Antenatal blood pressure for prediction of pre-eclampsia, preterm birth, and small for gestational age babies: development and validation in two general population cohorts

Study question Can routine antenatal blood pressure measurements between 20 and 36 weeks’ gestation contribute to the prediction of pre-eclampsia and its associated adverse outcomes? Methods This study used repeated antenatal measurements of blood pressure from 12 996 women in the Avon Longitudinal Study of Parents and Children (ALSPAC) to develop prediction models and validated these in 3005 women from the Southampton Women’s Survey (SWS). A model based on maternal early pregnancy characteristics only (BMI, height, age, parity, smoking, existing and previous gestational hypertension and diabetes, and ethnicity) plus initial mean arterial pressure was compared with a model additionally including current mean arterial pressure, a model including the deviation of current mean arterial pressure from a stratified normogram, and a model including both at different gestational ages from 20-36 weeks. Study answer and limitations The addition of blood pressure measurements from 28 weeks onwards improved prediction models compared with use of early pregnancy risk factors alone, but they contributed little to the prediction of preterm birth or small for gestational age. Though multiple imputation of missing data was used to increase the sample size and minimise selection bias, the validation sample might have been slightly underpowered as the number of cases of pre-eclampsia was just below the recommended 100. Several risk factors were self reported, potentially introducing measurement error, but this reflects how information would be obtained in clinical practice. What this study adds The addition of routinely collected blood pressure measurements from 28 weeks onwards improves predictive models for pre-eclampsia based on blood pressure in early pregnancy and other characteristics, facilitating a reduction in scheduled antenatal care. Funding, competing interests, data sharing UK Wellcome Trust, US National Institutes of Health, and UK Medical Research Council. Other funding sources for authors are detailed in the full online paper. With the exceptions of CM-W, HMI, and KMG there were no competing interests.


Prediction of blood pressure at different gestational ages from normograms
To prevent women who had many measurements of blood pressure during pregnancy from having too high an influence on the models, we randomly selected one blood pressure measurement per woman for any 2-week period where the woman had more than one measurement for inclusion in the models. This led to a median and interquartile range of 10 (9 to 11) blood pressure measurements per woman included in the models.
We included pre-pregnancy BMI in the multilevel models in four categories: underweight (<18.5 kg/m 2 ), normal weight (18.5-24.9 kg/m 2 ), overweight (25-29.9 kg/m 2 ) and obese (≥30 kg/m 2 ) as a covariate and also included an interaction between BMI category and each of the splines to allow the shape of the BP trajectory to differ by BMI category. Smoking (any smoking or never smoked) was also included as a categorical covariate and as an interaction with each of the splines. We fitted separate models for nulliparous and multiparous women.
The equation of the multilevel model for blood pressure change across pregnancy was:  spline  smoking  spline  smoking  spline  smoking  smoking   spline  obese  spline  overweight  spline  t  underweigh   spline  obese  spline  overweight  spline  t  underweigh   spline  obese  spline  overweight  spline  t  underweigh   spline  obese  spline  overweight  spline  t  underweigh   obese  overweight  t  underweigh   spline  u  spline  u  spline  u  spline  u where, y ij is the value of the i th MAP measurement on the j th individual, β 0 -β 24 describe the average trajectory of change, u 0j -u 4j describe how the j th individual's trajectory of MAP deviates from the average and GA ij is the gestational age in weeks of the i th measurement on the j th individual. It is centred at 12 weeks in order to set the intercept, β 0, to represent blood pressure at 12 weeks. The e 0ij and e 1ij terms describe the deviation of the i th measurement of MAP on the j th individual from the individual's trajectory. These are residual error terms.
The model was fitted separately for nulliparous and multiparous women rather than including parity as additional covariate, to allow for greater flexibility in the shape and variability of the trajectories for nulliparous and multiparous women.
The splines are defined as: (1) To calculate predictions conditional on the initial blood pressure measurement we used the multilevel models as above, and applied the method described by Tilling et al (2) and Pan and Goldstein. (3) The between-individual variance for individual j at the initial visit is: The within-individual variance for individual j at the initial visit is: The covariance between the deviations from the predicted curve at the initial visit (indexed 1) and g weeks for the jth individual is: V b1,gj= From these, the predicted blood pressure at gestational age g for individual j, conditional on the blood pressure value at the initial visit can be calculated as: Where y 1j is the observed value of the outcome (MAP) at the initial visit. Thus, the deviation of the initial measurement from the average trajectory was combined with the multilevel model information to predict what trajectory would be seen for the rest of gestation.

Multiple imputation of missing values in ALSPAC and the SWS
The same variables were included in multiple imputation models in both ALSPAC and SWS.
A separate imputation was done for blood pressure at each gestational age: 20, 25, 28, 31, 34 and 36 weeks, imputing only to the number of women who still had not delivered at each gestation. The variables included were either exposures/covariates to include in prediction models, outcomes or predictors of missingness. The variables included and information about how they were included in multiple imputation models are shown in the

Prediction Models for Preeclampsia:
The logistic regression models used to predict preeclampsia in ALSPAC are shown below. The risk of preeclampsia is calculated as where βX is as follows:

Prediction Models for Preterm Birth:
The logistic regression models used to predict preterm birth in ALSPAC are shown below. The risk of preterm birth is calculated as where βX is as follows: The logistic regression models used to predict small-for-gestational age in ALSPAC are shown below. The risk of small-for-gestational age is calculated as where βX is as follows:

Recalibration in SWS
Prediction models for preeclampsia at each gestational age were recalibrated for use in the SWS using the equation: Logit(probability of preeclampsia in SWS) = α recal + β recal (βX ALSPAC ) where βX ALSPAC are the parameters from the logistic regression in the ALSPAC cohort as detailed above and α recal and β recal represent the extent to which the intercept and the slope respectively of the regression model need to be altered for recalibration to the SWS cohort.
The values of α recal and β recal for each of the models at each gestational age are given in the table below.