A Risk Prediction Model for the Assessment and Triage of Women with Hypertensive Disorders of Pregnancy in Low-Resourced Settings: The miniPIERS (Pre-eclampsia Integrated Estimate of RiSk) Multi-country Prospective Cohort Study

Background:Pre-eclampsia/eclampsia are leading causes of maternal mortality and morbidity, particularly in low- and middle- income countries (LMICs). We developed the miniPIERS risk prediction model to provide a simple, evidence-based tool to identify pregnant women in LMICs at increased risk of death or major hypertensive-related complications.Methods and Findings:From 1 July 2008 to 31 March 2012, in five LMICs, data were collected prospectively on 2,081 women with any hypertensive disorder of pregnancy admitted to a participating centre. Candidate predictors collected within 24 hours of admission were entered into a step-wise backward elimination logistic regression model to predict a composite adverse maternal outcome within 48 hours of admission. Model internal validation was accomplished by bootstrapping and external validation was completed using data from 1,300 women in the Pre-eclampsia Integrated Estimate of RiSk (fullPIERS) dataset. Predictive performance was assessed for calibration, discrimination, and stratification capacity. The final miniPIERS model included: parity (nulliparous versus multiparous); gestational age on admission; headache/visual disturbances; chest pain/dyspnoea; vaginal bleeding with abdominal pain; systolic blood pressure; and dipstick proteinuria. The miniPIERS model was well-calibrated and had an area under the receiver operating characteristic curve (AUC ROC) of 0.768 (95% CI 0.735-0.801) with an average optimism of 0.037. External validation AUC ROC was 0.713 (95% CI 0.658-0.768). A predicted probability ≥25% to define a positive test classified women with 85.5% accuracy. Limitations of this study include the composite outcome and the broad inclusion criteria of any hypertensive disorder of pregnancy. This broad approach was used to optimize model generalizability.Conclusions:The miniPIERS model shows reasonable ability to identify women at increased risk of adverse maternal outcomes associated with the hypertensive disorders of pregnancy. It could be used in LMICs to identify women who would benefit most from interventions such as magnesium sulphate, antihypertensives, or transportation to a higher level of care.Please see later in the article for the Editors’ Summary. © 2014 Payne et al. The miniPIERS model was developed and validated on a prospective, multicentre cohort of women admitted to a participating centre with an HDP. Participating institutions were: the Colonial War Memorial Hospital, Suva, Fiji; Mulago Hospital, Kampala, Uganda; Tygerberg Hospital, Cape Town, South Africa; Maternidade Escola de Vila Nova Cachoeirinha, São Paulo, Brazil; Aga Khan University Hospital and its secondary level hospitals at Garden, Karimabad and Kharadar and Jinnah Post-graduate Medical College, Karachi, Pakistan; and Aga Khan Maternity & Child Care Centre, and Liaqat University of Medical Sciences, Hyderabad, Pakistan. Ethics approval for this study was obtained from each participating institution’s research ethics board as well as the clinical research ethics board at the University of British Columbia. All participating institutions had a hospital policy of expectant management for women with pre-eclampsia remote from term, and similar guidelines for treatment of women with regard to magnesium sulphate and antihypertensive agents. Institutions were chosen to participate on the basis of the consistency of these guidelines in order to achieve some level of homogeneity within the cohort and to reduce systematic bias that could result from differences in disease-modifying practices between institutions. Women were admitted to the study with any HDP defined as follows: pre-eclampsia, defined as (i) blood pressure (BP) ≥140/90 mmHg (at least one component, twice, ≥4 and up to 24 hours apart, after 20 weeks) and either proteinuria (of ≥2+ by dipstick, ≥300 mg/d by 24 hour collection, or ≥30 g/mol by urinary protein:creatinine ratio) or hyperuricaemia (greater than local upper limit of local non-pregnancy normal range); (ii) haemolysis, elevated liver enzymes, and low platelets (HELLP) syndrome even in the absence of hypertension or proteinuria [1]; or (iii) superimposed pre-eclampsia (clinician-defined rapid increase in requirement for antihypertensives, systolic BP [sBP] ≥170 mmHg or diastolic BP [dBP] ≥120 mmHg, new proteinuria, or new hyperuricaemia in a woman with chronic hypertension); or an “other” HDP defined as: (i) gestational hypertension (BP≥140/90 mmHg [at least one component, twice, ≥4 hours apart, ≥20+0 weeks] without significant proteinuria); (ii) chronic hypertension (BP≥140/90 mmHg before 20+0 weeks’ gestation); or (iii) partial HELLP (i.e., haemolysis and low platelets OR low platelets and elevated liver enzymes). All women participating in the study gave informed consent according to local ethics board requirements. Women were excluded from the study if they were admitted in spontaneous labour, experienced any component of the adverse maternal outcome before eligibility or collection of predictor variables, or had confirmed positive HIV/AIDS status with CD4 count <250 cells/ml or AIDS-defining illness. Candidate predictor variables for final model development were identified a priori as being those variables that: (i) would be available and easy to collect in all health care settings including the woman's home; (ii) have been shown to be associated with pre-eclampsia in previous studies [12]; and (iii) would be measurable using simple and reliable methods. These variables included demographics (maternal age, parity, and gestational age on admission); symptoms (headache, visual disturbances, chest pain/dyspnoea, right upper quadrant pain or epigastric pain, nausea, vomiting, and vaginal bleeding with abdominal pain); and signs (blood pressure and dipstick proteinuria). The values for these variables were collected prospectively from the woman's medical record as measured by the nurse or physician during regular antenatal, intrapartum, or postnatal care. If multiple measures of a candidate predictor were collected within the first 24 hours of admission, the worst predictor value obtained within that first 24 hours of admission was used. The value used was the worst in the clinical context, this could either be the highest or lowest value collected in the given 24 hour time period, depending on the measure in question. This method of using the worst value was chosen as it is consistent with clinical practice. Generally, clinicians will respond to the worst clinical value when making management decisions. The components of the composite adverse maternal outcome to be predicted by the model were determined by Delphi consensus [13] and include maternal mortality or one or more of serious central nervous system, cardiorespiratory, renal, hepatic, haematological, or other morbidity. The Delphi consensus process involved iterative review and feedback on the proposed outcome components from an expert group consisting of researchers and clinicians from both high- and low- or middle- income countries who have published work focused on HDPs. Representatives of the Delphi group brought expertise from medicine, obstetrics, pediatrics, anaesthesia, and critical care with sub-specialty expertise in maternal-fetal medicine, nephrology, haematology, and placental biology. Data were collected on the occurrence of all outcome components at any time during admission but for the purpose of the model, only those that occurred within 48 hours of admission were considered. All study sites were instructed to collect information on any “other” adverse events the woman experienced during pregnancy or immediately postpartum as part of the regular data collection process. This was done to ensure balanced reporting of events across all sites. Any reported “other” events were adjudicated by the study Working Group during regular meetings, at which time the decision was made whether to include the reported outcome as a study outcome, or not. A full description of data collected can be found on the study website (http://pre-empt.cfri.ca/OBJECTIVES/miniPIERS/Reference.aspx) and definitions of the adverse maternal outcome components are provided in Table S1. The external validation study was performed using data from the fullPIERS [11] dataset. The fullPIERS study was performed to develop and validate a prediction model for assessing risk in women with confirmed diagnoses of pre-eclampsia in high-resourced settings. This model includes gestational age at admission, chest pain/dyspnoea, oxygen saturation, platelet count, creatinine, and aspartate aminotransferase. Participating centres were tertiary academic hospitals located in Canada (six), the UK (two), New Zealand (one), and Australia (one). Only the fullPIERS data collected after 1 March 2008 were used for this study as this portion of the fullPIERS cohort was collected using the same protocol, inclusion and exclusion criteria, and data collection tools as later used for miniPIERS. Prior to this date, the fullPIERS cohort did not include abdominal pain, vaginal bleeding, or any headache. Any researcher interested in accessing the miniPIERS or fullPIERS data can do so through the Pre-eclampsia CoLaboratory (http://pre-empt.cfri.ca/OBJECTIVES/CoLaboratory.aspx). A summarized version of the study dataset is also available as supplementary Dataset S1. Data for the miniPIERS dataset were collected prospectively using standardized data collection forms and protocols for all sites and entered into a customized Microsoft Access database. As part of the study protocol, women were required to have at least one measure of proteinuria, blood pressure, and symptoms during the first 24 hours of admission. All data were reviewed for quality and consistency. When questions arose regarding data, these data were confirmed by re-review of the primary health record. Random review of 10% of cases was performed during the first year of the study to ensure data validity within and between study sites. The sample size required for model development was determined on the basis of the minimum standard of ten events per effective variable considered in the model according to the formula N = (n×10)/I where N is the sample size, n is the number of candidate predictor variables, and I is the estimated event rate in the population [14]. An estimated event rate of 15% based on our pilot data was used; for a model with 15 effective candidate predictor variables (i.e., dipstick proteinuria is counted three times to reflect inclusion of three indicator variables), the sample size required was 1,000 women [15],[16]. This sample size target was doubled to allow for subgroup analysis at the conclusion of the study after the finding of confounding by centre during the interim analysis. The relationship between each predictor variable and the combined adverse maternal outcome was first assessed by univariate logistic regression. Continuous variables were assessed for non-linearity, and were modeled as restricted cubic splines when appropriate [14]. Variables with a skewed distribution were log-transformed (natural log). Inclusion of the transformed variable in the final model was based on comparison to a model with the linear variable and selection of the model with the lowest Akaike information criterion (AIC) was automated during the model development process. To avoid co-linearity, correlation between variables was determined and only the more clinically relevant variable of a pair of highly correlated variables was retained. When a high degree of correlation existed between two symptoms (r>0.5) they were re-coded as a combined indicator variable. Stepwise backward elimination was used to build the most parsimonious model with a stopping rule of p<0·20. No interaction terms were included in the model as no interaction was hypothesized between candidate predictors prior to analysis. We assessed the potential for confounding by study site by examining the bivariate association of study site with predictor variables and with outcome rate. Dummy (indicator) variables for study site were included in the model to eliminate confounding of the predictor-adverse outcome relationship by study site. To make the final model generalizable to all study settings, the coefficients for site variables were excluded from the calculation of predicted probability, and the model's intercept was adjusted using previously published methods for updating a prediction model for a new setting [14]. Calibration ability of the model was assessed visually by plotting deciles of predicted probability of an adverse maternal outcome against the observed rate in each decile and fitting a smooth line [14],[17]. Discrimination ability was evaluated on the basis of AUC ROC [18]. The sensitivity, specificity, positive predictive value, negative predictive value, and likelihood ratios (LRs) of cut-offs for a positive test defined using the population within each risk group were calculated [19]. The following categories for interpretation of the LRs were used: informative (LR<0·1 or >10); moderately informative (LR 0·1–0·2 or 5–10); and non-informative (LR 0·2–5). A risk stratification table was generated to assess the extent to which the model’s predictions divided the population into clinically distinct risk categories [20]. Internal validation of the model was assessed using 500 iterations each of Efron’s enhanced bootstrap method [21]. Details of this approach have been described previously [11],[14]. The bootstrapping procedure involved (i) sampling with replacement from the original cohort to generate a bootstrap dataset of 2,081 women; (ii) redevelopment of the model including all model development steps; variable coding (transformations and categorizations), variable selection, and parameter estimation in the bootstrapped sample; (iii) estimation of the AUC ROC for the model in the bootstrap sample; (iv) application of this new model to the original dataset and estimation of AUC ROC. Model optimism is then calculated as the average difference between model performance in the bootstrap sample and the original dataset after 500 iterations of this procedure. The choice was made to use 500 iterations because previous studies have shown no benefit is achieved when using a higher number of repetitions [16]. A final assessment of calibration was performed using the Hosmer-Lemeshow goodness-of-fit test. A final assessment of model validity was performed by applying the miniPIERS model to the fullPIERS dataset and estimating the AUC ROC. Due to the marked difference in underlying rate of outcomes in the fullPIERS population (6.5% in fullPIERS versus 12.5% in miniPIERS), the model intercept (i.e., the baseline rate) was adjusted before estimating predictive performance [14]. This difference in outcome rate between the two cohorts is due to the difference in setting in which the data was collected, as noted in the description of the cohorts above, fullPIERS was completed in high-income country facilities only. Sensitivity analyses were performed to assess the generalizability of the model in various subsets of study data. In addition, sensitivity analyses were performed excluding the most common components of the adverse maternal outcome to ensure that model discriminatory ability was maintained. Generalizability of the model across study regions was further assessed based on the AUC ROC calculated for the model when applied to each region’s subset of the total miniPIERS cohort. All statistical analyses were performed using STATA v11·0 (StataCorp).