcare. The main topics of the MCPs include: (1) stressors and challenges of their clients; (2) clients’ mental health conditions; (3) impacts of COVID-19 on their care provision and (4) their views on health disparities caused by structural factors. The interviews will last 50 min and will be recorded with each participant’s consent. Audio recordings will be transcribed and coded using NVivo V.11.0. We will employ thematic analyses.56 The findings will complement the quantitative data in providing a comprehensive picture on how COVID-19 affects psychosocial well-being of the postpartum women of colour; offer in-depth interpretation and explanation of quantitative results; and explore the mediating pathways in which structural factors amplify existing disparities in maternal health in the context of the pandemic.

Machine learning-based predictive models

We will develop and evaluate machine learning-based predictive models to identify risk factors of SMMM and forecast progression of hypertensive disorder, pulmonary conditions, type 2 diabetes mellitus and cardiovascular diseases among postpartum women. The predictive models will synthesise individuals’ demographics, EHR, social contextual factors and community and healthcare system level data to make predictions of individuals’ clinical outcomes at key time points. Because data sources suggestive of these factors are variable and high-dimensional, and these factors are inherently interconnected over time,57 machine learning is a superior approach to predicting clinical outcomes and proactively detecting the associated risk factors for early intervention and treatment. Constructed models will demonstrate critical factors predictive of clinical outcomes and how these factors interact over time.

Supervised machine learning algorithm will be adopted for the predictive models. Using N3C, the algorithm will learn from input variables and predict SMMM and long-standing morbidities (eg, hypertension, pulmonary embolism, diabetes, cardiovascular diseases) over time. Input variables will include maternal characteristics, for example, sociodemographics, sociobehavioural data, social context variables, diagnoses, procedures, laboratory tests and medications. The prediction of SMMM and long-standing morbidities will take place at critical time points: <3, 6, 12 weeks, 6 and 12 months postpartum.58 We will develop deep learning algorithms because of their ability to integrate complex clinical data and social contexts from multiple sources with superior predictive performance, including: (1) convolutional neural network (CNN) for its ability to capture dynamic patterns among multilevel input variables; (2) recurrent neural network (RNN) with long short-term memory (LSTM) architecture for its ability of capturing temporal patterns of clinical events (eg, onset of preinfection conditions, viral infections and clinical events marked with gestational weeks, and the date of childbirths) and (3) Deep Boltzmann Machine (DBM) for its interpretable scoring mechanism for risk prediction.59

We will use a 10-fold cross validation. Specifically, S3C and N3C data will each be randomly partitioned into ten splits. In each of the 10 iterations, 7 splits of data will be randomly selected for model training, 2 splits of data will be used for internal validation (finetuning hyperparameters) and the 1oth split used for testing. We will use F measure, precision, recall and the area under the receiver operating characteristic curve, if unbalanced data, to measure the predictive performance of models. We will use support vector machine (SVM) as the baseline algorithm to compare against the performance of CNN, RNN (LSTM) and DBM. The best-performed model will be identified based on F measure.

We will rank input variables and/or clusters of input variables by calculating the importance scores57 (eg, mutual information, SVM-based recursive linear elimination). Two content coinvestigators will independently review the ranked results and identify clinical/social risk factors. Disagreement between two reviewers will be resolved by panel discussion. Development of sophisticated machine learning models for predicting long-standing morbidities will be used to identify important risk factors prenatally, which can be used for early intervention, treatment and community-wide interventions.

Power and sample size calculation

We estimate that there were 200 000 women who gave birth in S3C and 270 000 pregnant women in N3C for our study period. The primary outcome of interest is SMM and the main exposures of interest are race (white vs black) and pre–peri COVID-19 period. We assume that there are 64 000 (32%) black and 114 000 (57%) white in S3C60 and 38 070 (14.1%) black and 160 380 (59.4%) white in N3C.61 We also assume that with the same time length of pre and peri COVID-19, the prevalence of pregnancy will be similar (50%). For all aims, we consider the logistic regression with mixed effects. Based on 1.5%–2.5% incidence of SMM with 20%–50% increase due to the COVID-19 impact or race disparities, we conduct the power analysis based on the logistic regression using SAS and conclude that 200 000 sample size will be adequate to reach the power around 90%. Figure 3 illustrates the relationship of power and OR with setting of n=200 000, significance at 0.05, variable of interest with a ratio of 10:90. This will be the basic model considered, and it indicates strong power for the prediction model.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

Estimated power according to prevalence of outcomes and ORs.