Study population

This study is part of an ongoing prospective study being conducted by the Fetal Medicine Foundation, London, United Kingdom, for the first-trimester prediction of important fetal and obstetric disorders. Institutional review board project #02-03-033 approval was obtained on March 14, 2003. The details of patient evaluation and study methods have been extensively described in a prior report of metabolomic prediction of early-onset PE. A routine population of British women was prospectively screened from March 2003 through September 2009 and they all gave written consent to participate in the study, which was approved by the King’s College Hospital Research Ethics Committee. Briefly, women were recruited at 11 ⁺⁰ −13 ⁺⁶ weeks’ gestation. Maternal characteristics and medical history were documented and first-trimester ultrasound, including crown-rump length (CRL) and uterine artery Doppler pulsatility index (PI), was measured. Data collection was planned before laboratory testing. The lower of the left and right uterine artery Doppler PI value was used for PE prediction. Maternal serum samples were also obtained and stored at −80°C for subsequent laboratory analysis. The long-term objective of the project is to develop and evaluate new markers and existing biomarkers of PE. Apart from the previously published study on early-onset PE these cases represent the first metabolomic analysis of PE from this study population.

A total of 30 singleton pregnancies that subsequently developed late-onset PE requiring delivery >34 weeks formed the study group and were matched with 60 unaffected controls. These cases were not previously used in any prior publication. The late-onset PE cases were selected at random from our database of available stored samples. Each case of late-onset PE was matched with 2 controls who delivered a phenotypically normal neonate with appropriate birth weight for gestational age at term and did not develop any hypertensive disorder of pregnancy and who had blood collected within 3 days of assessment of the late-onset PE case. There was no evident source of bias in the selection of cases or controls. The definition of PE used was that proposed by the International Society for the Study of Hypertension in Pregnancy, namely systolic pressure ≥140 mm Hg or diastolic pressure ≥90 mm Hg on ≥2 occasions 4 hours apart >20 weeks of gestation, in women who were previously normotensive. Proteinuria was defined as a total of 300 mg in a 24-hour urine collection or 2 readings of at least 2 ⁺ proteinuria on a midstream or catheterized urine specimen in the absence of a 24-hour urine collection must also have been present in addition to the hypertension. Proteinuria must also have been present in addition to the hypertension for the diagnosis of PE. No HELLP syndrome or gestational hypertension cases were included.

Nuclear magnetic resonance (NMR) spectrometry was used for metabolite identification in the specimen samples. Sample preparation and NMR spectroscopy methods were performed as detailed previously using a 500-MHz Varian Inova NMR spectrometer (Varian Inc, Palo Alto, CA). Forty metabolites in each serum sample were identified and quantified in each case and control sample using commercial (NMR Suite 7.1; Chenomx Inc, Edmonton, Alberta, Canada) spectral fitting software containing an NMR spectral reference library of >200 compounds. On initial metabolomic analysis, readers were blinded to patient status.

Statistical analysis

Recommended statistical procedures for metabolomic analysis were followed. Log scaling was used for normalization of all metabolomic data. Principal component analysis (PCA) and partial least squares discriminant analysis (PLS-DA) were performed to identify patterns. PCA is an unsupervised classification technique for transforming a complex collection of data points such that the important properties of the sample can be more simply displayed along the X- and Y-axes. Two clusters on the PCA plot indicate that there are significant metabolite differences between the normal and control groups.

PLS-DA is used to enhance the separation between the groups by summarizing the data into a few latent variables that maximize covariance between the response and the predictors. To minimize the possibility that the observed separation on PLS-DA is due to chance, permutation testing was performed. This involved repeated (2000 times) data sampling, with different random labeling. A significant P value indicates that the separation observed between groups is very unlikely to be due to chance. The MetaboAnalyst computer program was used to perform (PCA and PLS-DA) analyses. A variable importance in projection (VIP) plot is a plot ranking the metabolites based on their importance in discriminating study from control groups. Metabolites with the highest values on X-and Y-axes are the most powerful group discriminators.

In comparing the concentrations of metabolites between groups, outlier testing was performed using Dixon Q test. The Dixon Q test is used for identification of outliers in the dataset and replaces that value with the one closest to it. Replacement of outliers helps to meet the assumption of normal distribution and equal variance between groups. Only a single value (for valine) was adjusted in this fashion, however. Kolmogorov-Smirnov and Shapiro-Wilk tests of normal distribution were performed. Metabolite concentrations in late-onset PE vs controls were compared using the 2-tailed t test. Mann-Whitney U test was used in comparing metabolite concentrations between groups that were not normally distributed. Other independent variables including fetal CRL, uterine artery Doppler, PI, and maternal age, parity, weight, ethnicity, smoking, and medical disorders were included in the genetic computing analyses along with the metabolite concentrations for PE prediction.

Genetic programming is a branch of evolutionary computing while genetic computing is a branch of genetic programming. The advantage of genetic computing lies in its ability to handle nonnormally distributed outcome measures and the large volume of data generated from “omics” laboratory analyses and it has previously been successfully used for metabolomic analysis. Genetic computing generates rules by which an optimal number of variables can be selected from a large number of exploratory variables for the optimal prediction of the outcomes of interest. The Gmax computer program, version 11.09.23 (Genetic Computing Consultants Limited, London, UK) was used for genetic computing analysis.

In addition, using a limited number of independent biochemical and demographic predictor variables, logistic regression was used to generate probability equations for the prediction of PE. Based on the regression equations generated, individual probability for developing late-onset PE was calculated. Receiver operating characteristics curves with sensitivity plotted against false-positive rates (1-specificity) were generated along with the 95% confidence interval and/or P values for the area under the curve. For the primary analysis, power analysis indicated that a minimum of 17 cases was needed in each group to have 80% power for a 2-sided P < .0001.

After the primary analysis, a normal group of 60 control cases previously reported for the early-onset PE metabolomic study derived from the same patient population was subsequently added to the current study group of 30 cases and 59 control patients to permit sufficient power to perform the regression analyses. Finally, we used PCA and PLS-DA analyses to determine whether late-onset PE could be differentiated from early-onset PE by directly comparing the 30 early-onset PE cases reported in a prior publication with the 30 late-onset PE cases in the current study.