Study population

Subjects for this nested case-control study were selected from a longitudinal birth cohort that was designed to identify predictors of preeclampsia in pregnant women who delivered at Brigham and Women’s Hospital in Boston, MA, from 2006-2008. Participants were recruited early in pregnancy (median, 10 weeks of gestation); each woman had completed demographic and medical history questionnaires and provided informed consent at enrollment. Gestational age was assessed by last menstrual period with verification by first-trimester ultrasound scan. Participants additionally provided urine samples from up to a total of 4 visits across gestation (median, 10, 18, 26, and 35 weeks of gestation). At delivery, detailed birth outcome and infant data were recorded. From this parent population, we selected all 130 women who delivered live singleton infants preterm or at <37 weeks completed gestation and 352 random control subjects. This study was approved by institutional review boards at the University of Michigan and Brigham and Women’s Hospital.

Women who delivered preterm were divided into subgroups. Women at delivery with spontaneous preterm labor or preterm premature rupture of the membranes (pPROM) were combined into a single group, because previous research has shown that women with these delivery precursors have similar patterns of placental inflammation. These births (n = 56) were considered spontaneous preterm. A second category included women whose preterm deliveries were determined to be a result of preeclampsia or intrauterine growth restriction, because these groups can be combined based on a similar causes of abnormal placentation. These births (n = 35) were considered placental preterm for analysis. Assignment of pregnancy outcomes was based on the criteria of the American College of Obstetricians and Gynecologists and standard clinical practice. The remaining PTBs (n = 39) did not fall into either causal-based subset (eg, followed repeated cesarean delivery or other maternal/fetal complications not listed earlier) and were not examined separately because of the lack of a hypothesized shared mechanism. Nevertheless, they were included in the primary analysis because unknown mechanisms may link oxidative stress to PTB overall, and this was an exploratory analysis.

Oxidative stress biomarker analysis

Urine samples (n = 1678 samples; n = 482 women) were stored at –80°C after collection until the time that oxidative stress biomarkers were measured. Both 8-OHdG and total 8-isoprostane were measured by Cayman Chemical (Ann Arbor, MI). For total 8-isoprostane, urine samples were hydrolyzed to deconjugate 8-isoprostane esterified to phospholipids and were passed through affinity columns for purification. Eluted samples were dried and resuspended in a buffer before measurement with enzyme immunoassay (EIA). The lower limit of detection was 3.9 pg/mL. For 8-OHdG, samples were diluted directly into buffer without purification. Concentrations were also measured with EIA, with a detection limit of 10.3 pg/mL. Levels of biomarkers below the limit of detection were replaced with the limit of detection divided by the square root of 2.

To account for urine dilution, specific gravity was measured in urine samples with the use of a digital handheld refractometer (Atago Co, Ltd, Tokyo, Japan). For examining biomarker distributions and variability, concentrations were corrected for specific gravity with the following formula: OS _c = OS([1.015 – 1]/[SG – 1]). OS _c represents the corrected biomarker concentration; OS is the uncorrected urinary concentration; 1.015 is the median specific gravity in all samples; and SG is the specific gravity of the sample. For regression analyses uncorrected concentrations were modeled, and specific gravity was included as a covariate. Extremely concentrated (specific gravity, >1.04) samples were excluded from all analyses (n = 4). Distributions of concentrations for both raw and corrected biomarkers were log-normal and ln-transformed for data analysis.

Statistical analysis

Analysis was performed using R version 3.0.2. Differences in biomarker levels by visit were tested with the use of linear mixed models with random intercepts only to adjust for intraindividual correlation, with biomarker regressed on visit of sample collection. To depict nonlinear trends in levels across gestation, generalized additive mixed models (GAMM; mgcv package in R) were created with the biomarker regressed on a smooth term for gestational age at urine sample collection, also with random intercepts only. Predicted values from GAMM models were plotted to show average trends over time. As an additional measure of variability in biomarker concentrations, we calculated intraclass correlation coefficients (ICCs), which represent the ratio of within-to-between individual variability.

Associations between oxidative stress markers and PTB were examined for overall PTB and also for the subtypes defined earlier. Odds ratios (ORs) were calculated with a geometric average biomarker concentration for each subject from levels measured at visits 1-3. Visit 4 concentrations were excluded from the average because the proportion of cases with a measure at this time point was low, because many women had already delivered (100% of cases had samples available at visit 1; 91% available at visit 2; 86% available at visit 3; 51% available at visit 4). Crude models were adjusted for specific gravity; full models included covariates that were associated with oxidative stress biomarkers in bivariate analysis that had been linked to PTB in previous studies. Covariates that were considered included maternal age, race/ethnicity, education level, health insurance provider, body mass index (BMI) at visit 1, use of tobacco or alcohol during pregnancy, parity, gender of the infant, and use of assisted reproductive technology.

Additionally, individual logistic regression models that examined urinary oxidative stress biomarkers at each visit in relation to PTB were constructed to investigate whether oxidative stress at a particular time point was more predictive of preterm delivery. These models were created with the use of the same covariates that were included in the average models.