Mini-combined test compared with NICE guidelines for early risk-assessment for pre-eclampsia: the SPREE diagnostic accuracy study

Article history

The research reported in this issue of the journal was funded by the EME programme as project number 14/01/02. The contractual start date was in February 2016. The final report began editorial review in September 2018 and was accepted for publication in June 2020. The authors have been wholly responsible for all data collection, analysis and interpretation, and for writing up their work. The EME editors and production house have tried to ensure the accuracy of the authors’ report and would like to thank the reviewers for their constructive comments on the final report document. However, they do not accept liability for damages or losses arising from material published in this report.

Copyright statement

© Queen’s Printer and Controller of HMSO 2020. This work was produced by Poon et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

2020 Queen’s Printer and Controller of HMSO

Background

Pre-eclampsia, which affects 2–3% of pregnancies, is a major cause of maternal and perinatal morbidity and mortality. 1–4 Pre-eclampsia can be subdivided into preterm pre-eclampsia, requiring delivery before 37 weeks’ gestation, and term-pre-eclampsia, with delivery at ≥ 37 weeks’ gestation. Preterm pre-eclampsia is associated with a higher incidence of adverse outcome. 5 Globally, 76,000 women and 500,000 babies die each year from pre-eclampsia.

Obstetricians managing women with preterm pre-eclampsia are faced with the challenge of balancing the need for achieving fetal maturation in utero with the risks to the mother and fetus from continuing the pregnancy longer. These risks include progression to eclampsia, development of placental abruption and haemolysis, elevated liver enzymes, low platelets (HELLP) syndrome. On the other hand, preterm delivery is associated with higher infant mortality rates and increased morbidity resulting from babies born small for gestational age (SGA), thrombocytopenia, bronchopulmonary dysplasia, cerebral palsy and an increased risk of various chronic diseases in adult life. Women who have experienced pre-eclampsia may also face additional health problems in later life, as the condition is associated with an increased risk of death from future cardiovascular disease, hypertension, stroke, renal impairment, metabolic syndrome and diabetes.

Objectives of the SPREE study

The objectives of the screening programme for pre-eclampsia (SPREE) study set out in our original application were as follows.

Description of the competing risk model

The competing risk model11 assumes that if the pregnancy was to continue indefinitely then all women would develop pre-eclampsia and whether or not they do so before a specified gestational age depends on competition between delivery before or after development of pre-eclampsia. 18,19

Each woman has a personalised distribution of gestational age at delivery with pre-eclampsia, as illustrated in Figure 1. The risks obtained from this distribution are probabilities of delivery with pre-eclampsia, assuming no other-cause delivery. These are given by the area under the probability density curve, as illustrated in Figure 1a, or the height of the cumulative distribution curve, as illustrated in Figure 1b. The cumulative distribution curve on Figure 1b is the graph of the areas on Figure 1a as functions of gestational age at delivery.

FIGURE 1.

Personalised distribution of gestational age at delivery with pre-eclampsia. Risk of delivery with pre-eclampsia before 37 weeks’ gestation is shown in the shaded area under the probability density (a) and the height of the cumulative distribution (b). The area shaded blue is 0.05 or 1 in 20 and this is the risk of pre-eclampsia with delivery before 37 weeks’ gestation.

There are many approaches that can be taken to fitting models to obtain personalised distribution. Our approach uses Bayes’ theorem to combine a prior distribution determined from maternal and pregnancy characteristics that are available at the first trimester of pregnancy with likelihoods from biomarkers measured during pregnancy. The benefit of this approach for risk assessment is that biomarkers can be added at different times and the posterior distribution can be updated with further information as the pregnancy progresses. If no biomarker information is available, then risks can be obtained from the prior model. From the model development perspective, new biomarkers can be incorporated by augmenting the likelihood model without the need for a completely new model. This facilitates extension of the model through the addition of new biomarkers. This report concerns biomarkers measured in the first trimester. The full model includes biomarkers measured in the second and third trimesters of pregnancy.

Prior model based on maternal factors

Figure 2 illustrates the prior distribution of gestational age at delivery with pre-eclampsia for three different scenarios. Figure 2a shows the reference distribution that applies to a white, nulliparous woman with no family history of pre-eclampsia, aged ≤ 35 years and with a weight of 69 kg and a height of 164 cm. Figure 2b shows the distribution for a woman with low-risk factors, resulting in a shift of the mean to the right by 6.7 weeks and, therefore, a reduction in risk of delivery with pre-eclampsia at < 40 weeks’ gestation. Figure 2c shows the distribution for a woman with high-risk factors, resulting in a shift of the mean to the left by 15.3 weeks and, therefore, an increase in risk of delivery with pre-eclampsia at < 40 weeks’ gestation. 11

FIGURE 2.

Prior distributions for three scenarios. The shaded areas show the risks of pre-eclampsia with delivery before 40 weeks’ gestation assuming no other cause for delivery. (a) Scenario 1 (age: 25 years, height: 164 cm, weight: 69 kg, race: white, family history of pre-eclampsia: no, parity: nulliparous, conception: spontaneous, medical conditions: none); (b) scenario 2 (age: 34 years, height: 170 cm, weight: 50 kg, race: white, family history of pre-eclampsia: no, parity: parous without pre-eclampsia, previous gestation at delivery: 40 weeks’ gestation, pregnancy interval: 2 years, conception: spontaneous, medical conditions: none); and (c) scenario 3 (age: 40 years, height: 140 cm, weight: 120 kg, race: black, family history of pre-eclampsia: yes mother, parity: parous with pre-eclampsia, previous gestational age at delivery: 37 weeks’ gestation, pregnancy interval: 1 year, conception: spontaneous, medical conditions: none).

Multiples of the median of biomarkers

The use of multiple of the median (MoM) values as standardised biomarker measures has a long history in screening that goes back over 40 years to early work on screening for anencephaly and spina bifida. 20 It is the established approach for standardisation incorporated in many software systems.

The purpose of this standardisation is to remove the effects of characteristics (e.g. gestational age, weight and ethnicity) associated with the individual being measured and characteristics associated with the measurement instrument. If y denotes the measurement of the biomarker of a particular individual and m the median value given the characteristics of this individual, then the MoM value is given by MoM = y/m. MoM values are generally obtained from regression models of log-transformed biomarker values. They are the antilogarithm of the errors or residuals from the regression model.

An important property of MoM values for our purposes is that, given the gestation of delivery with pre-eclampsia, MoM values can be assumed to be conditionally independent of the covariates included in the prior model. This is achieved by ensuring that MoM values are standardised for the covariates used to determine the prior model.

Likelihood functions for biomarker measurements

The likelihood is specified in terms of the distribution of log₁₀ transformed MoM values. Figure 3 shows the fitted model for the biomarker PAPP-A. For a given value of PAPP-A MoM, the likelihood of a particular value of gestational age at delivery with pre-eclampsia is given by the height of the Gaussian curve. This is illustrated for cases where PAPP-A MoM is 0.3 and 1.5 in Figure 4.

FIGURE 3.

Distribution of PAPP-A by gestational age at delivery with pre-eclampsia. The points are observations of gestational age at delivery with pre-eclampsia. The solid orange line is the broken stick model for mean log-MoM value of PAPP-A. The histogram on the right shows log-MoM values for pregnancies where birth occurred without pre-eclampsia. The Gaussian curves show the distribution of log-MoM PAPP-A conditionally on gestational age at delivery with pre-eclampsia at 26, 32 and 38 weeks’ gestation, and for those beyond 42.8 weeks’ gestation for which the mean log-MoM is zero. The data shown as points and in the histogram are from the SPREE study. The broken stick model is used to capture the increasing effect on markers with increasing disease severity and the fact that gestations of delivery with pre-eclampsia beyond term represent normality so therefore have a median of 1.0 as shown in the histogram on the right-hand side.

FIGURE 4.

Illustration of likelihood functions PAPP-A MoM values of 0.3 and 1.5. The heights shown in orange in parts (a) and (c) are the values of the likelihood function for deliveries with pre-eclampsia at 26, 32, 38 and beyond 41.2 weeks’ gestation. The likelihood function is shown in parts (b) and (d).

Application of Bayes’ theorem

The posterior distribution of gestational age at delivery with pre-eclampsia is obtained using Bayes’ theorem to combine the prior distribution from maternal and pregnancy characteristics with the likelihood function from biomarker MoM values.

The posterior distribution is obtained by multiplying the prior probability density by the likelihood function (Figure 5). Figures 5a–c show how the prior is modified by the likelihood for a MoM value of PAPP-A of 0.3. At each gestation, the prior density shown in Figures 5a and d is multiplied by the likelihood. For example, at 24 weeks’ gestation the prior probability is multiplied by about 6. In this case, densities at lower gestations are increased relative to those at higher gestations.

FIGURE 5.

Application of Bayes’ theorem in a case with PAPP-A MoM = 0.3 (a–c) and one with PAPP-A MoM = 1.5 (d–f). The prior distribution is the same in both cases with a prior risk of pre-eclampsia before 37 weeks’ gestation of 0.05 or 1 in 20. With a PAPP-A MoM of 0.3, the posterior risk is 0.122. With a PAPP-A MoM of 1.5, the posterior risk is 0.033.

To complete the posterior density, the area under the curve is made 1.0 by multiplying by a normalising constant to produce the posterior shown in Figures 5c and f. Figures 5d–f show how the prior is modified by the likelihood for a MoM value of PAPP-A of 1.5. In this case, densities at lower gestations are decreased relative to those at higher gestations.

The area defining the risk can be computed for other gestations to produce a cumulative distribution of risks that can be used as in individualised risk profile. This is illustrated in Figure 6, which shows the cumulative distribution for the prior for the posterior with PAPP-A MoM equal to 0.3 and for the posterior with PAPP-A MoM equal to 1.5.

FIGURE 6.

Prior (dark blue) and posterior risk profiles for PAPP-A MoM = 0.3 (solid light blue) and 1.5 (dashed light blue).

The examples shown in Figures 5 and 6 illustrate the elements of our model using univariate examples. The log-transformed biomarkers are assumed to follow a Gaussian distribution with a mean dependent on the gestational age at delivery with pre-eclampsia according to a broken stick model. In general, a multivariate Gaussian distribution is used and the association between biomarkers captured in the correlations between markers. Technical details and parameter estimates of the algorithm for risk calculation are given in Appendix 1.

Study design and population

This was a prospective multicentre cohort study carried out in seven NHS maternity hospitals in England, between 12 April 2016 and 15 December 2016.

The participating hospitals were King’s College Hospital, London; University Hospital Lewisham, London; Medway Maritime Hospital, Gillingham; Homerton University Hospital, London; North Middlesex University Hospital, London; Southend University Hospital, Essex; and The Royal London Hospital, London.

Study participants

During the study period, a total of 20,168 pregnant women attended one of the participating hospitals for assessment at 11–13 weeks’ gestation. Of the 18,089 women who provided written informed consent, 17,051 were eligible to participate in the study and were screened for pre-eclampsia. Outcome data were obtained from 16,747 women (Figure 7). The baseline characteristics of the participants are given in Table 1. Pre-eclampsia developed in 473 (2.8%) pregnancies; in 142 (0.8%) cases, this was preterm pre-eclampsia. 37

FIGURE 7.

Screening and follow-up.

Intake of aspirin

Aspirin from < 14 weeks’ gestation to delivery or 36 weeks’ gestation was taken by 749 (4.5%) of 16,747 women in the study population. The daily dose was 75 mg in 730 (97.5%) women and 150 mg in 19 (2.5%) women. Aspirin was taken by 400 (23.2%) women in the NICE screen-positive group and 349 (2.3%) women in the NICE screen-negative group. The reported reasons for treatment in the latter group were previous history of miscarriage (n = 153), stillbirth (n = 26), fetal growth restriction (n = 25), placental abruption (n = 8), thrombophilia (n = 18), cardiovascular surgery (n = 3), family history of pre-eclampsia (n = 6), current pregnancy conceived by in vitro fertilisation (n = 34), high body mass index (n = 21), low serum PAPP-A found at screening for fetal trisomies (n = 47), one episode of high blood pressure in the first trimester of pregnancy (n = 6), medical history of Lynch syndrome (n = 1) and Raynaud’s disease (n = 1).

Primary comparison

Classification into screen positive and screen negative by the NICE method and the mini-combined test stratified by outcome and aspirin treatment is shown in Table 2. With regard to Table 2, it is notable that when using a fixed overall SPR and the results are stratified by aspirin, there is a substantial imbalance between the SPRs obtained using the NICE method and those using the mini-combined test in the aspirin and no-aspirin groups. If the risk cut-off point is chosen so that the SPRs for the NICE method and the mini-combined test are the same, then the contingency results given in Table 3 are obtained. For the aspirin group, this shows a significant benefit for the mini-combined test over the NICE method (p = 0.003 without imputation for aspirin and p = 0.034 with imputation, assuming RRR = 0.3). For the no-aspirin group the mini-combined test is superior to the NICE method (p < 0.001). 38

Imputation

The impact of aspirin treatment on McNemar’s test is to prevent pre-eclampsia and, therefore, transfer events counts for aspirin in Table 2 from the pre-eclampsia to no pre-eclampsia. The data imputation methodology reverses this process and, for those treated with aspirin, some observations are transferred from the no pre-eclampsia to the pre-eclampsia data. The discordant cells positive with the mini-combined test and negative with the NICE method (n = 38) and negative with the mini-combined test and positive with the NICE method (n = 136) is the no pre-eclampsia data that have an impact on McNemar’s test. Owing to the larger number in the negative mini-combined test and positive NICE method cells, the effect of the imputation is to correct a negative bias against the NICE method. Imputation was undertaken using the Markov chain Monte Carlo method, as described in Appendix 2.

The results of multiple imputation of data on the incidence of all pre-eclampsia that would have occurred had it not been for the effect of treatment with aspirin are shown in Figure 8. After adjustment for the effect of aspirin (i.e. a 30% reduction in the rate of all pre-eclampsia) in those receiving this drug the DR of the NICE method was 31.6% (95% CI 27.3% to 35.9%) and that of the Bayes’ theorem-based method was 42.8% (95% CI 38.4% to 47.3%). The difference between the two methods was 11.2% (95% CI 6.9% to 15.6%) (see Table 4).

FIGURE 8.

Difference in DRs for pre-eclampsia with delivery at any gestational age [mini-combined test (i.e. Mat-CHs, MAP and PAPP-A) vs. NICE method] with relative risk reduction from aspirin of 30%. The first 10 rows give the results for 10 imputed samples.

Sensitivity analysis for assumptions regarding the effect of aspirin is given in Appendix 3, Figure 20. This shows that our conclusions are robust to the effect of aspirin and even in the most extreme case when aspirin is assumed to be 100% effective in preventing pre-eclampsia there is a substantial and overwhelmingly significant (p < 0.0001) improvement in DR using the mini-combined test (vs. the NICE method).

Key secondary comparisons

The three prespecified secondary comparisons were the NICE method compared with the competing risk model using maternal factors and the following biomarker combinations for the prediction of preterm pre-eclampsia (i.e. pre-eclampsia with delivery before 37 weeks’ gestation):

1. Mat-CHs, MAP and PAPP-A

2. Mat-CHs, MAP and PLGF

3. Mat-CHs, MAP, UTA-PI and PLGF.

There were 142 women who delivered with preterm pre-eclampsia. Screening performance of the competing risk model and the NICE method is summarised in Table 4 and shown in Figure 9. After adjusting for the effect of aspirin, the DR of the NICE method for preterm pre-eclampsia was 44.1% (95% CI 35.7% to 52.6%), which was lower than that of the Bayes’ theorem-based method using Mat-CHs, MAP and PAPP-A (53.5%, 95% CI 45.5% to 61.6%), Mat-CHs, MAP and PLGF (67.3%, 95% CI 59.6% to 75.0%), and Mat-CHs, MAP, PLGF and UTA-PI (79.6%, 95% CI 72.7% to 86.5%). The difference in DRs (i.e. competing risk model – NICE method) was 9.4% (95% CI 0.1% to 18.2%) using the mini-combined test, 23.2% (95% CI 13.2% to 33.3%) using Mat-CHs together with MAP and PLGF, and 35.5% (95% CI 25.2% to 45.8%) using Mat-CHs with MAP, UTA-PI and PLGF.

FIGURE 9.

Receiver operating characteristic curves for prediction of preterm pre-eclampsia by the competing risk model. The patient-specific risk is derived by a combination of Mat-CHs with the measurements of MAP and PAPP-A (orange curve), MAP and PLGF (light blue curve) and MAP, PLGF and UTA-PI (dark blue curve). Performance of risk assessment using NICE guidelines is shown as a horizontal interrupted black line.

The results of multiple imputation of data on the incidence of preterm pre-eclampsia that would have occurred had it not been for the effect of treatment with aspirin are shown in Figure 10. After adjustment for the effect of aspirin (i.e. a 60% reduction in rate of preterm pre-eclampsia) in those receiving this drug, the difference in DR of the three Bayes’ theorem-based methods from the NICE method was 10.5% (95% CI 2.3% to 18.8%), 24.0% (95% CI 14.3% to 33.7%) and 35.1% (95% CI 25.1% to 45.0%), respectively.

FIGURE 10.

Results of multiple imputation of data on the incidence of preterm pre-eclampsia that would have occurred had it not been for the effect of treatment with aspirin. The DR of risk assessment by the NICE guidelines is compared with that of screening by a combination of Mat-CHs with the measurements of (a) MAP and PAPP-A; (b) MAP and PLGF; and (c) MAP, PLGF and UTA-PI.

Additional data on performance of the competing risk model

Table 5 provides data on the DR of early, preterm and term pre-eclampsia at a fixed SPR of 10% in screening by various combinations of biomarkers. Table 6 presents an analysis of the incremental benefit in DR of individual biomarkers when added to a specific combination of markers. In all cases apart from the addition of PAPP-A, the addition of a biomarker improved the DR.

Additional figures of receiver operating characteristic curves for prediction of pre-eclampsia by the competing risk model are provided in Appendix 4.

Principal findings of this study

The SPREE study has demonstrated that risk assessment for pre-eclampsia as currently recommended by NICE guidelines8 identifies approximately 30% of women who would develop pre-eclampsia and about 40% of those that will develop severe pre-eclampsia leading to preterm birth, at a SPR of 10%.

Compliance with the NICE recommendation that women at high-risk for pre-eclampsia should be treated with aspirin from the first trimester to the end of pregnancy was only 23%. Such low compliance may, at least in part, be attributed to the generally held belief, based on the results of a meta-analysis in 2007,6 that aspirin reduces the risk of pre-eclampsia by only about 10%.

The performance of screening by competing risks model, which combines maternal factors with biomarkers,17,33 was superior to that of risk assessment by NICE guidelines. At the same SPR as for the NICE method, the DR for all pre-eclampsia in screening by Mat-CHs, MAP and serum PAPP-A was 42.5% and the DR for preterm pre-eclampsia by a combination of Mat-CHs, MAP, UTA-PI and PLGF was 82.4%, which is significantly higher than that of the NICE guidelines.

Strengths and limitations of this study

The strengths of the study include prospective examination of a large number of pregnant women in several maternity units covering a wide spectrum of demographic and racial characteristics. The results are therefore likely to be generalisable across the UK. More than 90% of patients attending for routine care agreed to participate in the study. Measurement of all biomarkers was recorded in all cases and complete follow-up was obtained from > 98% of participants. Consistency in data collection was maintained throughout the study period by ensuring adequate training for all investigators based on standardised protocols, regular UCL-CCTU monitoring, and external validation and quality assurance of biomarker measurements.

A potential limitation of the study is lack of formal health economic assessment concerning the implementation of combined screening for pre-eclampsia. Such assessment was beyond the scope of this study, but it is currently being carried out.

Comparison with results of previous studies

The performance of screening for preterm pre-eclampsia by the competing risks model, utilising Mat-CHs, MAP, UTA-PI and PLGF, observed in this study is comparable to that reported in several previous studies of singleton pregnancies at 11–13 weeks’ gestation. 17,19,39,40 The algorithm was originally developed from a study19 of 58,884 pregnancies. The DR of preterm pre-eclampsia was 77% at a FPR of 10%. 19 Subsequently, we used data from prospective screening in 35,948 pregnancies to update the original algorithm. The DR of preterm pre-eclampsia was 75% at a FPR of 10%. 17 The diagnostic accuracy of this algorithm was examined in a prospective multicentre study39 of 8775 pregnancies. The DR of preterm pre-eclampsia was 75% at a FPR of 10%. 39 In the screened population in the ASPRE trial,40 involving 25,797 pregnancies from 13 maternity hospitals in six countries, the DR of preterm pre-eclampsia after adjustment for the effect of aspirin was 77% at a FPR of 9.2%. 40 None of these studies found evidence that PAPP-A improved screening achieved by MAP, UTA-PI and PLGF.

Other first-trimester combined prediction models have been developed in different populations. Specifically, two Spanish cohort studies27,29 have developed models with the use of Mat-CHs, UTA-PI, MAP and biochemical markers that have demonstrated similar predictive performance in comparison with the Bayes’ theorem-based model. In contrast, three combined prediction algorithms, established from cohort studies in the USA, demonstrate lower predictive performance than the Bayes’ theorem-based model. 25,28,41

A recent systematic review has compared the performance of simple risk models (i.e. Mat-CHs only) with that of specialised models that include specialised tests (e.g. measurements of MAP, UTA-PI and/or biochemical markers) for the prediction of pre-eclampsia. 42 Seventy models from 29 studies were identified (17 models to predict pre-eclampsia of any gestation, 31 models to predict early-onset pre-eclampsia and 22 models to predict late-onset pre-eclampsia). Among them, 22 were simple risk models, whereas 48 were classified as specialised models. Comparing simple and specialised models, the latter performed better than the simple models in predicting both early- and late-onset pre-eclampsia. 42 The specialised models have been shown to increase the DR of pre-eclampsia by 18% (95% CI 0% to 56%) at a fixed FPR of 5% or 10%. 42 Such results further confirm that our approach to screening using a combination of various tests rather than a single test is better for the prediction of pre-eclampsia.

Implications on clinical practice

Recent evidence suggests that first-trimester risk assessment should focus on prediction of pre-eclampsia leading to preterm birth primarily with the aim of preventing preterm birth through treatment with aspirin from 11–13 weeks’ gestation. Aspirin is considerably more effective than previously thought in reducing the risk of preterm pre-eclampsia, provided the daily dose of the drug is ≥ 100 mg and the gestational age at onset of therapy is < 16 weeks. 1 In the ASPRE trial,7,43 use of aspirin (150 mg/day) starting from 11–14 weeks’ gestation reduced the risk of preterm pre-eclampsia by 62% and a secondary analysis of the trial reported that the reduction was even greater (75%) if the compliance was ≥ 90%. Against this background, there are ongoing debates about prediction and prevention of preterm pre-eclampsia centred on two questions: (1) whether or not aspirin should be recommended for all women or to a subpopulation of those women predicted to be at increased risk of developing pre-eclampsia and (2) if a strategy of prediction and prevention is to be used, what method should be used for prediction.

The arguments in favour of recommending aspirin to all women are that it avoids the need for prediction and the whole population benefits from the prophylactic treatment with aspirin. Arguments against this are that (1) compliance is likely to be worse when aspirin is applied to the whole population than when recommended to a subpopulation selected, and counselled, based on risk and (2) there is a need to balance the benefit from aspirin in prevention of preterm pre-eclampsia with potential harm from aspirin due to haemorrhagic and other adverse effects. Assuming that the entire population took aspirin, an incidence of 0.8% and a relative reduction in risk of preterm pre-eclampsia of 60%, 208 women would be exposed to aspirin treatment for every case of preterm pre-eclampsia prevented. Using risk stratification with Mat-CHs, MAP, UTA-PI and PLGF with the same SPR as NICE, 16 women would need to be exposed to aspirin to prevent one case compared with 30 women using the NICE guidelines.

Regarding the method of prediction, the debate centres around screening performance, costs and practical issues of implementation.

The main focus of this report has been on the DR achieved by using the competing risk model compared with that of the NICE method. For the same SPR as NICE, the DR for preterm pre-eclampsia achieved by combining Mat-CHs with MAP, UTA-PI and PLGF is 79.6% (95% CI 72.7% to 86.5%) compared with 44.1% (95% CI 35.7% to 52.6%) when using the NICE method. Using these estimates and with an incidence of preterm pre-eclampsia of 0.8% the positive predictive values are 1 in 16 compared with 1 in 29 for the competing risk model and NICE method, respectively. Among women who screen negative, the proportions with preterm pre-eclampsia (i.e. 1 – negative predictive value) are 1 in 550 compared with 1 in 200 for the competing risk model and NICE method, respectively.

The main argument against the use of risk algorithms, such as the competing risk model, is that they are too complex to use in practice. Simple methods, such as the NICE criteria or cut-off points applied to biomarker measurements or their ratios, should be preferred because they are easy to implement in practice. In fact, the essential features of our approach of using Bayes’ theorem to update likelihoods from biomarker MoM values to update a prior based on maternal factors have been used for many decades in screening for aneuploidies. These algorithms have been built into commercial software used extensively in practice. The commercial software suppliers have implemented the competing risk algorithm for pre-eclampsia screening into their software systems.

Regarding the approaches based on application of cut-off points to individual markers or ratios of different markers, the following points need to be considered. First, they do not provide individualised risks for decision-making. Second, their performance is inferior to approaches based on probability theory to make optimal use of the available information. Last, because biomarkers are affected by covariates such as ethnicity, they are likely to be inequitable in the way they perform across different groups within the population.

In the clinical implementation of the first-trimester combined test for preterm pre-eclampsia, recording Mat-CHs and medical history, measurement of blood pressure and hospital attendance at 11–13 weeks’ gestation for an ultrasound scan are an integral part of routine antenatal care. Measurement of UTA-PI can be carried out by the same sonographers and ultrasound machines used for the routine scan at 11–13 weeks’ gestation; however, the sonographers will require training to carry out this test and the measurement would add 2–3 minutes to the current 20–30 minutes used for the scan. Serum PLGF can be measured in the same blood sample and by the same automated platforms that are currently used for measurement of serum PAPP-A as part of routine clinical practice in screening for fetal trisomies in all maternity hospitals in England; however, there is an additional cost for the reagents. Extensive research has established reference ranges for MAP, PLGF and UTA-PI, has described the Mat-CHs that affect the measurements and has developed the infrastructure for auditing of results. 44–46

One decade ago, effective first-trimester screening for fetal trisomies was implemented in all maternity hospitals in the UK within a few months of the appropriate decision being taken by the UK National Screening Committee and NICE. 47 The same infrastructure can now be used to expand the aims of first-trimester screening to include identification of women at high risk of developing preterm pre-eclampsia and substantially reducing such risk through the prophylactic use of the appropriate dose of aspirin. 48

Conclusion

The SPREE study has demonstrated that the performance of first trimester screening for pre-eclampsia by a combination of maternal factors and biomarkers is superior to that achieved by the risk assessment method recommended by the current NICE guidelines.

Methods

Calibration was assessed visually through a series of figures showing the estimated incidence against that predicted from risk for pre-eclampsia at < 34 weeks’ gestation, < 37 weeks’ gestation and any pre-eclampsia for screening by maternal factors and various combinations of biomarkers. The plots were obtained by grouping the data into bins according to risk. The estimated incidence in each group was then plotted against the incidence predicted from the risks within each group (i.e. the mean risk). 35

Calibration, in the large, is a measure of whether the risks are generally too high or too low. This is quantified by the estimated intercept from a logistic regression of incidence on the logit of risk with the slope fixed at 1. The intercept is a measure of the deviation of the observed incidence from the predicted. For perfectly calibrated risks, the intercept should be zero. If there is a general tendency for underestimation, so that the observed incidence is larger than that predicted, the intercept will be positive. Conversely, for overestimation the intercept will be negative.

The calibration slope assesses the calibration across the range of risks and is the slope of the regression line of the logistic regression of incidence on the logit of risk. If the risk is well calibrated then the slope should be 1.0. A slope < 1 means that the relationship between risk and incidence is flatter than it should be. A calibration slope > 1 means that the relationship is steeper than it should be.

The risks produced from our competing risks model17,33 are for delivery with pre-eclampsia before a specific gestation, assuming no other cause for delivery. As other-cause deliveries are effectively censored observations, the actual incidence of pre-eclampsia would be expected to be lower than predicted. For early gestations, when there are few other-cause deliveries, the effects would be small. At later gestations, with many other-cause deliveries, the effect of censoring may be substantial. Consequently, we applied survival analysis and used a Kaplan–Meier49 estimate of incidence of delivery with pre-eclampsia treating deliveries from other causes as censored observations. The incidence counts were adjusted for the effect of censoring by multiplying the estimated incidence by the number of observations in each bin. The analysis was undertaken using the R statistical software with the survival package for Kaplan–Meier estimates of survival probabilities.

Results

Calibration of risks for the mini-combined test: Mat-CHs, MAP and PAPP-A in the SPREE study

Calibration plots of the predictive performance of the competing risk model for all pre-eclampsia using Mat-CHs, MAP and PAPP-A in the SPREE study are shown in Figure 11. Figures 11a and c show the incidence with no correction for censoring. Figures 11b and d show the Kaplan–Meier estimates of incidence obtained from a survival analysis treating births due to other causes as censored observations. The error bars are 95% CIs. The numbers in black shown above each interval are the number of pregnancies in each bin. Those in orange are the incidence counts in each of the bins. Statistics for the calibration-in-the-large intercept and the calibration slope are shown superimposed on Figure 11. The light blue line shows where the risks are perfect predictors of incidence. The vertical dashed line is the mean overall risk and the horizontal dashed line is the overall incidence. The histograms show the distribution of risks for pregnancies with pre-eclampsia (orange) and without pre-eclampsia (light blue). The figure at the bottom of each graph shows the estimated incidence/mean risk within each bin.

FIGURE 11.

Calibration plots for screening using the competing risk model for prediction of all pre-eclampsia by Mat-CHs, MAP and PAPP-A. (a) Estimated incidence with no adjustment for censoring for Mat-CHs, MAP and PAPP-A (all pre-eclampsia); (b) estimated incidence with adjustment for censoring for Mat-CHs, MAP and PAPP-A (all pre-eclampsia); (c) observed/expected incidence with no adjustment for censoring for Mat-CHs, MAP and PAPP-A (all pre-eclampsia); and (d) observed/expected incidence with adjustment for censoring for Mat-CHs, MAP and PAPP-A (all pre-eclampsia).

With no adjustment for censoring, risks were well calibrated and with adjustment for censoring the risks underestimated the incidence.

Calibration of risks for the preterm pre-eclampsia using Mat-CHs, MAP, UTA-PI and PLGF in the SPREE study

Calibration plots of the predictive performance of the competing risk model for preterm pre-eclampsia using Mat-CHs, MAP, UTA-PI and PLGF in the SPREE study are shown in Figure 12. Calibration of risks for the uncorrected and corrected incidence was good and the calibration slope was very close to 1.0. However, there was a tendency for the risks to underestimate the incidence of pre-eclampsia.

FIGURE 12.

Calibration plots for screening using the competing risk model for prediction of preterm pre-eclampsia by Mat-CHs, MAP, UTA-PI and PLGF. (a) Estimated incidence with no adjustment for censoring for Mat-CHs, MAP, UTA-PI and PLGF (pre-eclampsia before 37 weeks’ gestation); (b) estimated incidence with adjustment for censoring for Mat-CHs, MAP, UTA-PI and PLGF (pre-eclampsia before 37 weeks’ gestation); (c) observed/expected incidence with no adjustment for censoring for Mat-CHs, MAP, UTA-PI and PLGF (pre-eclampsia before 37 weeks’ gestation); and (d) observed/expected incidence with adjustment for censoring for Mat-CHs, MAP, UTA-PI and PLGF (pre-eclampsia before 37 weeks’ gestation). The histograms show the distribution of risks in pregnancies with pre-eclampsia before 37 weeks’ gestation (orange) and those without delivery with pre-eclampsia before 37 weeks’ gestation (light blue).

Additional plots for other combinations of biomarkers are provided in Appendix 5.

Calibration of risks in the ASPRE SQS

The data for the ASPRE SQS were derived from prospective screening for pre-eclampsia in 8775 women between February and September 2015 in 12 maternity hospitals in England, Spain, Belgium, Italy and Greece. 39 This study was carried out before the ASPRE trial7 and was primarily designed to examine the feasibility of multicentre screening and establish methods for quality assurance of the biomarkers. The algorithm used for screening was the same as in the SPREE study. 17,33 The diagnosis of pre-eclampsia was based on the previous criteria of the International Society for the Study of Hypertension in Pregnancy,34 whereas the new wider definition was used in the SPREE study. 21,34

Calibration plots of the predictive performance of the competing risk model for all pre-eclampsia using Mat-CHs, MAP and PAPP-A in the ASPRE SQS are shown in Figure 13, and calibration plots of the predictive performance of the model for preterm pre-eclampsia using Mat-CHs, MAP, UTA-PI and PLGF in the ASPRE SQS are shown in Figure 14.

FIGURE 13.

Calibration plots for screening using the competing risk model for prediction of all pre-eclampsia by Mat-CHs, MAP and PAPP-A for the ASPRE SQS. (a) Estimated incidence with no adjustment for censoring for Mat-CHs, MAP and PAPP-A (all pre-eclampsia); (b) estimated incidence with adjustment for censoring for Mat-CHs, MAP and PAPP-A (all pre-eclampsia); (c) observed/expected incidence with no adjustment for censoring for Mat-CHs, MAP + PAPP-A (all pre-eclampsia); and (d) observed/expected incidence with adjustment for censoring for Mat-CHs, MAP and PAPP-A (all pre-eclampsia).

FIGURE 14.

Calibration plots for screening using the competing risk model for prediction of preterm pre-eclampsia by Mat-CHs, MAP and PAPP-A for the ASPRE SQS. (a) Estimated incidence with no adjustment for censoring for Mat-CHs, MAP and PAPP-A (pre-eclampsia before 37 weeks); (b) estimated incidence with adjustment for censoring for Mat-CHs, MAP and PAPP-A (pre-eclampsia before 37 weeks); (c) observed/expected incidence with no adjustment for censoring for Mat-CHs, MAP and PAPP-A (pre-eclampsia before 37 weeks); and (d) observed/expected incidence with adjustment for censoring for Mat-CHs, MAP and PAPP-A (pre-eclampsia before 37 weeks’ gestation).

In screening by the mini-combined test, the risks overestimate the unadjusted incidence and are relatively well calibrated for the adjusted incidence. In the case of screening by Mat-CHs, MAP, UTA-PI and PLGF for preterm pre-eclampsia, adjustment for censoring makes very little difference and calibration is generally good.

Discussion

This study has examined the predictive performance of the competing risk model17,33 in two prospective screening studies (i.e. the SPREE study and the ASPRE SQS). The results demonstrate that in both the SPREE study and the ASPRE SQS calibration of risks for pre-eclampsia was generally good, with the calibration slope very close to 1.0. However, in the case of the SPREE study there was a tendency for the risks to underestimate the incidence of pre-eclampsia. This was made worse when adjustments were made for censoring. With the ASPRE SQS, calibration was improved with adjustment for censoring, with no underestimation of risks.

Calibration refers to how well the predictions from the model agree with the observed outcomes. Deviations between the predicted and observed outcome not only reflect on the accuracy of a given model but could also be the consequence of differences between the studies used for development of the model and the studies used for validation in terms of the below:

• methodology and accuracy of recording Mat-CHs and medical history, and the measurement of biomarkers

• ascertainment and definition of the outcome measure.

Possible explanations for the differences in findings between the SPREE study and the ASPRE SQS are that in the SPREE study we used the new criteria for defining pre-eclampsia, which results in a higher incidence. In addition, the ascertainment was higher because the study was specifically designed for this purpose. The ASPRE SQS focused on quality assurance of biomarkers in preparation for the ASPRE trial.

The strengths of this study include that (1) it is a large prospective evaluation of the calibration of risks from a prespecified algorithm, (2) the assessment of calibration allowed for the effect of censoring due to births from causes other than pre-eclampsia and (3) it is a multicentre study with a diverse population, making the results generalisable.

Methods

Decision curves are graphical displays of net benefit as a function of threshold probabilities. Each threshold represents a risk at which a true positive is equivalent to avoiding a false positive. At one extreme, a threshold of zero represents the position where true positive outweighs any number of false positives. At the other extreme, a threshold of 1 represents the point at which avoidance of a false positive outweighs any number of true positives. A threshold of 0.5 means that the avoidance of a false positive equates to a true positive, so that the probability threshold that determines the decision is 0.5. For a given threshold probability, p_t, the net benefit is given by:

where n is the number of participants in the study.

The higher the net benefit the better. The maximum net benefit occurs when the DR for a given test is 100%, the FPR is zero and the net benefit is the same as the incidence.

If there is no screening test and all individuals are considered as negative then the net benefit is zero. If there is no screening test and all individuals are considered as positive then the net benefit is given by:

In prediction and prevention of preterm pre-eclampsia this would represent the policy of treating everyone with aspirin.

Net benefits were computed and plotted using the R statistical software. A log-scale was used for the probability threshold so that the lower values of most interest occupied a greater proportion of the scale.

The decision curves shown in dark blue in Figure 15 are for prediction of pre-eclampsia before 34 weeks’ gestation, before 37 weeks’ gestation and at any gestational age using the competing risk with different combinations of markers. Screening for pre-eclampsia with delivery before 34 and 37 weeks’ gestation was chosen for the purposes of examining policies of treatment with aspirin.

Results

Decision curves for pre-eclampsia with delivery before 34 and 37 weeks’ gestation for the SPREE study data using the competing risk model with the following combinations of markers are shown in Figure 15:

• Mat-CHs and MAP

• Mat-CHs, MAP and PLGF

• Mat-CHs MAP and UTA-PI

• Mat-CHs, MAP, UTA-PI and PLGF.

FIGURE 15.

Decision curves for screening for pre-eclampsia with delivery before (a) 34 weeks’ gestation; and (b) 37 weeks’ gestation. The light blue curve is the decision curve for the policy of screening everyone positive. The pink line is the decision curve for screening everyone negative.

Decision curves for other marker combinations and for pre-eclampsia with delivery at any gestational age are given in Appendix 5, Figure 87.

Discussion

In general, the decision curves for the combined test are superior to the policies of screening everyone positive or screening everyone negative for the range of threshold probabilities that would seem reasonable. The best performance is provided by the competing risk model incorporating Mat-CHs and the triple test. It is notable that the net benefit of screening all positive is lower than that of screening using the competing risk model for probability thresholds as low as 0.001.

Methods

In all three populations there was prospective screening for pre-eclampsia in singleton pregnancies at 11⁺⁰ to 13⁺⁶ weeks’ gestation. Women aged > 18 years with a singleton pregnancy and a live fetus at the 11- to 13-week scan were included in the study. Women who were unconscious or severely ill at the time of recruitment, those with learning difficulties or serious mental illness and those with major fetal abnormality identified at the 11- to 13-week scan were excluded from the study.

The visit at 11–13 weeks’ gestation included (1) recording of Mat-CHs and obstetric and medical history,33 (2) measurement of maternal weight and height, (3) measurement of MAP by validated automated devices and standardised protocol,32 (4) measurement of the left and right UTA-PI by transabdominal colour Doppler ultrasound and calculation of the mean pulsatility index,12 and (5) measurement of serum concentration of PLGF and PAPP-A (using a DELFIA Xpress analyser or BRAHMS KRYPTOR analyser). Gestational age was determined from the fetal crown–rump length. 31 The women gave written informed consent to participate in the studies, which were approved by the relevant research ethics committee in each participating centre.

In this study and from each of the three data sets (i.e. the SPREE study, the ASPRE SQS and AJOG) we included only pregnancies delivering a phenotypically normal live birth or stillbirth at ≥ 24 weeks’ gestation, and we excluded pregnancies with aneuploidies and major fetal abnormalities and those ending in termination, miscarriage or fetal death before 24 weeks.

Results

Maternal and pregnancy characteristics in AJOG, the ASPRE SQS and the SPREE study populations are provided and compared in Table 10. The population in AJOG appears to be at higher risk of pre-eclampsia than in the other two data sets. There were also significantly more black women and fewer white women in the AJOG data set, as well as more women with chronic hypertension, diabetes or a family history of pre-eclampsia and more nulliparous women and parous women with a previous history of pre-eclampsia. Despite this increased risk in the AJOG data set, the observed incidence of pre-eclampsia was decreased, raising the possibility of underascertainment.

The DR for preterm pre-eclampsia at a fixed SPR of 10% in the three data sets is summarised in Figure 16. The DR for pre-eclampsia at < 32, < 37 and ≥ 37 weeks’ gestation and all pre-eclampsia at a fixed SPR of 10% by various combinations of biomarkers is given in Tables 11–14.

FIGURE 16.

Detection rate for preterm pre-eclampsia at a fixed SPR of 10% in the three databases.

Receiver operating characteristics curves for the performance of screening for preterm pre-eclampsia are given in Appendix 7.

Discussion

Differences in screening performance between studies, reflecting differences in study populations, study design and measurement quality, are inevitable. However, the results of this analysis on performance of screening show that there is little evidence of substantive differences in DRs across the three data sets at a SPR of 10% (equivalent to that of NICE guidelines).

The combined results demonstrate that (1) the performance of screening is substantially better for early pre-eclampsia and preterm pre-eclampsia than for term pre-eclampsia (DR by the triple test at a 10% SPR of 90% and 75% vs. 41%, respectively); (2) in preterm pre-eclampsia screening by Mat-CHs is sequentially improved by the addition of one, two and three biomarkers; and (3) addition of PAPP-A to any combination of biomarkers that includes PLGF has little benefit.

Our analysis uses two independent test data sets for assessment of screening performance: the ASPRE SQS and the SPREE study. The AJOG data were used as a training data and their inclusion in the analysis is questionable. However, this is a very large data set relative to the number of parameters involved in model fitting. Moreover, the results from AJOG are very similar to those from the ASPRE SQS and the SPREE study. For this reason, and to provide evidence based on large number of cases, we decided to include the AJOG data.

Methods

Results

Model-based performance of two-stage screening

The model-based DR of preterm pre-eclampsia and early pre-eclampsia in women of white racial origin at a SPR of 10% and 20% is shown in Tables 15 and 16. When screening the whole population by the triple test at a SPR of 10%, the DR of preterm pre-eclampsia was 67% and of early pre-eclampsia was 85%. The corresponding values at a SPR of 20% were 81% and 93%, respectively.

TABLE 15 - Empirical and model-based performance of two-stage screening at 11–13 weeks’ gestation for pre-eclampsia with delivery at < 32 and < 37 weeks’ gestation at a fixed overall SPR of 10% for white women in the population subdivided according to racial origin

Method of first-stage screening	Racial group	Need for second-stage screening (%)	SPR, % (95% CI)	DR for pre-eclampsia
				< 37 weeks’ gestation	< 32 weeks’ gestation
AJOG
Mat-CHs	White (n = 25,879)	72.3	12.3 (11.9 to 12.7)	66.9 (58.4 to 74.6)	81.0 (58.1 to 94.6)
	Black (n = 6681)	99	36.6 (35.5 to 37.8)	92.7 (86.6 to 96.6)	100.0 (90.5 to 100.0)
	South Asian (n = 1623)	88.7	21.3 (19.3 to 23.4)	95.0 (75.1 to 99.9)	100.0 (54.1 to 100.0)
Mat-CHs + MAP + UTA-PI	White (n = 25,879)	36.7	12.2 (11.8 to 12.6)	67.6 (59.2 to 75.3)	81.0 (58.1 to 94.6)
	Black (n = 6681)	70.5	36.2 (35.0 to 37.3)	92.7 (86.6 to 96.6)	100.0 (90.5 to 100.0)
	South Asian (n = 1623)	50.6	20.8 (18.9 to 22.9)	95.0 (75.1 to 99.9)	100.0 (54.1 to 100.0)
Mat-CHs + MAP + PLGF	White (n = 25,879)	25.2	12.2 (11.8 to 12.6)	67.6 (59.2 to 75.3)	85.7 (63.7 to 97.0)
	Black (n = 6681)	56.8	36.4 (35.2 to 37.6)	92.7 (86.6 to 96.6)	100.0 (90.5 to 100.0)
	South Asian (n = 1623)	36.8	21.0 (19.1 to 23.1)	95 (75.1 to 99.9)	100.0 (54.1 to 100.0)
ASPRE SQS
Mat-CHs	White (n = 6883)	70.6	7.7 (7.1 to 8.4)	65.8 (48.6 to 80.4)	100.0 (63.1 to 100.0)
	Black (n = 1090)	98.1	28.6 (26 to 31.4)	92.9 (66.1 to 99.8)	100.0 (63.1 to 100.0)
	South Asian (n = 462)	86.8	10.8 (8.1 to 14)	100.0 (29.2 to 100.0)
Mat-CHs + MAP + UTA-PI	White (n = 6883)	33	7.7 (7.0 to 8.3)	65.8 (48.6 to 80.4)	100.0 (63.1 to 100.0)
	Black (n = 1090)	67.9	28.6 (26 to 31.4)	92.9 (66.1 to 99.8)	100.0 (63.1 to 100.0)
	South Asian (n = 462)	42.9	10.8 (8.1 to 14)	100.0 (29.2 to 100.0)
Mat-CHs + MAP + PLGF	White (n = 6883)	17.2	7.7 (7.1 to 8.4)	63.2 (46 to 78.2)	87.5 (47.3 to 99.7)
	Black (n = 1090)	49.3	28.6 (26.0 to 31.4)	92.9 (66.1 to 99.8)	100.0 (63.1 to 100.0)
	South Asian (n = 462)	23.6	10.6 (7.9 to 13.8)	100.0 (29.2 to 100.0)
SPREE study
Mat-CHs	White (n = 11,922)	69.8	7.0 (6.6 to 7.5)	69.6 (58.2 to 79.5)	78.9 (54.4 to 93.9)
	Black (n = 2337)	98.6	28.8 (26.9 to 30.6)	91.3 (79.2 to 97.6)	100.0 (71.5 to 100.0)
	South Asian (n = 1361)	87.6	12.8 (11.1 to 14.7)	100.0 (78.2 to 100.0)	100.0 (29.2 to 100.0)
Mat-CHs + MAP + UTA-PI	White (n = 11,922)	31.5	7.0 (6.5 to 7.4)	72.2 (60.9 to 81.7)	84.2 (60.4 to 96.6)
	Black (n = 2337)	64.4	28.3 (26.5 to 30.2)	91.3 (79.2 to 97.6)	100.0 (71.5 to 100.0)
	South Asian (n = 1361)	45.5	12.5 (10.8 to 14.4)	100.0 (78.2 to 100.0)	100.0 (29.2 to 100.0)
Mat-CHs + MAP + PLGF	White (n = 11,922)	16.2	6.9 (6.5 to 7.4)	68.4 (56.9 to 78.4)	84.2 (60.4 to 96.6)
	Black (n = 2337)	49.8	28.5 (26.7 to 30.4)	89.1 (76.4 to 96.4)	100.0 (71.5 to 100.0)
	South Asian (n = 1361)	24.8	12.6 (10.8 to 14.4)	100.0 (78.2 to 100.0)	100.0 (29.2 to 100.0)

TABLE 16 - Empirical and model-based performance of two-stage screening at 11–13 weeks’ gestation for pre-eclampsia with delivery at < 32 and < 37 weeks’ gestation at a fixed overall SPR of 20% for white women in the population subdivided according to racial origin

Method of first-stage screening	Racial group	Need for second-stage screening (%)	SPR, % (95% CI)	DR for pre-eclampsia
				< 37 weeks’ gestation	< 32 weeks’ gestation
AJOG
Mat-CHs	White (n = 25,879)	72.3	22.6 (22.0 to 23.1)	79.9 (72.2 to 86.2)	90.5 (69.6 to 98.8)
	Black (n = 6681)	99.0	53.2 (52.0 to 54.4)	98.4 (94.2 to 99.8)	100.0 (90.5 to 100.0)
	South Asian (n = 1623)	88.7	34.7 (32.4 to 37.1)	95.0 (75.1 to 99.9)	100.0 (54.1 to 100.0)
Mat-CHs + MAP + UTA-PI	White (n = 25,879)	44.6	22.4 (21.9 to 22.9)	81.3 (73.8 to 87.4)	95.2 (76.2 to 99.9)
	Black (n = 6681)	77.2	52.3 (51.1 to 53.5)	97.6 (93.0 to 99.5)	100.0 (90.5 to 100.0)
	South Asian (n = 1623)	57.9	33.9 (31.6 to 36.2)	95.0 (75.1 to 99.9)	100.0 (54.1 to 100.0)
Mat-CHs + MAP + PLGF	White (n = 25,879)	34.7	22.3 (21.7 to 22.8)	79.9 (72.2 to 86.2)	90.5 (69.6 to 98.8)
	Black (n = 6681)	67.0	52.3 (51.1 to 53.5)	97.6 (93 to 99.5)	100.0 (90.5 to 100.0)
	South Asian (n = 1623)	46.6	33.6 (31.3 to 36)	95 (75.1 to 99.9)	100.0 (54.1 to 100.0)
ASPRE SQS
Mat-CHs	White (n = 6883)	70.6	16.1 (15.3 to 17)	78.9 (62.7 to 90.4)	100.0 (63.1 to 100.0)
	Black (n = 1090)	98.1	45 (42.1 to 48.1)	100.0 (76.8 to 100.0)	100.0 (63.1 to 100.0)
	South Asian (n = 462)	86.8	22.1 (18.4 to 26.1)	100.0 (29.2 to 100.0)
Mat-CHs + MAP + UTA-PI	White (n = 6883)	40.6	16.2 (15.4 to 17.1)	84.2 (68.7 to 94)	100.0 (63.1 to 100.0)
	Black (n = 1090)	74.7	44.5 (41.5 to 47.5)	100.0 (76.8 to 100.0)	100.0 (63.1 to 100.0)
	South Asian (n = 462)	52.2	22.3 (18.6 to 26.4)	100.0 (29.2 to 100.0)
Mat-CHs + MAP + PLGF	White (n = 6883)	25.5	15.8 (15 to 16.7)	81.6 (65.7 to 92.3)	100.0 (63.1 to 100.0)
	Black (n = 1090)	57.7	44.6 (41.6 to 47.6)	100.0 (76.8 to 100.0)	100.0 (63.1 to 100.0)
	South Asian (n = 462)	35.3	21.4 (17.8 to 25.5)	100.0 (29.2 to 100.0)
SPREE study
Mat-CHs	White (n = 11,922)	69.8	14.9 (14.3 to 15.6)	82.3 (72.1 to 90.0)	84.2 (60.4 to 96.6)
	Black (n = 2337)	98.6	44.6 (42.6 to 46.7)	97.8 (88.5 to 99.9)	100.0 (71.5 to 100.0)
	South Asian (n = 1361)	87.6	23.6 (21.4 to 25.9)	100.0 (78.2 to 100.0)	100.0 (29.2 to 100.0)
Mat-CHs + MAP + UTA-PI	White (n = 11,922)	38.4	14.7 (14.1 to 15.4)	84.8 (75.0 to 91.9)	89.5 (66.9 to 98.7)
	Black (n = 2337)	72.5	43.9 (41.8 to 45.9)	97.8 (88.5 to 99.9)	100.0 (71.5 to 100.0)
	South Asian (n = 1361)	55.8	23.6 (21.4 to 25.9)	100.0 (78.2 to 100.0)	100.0 (29.2 to 100.0)
Mat-CHs + MAP + PLGF	White (n = 11,922)	23.8	14.4 (13.8 to 15)	83.5 (73.5 to 90.9)	84.2 (60.4 to 96.6)
	Black (n = 2337)	60.2	43.9 (41.9 to 46)	95.7 (85.2 to 99.5)	100.0 (71.5 to 100.0)
	South Asian (n = 1361)	34.8	22.9 (20.7 to 25.3)	100.0 (78.2 to 100.0)	100.0 (29.2 to 100.0)

In two-stage screening with Mat-CHs as the method of screening in the first stage and reserving measurements of MAP, UTA-PI and PLGF for the second stage to only 70% of the population, a similar DR was achieved as when screening the whole population by the triple test, irrespective of whether the SPR was 10% or 20% (Figure 18). In the case of first-stage screening by Mat-CHs, MAP and UTA-PI, the DR was similar to that achieved by screening the whole population with the triple test by limiting measurement of PLGF in the second stage to approximately 30% of the population for an overall SPR of 10% and 40% for a SPR of 20%. In the case of first-stage screening by Mat-CHs, MAP and PLGF, a similar DR was achieved as in screening the whole population with the triple test by limiting measurement of UTA-PI in the second stage to approximately 20% of the population for an overall SPR of 10% and 30% for SPR of 20%.

FIGURE 18.

Relationship between model-based DR of pre-eclampsia with delivery at < 37 weeks’ gestation (dark blue curve) and < 32 weeks’ gestation (light blue curve) and percentage of the population requiring second-stage screening at a fixed overall SPR of 20% in women of white racial origin. Screening at first stage (a) Mat-CHs; (b) Mat-CHs, MAP and UTA-PI; and (c) Mat-CHs, MAP and PLGF.

Discussion

The findings of this study demonstrate that a similar SPR and DR can be achieved with a two-stage strategy of screening at substantially lower costs than with carrying out screening with all biomarkers in the whole population. If the method of first-stage screening is Mat-CHs, then measurement of biomarkers can be reserved for only 70% of the population and if some of the biomarkers are included in first-stage screening then the need for the complete triple test can be reduced to 30–40% of the population.

Screening by the triple test in a population of white women resulted in a DR of preterm pre-eclampsia of 67% at a SPR of 10%, and this increased to 81% at a SPR of 20%. Randomised trials on the use of aspirin have reported that the drug is not associated with increased risk of adverse events, and in the case of antepartum haemorrhage the risk may actually be reduced. 51 In this respect, it may be acceptable that in screening for pre-eclampsia the SPR could be about 20% so as to maximise the DR.

The inevitable consequence of fixing a risk cut-off point aiming to achieve a given SPR in a white population is that the rate would be considerably higher for women of black or South Asian racial origin. An alternative strategy in screening is to fix the SPR to be the same for all racial groups and to use different risk cut-off points for each group. However, in a multiracial society such strategy would not be easy to implement, and in any case it would be wrong because it would merely mask the increased risk for pre-eclampsia in certain racial groups. Inevitably, the overall performance of screening in a racially mixed population will depend on the proportion of the various racial groups. This is analogous to screening for Down syndrome, for which the maternal age-derived prior risk is combined with the measurement of first- and/or second-trimester biomarkers to derive the posterior risk. At a fixed-risk cut-off point both the SPR and the DR increase with maternal age and, therefore, the overall performance of screening depends of the maternal age distribution of a given study population.

The strengths of this screening study are the (1) large number of patients examined; (2) accurate recording of maternal factors and biomarkers; (3) use of Bayes’ theorem to combine the prior distribution of gestational age at delivery with pre-eclampsia from maternal factors with biomarkers to estimate patient-specific risks and the performance of screening for pre-eclampsia delivering at different stages of pregnancy; and (4) comparison of model-based results and empirical results on performance of screening.

The observed performance of two-stage screening applies to our study population and comparison between studies requires the appropriate adjustments for the characteristics of the population under investigation. In the application of screening in different countries it is likely that adjustments would be necessary for the calculation of MoM values for the biomarkers and establishment of a system for quality assurance of the measurements.

Methods

Results

Case–control study

In women who developed pre-eclampsia, compared with those without pre-eclampsia, the MoM values of inhibin A (Figure 19), UTA-PI and MAP were increased and PLGF was decreased, and the deviation from normal was greater for early pre-eclampsia than for late pre-eclampsia for all four biomarkers.

FIGURE 19.

Scatter diagram and regression line for the relationship between inhibin A MoM values and gestational age at delivery in pregnancies with pre-eclampsia.

The empirical and model-based performance of screening for pre-eclampsia by maternal factors with combinations of biomarkers is shown in Table 17. Addition of inhibin A was associated with a non-significant improvement in the performance of screening by (1) Mat-CHs and PLGF, (2) Mat-CHs, MAP and PLGF and (3) Mat-CHs, MAP, UTA-PI and PLGF. The empirical results were comparable to the model-based results. However, in view of the small number of cases examined and the wide CIs, there is considerable uncertainty concerning the additional value of inhibin A in improving the performance of screening achieved by the other biomarkers.

TABLE 17 - Empirical (with 95% CI) and model-based DRs at a 10% FPR in screening for pre-eclampsia with delivery at < 37 and < 32 weeks’ gestation by maternal factors and combinations of biomarkers

Method of screening	Pre-eclampsia with delivery at < 37 weeks’ gestation		Pre-eclampsia with delivery at < 32 weeks’ gestation
	Empirical (95% CI)	Modelled	Empirical (95% CI)	Modelled
Mat-CHs	38.3 (24.5 to 53.6)	46.6	33.3 (7.5 to 70.1)	52.6
Mat-CHs + PLGF	46.8 (32.1 to 61.9)	64.2	66.7 (29.9 to 92.5)	74.1
Mat-CHs + inhibin A	46.8 (32.1 to 61.9)	52.1	66.7 (29.9 to 92.5)	64.1
Mat-CHs + PLGF + inhibin A	55.3 (40.1 to 69.8)	67.1	77.8 (40.0 to 97.2)	80.9
Mat-CHs + MAP	46.8 (32.1 to 61.9)	53.4	66.7 (29.9 to 92.5)	65.5
Mat-CHs + MAP + UTA-PI	70.2 (55.1 to 82.7)	69.6	88.9 (51.8 to 99.7)	83.6
Mat-CHs + MAP + PLGF	59.6 (44.3 to 73.6)	68.8	66.7 (29.9 to 92.5)	81.9
Mat-CHs + MAP + inhibin A	55.3 (40.1 to 69.8)	59.2	77.8 (40.0 to 97.2)	74.1
Mat-CHs + MAP + PLGF + inhibin A	68.1 (52.9 to 80.9)	72.1	77.8 (40.0 to 97.2)	85.8
Mat-CHs + MAP UTA-PI + PLGF	74.5 (59.7 to 86.1)	76.7	88.9 (51.8 to 99.7)	90.5
Mat-CHs + MAP + UTA-PI + PLGF + inhibin A	72.3 (57.4 to 84.4)	77.8	88.9 (51.8 to 99.7)	92.8

It was concluded that inclusion of serum inhibin A could potentially improve the performance of first-trimester screening for pre-eclampsia, but this needs to be investigated further by prospective screening studies.

Discussion

This study has demonstrated the feasibility of utilising the competing risk model to investigate the potential value of new biomarkers of pre-eclampsia.

The findings of the case–control study demonstrated that in women who developed pre-eclampsia, inhibin A is increased and the deviation from normal is greater for early pre-eclampsia than late pre-eclampsia. On the basis of modelling, it was suggested that addition of inhibin A may improve the performance of screening by Mat-CHs and combinations of MAP, UTA-PI and PLGF. Although the empirical results were compatible with the model-based results, there was considerable uncertainty concerning the additional value of inhibin A in improving the performance of screening achieved by the other biomarkers. The results of the screening demonstrated that inhibin A is unlikely to be a useful first trimester biomarker of pre-eclampsia.

Methods

The data for this study were derived from the SPREE study. 38 We investigated the performance of screening for SGA by a combination of Mat-CHs and biomarkers at 11–13 weeks’ gestation. In the ASPRE trial, women with singleton pregnancies identified by combined screening as being at high risk for preterm pre-eclampsia (> 1 in 100) participated in a trial of aspirin (150 mg/day from 11 to 14 weeks’ gestation until 36 weeks’ gestation) compared with placebo. 40 We used the data from the SPREE study to estimate the proportion of SGA neonates born at ≥ 37, < 37 and < 32 weeks’ gestation, with first-trimester combined risk for preterm pre-eclampsia (calculated by the competing risk model through a combination of Mat-CHs, MAP, UTA-PI and PLGF) of > 1 in 100. 17,33

Data on pregnancy outcome were collected from the hospital maternity records or the women’s general medical practitioners. The diagnosis of SGA < 10th, < 5th or < 3rd percentile was based on a reference range of birthweight with gestational age in our population. 61 The pregnancies with SGA fetuses were subdivided according to the presence or absence of pre-eclampsia. 21,34

Results

First-trimester screening was carried out in 17,051 pregnancies, but 304 were lost to follow-up and 296 resulted in miscarriage or pregnancy termination at < 24 weeks’ gestation. Therefore, the study population comprised 16,451 pregnancies. 60

Pre-eclampsia developed in 473 (2.9%) pregnancies, including 33 with early pre-eclampsia, 142 with preterm pre-eclampsia and 331 with term pre-eclampsia (Table 19). In the early pre-eclampsia group, 84.8%, 84.8% and 81.8% of babies were SGA < 10th, < 5th and < 3rd percentile, respectively. In the preterm pre-eclampsia group, 70.4%, 60.6% and 54.9% of babies were SGA < 10th, < 5th and < 3rd percentile, respectively. In the term pre-eclampsia group, 19.3%, 13.9% and 11.5% of babies were SGA < 10th, < 5th and < 3rd percentile, respectively. 60

Birth at < 32, < 37 and ≥ 37 weeks’ gestation occurred in 172 (1.0%), 950 (5.8%) and 15,501 (94.2%) pregnancies, respectively. The proportion of babies who were SGA and the contribution of pre-eclampsia to the incidence of SGA are shown in Table 20. Essentially, 41.3% of babies born at < 32 weeks were SGA < 10th percentile and in 39.4% of such pregnancies there was pre-eclampsia. Similarly, a high proportion of babies born at < 37 weeks were SGA < 10th percentile (36.1%) and in 29.2% of such pregnancies there was pre-eclampsia. In contrast, the incidence of SGA < 10th in babies born at term was 11.4%, and only 3.6% of these babies were from pregnancies with pre-eclampsia. 60

In screening by a combination of Mat-CHs, MAP, UTA-PI and PLGF and a risk cut-off point for preterm pre-eclampsia of 1 in 100, the SPR was 12.2% (2014/16,451) and the high-risk group included 30 (90.9%) of early pre-eclampsia, 118 (83.1%) of preterm pre-eclampsia and 162 (48.9%) of term pre-eclampsia. The proportion of SGA neonates from all pregnancies and those with and without pre-eclampsia with a first-trimester combined risk for preterm pre-eclampsia of > 1 in 100 is shown in Table 21. 60

In the high-risk group, the proportions of neonates born at ≥ 37 weeks' gestation who were below the 10th, 5th and 3rd weight percentile were 20.7%, 25.2% and 26.0%, respectively. The corresponding percentages for those born at < 37 weeks’ gestation were 45.8%, 48.9% and 51.8% and for those born at < 32 weeks’ gestation were 56.3%, 60.0% and 63.2%. In the group of SGA < 10th percentile neonates from pregnancies with pre-eclampsia, the high-risk group included about 81% of those born at < 37 weeks’ gestation and 89% of those born at < 32 weeks’ gestation. The corresponding values from pregnancies without pre-eclampsia were approximately 31% and 35%. In the group of SGA < 5th percentile neonates from pregnancies with pre-eclampsia, the high-risk group included about 79% of those born at < 37 weeks’ gestation and 89% of those born at < 32 weeks’ gestation. The corresponding values from pregnancies without pre-eclampsia were approximately 34% and 38%. In the group of SGA < 3rd percentile neonates from pregnancies with pre-eclampsia, the high-risk group included about 78% of those born at < 37 weeks’ gestation and 89% of those born at < 32 weeks’ gestation. The corresponding values from pregnancies without pre-eclampsia were approximately 38% and 40%. 60

Discussion

The findings of this study demonstrate that (1) neonates are SGA < 10th percentile in 70% of pregnancies with preterm pre-eclampsia and in 85% of those with early pre-eclampsia; (2) in women who deliver preterm and early SGA neonates there is pre-eclampsia in 29% and 39% of cases, respectively; and (3) screening for preterm pre-eclampsia by a combination of Mat-CHs, MAP, UTA-PI and PLGF at 11–13 weeks’ gestation identifies a high-risk group that contains about 46% of cases of preterm SGA and 56% of cases of early SGA.

Analysis of data from the SPREE study and the ASPRE trial showed that in pregnancies identified by first-trimester combined screening as being at high-risk for preterm pre-eclampsia, use of aspirin reduces the incidence of preterm SGA by about 40% and that of early SGA by 73%. 60 The aspirin-related decrease in incidence of SGA was mainly due to a decrease in the pregnancies with pre-eclampsia, the decrease being approximately 70% in babies born at < 37 weeks’ gestation and approximately 90% in babies born at < 32 weeks’ gestation. In the pregnancies without pre-eclampsia use of aspirin was not associated with a significant reduction in incidence of SGA. On the basis of these results, it was estimated that first-trimester screening for preterm pre-eclampsia and use of aspirin in the high-risk group would potentially reduce the incidence of preterm and early SGA by about 20% and 40%, respectively. 60

Preterm pre-eclampsia is, to a great extent, predictable by first-trimester combined screening and preventable by use of aspirin (i.e. 150 mg/day from the first to the third trimester). 7,17,33 A beneficial consequence of such strategy is the prevention of a high proportion of cases of preterm SGA, because (1) preterm pre-eclampsia is commonly associated with SGA and (2) a high proportion of cases of preterm SGA are associated with pre-eclampsia.

The ASPRE trial demonstrated that, in women identified by means of first-trimester screening as being at high risk for pre-eclampsia, use of aspirin reduces the incidence of preterm pre-eclampsia and early pre-eclampsia by approximately 60% and 90%, respectively. 7 A secondary analysis of the trial demonstrated that use of aspirin reduces the length of stay in the neonatal intensive care unit by approximately 70%. 62 More than 80% of the length of stay was contributed from babies born at < 32 weeks’ gestation and the aspirin-related reduction in length of stay could essentially be attributed to prevention of early pre-eclampsia. The consequence of reduction in length of stay in the neonatal intensive care unit is a substantial saving in health-care cost, which is well in excess of the cost of population screening and treatment of the high-risk group with aspirin. Reduction in the risk of birth at < 32 weeks’ gestation is also likely to be associated with reduction in risk of infant death, cerebral palsy and long-term use of specialised health-care resources. 63–65

In conclusion, first-trimester screening for pre-eclampsia by the combined test identifies a high proportion of cases of preterm SGA that can potentially be prevented by the prophylactic use of aspirin. Most cases of SGA, particularly those delivering at term, are neither predictable in the first trimester nor preventable by prophylactic use of aspirin.

Contributions of authors

Liona C Poon (https://orcid.org/0000-0002-3944-4130) (Professor, Maternal Fetal Medicine Specialist) conceptualised the study and designed the research, performed the research and approved the report.

David Wright (https://orcid.org/0000-0003-4800-3190) (Professor, Biostatistician) conceptualised the study and designed the research, analysed and interpreted data, and approved the report.

Steve Thornton (https://orcid.org/0000-0001-7613-0408) (Professor, Obstetrics Specialist) conceptualised the study and designed the research, and approved the report.

Ranjit Akolekar (https://orcid.org/0000-0001-7265-5442) (Professor, Maternal Fetal Medicine Specialist) performed the research and approved the report.

Peter Brocklehurst (https://orcid.org/0000-0002-9950-6751) (Professor, Epidemiologist) managed the study and approved the report.

Kypros H Nicolaides (https://orcid.org/0000-0003-3515-7684) (Professor, Fetal Medicine specialist) conceptualised the study and designed the research, performed the research and approved the report.

Publications

Tan MY, Poon LC, Rolnik DL, Syngelaki A, de Paco Matallana C, Akolekar R, et al. Prediction and prevention of small-for-gestational-age neonates: evidence from SPREE and ASPRE. Ultrasound Obstet Gynecol 2018;52:52–9.

Tan MY, Syngelaki A, Poon LC, Rolnik DL, O’Gorman N, Delgado JL, et al. Screening for pre-eclampsia by maternal factors and biomarkers at 11–13 weeks’ gestation. Ultrasound Obstet Gynecol 2018;52:186–95.

Tan MY, Wright D, Syngelaki A, Akolekar R, Cicero S, Janga D, et al. Comparison of diagnostic accuracy of early screening for pre-eclampsia by NICE guidelines and a method combining maternal factors and biomarkers: results of SPREE. Ultrasound Obstet Gynecol 2018;51:743–50.

Data-sharing statement

We shall make data available to the scientific community with as few restrictions as feasible, while retaining exclusive use until the publication of major outputs. Anonymised data will be deposited here at URL: www.fetalmedicine.org to encourage wider use.

Patient data

This work uses data provided by patients and collected by the NHS as part of their care and support. Using patient data is vital to improve health and care for everyone. There is huge potential to make better use of information from people’s patient records, to understand more about disease, develop new treatments, monitor safety, and plan NHS services. Patient data should be kept safe and secure, to protect everyone’s privacy, and it’s important that there are safeguards to make sure that it is stored and used responsibly. Everyone should be able to find out about how patient data are used. #datasaveslives You can find out more about the background to this citation here: https://understandingpatientdata.org.uk/data-citation.

Disclaimers

This report presents independent research. The views and opinions expressed by authors in this publication are those of the authors and do not necessarily reflect those of the NHS, the NIHR, the MRC, NETSCC, the EME programme or the Department of Health and Social Care. If there are verbatim quotations included in this publication the views and opinions expressed by the interviewees are those of the interviewees and do not necessarily reflect those of the authors, those of the NHS, the NIHR, NETSCC, the EME programme or the Department of Health and Social Care.

The prior model

The prior model assumes that g has a Gaussian distribution with mean µ_g and standard deviation σ_g. The mean µ_g is obtained from the regression model with the parameters given in Table 22.

Likelihood

The likelihood is a multivariate Gaussian density for the distribution of biomarker log₁₀ MoM values conditional on g. The elements of the mean vector (µ_x) of this distribution are given by the broken stick model:

0 else.

The estimated regression parameters and biomarker likelihood covariance matrix are given in Tables 23 and 24, respectively.

Posterior probability density

The posterior probability density p(g), up to a normalising constant, is calculated by multiplying the likelihood by the prior:

In Equation 4, dmvnorm is the multivariate Gaussian density for x, with mean µ_x and covariance matrix Σ_x. dnorm is the univariate Gaussian density for g, with mean µ_g and standard deviation σ_g according to the history model.

Posterior risks

The risk of pre-eclampsia requiring delivery before gestation G (weeks), assuming no other-cause delivery is given by:

where g* = max(24, Current gestation).

Data

1. n: 16,747.

2. n.centre: 7.

3. centre: vector of 16,747 centre labels.

4. event: vector of 16,747 observed pre-eclampsia events (0 = no event; 1 = pre-eclampsia).

5. aspirin: vector of 16,747 aspirin indicators (0 = no aspirin; 1 = aspirin).

6. nice: vector of 16,747 indicators (0 = NICE negative, 1 = NICE positive).

7. risk: vector of 16,747 indicators (0 = risk negative, 1 = risk positive).

5000 Markov chain Monte Carlo iterations were carried out to ensure convergence and checked visually using parameter trace plots.

A further 1000 iterations were carried out and values of the counts s1, s2, s3 and s4 were recorded for observation 100, 200, . . . , 1000, providing 10 imputed data sets. The observed event counts and imputed counts, assuming a relative risk reduction with aspirin of 0.3, allowing for the effects of aspirin are shown in Table 25. It is notable that the number of events imputed in the NICE positive and risk negative cell exceeds that in the NICE negative and risk positive group. The reflects the different numbers treated by aspirin in the two groups. This reduces the effect size (see Figures 8 and 9).

Estimates and their standard errors of the difference in DRs were computed for each of the imputed data sets. These were then pooled using the formula (3.1.2) of Rubin. 66

All pre-eclampsia

Difference in DRs (i.e. mini-combined test – NICE method) of pre-eclampsia at any gestational age for a fixed SPR of 10.3% defined by NICE (Figure 20).

FIGURE 20.

Difference in DRs for pre-eclampsia with delivery at any gestational age [mini-combined test (i.e. Mat-CHs, MAP + PAPP-A) vs. the NICE method] with a relative risk reduction from aspirin of (a) 30%; (b) 50%; and (c) 100%.

Pre-eclampsia with delivery before 37 weeks

Difference in DRs (i.e. mini-combined test vs. NICE method) of preterm pre-eclampsia (< 37 weeks’ gestation) for a fixed SPR of 10.3% defined by NICE (Figure 21).

FIGURE 21.

Difference in DRs (Mat-CHs + MAP + PAPP-A vs. the NICE method) with a relative risk reduction from aspirin of (a) 60%; (b) 80%; and (c) 100%.

Pre-eclampsia with delivery before 37 weeks

Difference in DRs (i.e. Mat-CHs, MAP and PLGF vs. the NICE method) of preterm pre-eclampsia (< 37 weeks’ gestation) for a fixed SPR of 10.3% defined by NICE (Figure 22).

FIGURE 22.

Difference in DRs (Mat-CHs + MAP + PLGF vs. the NICE method) with a relative risk reduction from aspirin of (a) 60%; (b) 80%; and (c) 100%.

Pre-eclampsia with delivery before 37 weeks

Difference in DRs (i.e. Mat-CHs, MAP, UTA-PI and PLGF vs. the NICE method) of preterm pre-eclampsia (< 37 weeks’ gestation) for a fixed SPR of 10.3% defined by NICE (Figure 23).

FIGURE 23.

Difference in DRs (Mat-CHs + MAP + PLGF vs. the NICE method) with a relative risk reduction from aspirin of (a) 60%; (b) 80%; and (c) 100%.