Examining the role of neighborhood-level sociostructural indices in predicting spontaneous preterm birth in the New York City metropolitan area

Abstract

BACKGROUND:

Prior predictive models of spontaneous preterm birth that include cervical length and other individual covariates have achieved only moderate discriminative strength. Indices which reflect the sociostructural environment have been associated with preterm birth overall, but the extent to which these measures may improve prediction of spontaneous preterm birth among individuals with measured cervical length has not been tested.

OBJECTIVES:

To determine whether a cervical length model including measures of the neighborhood-level social environment increases spontaneous preterm birth discrimination compared with a cervical length model without such indices.

METHODS:

This was a retrospective cohort study of all gestations with (1) delivery between January, 2013 and August, 2023 and (2) at least one transvaginal cervical length measurement between 16 weeks and 0 days’ and 24 weeks and 6 days’ gestation. The primary outcome was spontaneous preterm birth. We ascertained neighborhood-level sociostructural determinants of health from the Index of Concentration at the Extremes (ICE) and Child Opportunity Index (COI). We used logistic regression to predict the probability of spontaneous preterm birth in models with (1) cervical length alone, (2) cervical length plus social indices, and (3) cervical length, social indices, and individual characteristics. We examined model discrimination using the area under the receiver operating characteristics curve (AUC) and 95% confidence intervals. We used the net reclassification index (NRI) to quantify the improvement in predictive ability between nested models.

RESULTS:

1884 pregnancies were included, of which 119 (6.3%) resulted in spontaneous preterm birth. A model with cervical length alone achieved an AUC of 0.66 (0.60, 0.72), increasing to 0.75 (0.70, 0.79) with individual covariates. Adding social indices yielded a model with an AUC of 0.75 (0.71, 0.80). The model with cervical length, individual covariates, and social indices had an NRI of 0.04 (0.01, 0.08), indicating improved performance compared with the previous models.

CONCLUSIONS:

Among individuals with cervical length screening, the inclusion of social indices improved the prediction of spontaneous preterm birth compared to a model with only individual-level covariates.

Video Abstract

Download video file ^{(4.7MB, mp4)}

Introduction

Preterm birth (PTB), defined as birth prior to 37 completed weeks’ gestation, occurs in 10.2% of all pregnancies.¹ PTB and low birthweight comprise the second leading cause of neonatal mortality after birth defects,² thus prediction and prevention are tantamount to reducing neonatal mortality. Nearly half of all PTBs are due to spontaneous preterm labor, and another 25% due to preterm prelabor rupture of membranes (PPROM).³ Multiple risk factors have been identified including low body mass index (BMI), short interval pregnancy, multifetal gestations, history of PTB, vaginal bleeding related to placental abruption or previa, smoking, and intrauterine infection.³ Specifically, a history of spontaneous preterm birth (sPTB) is associated with more than a three-fold increase in a person’s odds of sPTB in a subsequent pregnancy.⁴ Cervical length (CL) screening has been shown to accurately identify some, but not all, patients at greater risk of sPTB⁵ and is the strongest predictor of PTB currently available.³ The American College of Obstetricians and Gynecologists (ACOG) recommends routine CL assessment at 18w0d-22w6d for all patients using either a transabdominal or transvaginal ultrasound approach.⁶

Prior attempts to develop risk prediction models beyond CL have not yet achieved sufficient discrimination for clinical application,^6,7 nor have attempts to include other biological markers.⁸ Indices reflecting the built and social environment may aid in this prediction. The Index of Concentration at the Extremes (ICE), which reflects racialized and economic segregation, has previously been associated with risk of PTB.⁹ The Child Opportunity Index (COI) reflects the quality of structural resources and conditions in a neighborhood-level area and has been associated with PTB.^10,11 These indices are publicly available, geospatially-linked measures that could be integrated into clinical algorithms with simple inputs such as the patient’s residential address. We therefore aimed to understand their utility in sPTB risk prediction. Specifically, our goal was to examine if the inclusion of ICE and COI improved accurate classification of sPTB risk compared with baseline models using CL and other individual-level patient characteristics.

Materials and methods

Cohort selection

This was a retrospective cohort study of all singleton gestations with (1) delivery at a tertiary, urban academic hospital between January 2013 and August 2023, (2) at least one transvaginal CL measurement between 16w0d and 24w6d gestation, and (3) residential address in the New York City (NYC) metropolitan area and census tract for which both ICE and COI are available.

Exposures

Each patient’s ultrasound images were reviewed by the study team, during which transvaginal CL between 16w0d and 24w6d gestation were verified. If the CL on the images did not perfectly match what was recorded but the difference could be attributed to a rounding error, then the CL that most closely matched the recorded length was used. If the CL on transvaginal images otherwise did not match what was in the report, then the highest quality, shortest length was used for analysis, as is standard clinical practice. When multiple CL measurements were obtained during the course of a single pregnancy, the CL measured closest to 20w0d gestation was used for all analyses. All patients at our institution are offered transvaginal CL assessment at the time of the second trimester anatomy ultrasound regardless of risk factors for sPTB, providing a best point of comparison. Patients with a short cervix were managed with the usual standard of care based on their individual risk factors.

Abstracted individual-level data for inclusion in analyses included medical history, gestational age at delivery, progesterone therapy, and cerclage placement, abstracted from electronic medical records (EMR) by clinicians on the study team. Medical history specifically included maternal age, pre-pregnancy BMI, insurance type, substance use (tobacco, alcohol, and/or recreational drug use), progesterone prescription during pregnancy (vaginal suppository and/or 17-hydroxyprogesterone for intramuscular injection), the placement of a cervical cerclage, and history of prior sPTB.

We ascertained neighborhood-level social determinants of health (SDOH) from nationally validated, publicly available databases of social indices: Index of Concentration at the Extremes (indicator of racial-economic segregation, ICE)¹² and Child Opportunity Index (neighborhood resources, COI) 3.0,¹³ both based on patient address and 2020 census tract. ICE is a proxy for structural racism based on the American Community Survey that measures the proportion of wealthy White households relative to poor Black households in a given census tract. It is a continuous measure ranging from −1 to 1 (−1 represents neighborhoods with entirely poor, Black households; 1 indicates entirely wealthy, White households).¹⁴ The COI is a composite index of children’s neighborhood opportunity that contains data for every census tract in the United States (US) from 2012 to 2021. It includes 44 indicators across 3 domains: education, health and environment, and social and economic; scores range from 1 to 100.¹³ We defined the NYC metropolitan area as most of downstate New York, northern and central New Jersey, and western Connecticut. Patients with addresses outside this metropolitan area were excluded from our analyses. A list of all included zip codes is available in Supplemental Table 1.

Outcomes

The primary outcome was sPTB, which was defined as a vaginal or cesarean delivery occurring before 37 completed weeks’ gestation in the setting of preterm labor, PPROM, or prelabor placental abruption. This is used to define whether an individual has a history of prior sPTB. These outcomes were abstracted by the study team through review of the admission history and physical, delivery note, and other notes in the EMR.

Statistical analysis

The Mann Whitney U test was used to look for baseline difference between the mean CL, COI, and ICE value when stratified by the presence or absence of sPTB. We used logistic regression models to predict the odds of sPTB in four models: (1) CL only, (2) CL with individual characteristics, (3) CL plus social indices, and (4) CL, social indices, and individual characteristics. Individual-level predictors included the ones described above. We compared model discrimination using the area under the receiver operating characteristics curve (AUC). The classification table for the first model (CL alone) was reviewed and a predicted risk of sPTB of 5% was selected based on maximized sensitivity and specificity as the cutoff to classify patients as high or low risk for sPTB.

We used the net reclassification index (NRI) to quantify improvement in ability of risk classification between nested models. NRI is a statistical method developed from more general AUC comparisons to assess the relative predictive quality of different models.^15–17 Specifically, NRI applies each model to the dataset and compares how well a new model reclassifies cases (exposed) and controls (unexposed) relative to the outcome. Positive NRI values indicate that the new model improves classification, negative values indicate worsened classification, and NRI=0 indicates no net change in classification accuracy comparing nested models. The NRI used in conjunction with other discrimination measures can facilitate a more granular comparison of clinical prediction models.^16,18

A calibration decile plot was generated for our final model to look for evidence of systematic over- or underestimation of sPTB risk. Discrimination metrics such as the AUC are rank-order statistics that reflect a model’s ability to distinguish individuals at higher versus lower risk and are insensitive to systematic inaccuracies in absolute risk prediction. Calibration, on the other hand, refers to the agreement between model-estimated and observed event rates.¹⁹ We visualized calibration by dividing participants into deciles of predicted probabilities from the logistic regression model and plotted the mean predicted probability against the mean observed event rate within each decile group.²⁰

All analyses were performed using R version 4.4 (24 April 2024).

Missing data

Of the initial sample of 1,916 pregnancies, 21 (1.1%) were excluded due to missing data (either social indices or pre-pregnancy BMI). There were no missing data for the pregnancies included in the analytic sample.

Sensitivity analyses

Two prespecified sensitivity analyses were performed. The first examined CL as a dichotomous variable of short cervix or not, defined as CL <2.50 cm.^21–23 The second restricted the primary analyses to a high-risk subgroup of pregnancies with history of prior sPTB.

Ethics approval

This study was approved by our Institutional Review Board, the Program for the Protection of Human Subjects, on 18 November 2023 along with a waiver of informed consent and HIPAA authorization given the retrospective nature of the study.

Results

Figure 1 depicts the study enrollment process. We reviewed medical records and abstracted data for 1,916 unique pregnancies. We excluded seven cases with addresses outside the NYC Metropolitan Area, and four cases with invalid addresses that could not be geocoded. This yielded an analytic dataset of 1,905 unique pregnancies. We then excluded 20 cases without a matched neighborhood-level index value (ICE and/or COI), and one case without a pre-pregnancy BMI value. This resulted in 1,884 unique pregnancies in the analysis.

FIGURE 1. Cohort selection flow diagram.

Table 1 illustrates the demographic information of the study sample. The majority of individuals (>70%) were aged 35 years or less at the time of delivery. The majority of individuals (>70%) identified their race and ethnicity as non-White. 56% of cases were covered by private health insurance, compared with 44% who had public insurance coverage or no insurance recorded. The majority of cases (>60%) had a pre-pregnancy BMI below 30 kg/m². 15% of patients specifically had a prior sPTB. The rate of PTB (spontaneous or medically-indicated) was 11%, and the rate of sPTB was 6%. The rate of cesarean delivery was 34%.

TABLE 1.

Demographic and clinical covariates

Characteristic	Total=1884
	N	%
Age at Delivery, years
≤25	328	17.41
26–30	450	23.89
31–35	617	32.75
36–40	382	20.28
>40	107	5.68
Race and Ethnicity
Asian, Native Hawaiian, Pacific Islander	141	7.48
Black or African American	374	19.85
Hispanic or Latina	692	36.73
White	516	27.39
Other	138	7.32
Not Reported	23	1.22
Insurance Coverage
Private	1063	56.42
Public or None	821	43.58
Pre-Pregnancy Body Mass Index, kilograms/meter2
<18.5	43	2.28
18.5–24.9	676	35.88
25.0–29.9	579	30.73
30.0–34.9	337	17.89
35.0–39.9	141	7.48
≥40	108	5.73
Substance Use in Pregnancy	73	3.87
History of Preterm Birth	343	18.21
History of Spontaneous Preterm Birth	275	14.60
Prescribed Progesterone in Pregnancy	257	13.64
IVF Pregnancy	98	5.20
Cerclage Placed during Pregnancy	70	3.72
Delivery Outcomes
Gestational Age at Delivery, weeks
≤29	22	1.17
30–36	188	9.98
37–40	1244	66.03
>40	430	22.82
Preterm Birth	210	11.15
Spontaneous	119	6.32
Medically Indicated	91	4.83
Preterm Prelabor Rupture of Membranes	76	4.03
Mode of Delivery
Vaginal	1251	66.40
Cesarean	633	33.60
Placental Abruption	35	1.86

Table 2 shows the mean and standard deviation for CL, ICE, and COI, stratified by the outcome of sPTB. Patients who had a sPTB had a shorter average CL, a smaller average COI value, and a larger average ICE value.

TABLE 2.

Cervical length, child opportunity index, and index of concentration at the extremes stratified by spontaneous preterm birth

Characteristic	No sPTB mean (SD)	sPTB mean (SD)	P
CL (cm)	3.95 (0.69)	3.41 cm (1.01)	<.001
COI (scale 0 to 100)	47.68 (35.72)	39.55 (33.12)	0.01
ICE (scale −1 to 1)	−0.03 (0.38)	−0.13 (0.36)	.04

CL, cervical length; COI, Child Opportunity Index; ICE, Index of Concentration at the Extremes; sPTB, spontaneous preterm birth Range of CL: 0.00 cm−7.85 cm.

Table 3 shows the results for the various logistic regression models. CL was inversely associated with the risk of sPTB in all models. In the model with CL alone, the AUC was 0.66 (0.60, 0.72). The addition of ICE and/or COI, without individual covariates, marginally increased the AUC. The addition of individual covariates improved the AUC to 0.75 (0.70, 0.79) without social indices and to 0.75 (0.71, 0.80) with social indices. We examined collinearity among CL and individual-level predictors to ensure our findings were not inflated; the Variance Inflation Factors were <3.0 in all cases, suggesting low collinearity between predictor variables.

TABLE 3.

Area under the receiver operating curve for logistic regression models predicting spontaneous preterm birth, n=1884

	AUC	95% CI
Model 1: CL	0.66	0.60, 0.72
Model 2: CL + Individual-Level Predictorsa	0.75	0.70, 0.79
Models 3–5: CL + Social Indices
3: CL + COI	0.66	0.61, 0.72
4: CL + ICE	0.67	0.61, 0.72
5: CL + COI + ICE	0.67	0.61, 0.72
Model 6: CL + Individual-Level Predictors + Social Indices	0.75	0.71; 0.80

CL, cervical length; COI, Childhood Opportunity Index; ICE, Index of Concentration at the Extremes; AUC, area under the receiver operator curve.

^aIncludes: maternal age; pre-pregnancy BMI, insurance type, substance use, prescribed progesterone, cervical cerclage, and history of sPTB.

Net reclassification analysis showed that, in models with CL and additional individual-level covariates, the inclusion of ICE and COI improved classification by 4% (1.0%–8.0%), with a 3% (0.0–8.0%) increase in those correctly classified as high risk and no difference in those classified as low risk (1.0% [0.0%–2.0%]) (Table 4). We did not find evidence of systematic over- or under-prediction of sPTB risk when reviewing the decile calibration plot for model 6 (Supplemental Figure 1).

TABLE 4.

Net reclassification indices between selects models

	Reclassification Index	NRI (95% CI)
Model 1: CL alone vs Model 3: CL + COI	NRI	0.03 (0.01, 0.06)
	NRI+	0.03 (0.01, 0.06)
	NRI−	<0.01 (−0.01, 0.01)
Model 1: CL alone vs Model 4: CL + ICE	NRI	0.02 (0.01, 0.05)
	NRI+	0.02 (0.01, 0.04)
	NRI−	0.01 (0.00, 0.02)
Model 1: CL alone vs Model 5: CL + COI + ICE	NRI	0.03 (−0.01, 0.07)
	NRI+	0.03 (−0.01, 0.07)
	NRI−	<0.01 (−0.01, 0.02)
Model 2: CL + Individual-Level Predictors vs Model 6: CL + Individual-Level Predictors + COI + ICE	NRI	0.04 (0.01, 0.08)
	NRI+	0.03 (0.01, 0.08)
	NRI−	0.01 (<0.01, 0.02)

NRI, Net reclassification index; CI, confidence interval; COI, Child Opportunity Index; ICE, Index of Concentration at the Extremes; sPTB, spontaneous preterm birth; CL, cervical length.

Classification cut-off set at 5% risk of sPTB.

Results in bold are statistically significant.

The sensitivity analysis for short cervix as a dichotomous variable showed an AUC of 0.74 (0.68, 0.79) for the model including COI, ICE, and individual covariates (Supplemental Table2a). We examined the incremental benefit of social indices compared to short cervix alone and found that, overall, including social indices significantly improved classification of sPTB risk (NRI:13.0 –14.0%, [5.0–21.0%]). However, patterns differed by risk group, with inclusion of social indices improving classification of individuals at lower risk (NRI: 35.0–36.0%, [33.0–39.0%]) but resulting in poorer classification of those at high risk of sPTB (NRI: −22.0 to −23.0%, [−31.0 to −14.0%]) (Supplemental Table 2b).

In the subgroup analysis of gestations with history of prior sPTB, the AUC for the model including COI, ICE, and individual covariates was 0.68 (0.58, 0.76) (Supplemental Table 3a). The predictive models that included social indices showed improved classification of sPTB risk in this population when compared with those without social indices (NRI: 1.0–3.0% [0.0–5.0%]) (Supplemental Table 3b).

Comment

Principal findings

We developed prediction models using individual- and neighborhood-level predictors that showed good discrimination (AUC=0.75) of sPTB risk among births to women with a transvaginal ultrasound-measured CL in pregnancy. In net reclassification analyses, we observed a modest improvement in the proportion of individuals correctly classified as high-risk when models included social indices compared with models using individual-level predictors alone, despite an unchanged AUC.

Results in the Context of What is Known

Meertens et al. conducted a systematic review of potentially viable models to predict sPTB and attempted to externally validate viable ones. In that study, AUCs for sPTB ranged from 0.54–0.67 and models still performed poorly with evidence of overfitting even after recalibration.⁷ Our model and its predictive ability may benefit from the use of patient-level data that accounts for multiple risk factors for sPTB. The use of history of prior sPTB is particularly important since it is associated with a 3.6 times odds of recurrence in subsequent pregnancies.⁴ The model was strengthened by the inclusion of both COI and ICE. COI aiding in prediction aligns with work done by Belanoff et al., which showed an association when examining over 250,000 births among non-Hispanic Black and Hispanic patients in Boston, MA.¹⁰ Similarly, COI was shown to be associated with other adverse pregnancy outcomes beyond PTB, including hypertensive disorders of pregnancy, small for gestational age at birth, stillbirth, and gestational diabetes mellitus in a secondary analysis of the nuMoM2b cohort.¹¹ Phillips-Bell et al. published work in 2024 using birth certificate data in Florida showing ICE was also associated with PTB. Interestingly, when looking at subtypes of PTB, ICE was associated with early and medically-indicated PTB but neither late PTB nor sPTB.⁹ Inclusion of ICE in our model improved predictability of sPTB, which may reflect regional differences in obstetric populations or health care systems between NYC and Florida. ICE may be a more valid predictor of health outcomes in densely populated, racially diverse areas, such as NYC, compared with other regions of the US. Similarly, COI reflects resources in a specific built environment, which may indicate a relationship between stress secondary to poor resources and gestational outcomes.

Clinical implications

These findings are practically useful for policy and clinical stakeholders interested in implementing prediction models in routine clinical practice. We found that the addition of neighborhood-level social indices improved the classification of sPTB risk beyond what is achieved by individual-level clinical predictors. This reinforces the importance of accounting for sociostructural determinants in predicting adverse obstetric health outcomes. We also demonstrated that parsimonious prediction models including one neighborhood-level index or the other (either COI or ICE), in addition to CL, improved the classification of patients at high risk for sPTB. This finding is important since neighborhood-level indices can be easily ascertained from the patient address via publicly available data. These are variables that are readily available at the time of midtrimester CL screening and therefore may present a low touch means of incrementally improving sPTB risk stratification. From a population standpoint, our results provide policy makers an opportunity to design social programs that target the socioeconomic factors reflected in COI and ICE.

More precise identification of sPTB risk factors may improve counseling provided to patients. Future work is needed to validate this model prospectively in external populations and determine interventions additional to the current standard of care that may help reduce risk of PTB in high-risk patients. Future work should also distinguish between relationships with different etiologies of sPTB; our study did not differentiate assess sPTB due to PTL vs PPROM vs placental abruption, so our results cannot speak to the biomechanical processes underlying this risk.

Research implications

Prediction studies often yield high-dimensional, complex models that are difficult to deploy at the clinical bedside. Ascertaining multiple patient-level characteristics for prediction model algorithms requires time and personnel for chart abstraction or patient interview in busy clinical environments. Further, there are often important differences in predictor variable definitions and availability comparing model development data sets to the data elements available in a given clinic or hospital setting, which can compromise predictive performance when models are deployed in real-world practice.²⁴ Future work in this regard should optimize the integration of social indices into clinical prediction and practice, thereby allowing providers to more easily employ the improved predictive strength from social indices. Further study could directly compare area-level social indices with specific individual-level indicators of socioeconomic position such as type of insurance to identify parsimonious strategies to account for social determinants of health in obstetric prediction tasks.

Strengths and limitations

The incidence of sPTB in our study population was close to that of the national rate, supporting generalizability to other US settings. CL was abstracted manually from images of transvaginal ultrasonographic measurements–overseen by clinicians in maternal-fetal medicine–minimizing errors that could stem from transcription of reports and ensuring quality of the images themselves. Similarly, the manual chart review for clinical characteristics and outcomes bolstered the accuracy of individual-level clinical data. The use of net reclassification analysis allowed us to compare nested models and quantify the incremental benefit of additional predictors as opposed to comparing model discrimination based on AUC alone.

We used a single CL measurement in prediction models and did not account for the number of CL measurements, including those who had serial CL measurements due to a history of sPTB. Further research should investigate improvements in ascertaining sPTB risk from longitudinal CL measures across pregnancy. Our sample comprised patients seen at a single institution in New York City, which limits the generalizability to other settings. Our sample was also majority non-White, which may limit the generalizability to other populations here in the US. However, our results reflect a population enriched for higher PTB risk, given racial and ethnic disparities in PTB²⁵ and the likelihood that patients with a history of sPTB receive routine transvaginal cervical length screening. Additionally, though transvaginal CL screening is offered to all patients at our institution, not all patients consent to it, so there might be differences between those patients and those who do for which we cannot account. During our study period 17-hydroxyprogesterone was both introduced into the care of patients with history of sPTB and subsequently removed from the market. We included all types of progesterone when collecting and analyzing our data but we are unable to determine if progesterone type would impact these results. Similarly, ACOG released its guidance on cerclage placement in 2014, which could impact which patients were offered cerclage in the first year of our cohort vs the other nine.²⁶

We acknowledge that the NRI value was small comparing models including social indices to those with individual-level predictors alone, suggesting modest clinical significance of the model improvement. However, we believe that small changes in predictive accuracy for this adverse outcome are notable, and we evaluated categorical reclassification improvement with respect to a clinically meaningful risk threshold and report both event and non-event NRI components in conjunction with AUC for full description of discrimination across our nested models.²⁷ Finally, our models were not externally validated with an independent cohort of pregnancies; such validation will further improve the understanding of sPTB risk.

Conclusions

We built a novel prediction model for sPTB with an acceptable AUC that was improved by the addition of COI and ICE. Prospective validation of this model is needed to show consistent performance prior to clinical application. Future research should explore additional, diverse sociostructural indices alongside individual clinical characteristics to improve sPTB risk prediction.

Supplementary Material

Supplementary material associated with this article can be found in the online version at doi:10.1016/j.ajogmf.2025.101837.

AJOG MFM at a Glance.

Why was this study conducted?

To test whether the inclusion of neighborhood-level social indices improves the strength of cervical length in predicting spontaneous preterm birth.

Key findings?

This study yielded a predictive model which incorporated neighborhood-level social indices, adequately predicted risk of spontaneous preterm birth, and improved on existing models that use cervical length surveillance.

What does this add to what is known?

This study identifies publicly available, geographically linked measures of social determinants of health that improve the capacity of cervical length-based prediction to identify risk of spontaneous preterm birth.