Performance of a validated spontaneous preterm delivery predictor in South Asian and Sub-Saharan African women: a nested case control study

Abstract

Objectives

To address the disproportionate burden of preterm birth (PTB) in low- and middle-income countries, this study aimed to (1) verify the performance of the United States-validated spontaneous PTB (sPTB) predictor, comprised of the IBP4/SHBG protein ratio, in subjects from Bangladesh, Pakistan and Tanzania enrolled in the Alliance for Maternal and Newborn Health Improvement (AMANHI) biorepository study, and (2) discover biomarkers that improve performance of IBP4/SHBG in the AMANHI cohort.

Study design

The performance of the IBP4/SHBG biomarker was first evaluated in a nested case control validation study, then utilized in a follow-on discovery study performed on the same samples. Levels of serum proteins were measured by targeted mass spectrometry. Differences between the AMANHI and U.S. cohorts were adjusted using body mass index (BMI) and gestational age (GA) at blood draw as covariates. Prediction of sPTB < 37 weeks and < 34 weeks was assessed by area under the receiver operator curve (AUC). In the discovery phase, an artificial intelligence method selected additional protein biomarkers complementary to IBP4/SHBG in the AMANHI cohort.

Results

The IBP4/SHBG biomarker significantly predicted sPTB < 37 weeks (n = 88 vs. 171 terms ≥ 37 weeks) after adjusting for BMI and GA at blood draw (AUC= 0.64, 95% CI: 0.57–0.71, p < .001). Performance was similar for sPTB < 34 weeks (n = 17 vs. 184 ≥ 34 weeks): AUC = 0.66, 95% CI: 0.51–0.82, p = .012. The discovery phase of the study showed that the addition of endoglin, prolactin, and tetranectin to the above model resulted in the prediction of sPTB < 37 with an AUC= 0.72 (95% CI: 0.66–0.79, p-value < .001) and prediction of sPTB < 34 with an AUC of 0.78 (95% CI: 0.67–0.90, p < .001).

Conclusion

A protein biomarker pair developed in the U.S. may have broader application in diverse non-U.S. populations.

Keywords:

• Preterm birth
• biomarkers
• low- and middle-income countries
• IBP4
• SHBG

Introduction

Preterm birth (PTB) affects approximately 15 million infants annually, about 11% of all live births worldwide [1]. Globally, PTB and related complications are the leading causes of neonatal deaths (35%) [2,3] and of deaths in children under five years [4]. Surviving preterm infants may experience significant morbidities such as chronic lung disease, hearing and visual impairments, neurodevelopmental disabilities [5], and chronic disease in adulthood [6,7]. The familial and economic burden of PTB is substantial [8].

The incidence of PTB ranges from approximately 5% in some European countries to 18% in certain African countries [9]. Worldwide, more than 60% of all PTBs occur in sub-Saharan Africa and South Asia [9]. The majority of studies identifying biomarkers predictive of spontaneous PTB (sPTB) may lack sufficient clinical performance, suffer from small subject numbers, or both [10], and have been conducted in high-income countries [11,12]. However, the underlying risk factors and causes of sPTB may differ in low-and middle-income countries (LMICs). Nevertheless, despite different upstream causes, there may be downstream pathway convergence, which may enable predictive tests developed in high-income countries to perform well in LMICs.

The World Health Organization coordinated the Alliance for Maternal and Newborn Health Improvement (AMANHI) study, involving about ten thousand pregnant women in three sites of South Asia and sub-Saharan Africa (Sylhet, Bangladesh; Karachi, Pakistan; and Pemba Island, Tanzania) [13]. The objectives of the AMANHI study were to establish a biobank and to identify biomarkers of adverse pregnancy outcomes in developing countries [13]. The study plan included steps to evaluate candidate biomarkers identified in high-income countries and to conduct novel discovery studies [13].

A maternal serum-based proteomics predictor of sPTB < 37 weeks was validated in a United States (U.S.) cohort with maximal performance in serum collected 19^1/7–20^6/7weeks gestation [12]. The predictor was comprised of a ratio of two proteins, insulin-like growth factor-binding protein 4 (IBP4) and sex hormone-binding globulin (SHBG) and demonstrated better performance in a stratified BMI range of > 22 – ≤ 37 [12]. Recently, these biomarkers were shown to predict very early PTB (< 32 weeks) from spontaneous and medically indicated causes [14]. IBP4/SHBG also predicted neonatal morbidity and length of neonatal hospital stay, suggesting sensitivity to determinates of neonatal outcomes [14].

The objectives of this study were to: (1) verify the performance of the IBP4/SHBG biomarker in the AMANHI study cohort, and, (2) discover novel classifier proteins that improve the performance of IBP4/SHBG in the AMANHI study cohort. Exploratory analyses were conducted to identify additional novel proteomic and clinical variable biomarkers of sPTB across all three geographies combined, with validation planned utilizing future cohorts.

Materials and methods

Study design, settings, and participants

Between 2014 and 2018, the AMANHI biobanking study prospectively enrolled 10,001 pregnant women, identified through population-based surveillance in three countries: Bangladesh, Pakistan, and Tanzania [13]. Trained community health workers visited all women of child-bearing age in the study areas every two to three months to identify pregnancies and obtain informed consent. GA was determined using ultrasound before 20 weeks of gestation using standardized measurements [13]. Community health workers made four antepartum (8–19 weeks, 24–28 weeks, 32–36 weeks, and 38+ weeks of gestation) and two postpartum home visits to collect background characteristics, previous medical history, risk factors, exposures, outcomes, and morbidity for the index pregnancy. BMI was calculated from maternal height and weight measured at the enrollment visit. Maternal blood was collected and processed using a standardized protocol and serum samples were stored at −80^°C. De-identified samples were shipped to the U.S. via courier in a liquid nitrogen dry shipper.

Selection of cases and controls

In developing the protocol for this study, a power and sample size analysis determined that 32 cases and 64 controls per site achieved 91% power to distinguish an AUC = 0.7 from AUC = 0.5 (random performance). Combining three sites, the case-control study comprised 300 subjects (100 sPTB cases < 37 weeks of gestation and 200 control term deliveries ≥ 37 weeks) enrolled in 2014–2016. Inclusion criteria included the ability to consent, singleton pregnancy, and serum collection within 17^0/7 and 19^6/7weeks. Exclusion criteria included signs/symptoms of preterm labor at the time of specimen collection, major fetal anomaly, blood transfusion during the current pregnancy, use of progesterone after 12^6/7weeks gestation, use of heparin, or serum hemolysis > 100 mg/dl. Two-term births per case matched by the gestational week of blood draw and site were selected randomly from qualifying and available samples: Bangladesh (36 sPTB / 72 term), Pakistan (23 sPTB / 46 term), and Tanzania (40 sPTB / 80 term). One case and one control from Bangladesh were excluded from analyses because the case sample did not show pregnancy-specific proteins, and the control sample was drawn in week 16.

Laboratory methods

De-identified samples were received blinded, randomized, and processed in a CLIA-certified laboratory according to a standard operating procedure [12,15]. Briefly, serum samples were depleted of high abundance proteins, trypsinized, fortified with stable isotope-labeled internal standard (SIS) peptides, desalted, and analyzed by coupled liquid chromatography-multiple reaction monitoring mass spectrometry (LC-MRM-MS) measuring 122 proteins associated with pregnancy, of placental origin, or for quality control. Peptides were quantified as the response ratio between endogenous and SIS peak area counts. Quality was assessed for each batch [12,15] and overall.

Statistical analyses

Significant differences (p < .05) in demographics and clinical variables between the U.S. validation and the AMANHI cohorts were determined using a t-test (means) or a Wilcoxon test (medians) for continuous variables and the Fisher’s Exact test for categorical variables, with missing values excluded from analyses [16,17].

IBP4/SHBG biomarker scores were calculated as described [12,14]. As prespecified in the data analysis plan, because subjects were largely outside of the intended use in geography, anthropometrics and GA at blood draw, emphasis was on confirmation of the IBP4/SHBG biomarker after controlling for these differences. For validation in the AMANHI cohort, we tested the prediction of sPTB using logistic regression with models comprised of the IBP4/SHBG biomarker with and without adjustment for GA at blood draw and BMI. The appropriateness of the assumption of linear relationships was assessed by calculating average marginal effects [18]. Predictive performance was reported by the area under the receiver operator curve (AUC) with prespecified direction (cases > controls) and 95% confidence intervals (CI) calculated by DeLong’s method [19]. A Wilcoxon one-sided test was used to calculate p-values. Subjects with missing BMI values were omitted, although imputation of missing BMI values using Multivariate Imputation by Chained Equations [20] yielded similar results.

Table 1. Comparison of AMANHI and U.S. validation cohorts.

	AMANHI (N = 298)	U.S. Validation Cohort [12] (N = 54)	p-value
Maternal age (years)
Mean (SD)	26.1 (5.97)	26.2 (6.25)	.90
Median [Q1, Q3]	25.0 [21.0, 30.0]	24.5 [21.3, 30.0]	.94
Body mass index (kg/m^2)
Mean (SD)*	21.8 (4.19)	27.7 (6.22)	<.001
Median [Q1, Q3]*	20.8 [18.6, 24.0]	26.5 [22.3, 31.3]	<.001
Gravida
Multigravida	235 (78.9%)	41 (75.9%)	.60
Primigravida	63 (21.1%)	13 (24.1%)
Prior PTB^+ (Multigravida)
No	225 (95.7%)	34 (82.9%)	.006
Yes	10 (4.2%)	7 (17.0%)
Gestational age at blood draw (days)
Mean (SD)	128 (6.39)	140 (4.20)	<.001
Median [Q1, Q3]	127 [122, 134]	139 [135, 144]	<.001
Gestational age at birth (days)
Mean (SD)	267 (18.4)	265 (24.2)	.44
Median [Q1, Q3]	271 [256, 281]	273 [256, 281]	.90

*Missing BMI values omitted from calculations: AMANHI cohort, n = 39; U.S. cohort, n = 1.

⁺Prior PTB is restricted to prior sPTB in the U.S. validation cohort. Prior PTB is not limited by the timing of initiation of delivery in the AMANHI cohort.

To improve the performance of the baseline predictor (IBP4/SHBG + GA at blood draw * BMI), causal inference network analysis [21,22] was used to select additional proteins (log-transformed response ratios) and clinical variables as nodes directly causal of sPTB (Supplemental). Candidate proteomic and clinical variables were combined with the baseline predictor in logistic regression models as above. Significant classification performance improvement was defined as an AUC greater than the upper 95% confidence bound of the AUC of the baseline predictor (base R, stats package ≥ 4.0.3, [23]).

Prediction of early sPTB was assessed without exclusion or overrepresentation of late sPTBs or early-term births [24]. For example, to predict sPTB < 34 weeks, subjects delivering ≥ 34 weeks were defined as controls and adjusted to their natural rate in the AMANHI population. To minimize bias, adjustment was repeated 100 times and median AUCs were reported.

For Kaplan-Meier analysis, subjects were divided into lower and higher risk groups based on percentile thresholds from 5^th to 95^th, in 5% increments. GA at birth was used as the time variable, and significance was assessed by the log-rank test.

Ethics approval

The study protocol was approved by the following ethics committees: WHO Institutional Review Board (IRB), International Centre for Diarrhoeal Disease Research Bangladesh (icddr,b) in Bangladesh, Aga Khan University in Pakistan, Zanzibar Medical Research and Ethics Committee in Tanzania, and the IRB of Johns Hopkins Bloomberg School of Public Health, U.S.A.

Results

Comparison of AMANHI and U.S. Validation cohorts

Clinical characteristics of the AMANHI cohort were compared to the original U.S. validation study cohort [12] (Table 1). The mean GA at blood draw was significantly different between the two studies (128 vs. 140 d, p < .001). The optimal blood draw window for the IBP4/SHBG biomarker was 19^1/7−20^6/7weeks gestation, while the AMANHI samples spanned weeks 17^0/7–19^6/7. The mean BMI of the AMANHI cohort was significantly lower than the U.S. cohort (21.8 kg/m² vs. 27.7 kg/m², p < .001), with 155 of 259 subjects with recorded BMI falling below the optimal U.S. BMI (> 22 to ≤ 37 kg/m²) [12]. The proportion of AMANHI subjects with a prior PTB was lower than in the U.S. cohort (Table 1). However, because AMANHI data collection for prior preterm birth was based on recall, the prevalence may be underestimated. There were no significant differences in maternal age, gravidity, or GA at delivery between the cohorts (Table 1).

Performance of the validated IBP4/SHBG sPTB predictor in the AMANHI cohort

IBP4/SHBG is influenced by GA and BMI [12], and SHBG blood levels are associated with BMI [25]. Thus, because the U.S. and AMANHI populations had significantly different blood draw windows and BMI, without a population adjustment for these variables the IBP4/SHBG biomarker score did not reach statistical significance (p = .069, Table 2). However, with adjustment for GA at blood draw and BMI, IBP4/SHBG significantly classified sPTB subjects (AUC = 0.64, 95% CI: 0.57–0.71, p < .001, Table 2).

Table 2. Comparison of three sPTB predictive models containing IBP4/SHBG.

	sPTB < 37					sPTB < 34
Model	AUC	95% CI	p-value	Cases (n)	Controls (n)	AUC	95% CI	p-value	Cases (n)	Controls (n)
U.S. validated predictor: IBP4/SHBG	0.55	0.48–0.62	.069	99	199	0.62	0.49–0.75	.044	19	214
Adjusted for demographic differences: IBP4/SHBG + GABD*BMI	0.64	0.57–0.71	<.001	88	171	0.66	0.51–0.82	.012	17	184
Discovery phase model: IBP4/SHBG + GABD*BMI + (EGLN x PRL) / TETN	0.73	0.66–0.79	<.001	88	171	0.78	0.67–0.90	<.001	17	184

GABD – gestational age at blood draw; GABD*BMI denotes the addition of the main effects of GABD and BMI plus their interaction (product): GABD + BMI + GABD:BMI.

To test performance in classifying early sPTB, subjects with sPTB < 34 weeks were compared to control births (≥ 34 weeks). Importantly, the baseline predictor (IBP4/SHBG, GA at blood draw and BMI) also significantly classified early (< 34 weeks) sPTB (AUC = 0.66, 95% CI: 0.51–0.82, p = .012).

Discovery of novel predictors for the AMANHI cohort

To discover improved sPTB prediction in AMANHI geographies, we used artificial intelligence network techniques to select new features. This conditional correlation network analysis identified direct antecedents of sPTB: primigravida, prior PTB, and twelve proteins, in addition to IBP4/SHBG, GA at blood draw, and BMI (Supplemental Figure). These top features were added to the baseline IBP4/SHBG predictor individually, in pairs, or in triplets. AUC was significantly improved over the baseline predictor only with the addition of three proteins. Extra clinical variables did not significantly improve performance.

The top predictor ranked by AUC included three new analytes: endoglin (EGLN) and prolactin (PRL) were positively associated with sPTB, and tetranectin (TETN) was negatively associated, forming a second ratio in the algorithm, (EGLN x PRL)/TETN (Table 2). Performance for predicting sPTB < 37 or sPTB < 34 in this discovery phase predictor was AUC = 0.72, p < .001 and AUC = 0.78, p < .001, respectively (Table 2).

The subjects were then stratified into low- and high-risk groups at an 85^th percentile threshold, where 15% of subjects would be deemed higher risk. A Kaplan-Meier analysis indicated that subjects in the high-risk group (the top 15%) delivered significantly (p < .001) earlier than those in the lower-risk group (Figure 1). Significant separation was also seen from 95^th (5% at higher risk) to 15^th (85% higher risk) percentile thresholds (p < .001–.028).

Figure 1. Kaplan-Meier analysis of GA at delivery for lower- and higher-risk groups. Time-to-event analysis shows the rate of births for the top 15% of scores (higher risk) in the discovery phase predictor (IBP4/SHBG + GA at blood draw * BMI + [EGLNxPRL]/TETN) vs. low scores (lower risk). Vertical lines indicate delivery at 34 and 37 weeks.

Discussion

We confirmed that a U.S.-validated proteomics predictor of sPTB could be applied in LMICs, after adjusting for expected demographic differences between the populations. In the subsequent discovery phase of this study, a predictor of sPTB < 37 weeks including three new proteins showed improved predictive performance.

The AMANHI cohort is different from the U.S. cohort. The mean BMI for AMANHI subjects was below the optimal BMI range identified in the U.S. cohort [12]. AMANHI blood samples were drawn in weeks 17–19 of gestation, whereas the U.S. test was validated for weeks 19–20 gestation and demonstrated dependence on the blood draw period [12]. Nevertheless, the IBP4/SHBG biomarker, when adjusted for these differences, significantly identified sPTB subjects. Importantly, the adjusted predictor performed well for the prediction of sPTB < 37 and < 34 weeks. As adverse outcomes are inversely related to GA [26], it is critical that a sPTB predictor be able to identify those patients destined for early delivery.

An AMANHI discovery phase predictor comprised of IBP4/SHBG, GA at blood draw, BMI, EGLN, PRL, and TETN demonstrated improved classification performance for both sPTB < 37 and < 34 weeks. The improved predictor significantly stratified patients by GA at delivery over a wide range of percentiles thresholds, demonstrating potential flexibility in implementation.

Optimal timing of prognostic test administration requires balancing the need for timely intervention with the ability to access women seeking care. Administration of a second-trimester test may appropriately address this balance in LMICs, where few women seek prenatal care in the first trimester [27–29]. Together, flexibility in stratification and timing of blood draw demonstrates that these biomarkers may be suitable for the development of a clinically useful sPTB diagnostic test applicable across LMIC geographies.

The biological plausibility of the IBP4/SHBG biomarker has been discussed elsewhere [12]. Briefly, SHBG’s decreased abundance in the second trimester in women who subsequently develop sPTB [12] or preeclampsia [30], a major medical indication for PTB, may result from pro-inflammatory signals [31]. Decreased SHBG levels would be predicted to result in increased levels of free estrogens that oppose progesterone and deliver pro-labor signals [12]. Insulin-like growth factor (IGF) signaling pathways have been implicated as key regulators of placental development and fetal nutrient programming [32]. Higher maternal serum levels of IBP4, a key regulator of IGF2 bioavailability in the placenta bed [33], are associated with growth-restricted fetuses [34] and sPTB [12].

In the AMANHI cohort, we observed that median levels of both EGLN and PRL were elevated in sPTB relative to term, whereas TETN levels were decreased. Endoglin, a transmembrane coreceptor for transforming growth factor-beta (TGFβ) [35], regulates differentiation, cell migration, and angiogenesis [36,37]. EGLN is expressed in the placenta [38], where it inhibits trophoblast migration and invasion [39]. Elevated circulating levels of soluble EGLN (sEGLN) are associated with preeclampsia (PE) [40,41] and may serve as a biomarker to predict PE, particularly in combinations with other angiogenesis factors [42] and uterine artery doppler lowest pulsatility index [43,44]. Soluble EGLN levels were also elevated in women delivering infants who are small for gestational age [44–46], preterm [46,47], or in amniotic fluid from pregnancies complicated by intraamniotic infection [48]. However, the significance of sEGLN to predict sPTB, as opposed to PE, from serum at 17–19 weeks gestation is an unexpected finding of this study.

Prolactin, a pituitary growth factor responsible for the development of mammary glands and milk production, increases 10–20-fold in pregnancy [49]. Its expression in decidua during pregnancy [50], and reported pleiotropic activities including immune-modulation [51], regulation of insulin resistance by facilitating the transport of glucose and other nutrients across the placenta [52], and placental angiogenesis [53], suggest important roles in pregnancy health. Circulating and urine PRL levels (full length and anti-angiogenic fragment) were higher in severe vs mild PE, and predicted adverse maternal and fetal outcomes such as small for gestational age [54]. Cervicovaginal fluid PRL was more detectable in women symptomatic of preterm delivery than in asymptomatic women [55]. A systematic review and meta-analysis of potential sPTB biomarkers found that cervicovaginal PRL was one of three biomarkers out of 30 meeting inclusion criteria with a high (> 10) positive likelihood ratio [10]; however, its utility to predict sPTB as a blood-based biomarker has not been reported to our knowledge.

TETN has been implicated in extracellular matrix remodeling and fibrinolysis via interactions with plasminogen [56] and fibrin [57]. Lower levels of serum/plasma TETN are associated with various disease states including cancer, particularly metastatic disease [58], arthritis [59,60], heart failure [61], and PE [62]. Exosomes derived from tumor cells over-expressing TETN reduced VEGF secretion and inhibited angiogenesis [63]. In both amniotic fluid and fetal serum, the correlation between TETN levels and gestational age was seen, suggesting a role in fetal maturation [64]. TETN is reported to be negatively regulated by TGFβ [65], a pathway of importance in decidualization and placentation [66–68]. By extension, TETN may be involved in trophoblast invasion. However, direct evidence for the role of TETN in preterm birth is lacking.

Studies examining the proteomics of sPTB in LMICs with a high burden of prematurity are limited. A recent study by Jehan et al. utilized a multi-omics approach to discover plasma and urine biomarkers of sPTB from AMANHI and Global Alliance to Prevent Prematurity and Stillbirth (GAPPS) studies in a combined cohort of 81 subjects [69]. Interestingly, even though some of the geographies differed between the studies, both highlight the prognostic potential of proteomics and the importance of inflammatory and glucose homeostasis pathways.

Our study had many strengths. The AMANHI study design included a large population-based cohort with early gestational dating conducted by trained professionals. Additionally, serum, sociodemographic, and pregnancy characteristics were collected in a harmonized manner across all three sites. In the completely independent AMANHI cohort, following a pre-specified process, we tested a previously discovered and validated proteomics predictor and established its validity in this LMIC population.

Importantly, the analyses of prediction of early sPTB < 34 weeks should be interpreted with caution, due to small subject numbers. As well, the discovery phase predictor encompassing novel proteins requires further validation. This study was not sufficiently powered to allow for subset analyses. Future studies include exploration of the pathways leading to sPTB in different geographies.

Conclusions

We demonstrated that a serum protein predictor that was discovered, verified, and validated in the United States can predict sPTB in LMICs. Patient characteristics and timing of blood draw may be useful considerations when developing and applying a predictive test to a new geography.

Geolocation

The locations in this study included: Sylhet, Bangladesh (24.89904° N, 91.87198° E); Karachi, Pakistan (24.8607° N, 67.0011° E); and Pemba Island, Tanzania (5.0319° S, 39.7756° E).

Acknowledgments

We acknowledge the study participants for their essential contributions of time, care, samples and data; the dedicated field and data teams for implementing the AMANHI study; the Sera Clinical Laboratory for data generation; Max T. Dufford, BS, for data management; ChienTing Hsu, MS, Ryan M. Treacy, BS, and Jeff S. Flick, PhD, for preliminary analyses; and Todd L. Randolph, MD and Gregory C. Critchfield, MD, MS for critical review.

Disclosure statement

JJB, JB, TCF, ACF, MBB, ADP, and DEH are stockholders and employees or consultants of Sera Prognostics. JJB, JB, ADP, TCF, MBB and DEH have patents issued and pending related to this work. All other authors report no conflict of interest.

Data availability statement

Deidentified individual participant data will be made available upon approval of the manuscript, including data dictionaries and data that underlie the results reported in this article. Data will be shared with researchers who contact either of the corresponding authors requesting use of the data for research on preterm birth in LMICs. Requestors will need to sign a data access agreement.

Additional information

Funding

This work was funded by the Bill and Melinda Gates Foundation, including a grant to Sera Prognostics (Contract ID OPP1127876) and coordinated by the World Health Organization. The funders had no roles in designing of the biobank study, data collection, and data analysis.