Prediction of complications in early-onset pre-eclampsia (PREP): development and external multinational validation of prognostic models

We approached potentially eligible mothers from 53 obstetric units within secondary and tertiary care hospitals in the UK. Information about the study was provided when women attended the hospital for obstetric care. We recruited consecutive women with suspected or confirmed pre-eclampsia before 34 weeks’ gestation from December 2011 to April 2014. Women with confirmed early-onset pre-eclampsia contributed to model development (Additional file 1: Table S1). Women were excluded from the study if the outcome of interest had occurred prior to the assessment of predictors, there was insufficient time to obtain informed consent or there was a lack of translators for non-English speaking mothers. All women in the study were managed according to National Institute for Health and Care Excellence guidelines on Hypertension in Pregnancy [ 18].

An initial list of 33 candidate predictors was identified from published systematic reviews [ 19– 24] and primary studies [ 4, 25– 30], and subsequently (before any data analysis) reduced to 22 after prioritising their importance by a Delphi consensus among experts [ 22]. Candidate predictors included maternal characteristics, relevant medical history, symptoms, signs, investigations and interventions that have the potential to modify the probability of outcome (Table  1). In cases of multiple collected values for the same predictor, we chose the worst value within the first 24 hours. We also considered predictors for fetal outcomes shown in Additional file 1: Table S2.

An independent panel of experts ranked the outcomes for their importance to clinical practice [ 17]. Based on the PIERS study, the components of the outcome were identified by a Delphic consensus [ 3]. We defined the primary outcome as maternal complication that included maternal death, neurological, hepatic, cardiorespiratory, renal or haematological complications, or delivery before 34 weeks (Additional file 1: Table S3a). The panel agreed that delivery before 34 weeks is often offered for medical reasons to avoid complications as per the national guidelines; thus, excluding this as an adverse outcome would underestimate the true incidence of adverse outcomes and lead to prognostic models that yield too low risk predictions of actually developing an adverse outcome [ 17]. Hence, deliveries before 34 weeks’ gestation were added as a component of the primary outcome before data analysis. The secondary outcomes included composite perinatal outcomes by discharge, which were also developed by Delphi consensus (Additional file 1: Table S3b) [ 3].

Based on systematic reviews on accuracy of tests in predicting complications in women with pre-eclampsia [ 3, 19, 23, 24], we expected 20% of women with early-onset pre-eclampsia to develop adverse maternal outcome at any time point, and about 10 events were required for each candidate predictor to reduce model overfitting issues [ 31]. We revised our initial sample size of 500 women, to continue recruitment until 100 adverse maternal events were reached. Prior to analysis, discussions with the independent Steering Committee resulted in extension of the primary outcome to also include delivery before 34 weeks’ gestation [ 17] (Additional file 1: Table S4).

We externally validated the developed PREP models in two independent datasets, namely the PIERS (Pre-eclampsia Integrated Estimate of RiSk) study [ 4], a prospective cohort on prediction of adverse maternal outcomes in women with any-onset pre-eclampsia (634 women recruited in Canada, New Zealand, Australia and the UK), and the PETRA (Pre-Eclampsia TRial Amsterdam) study, a randomised trial on effectiveness of plasma expansion in the management of early-onset hypertensive disease in pregnancy (216 women) [ 11].

A logistic regression model was used to predict the risk of adverse perinatal outcomes, as for the maternal outcomes. In addition to the maternal predictors, we evaluated cardiotocography and uterine artery Doppler. Pregnancy was the unit of analysis and non-linear terms were not considered. A similar variable selection process on the full list of maternal and fetal candidate predictors was done.

All analyses were carried out using Stata version 12.0 [ 35].

Over 90% (866/954) of women with early-onset pre-eclampsia had new onset disease. More than half of all women were nulliparous (551/954, 58%) and the mean gestational age of diagnosis was 30.5 (SD 2.9) weeks. More than three-quarters of women were on anti-hypertensives at baseline (753/948, 79%). Of the 15% of mothers who received magnesium sulfate at baseline in the PREP study, the rationale was documented as severe pre-eclampsia in 72% (104/144), and not reported in the rest (Table  1).

About a fifth of mothers diagnosed with the condition suffered a complication (169/946, 18%) by 48 hours after diagnosis of early onset pre-eclampsia, and two-thirds (633/946, 66.9%) by postnatal discharge. The most frequent outcome was early preterm delivery before 34 weeks gestation (580/946, 61.3%), followed by postpartum haemorrhage (74/946, 7.8%) (Additional file 1: Table S5a). Three-quarters of all neonates (702/945, 74%) had suffered at least one complication by discharge, with admission to the neonatal intensive care unit being the most common outcome in 72% (681/945) (Additional file 1: Table S5b).

In the external datasets, data were available for 636 (PIERS study) and 216 women (PETRA study) with early-onset pre-eclampsia (Additional file 1: Table S6). The reduced version of the PREP-S (without serum urea and deep tendon reflex) and PREP-L (without serum urea) models were evaluated in the PIERS and PETRA cohorts.

An increase in the gestational age at diagnosis of pre-eclampsia and any pre-existing medical history were associated with a reduction in the odds of perinatal complications (OR, 0.91; 95% CI, 0.39–0.99 and OR, 0.65; 95% CI, 0.44–0.98, respectively). Raised urine protein:creatinine ratio (OR, 1.29; 95% CI, 1.11–1.50), serum urea (OR, 1.72; 95% CI, 1.07–2.76), management with anti-hypertensives (OR, 1.56; 95% CI, 1.04–2.37), treatment with magnesium sulfate (OR, 2.40; 95% CI, 1.04–5.57), abnormal uterine artery Doppler (OR, 1.94; 95% CI, 1.08–3.51), and when the expected fetal weight was less than 10th centile by ultrasound (OR, 2.54; 95% CI, 1.46–4.40) were all associated with an increase in the odds of complications (Additional file 1: Table S2).

We developed the PREP models by using a large, well-defined and prospectively collected cohort of mothers from several centres in the UK. We clearly defined the predictors, and only used values within 24 hours of diagnosis of early onset pre-eclampsia, to allow application of the model at the point of diagnosis. The final models included clinically relevant predictors used in practice to assess disease severity, such as blood pressure, thereby providing face validity. Missing values of predictors were dealt with multiple imputation to avoid loss of useful information. Overfitting issues were reduced by a large sample size and through optimism-adjustment methods.

We identified the components of the composite outcome by a Delphi survey of experts, and prioritised for their clinical importance, thereby ensuring that only the clinically most relevant outcomes were included. An a priori expert workshop provided guidance to minimise bias from treatment paradox. We addressed this problem by including administration of anti-hypertensives and magnesium sulfate as predictors, and delivery before 34 weeks as an outcome [ 17]. In women with severe pre-eclampsia, magnesium sulfate is administered either for prevention of eclamptic seizures, or more recently, for neuroprotection of the preterm infant [ 37]. However, we expect the proportion of women who received it solely for fetal neuroprotection to be small, and less likely to significantly alter the performance of the models. We had sufficient sample size for the chosen number of candidate predictors [ 12]. Our models, particularly PREP-L and also PREP-S by 48 hours, validated well in external populations despite variations in country, setting and management. This suggests the models are transportable and potentially useful for other non-UK populations.

We only assessed predictors that are available in current everyday practice in high-resource settings, which may be seen as a weakness. Differences in management could affect the performance of the model. Although the models included predictors of composite outcome, we were unable to identify those associated with specific outcomes due to the small numbers of individual complications. The logistic model does not allow for predictions over time and therefore we also developed a survival model to provide risks at various time points including at 48 hours. However, we censored at 34 weeks, as one of the components of the outcome was delivery by 34 weeks. We were also unable to assess the added predictive contribution of biomarkers (sFlt1, sEng, PIGF) in maternal blood or urine, because the planned ASTRONAUT study which was going to provide data on biomarkers did not commence. We also could not differentiate iatrogenic premature births from spontaneous births. However, it is likely that pre-eclampsia and its complications, such as abruption, could predispose to spontaneous preterm labour [ 38]. Although admission to neonatal intensive care unit was considered an indication of morbidity by Delphi consensus [ 13, 39], its inclusion as a perinatal outcome may have weakened the prediction of other more clinically relevant outcomes. Due to the non-availability of all predictors in external datasets, we were unable to validate the original PREP models. However, our validation of reduced PREP-L and PREP-S models gives an insight into how the original models might perform. For example, a reduced PREP-L model showed good discrimination and calibration for predicting overall risk by discharge, and thus we expect the original PREP-L to have similar, if not better, performance when fully externally validated.

Current prediction models, such as PIERS and mini PIERS, included women with any onset pre-eclampsia, and not specifically those with early onset pre-eclampsia [ 4, 40]. Furthermore, existing models do not have sufficient sample size for predicting complications at 48 hours in women with early onset pre-eclampsia. They evaluated a large number of predictors for relatively fewer outcomes observed at 48 hours, and used the worst values of the tests, which may have overestimated the performance of the model [ 12]. Effective interventions such as delivery and anti-hypertensives could have contributed to the observed absence of traditional risk factors such as blood pressure, a crucial part of clinician’s risk assessment, in these models [ 4, 40].

The performance of the PIERS model was dominated by haematological complications, such as transfusion of any blood product or low platelet count, which contributed to about half of all complications. Transfusion of blood products scored the lowest for clinical importance in the management of women with early-onset pre-eclampsia, in a survey of experts [ 41]. Our PREP models’ predictions were heavily influenced by risk of early delivery before 34 weeks, a clinically more important outcome which helps inform mothers make informed decisions. Compared to the current PIERS model that had a C-statistic of between 0.7 and 0.8 at one week after admission, the PREP-L model showed discrimination of over 0.8 for a longer period until postnatal discharge from the hospital.

Currently, mothers diagnosed with early onset pre-eclampsia in secondary care are transferred to tertiary hospitals. However, with only a third of these mothers suffering a complication by 48 hours, in current practice, many mothers diagnosed with the condition in secondary care would have been transferred to a tertiary centre. The promising discrimination and calibration performance of the PREP-S at 48 hours makes it suitable for use as a triage tool, in accurately identifying mothers for in utero transfer to tertiary care unit. It will also determine use of corticosteroids depending on the predicted probability of complications. With good agreement between the predicted and observed risk of complications in the PREP-S at 48 hours in the low- and intermediate-risk groups, women with a predicted probability of complications below 50% can avoid unnecessary transfer to tertiary units. Women categorised to be low risk by the PREP-L model could be followed-up in an outpatient setting, with high- and very high-risk women monitored as inpatients with regular intensive monitoring.

The model is available for everyday use in clinical practice as an excel sheet, and is being developed as an app to obtain individualised risk estimates for decisions on further management. Provision of personalised risk information allows parents to have the opportunity to discuss the expected outcomes. It is important to recognise that all prediction models in this field, including our PREP models, provides risk estimates in the context of current care and clinical management decisions. The models are not designed to guide clinicians’ decisions on choice of management such as timing of delivery, administration of anti-hypertensives and magnesium sulfate. A woman with a low predicted risk should be viewed as an individual with low outcome risk if current care pathways are used, as it may be the clinical care that results in her low-risk status.

The next stage in this prediction research is evaluation of the impact of using PREP models in clinical practice, in terms of whether they help improve subsequent maternal and perinatal outcomes; this requires well-planned robust randomised trials. The threshold for intervention, such as transfer to a tertiary unit or admission to the hospital and subsequent delivery, needs to be established by working with Patient and Public Involvement groups. The design for such clinical trials should be to randomise centres on use of PREP models as triaging tools to plan management in women with early onset pre-eclampsia. Evaluating the added value of biomarkers to our models may also be a subject of future research, though they have not been found useful in recent research [ 42]. This may be supported by individual patient data meta-analyses to improve power to identify important predictors.