Prediction of pre-eclampsia by maternal characteristics : A case-controlled validation study of a Bayesian network model for risk identification of pre-eclampsia

A. Meurs, van, M. Velikova, M. Beatrijs van der Hout-van der Jagt, Barbara Vermeulen-Giovagnoli, S G Oei

Research output: Contribution to journalArticleAcademicpeer-review

2 Downloads (Pure)


Brief Introduction: Pre-eclampsia (PE) is worldwide a leading and rising cause of maternal and perinatal morbidity and mortality. As PE remains a serious and poorly understood complication of pregnancy, it is necessary to recognize the disease before it threatens the survival of mother and fetus. A validated tool that allows real-time maternal risk stratification is needed to guide care. A recent advanced model for pre-eclampsia presented in Velikova et al. 2014 provides that potential. In contrast to the previous study where the testing with this model was done with data for high-risk pregnancies, in this study we aimed at evaluating the capability of the BN model for pre-eclampsia, by looking at the predictions for normal and pre-eclamptic pregnancies. Materials & Methods: The model is based on a Bayesian network methodology, which has been successfully applied for clinical problems. A Bayesian network (BN) is a statistical model that represents a set of variables (e.g. risk factors, diseases and symptoms) and their dependencies by means of a graph and probability distributions. The advantage of a Bayesian network is that it can be used to make personalized predictions, for example for the development or presence of a disease, by entering patient-specific data. We validated the BN model for PE (PE model) in a retrospective case-control study. 10 women diagnosed with PE admitted to the obstetric high care ward of a tertiary care center were enrolled. Their characteristics were matched with 10 pregnant women without any illness. We collected pregnancy data that was relevant for the model, including: (i) risk factors: age; BMI; smoking; parity; twin pregnancy; family history of PE; previous history of PE; preexisting vascular disease; preexisting renal disease; anti-phospholipid syndrome; diabetes mellitus, (ii) medication and measurements from 10 standard check-ups during the pregnancy: blood pressure, protein-to-creatinine ratio, serum creatinine and hemoglobin and medication and (iii) the outcome variable: whether or not preeclampsia is present. PE risk estimation from the model for each patient was compared to PE development. Model performance was assessed by means of the area under the receiver operating characteristic (ROC) curve (AUC). Clinical Cases or Summary Results: Throughout pregnancy, the PE model predicted a high absolute risk for PE in 9 out of 10 PE patients versus 2 out of 10 non-preeclamptic women. This is shown in Figure 1. Each pregnancy week shows the number of patients that has not delivered yet, the color indicating whether they are ultimately diagnosed with PE (red) or not (green). The predicted risk by the model is indicated with a percentage (y-axis) for each patient at each gestational week (x-axis). ROC curves and PE-risk cut-offs were calculated for different gestational weeks. This resulted in AUC score at 24 weeks of 0.895, at 28 weeks of 1.000, at 32 weeks of 0.986, at 36 weeks 1.000, at 38 weeks of 0.375, at 40 weeks of 1.000. Concluding, we find high AUC scores, except for the prediction at 38 weeks of gestation, due to missing data. Therefore no cut-off value was calculated for week 38. Sensitivity, specificity, the positive predictive value (PPV) and negative predictive value (NPV) of the PE model are as well calculated per week of gestation, except for week 38. The sensitivity of the PE model is 100% at each pregnancy week. We find a specificity of 83%, 100%, 91%, 100% and 100% at 24, 28, 32, 36 and 40 weeks, respectively. Thus, 8-17% of women without PE will be screened as having an increased risk. The PPV is calculated as 33%, 100%, 83%, 100% and 100% at 24, 28, 32, 36 and 40 weeks, respectively. The NPV is calculated as 100% at each pregnancy week. Conclusions: When data is available in early pregnancy, the PE model is able to distinguish between PE and non-PE pregnant women and able to predict a higher risk for the diagnosed patients. In particular, at gestational week 12 the chance for PE was twice or higher for PE patients than for 8 of the non-PE pregnant women, and for weeks 16-24 this chance for PE patients was up to eight times higher for the PE patients. This is a particularly important result given the aim of a timely identification of women at risk, which is to facilitate much targeted monitoring. Despite the fact not all data was available for all pregnancy checkups, the PE model was still able to compute the individual, absolute risk for pre-eclampsia. Although for some patients PE was only predicted late in pregnancy, this was for all patients before or at latest at the same moment of clinical diagnosis. However, we expect that prediction will improve when all measurements are available from the pregnancy checkups. From the model it follows that a dynamic cut-off is needed that increases with pregnancy duration. Current results are promising. We propose to perform an RCT with a larger number of patients to establish this cut-off curve with more accuracy and to validate the PE model prospectively. Once validated, the model can assist in early PE diagnosis and thus allow early treatment of PE. The PE model can be integrated in e-health applications to allow real-time monitoring of pregnant women anywhere. By this way we can personalize the healthcare during pregnancy. (Figure presented)
Original languageEnglish
Pages (from-to)351-352
Number of pages2
JournalJournal of Maternal-Fetal and Neonatal Medicine
Issue numbersuppl. 1
Publication statusPublished - 2014


  • Antiphospholipid Syndrome
  • Blood Pressure
  • Case-Control Studies
  • Creatinine
  • Diabetes Mellitus
  • Family
  • Female
  • Fetus
  • Morbidity
  • Netherlands
  • Obstetrics
  • Parity
  • Perinatal Care
  • Pre-Eclampsia
  • Pregnancy
  • Pregnant Women
  • Probability
  • Risk
  • Risk Factors
  • Risk stratification
  • Serum
  • Smoking
  • Syndrome
  • Time
  • accuracy
  • age
  • aspiration needle
  • blood
  • case control study
  • color
  • complication
  • creatinine blood level
  • development
  • diagnosis
  • diseases
  • drug therapy
  • family history
  • first trimester pregnancy
  • health care
  • hemoglobin
  • high risk pregnancy
  • history
  • human
  • implantable cardioverter defibrillator
  • increased
  • kidney disease
  • materials
  • measurement
  • medical care
  • methodology
  • methods
  • model
  • monitoring
  • mortality
  • mother
  • outcome variable
  • patient
  • perinatal morbidity
  • phospholipid
  • prediction
  • predictive value
  • preeclampsia
  • pregnant woman
  • pressure
  • protein
  • protein to creatinine ratio
  • receiver operating characteristic
  • risk factor
  • standard
  • statistical model
  • stratification
  • study
  • survival
  • symptom
  • telehealth
  • tertiary care center
  • twin pregnancy
  • validation study
  • vascular disease
  • ward


Dive into the research topics of 'Prediction of pre-eclampsia by maternal characteristics : A case-controlled validation study of a Bayesian network model for risk identification of pre-eclampsia'. Together they form a unique fingerprint.

Cite this