Study population

We conducted a secondary analysis of data from the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be (nuMoM2b) prospective cohort [29, 30]. The study recruited pregnant women who would be delivering for the first time from hospitals affiliated with eight clinical centers and collected data on each participant over the course of four study visits during their pregnancy via in-clinic interviews, self-administered questionnaires, clinical measurements, and chart abstractions. The cohort included 9,289 women who enrolled and had their first study visit in the first 6 to 14 weeks of their pregnancy and who consented to the release of their data. All local institutional review boards approved the study protocol, and participants provided written informed consent prior to enrollment. The Georgia Institute of Technology Internal Review Board and the Eunice Kennedy Shriver National Institute of Child Health and Human Development both approved this secondary analysis.

We accessed the nuMoM2b data on December 8, 2021. We did not have access to information that could identify individual participants during or after data collection.

Data preparation

For this secondary data analysis, we chose to exclude women that delivered at gestational age < 22 weeks or > 43 weeks, those with terminations, and those with unknown outcomes. We also excluded women with missing race/ethnicity and pre-pregnancy conditions data (Fig 1). To address missing data among confounders, we created 5 imputed datasets using multivariate imputation by chained equations (MICE) [31]. Please see S1 Appendix for more details about the exclusion protocol and S2 Appendix for more details about the data structure and preparation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Exclusion criteria.

https://doi.org/10.1371/journal.pone.0306206.g001

Main outcome measures

The primary outcomes of this study were severe preeclampsia, blood transfusion, and postpartum readmission (up to 14 days). The outcome of severe preeclampsia is indicated by one of the following diagnoses: preeclampsia with severe features, HELLP syndrome, superimposed preeclampsia, or eclampsia. The nuMoM2b definitions for these new onset hypertensive disorders are based on clinical criteria can be found in S2 Appendix of Facco et al. (2017) [32]. We selected these indicators due to their strong association with severe maternal morbidity as defined by the Centers for Disease Control and Prevention, the quality of measurement in the nuMoM2b dataset, and their impact on acute and long-term maternal health. Severe preeclampsia is a hypertensive disorder of pregnancy often presenting as new-onset hypertension and proteinuria [10]. Severe preeclampsia is a major cause of maternal morbidity, [33] and can rapidly progress to serious complications, including death of both mother and fetus [10]. Blood transfusion is included as an indicator of severe maternal morbidity (SMM) by the Centers for Disease Control and Prevention (CDC) [34]. Postpartum readmission represents a costly, rare, and heterogeneous postpartum adverse event. There is large variation in the indications for postpartum readmission in the U.S. The three most frequent indications are hypertensive disorders, postdelivery infections, and psychiatric disease [35]. (S2 Tables 1 and 2 in S2 Appendix) provides details about the definitions and collection of the primary outcomes in the nuMoM2b study. We fit regression models to each of these outcomes individually rather than create a composite outcome to preserve clinical interpretability.

Pre-pregnancy conditions

Pre-pregnancy conditions were collected by the nuMoM2b study team at each of the first three visits through an in-person interview with the participant, who indicated the presence of each condition anytime throughout their life. The nuMoM2b study team reconciled this data with chart abstraction to finalize the indication of pre-pregnancy conditions.

The nuMoM2b study team collected 41 pre-pregnancy conditions, and we categorized them into 12 condition types according to the framework of Tang et al. [36] (S2 Table 3 in S2 Appendix). We grouped the pre-pregnancy conditions based on similarities in treatments, clinical manifestation, or organization in the health care system. For example, we grouped cervical dysplasia, fibroids, and PCOS under the condition type “Gynecological” because they are conditions that affect the function of female reproductive organs and are treated by the same gynecologist specialty branch. The condition types included: Autoimmune, Cardiovascular, Endocrine, Gastrointestinal, Gynecological, Hematologic, Kidney, Lung, Mental, and Neurological. This classification was developed and validated by three of our authors (KS, SB, MP) with extensive medical and specifically obstetric knowledge.

To investigate the effects of specific multimorbidity patterns on adverse maternal outcomes and racial/ethnic disparities, our analysis considered pairs of co-occurring pre-pregnancy condition types. For example, consider a patient with the following three pre-pregnancy conditions: anemia (hematologic), hypertension (cardiovascular), and migraine headaches (neurological). This patient was indicated to have the following co-occurring condition types: hematologic & cardiovascular, cardiovascular & neurological, and hematologic & neurological.

Confounders

We adjusted for confounding factors including maternal age, insurance, body mass index (BMI), and sociodemographic information including income and education. Each of these factors were collected during Visit 1, which occurred between gestational age 6 and 14 weeks. Mode of delivery and type of labor were not included as confounders as they are considered to be part of the causal pathway leading to adverse maternal outcomes [37].

Statistical analysis

We used Poisson regression models with robust standard errors to analyze the associations between pre-pregnancy condition types, self-reported race/ethnicity, and adverse maternal outcomes with the results reported as crude and adjusted risk ratios (RRs) and 95% confidence intervals (CIs). We used a class-weighting method which gives a higher priority to correctly classifying the subgroup of women who experienced an adverse maternal outcome (see S2 Appendix).

Potential mediators were selected for each adverse maternal outcome in a stepwise process to avoid multicollinearity. Each adverse maternal outcome was regressed by race/ethnicity, confounders, and all potential mediators (all pre-pregnancy condition types and all their combined effects) (hereafter referred to as the “Outcome Model”). Only condition types and combined effects reaching statistical significance (p-value ≤ 0.05) in this model were included in further analysis. Additionally, only combined effects with at least 1% prevalence in the dataset were considered. Next, we used a model-building approach starting with a standard individual-level Poisson model (Model 1, crude), followed by a Poisson model that adjusted for confounders (Model 2, adjusted). We additionally controlled for significant pre-pregnancy condition types (Model 3), and finally, we additionally controlled for significant combined effects between co-occurring pre-pregnancy condition types (Model 4).

Next, we sought to determine if pre-pregnancy condition types and their combined effects contributed to the racial disparities in adverse maternal outcomes using mediation analysis. To analyze mediation, the potential mediator condition types and their combined effects were individually regressed by race/ethnicity and adjusted for confounders (“Mediator Model”). The Outcome and Mediator models were combined to compute the mediation proportion, which estimates the proportion of the risk factor’s impact (race/ethnicity) on the outcome that is attributable to the mediator (pre-pregnancy condition type or combined effect) [38]. This analysis was conducted for each adverse maternal outcome and results were averaged across the five imputed datasets.

Performance analysis

Finally, we analyzed the predictive performance of our regression models (Model 1, Model 2, Model 3, and Model 4) to understand which pre-pregnancy patient characteristics accurately classified adverse maternal outcomes. Performance of each model was evaluated for each racial/ethnic subgroup using the average area under the ROC curve (AUC) across 10-fold cross validation and 95% confidence intervals were computed with 2000 stratified bootstrap replicates [39]. Similar modeling was performed by Khanna et al. (2019) in which regression models were used to predict hospitalization using risk factors including the co-occurrence of two factors [40]. They report both the risk factors’ associations with hospitalization and the regression model’s predictive performance (AUC).

We used R, version 4.2.1, for all analysis, caret in R to train and test our prediction models, [41] pROC in R to compute performance metrics, [39] and mediation in R to conduct mediation analysis [42]. We use p-value threshold of 0.05 for statistical significance.

Main findings

Our findings suggest accounting for pre-pregnancy condition types explained some of the association between non-Hispanic Black race/ethnicity and severe preeclampsia. We found that including pre-pregnancy condition types improved diagnostic ability (AUC) in predictive models for severe preeclampsia in the overall cohort. These results highlight the potential for risk prediction using pre-pregnancy conditions in a diverse, but low-risk population.

Consistent with previous studies, [4, 21, 43–45] we found that identification as a minority race and ethnicity was associated with a higher risk of adverse maternal outcomes among a cohort of nulliparous individuals. Adding to some studies that have examined the impact of specific co-occurring condition combined effects, [20, 46, 47] our study focused on exploring all potential pre-pregnancy condition type combined effects to understand their association with adverse maternal outcomes (specifically severe preeclampsia, postpartum readmission, and blood transfusion) and quantified the value of including combined effects in predictive models. No pre-pregnancy condition combined effects were significantly associated with adverse maternal outcomes.

We also quantified the association between individual pre-pregnancy condition type and adverse maternal outcomes. Each adverse maternal outcome was associated with different pre-pregnancy condition types, except the autoimmune condition type, which was associated with increased risk for both blood transfusion and postpartum readmission. It is well known that pregnancy is associated with changes in the maternal immune system, and specifically the postpartum period is associated with autoimmune disease flares [48].

Our work adds to previous studies on the use of prediction models to predict various adverse outcomes, which have ranged from using symptoms and signs, [17, 19] laboratory tests and biomarkers, [49, 50] and demographics and medical history [51, 52]. With the inclusion of pre-pregnancy condition types, our model yields an AUC of 0.65 for predicting severe preeclampsia. This is comparable to the AUC of previously published obstetric comorbidity indices predicting severe maternal morbidity, which includes severe preeclampsia (Bateman, et al. [18]: AUC, 0.65 and Easter, et al. [17]: AUC, 0.70 as reported by Leonard et al. [19]). The prediction models for severe preeclampsia improved with the addition of pre-pregnancy condition types. Together, our findings and prior studies suggest that maternal morbidity cannot be accurately predicted using medical conditions alone as these models cannot account for structural factors that contribute to health inequities.

Clinical implications

Understanding the potential reasons for adverse maternal outcomes is an important pathway to understanding and reducing racial and ethnic disparities and high rates of maternal morbidity and mortality.

Our findings suggest that the presence of pre-pregnancy conditions and confounders may explain some of the observed association between non-Hispanic Black race/ethnicity and severe preeclampsia. In a prior study, Black race was associated with increased odds of pregnancy-induced hypertension after adjustment for preexisting conditions and demographic factors [53]. However, our study adjusts for a different set of preexisting conditions. The prevention and management of these pre-existing conditions, specifically cardiovascular conditions, before, during, and between pregnancies could be an important consideration to decrease racial and ethnic disparities in adverse maternal outcomes. Although the nuMoM2b cohort study collected data on a wide range of pre-pregnancy conditions, we could not explain a large proportion of the disparities in maternal morbidity in this cohort. Adverse maternal outcomes and pre-pregnancy conditions have been linked to the environmental context and social conditions in which people exist [25]. For example, a study by Lueth et al. (2022) found that a high allostatic load, as an estimate of chronic stress, was significantly associated with adverse pregnancy outcomes and partially mediated the association between self-reported race and adverse pregnancy outcomes in the nuMoM2b dataset [54]. A wider array of determinants (structural racism, socioeconomics, political context, etc.) should be evaluated to explain racial/ethnic disparities in adverse maternal outcomes and to develop comprehensive interventions to promote health equity. Future work may also benefit from the inclusion of additional data such as the severity and management of morbidity and multimorbidity during the preconception, interconception, and postpartum periods to determine whether differential management or access to healthcare systems reduces the observed disparities.

Our work suggests that information about pre-pregnancy conditions can be useful in improving the ability to risk-stratify individuals. Predictive modeling may be helpful in further exploring the complex relationships of co-occurring conditions. However, more research is necessary to inform best clinical practices for use of predictive models with a focus on mitigating unintended consequences and preventing the exacerbation of disparities [55].

Strengths and limitations

Our study has several strengths. Using a large and comprehensive dataset, we evaluated the association between race and ethnicity, pre-pregnancy conditions, and adverse maternal outcomes in a cohort of nulliparous women in the United States. This dataset contains thoroughly collected data that goes beyond a typical electronic health record including health history and conditions, demographics, and survey questionnaires. Our condition type groupings allowed for larger sample sizes of conditions and more accurate estimates of risk ratios. In addition, the use of feature selection algorithms allowed for the exploration of combined effects to improve model performance for predicting adverse maternal outcomes.

Our study has several important limitations. First, the analysis was limited to the data collected in the nuMoM2b dataset, which only includes nulliparous, predominantly non-Hispanic white women (61.0%) who received regular prenatal care at academic medical centers beginning in their first trimester; thus, our findings may not be generalizable to different birthing populations. Second, the nuMoM2b dataset also does not indicate the severity or management of pre-pregnancy conditions. Third, our dataset only collected data up to 14 days postpartum, although many postpartum readmissions occur after this timespan [56]. Finally, adverse pregnancy outcomes are rare and occur infrequently in the nuMoM2b dataset, particularly as we stratified by racial subgroups and by multimorbidity. A larger sample size may be able to identify more clinically relevant associations between pre-pregnancy conditions and maternal outcomes that occur infrequently.