Research Design and Population
This retrospective cross-sectional study used CDC's pregnancy risk assessment and monitoring system (PRAMS) data (2016-2021 phase 8-2021) and social vulnerability index (SVI) data (2014–2018 and 2016–2020).
baby carriage [23] is an ongoing population-based surveillance project to monitor mother attitudes and experiences before, during and immediately after pregnancy. Each month, participating countries sample samples of individuals who have recently given birth to living infants (within 2-6 months, within 2-6 months) identified through birth certificate files for data collection via mailed surveys and follow-up telephone interviews. [24]. Participating states choose to oversample the preferred subpopulations, and can usually sample 1,000-3,000 people each year [24]. The final PRAMS dataset is weighted to sample design, non-response, and non-coverage to produce estimates of representative populations [24]. To reduce non-response bias, the CDC PRAMS Working Group sets a response rate threshold of 55-70%, depending on the year of survey. [24].
SVI [25] We assess the relative vulnerability of each county within the state by ranking census regions on 15-16 social factors (depends on year) and (1) socioeconomic status, (2) household characteristics, (3) racial/ethnic minority status, and (4) housing type and transport. Tract Ranking is percentile based, ranging from zero to 1, with higher values indicating higher vulnerability. Tracts below the 90th percentile are assigned a value of 1 (“flagged”) to indicate a high vulnerability, while routes below the 90th percentile are assigned a value of zero. [26].
This study was considered exempt from ethical review by the Emory University Institutional Review Board (Study00003674) because of the use of published identification data.
Inclusion criteria
The study population consisted of novice women who gave birth to a living singleton infant without birth defects and self-identified as NHB or NHW. Originally, the dataset contained 229,697 people. We excluded people (1) aboutn = 143,370), (2) has multiple pregnancies (n = 1,480), or (3) delivered infants with congenital defects (n = 814). Next, we excluded people who were not NHB or NHW (n = 23,278). This continuous removal process reduced the initial sample size to 60,755. Additionally, we subset data into nine states (FL, GA, IA, IN, IN, IN, MO, NC, WI, and WY) using available data on racism, stressful life events, and SVIs. The final sample size was 9,595.
countermeasure
result
Gestational age was grouped into PTB (<37 weeks) and periodic births (37+ weeks) at the weeks of completion (≤27, 28–33, 34–36, 37–39, and 40+). This decision was made with a very large number of cases of PTB (less than 28 weeks; n = 182) and very PTB (28-31 weeks, n= 535) were small, especially after stratification by race and ethnicity. Limited sample sizes do not provide sufficient data to reliably train and predict these individual results.
Predictors
The variables in the analysis were based on previous literature on risk/protective factors for PTB. Most of our predictors were SDOH, defined by Healthy People 2030, representing economic stability, educational access and quality, access and quality to healthcare, neighborhood and built environment, and social and community contexts. [27]. The study year variables and US states were also included, and we considered potential differences in characteristics of cohorts of maternal births at a particular year and location (Table S1 for details).
Individual factors include socioeconomic, psychological, medical and behavioral factors. in particular, Socioeconomic factorsIt incorporates maternal race and ethnicity, mother's age, marriage status status, total household income per year, and the number of dependents who depend on that income, and receives special supplementary nutrition programs for women, infants, children (WIC), maternal education, health insurance and residential areas. Psychological factorsIt incorporated 14 stressful life events, depression, physical violence by current or ex-husband/partners, and perceptions of racism. Medical Factorsincorporated the number of previous other pregnancy outcomes that ended up in pregnancy terminations, diabetes before pregnancy, gestational diabetes, hypertension before pregnancy, gestational hypertension, pre-pregnancy body mass index (BMI) (categorical), weight gain during pregnancy (categorical), fever during pregnancy, premature rupture of membrane (PROM), and medical risk factors.Variables for weight gain during pregnancy were developed based on medical laboratory guidelines. [28]specify the recommended range of weight gain by prepregnancy BMI category. Behavioral factorsNumber of prenatal care visits (PNC), initiation of PNC in early pregnancy, validity of Kotelchuk in the prenatal care utilization index, multivitamin intake, pregnancy intent, smoking cigarettes, e-cigarette smoking, drinking alcohol. lastly, Context FactorsTotal of 31 (15 [2014–2018] and 16 [2016–2020]) Social factors from SVI. As mentioned earlier, census regions below the 90th percentile of a given factor are assigned a value of 1 (flag) to indicate the highest vulnerability area, while all others are assigned a value of 0. More flags suggest more socially vulnerable counties in the state.
Data analysis
Variable selection and processing of missing data
Of the 100 predictors previously determined for each literature review, 13 were excluded due to the degree of missing ≥10%. [29] (Table S2). As required in ML for prediction, missing observations were removed for results. For predictors, a central assignment was made to the numerical variable, and the code ('Na') was assigned to missing observations of categorical variables, expecting that the model would automatically learn that the missing pattern was implied [30].
Descriptive statistics
Characteristics of study populations with frequency, percentage, mean, and standard deviation were summarized. We reported black and white differences in these properties and associations with PTB using Pearson's Chi-square test and Wilcoxon's rank total test. Statistical significance was set to α=0.05. In particular, we only selected predictors that were significantly related to PTB for modeling that differed from NHB and NHW women.
ml
To predict PTB, we constructed elastic nets, random forests (RF), and extreme gradient boost (XGB) models. These algorithms were chosen for a wide range of applications, relatively easy to use, and diverse learning capabilities. [31]. Data were split into training and test sets (70/30), each maintaining the same PTB:term birth ratio to reduce data imbalance. Predictors were preprocessed by scaling continuous variables and dummy coding of categorical variables. Hyperparameter tuning was performed using random grid search via Latin Hypercube sampling (grid size = 500). Individual models for each combination of hyperparameters were assessed for performance with a 5x cross validation and area under the curve (AUC) score as the evaluation matrix.
We increased the explainability of the model by using SHAP (Shapley Additive Description) values to identify key predictors of PTB among women with NHB and NHW respectively. [32]. Rooted in collaborative game theory, SHAP ensures a fair attribution of the importance of features by calculating the average marginal contribution of each feature across a subset of all features in the model. The absolute SHAP values reflect the magnitude of the impact of functions within the model, but this study does not allow direct comparisons across NHB and NHW models. [33]. By quantifying how much each individual and contextual factor contributes to PTB risk within each racial and ethnic group, Shap provides practical insights to increase confidence in ML predictions and guide tailored prevention strategies.
Sensitivity analysis
Two sensitivity analyses were performed. First, we retrained the best performance model on all 85 candidate predictors, omitting the first univariate. p– Value screen. Second, after aiding in the definition of spontaneous PTB and removing PROM, a variable that could precede medically indicated PTB, the analysis was repeated with 84 predictors. These tests addressed two concerns. (i) the original screen may rule out clinically relevant but not critical predictors, and (ii) some algorithms (elastic nets) perform their own function selection.
The final model using the selected hyperparameters (Table S3) was evaluated in the test set. All data analyses were performed using R version 4.0.2 (2020-06-22).
