Clinical risk prediction for pre-eclampsia in nulliparous women. Power calculation
Power calculation
The number of women required to be screened was based on achieving suitable screening test characteristics and precise estimates of their values. Given a pretest probability (prevalence) of pre-eclampsia of 5%, then a post-test probability of 30% or greater would make this a useful test, based on current clinical practice. The algorithm must therefore have sufficient ability such that it is unlikely the post-test probability will fall below 0.30 (30%) for pre-eclampsia. This can be attained, with a power of 80%, in a cohort of 3000 if the true positive likelihood ratio of the screening test is 9.2 to 10.0. Given a prevalence of 5%, if we observe a sensitivity of 90% this cohort size will give a 95% confidence interval for this sensitivity of 84.0 to 94.3, and a specificity of 91%.
Datasets
We used two datasets to construct the predictive models for pre-eclampsia. The first comprised clinical variables obtained at 16-16 weeks’ gestation and the second comprised clinical data at 14-16 weeks combined with variables from the 19-21 week ultrasound scan. Of the 933 original and derived variables recorded, we excluded variables added after recruitment commenced (n=76), paternal variables (n=48), variables not applicable to prediction of pre-eclampsia (n=246), variables with more than 10% missing data (clinical laboratory data and work variables not applicable to women not working, n=27), and 402 variables with P>0.10 on univariable comparison of women with and without pre-eclampsia. Of the remaining 134 variables, we selected 38 as candidate predictors on the following criteria: known potential risk factors for pre-eclampsia, ease of collection in the clinical setting, and potential applicability to future populations (see table A in appendix 1 on bmj.com for a full list of variables). With this approach the only established “risk” factor not included in the candidate predictors was cigarette smoking, as in our dataset this was not associated with pre-eclampsia. We added the number of cigarettes smoked a day at 15 weeks’ gestation as a candidate predictor, giving a total of 39 variables for the multivariable analysis. Variables were not included as candidate predictors because of colinearity (n=61), a low cell count (<5) in the χ2 test (n=11), lack of a consistent relation with pre-eclampsia in literature (n=4), or not readily applicable to a future obstetric population (n=20). Among the 39 candidate predictors, data were complete in 32, missing in <1% for six variables, and missing in 6% for participant’s birth weight. We imputed missing continuous data (n=4) with expectation maximization and used the median for three variables unrelated to other observed data. The expectation maximisation algorithm was implemented in the “mix” package in R, version 2.9.1. To evaluate its imputation, we used a permutation technique on the complete dataset. For each variable, we systematically removed each data point and imputed the “missing” value using expectation maximisation. We calculated the ratio of mean absolute error between imputed and original values to the mean value for that variable. The mean ratio for the variables imputed with expectation maximisation was 10.8%.
Clinical risk prediction for pre-eclampsia in nulliparous women. Part 5
Information was collected on vaginal bleeding early in pregnancy (gestation, severity and duration of bleeding, and recurrent bleeds), hyperemesis, and infections during pregnancy. Vegetarian status was recorded, and other dietary information before conception and during pregnancy was obtained from food frequency questions for fruit, green leafy vegetables, oily and other fish, and fast foods. Use of folate and multivitamins, cigarettes, alcohol (including binge drinking), and recreational drugs (including marijuana, amphetamine, cocaine, heroin, ecstasy, LSD (lysergide)) was recorded for before conception, first trimester, and at 15 weeks. A lifestyle questionnaire was completed on work, exercise and sedentary activities, snoring, domestic violence, and social supports. Psychological scales were completed to measure perceived stress, depression, anxiety, and behavioural responses to pregnancy (adapted from the behavioural responses to illness questionnaire23). Two consecutive manual blood pressure measurements (mercury or aneroid sphygmomanometer, with a large cuff if the arm circumference ≥33 cm and Korotkoff V for diastolic blood pressure) were recorded. Other maternal measurements included maternal height and weight and waist, hip, arm, and head circumference. Proteinuria in a midstream urine specimen was measured by dipstick or a protein:creatinine ratio. Random whole blood glucose and serum lipid concentrations (triglycerides, total cholesterol, high density lipoprotein cholesterol, low density lipoprotein cholesterol, total cholesterol:high density lipoprotein cholesterol ratio) were also measured. Ultrasound examination at 19-21 weeks’ gestation included measurements of the fetus (biparietal diameter, head circumference, abdominal circumference, and femur length) and Doppler studies of the umbilical and uterine arteries.24 All fetal measurements were adjusted for gestational age by calculating the multiple of the median for each gestational week. Mean uterine resistance index (RI) was calculated from the left and right uterine resistance index. If only a left or right uterine resistance index was available, this was used as “mean resistance index” (n=20). Notching of each uterine artery was recorded. An abnormal uterine artery Doppler result was defined as a mean resistance index >90th centile (>0.695). Participants were followed prospectively, and research midwives collected data on pregnancy outcome and measurements of the baby. Data monitoring included individual checks of all data for each participant, including checks for any transcription errors of the lifestyle questionnaire, and detection of illogical or inconsistent data and outliers with customised software.
Primary outcome
Our primary outcome was pre-eclampsia defined as systolic blood pressure ≥140 mm Hg or diastolic blood pressure ≥90 mm Hg, or both, on at least two occasions four hours apart after 20 weeks’ gestation but before the onset of labour, or postpartum, with either proteinuria (24 hour urinary protein ≥300 mg or spot urine protein:creatinine ratio ≥ 30 mg/mmol creatinine or urine dipstick protein ≥++) or any multisystem complication of pre-eclampsia.19 25 Multisystem complications included any of acute renal insufficiency defined as a new increase in serum creatinine concentration ≥100 μmol/L antepartum or >130 μmol/L postpartum; effects on liver, defined as raised aspartate transaminase or alanine transaminase concentration, or both, >45 IU/L and/or severe right upper quadrant or epigastric pain or liver rupture; neurological effects included eclampsia, imminent eclampsia (severe headache with hyper-reflexia and persistent visual disturbance), or cerebral haemorrhage; and haematological effects included thrombocytopenia (platelets <100×109/L), disseminated intravascular coagulation, or haemolysis. The reference group was women who did not develop preeclampsia.
Other definitions
The estimated date of delivery was calculated as follows: if the woman was certain of the date of her last menstrual period (LMP), the estimated date of delivery was adjusted only if a scan at <16 weeks’ gestation found a difference of seven or more days between the scan gestation and that calculated by the LMPor a scan at 20±1 weeks found a difference of 10 or more days. If her date was uncertain, scan dates were used to calculate the estimated date of delivery. Preterm pre-eclampsia was pre-eclampsia resulting in delivery before 37 weeks’ gestation. Small for gestational age was defined as a birth weight below the 10th customised centile, adjusted formaternal height, bookingweight, ethnicity, and delivery gestation and infant’s sex.
Clinical risk prediction for pre-eclampsia in nulliparous women. Part 4
Before preventive treatment and stratified antenatal care can be offered to nulliparous women, we need to identify those at high risk of pre-eclampsia. At present there is no method to accurately stratify healthy nulliparous women according to their risk profile for preeclampsia. This study is part of the SCOPE (Screening for Pregnancy Endpoints) study, a prospective, multicenter cohort study of “healthy” nulliparous women with the primary aim of developing screening tests to predict pre-eclampsia, infants who are small for gestational age, and spontaneous preterm birth. The study design incorporates prospective collection of information on all known clinical risk factors for pre-eclampsia. The objectives are to develop multivariable predictive models for pre-eclampsia (based on clinical risk factors present in early pregnancy alone or in combination with ultrasound estimates of uteroplacental perfusion and fetal measurements at 19-21 weeks’ gestation) and determine their performance to predict pre-eclampsia as a baseline for future external validation; identify the rate of pre-eclampsia associated with specific combinations of clinical risk factors and ultrasound scan variables; and develop a proposal for risk stratification of “healthy” nulliparous women, based on combinations of key clinical risk factors and scan indices, to identify a subgroup at increased risk of pre-eclampsia for whom specialist referral might be indicated.
METHODS
Five centres (Auckland, New Zealand; Adelaide, Australia; London and Manchester, UK; and Cork, Ireland) recruited nulliparous women with singleton pregnancies to the SCOPE study between November 2004 and August 2008. Women (n=4961) attending hospital antenatal clinics, obstetricians, general practitioners, or community midwives before 15 weeks’ gestation were invited to participate. Exclusion criteria included recognized as high risk of pre-eclampsia, small for gestational age baby or spontaneous preterm birth because of underlying medical conditions (chronic hypertension requiring antihypertensive drugs, diabetes, renal disease, systemic lupus erythematosus, antiphospholipid syndrome, sickle cell disease, HIV), previous cervical knife cone biopsy, three or more abortions or three or more miscarriages, current ruptured membranes; known major fetal anomaly or abnormal karyotype; or intervention that could modify the outcome of pregnancy (such as aspirin, cervical suture). A research midwife interviewed and examined women at 14-16 and 19-21 weeks’ gestation. Women underwent an ultrasound scan at 19-21 weeks. At the time of interview, data were entered on an internet accessed central database with a complete audit trail (MedSciNet). At 14-16 weeks’ gestation the following data were collected: demographic information including age, ethnicity, immigration details, education, work, socioeconomic index, income level, living situation; the woman’s birth weight and gestation at delivery and whether it was a singleton or multiple pregnancy; previous miscarriages, abortions, or ectopic pregnancies and whether these pregnancies were with the same partner as the current pregnancy or not; history of infertility, use of assisted reproductive technologies, duration of sexual relationship, and exposure to partner’s sperm; gynaecological (including polycystic ovarian syndrome) and medical history, including hypertension while taking combined oral contraception, asthma, urinary tract infection, inflammatory bowel disease, thyroid disease, and thromboembolism; and family history (in mother and sisters) of obstetric complications (miscarriage, pre-eclampsia, eclampsia, gestational hypertension, spontaneous preterm birth, any preterm birth, gestational diabetes, stillbirth, and neonatal death) and family history (mother, father, sibling) of medical conditions (hypertension, coronary artery heart disease, cerebrovascular accident, type 1 and 2 diabetes, and venous thromboembolism).
Clinical risk prediction for pre-eclampsia in nulliparous women. Part 3
If it was implemented into clinical practice, the clinician would derive the mean arterial pressure from systolic and diastolic blood pressure entered into the algorithm. A history of coronary artery disease in the woman’s father was associated with a 1.9-fold increase in the risk of preeclampsia, consistent with a previous report and the association between pre-eclampsia and subsequent ischaemic heart disease. Confirming the results of a case-control study, a lower maternal birth weight was associated with an increased risk of pre-eclampsia, with an even greater risk when low maternal birth weight coexisted with other key risk factors. Prolonged vaginal bleeding in early pregnancy was associated with a twofold increase in risk of pre-eclampsia. As reported by others, most of these bleeds were mild in severity, suggesting that a discrete bleeding pattern could be associated with later pre-eclampsia. Several factors were associated with a reduced risk of pre-eclampsia.
A single early miscarriage with the same partner, eating a lot of fruit, and smoking were protective, again reassuringly consistent with previous reports. The protective influence of cigarette smoking in our cohort was less than previously reported, and cigarettes did not remain in the model when we added uterine artery Doppler indices. Alcohol use in the first trimester was protective but requires confirmation in other cohorts. Obese women are reported to drink less alcohol, possibly because food fulfils their addictive behaviour. Obesity is unlikely to be the only explanation, however, as the protective effect of alcohol is retained with BMI in the model, and there was no interaction between BMI and alcohol. A recent series of publications reported algorithms to predict pre-eclampsia based on clinical risk factors in a general population comprising high risk women (previous pre-eclampsia and medical conditions), nulliparous women, and low risk women (multiparas with previous uncomplicated pregnancies).
A model is fitted to the population in which it was developed, using the available candidate predictors. A general antenatal population constructed of subgroups with different risk profiles is difficult to replicate and future “general populations” are likely to comprise a different case-mix. The importance of population differences is evident in the failure of one proposed algorithm to validate in a high risk population, raising questions as to more general applicability to other populations such as “healthy” nulliparous women. Poor performance on validation might also occur because key predictors are missing from the model. When the list of candidate predictors includes strongly predictive factors, such as previous pre-eclampsia, renal disease, and chronic hypertension, these will take precedence, replacing other factors that might be more relevant to healthy nulliparous women. In contrast, in SCOPE, we investigated candidate predictors applicable to healthy nulliparous women.
Clinical risk prediction for pre-eclampsia in nulliparous women. Strengths and weaknesses
A major strength of this study is its large multicenter prospective design with excellent follow-up. As the focus of the SCOPE study is development of tests to predict pregnancy outcome with potential to translate into clinical care, we recruited a clearly defined population of nulliparous women, enabling identification of similar populations for external validation. This is critical for generalisability of a risk assessment algorithm; the population in which the algorithm is developed needs to be identifiable if a screening test is to be used in clinical care. We obtained high quality data for all known risk factors for pre-eclampsia from questionnaires administered at interviews, along with detailed standard operating procedures. Use of a real time database, with automated checking procedures, reduced data entry errors and transcription errors. For a dataset of this size, the rate of missing data was minimal. The intensive two stage data monitoring adds confidence in data integrity. Potential measurement errors, such as in self reported family history, could have occurred, but as the goal was to develop a prototype algorithm ultimately for clinical use, this limitation was accepted. Principal investigators reviewed outcome data for cases, ensuring accurate diagnosis. One of the challenges when predicting rare events in prospective cohorts, such as SCOPE, is the relatively low numbers of cases compared with studies based on huge epidemiological databases. While the latter might have a substantially greater number of events, their interpretation is restrained by less accurate diagnosis. There is no consensus as to the best method for selection of variables. Given the rich dataset of potential predictors for pre-eclampsia, we used a pruning step based on significance testing and then selected a subset of candidate variables on a priori knowledge. This could have introduced variable selection bias, but it is reassuring that the clinical risk factors and their strength of association with pre-eclampsia are consistent with the literature. While we could undertake only internal validation at this stage, external validation is planned.
Comparisons with other studies
Previous studies investigating risk factors for preeclampsia have used birth registries or hospital databases, randomised trials with negative results (that is, no treatment effect shown), and, in a few studies, prospective cohorts (usually general obstetric populations) designed to investigate outcomes of pregnancy. Consistent with other contemporary studies, the women who developed pre-eclampsia were younger, more obese, and more likely to have lower socioeconomic status. Many of the risk factors included in the algorithm presented here are associated with a similar degree of risk to that previously reported, giving confidence regarding the potential applicability of the algorithm to other populations. Higher blood pressure within the normal range, a higher BMI, and a family history of pre-eclampsia had similar predictive characteristics to those observed in other studies. In our algorithm, blood pressure was the most important risk factor driving the estimated probability of pre-eclampsia. Mean arterial pressure, rather than systolic or diastolic blood pressure blood, was selected by stepwise logistic regression and included in the model.
Clinical risk prediction for pre-eclampsia in nulliparous women. Part 2
Overall, application of this framework for specialist referral would result in 34% of women with preeclampsia (63 of 186) and 53% of those who develop preterm pre-eclampsia (25 of 47) being referred. If the referral criteria did not include uterine Doppler, fewer cases of preterm pre-eclampsia (16 of 47, 34%) would be detected. These results also show that negative prediction based on clinical risk assessment, with or without Doppler ultrasonography, is too inaccurate to allow a reduction in antenatal care.
DISCUSSION
In this large prospective international cohort of “healthy” nulliparous women an algorithm for preeclampsia, which included clinical risk factors at 15 weeks’ gestation, had moderate predictive performance. The algorithm included well recognised risk factors (blood pressure, BMI, and a family history of pre-eclampsia) along with less established factors, such as prolonged vaginal bleeding, maternal low birth weight, and the woman’s father having coronary artery disease. The algorithm had moderate predictive performance— the area under the receiver operating characteristics curve (AUC) was 0.76—and detected 37% and 61% of women who developed pre-eclampsia with a false positive rate of 10% and 25%, respectively. Addition of information from ultrasonography did not significantly improve performance of the algorithm, with an AUC of 0.77. We would expect poorer screening performance of the algorithm in other nulliparous populations, as evident by theAUCof 0.71 on internal validation. Given the prospective design, cohort size, comprehensive range of candidate predictors, high quality data, and completeness of follow-up, this is likely to be indicative of the best performance achievable using clinical and ultrasonography data to predict pre-eclampsia in a “healthy” nulliparous population. To considerably improve prediction performance will require either the development of specific clinical risk algorithms for disease subtypes, such as preterm and term pre-eclampsia, or the addition of biomarkers, or both.
The concept of a personalised clinical risk estimate for disease, to which biomarkers can be added, is established in several areas of medicine. The algorithm to predict pre-eclampsia reported here provides a first step towards a personalised risk score for pre-eclampsia among nulliparous women. It is inevitable the model will be overfitted to our population and external validation of the algorithm in other nulliparous populations is essential.Weplan to evaluate its performance in the next 3000 women recruited into SCOPE, nearly all of whom will be recruited in different centres than the initial 3500 women. Validation should also be performed in other study populations of nulliparous women.
Clinical risk prediction for pre-eclampsia in nulliparous women. Impact of definition of pre-eclampsia
The predicted probability of pre-eclampsia can then be calculated from 1/(1+e−riskscore). For example, for a 28 year old nulliparous woman whose birth weight was 2400 g, with a mean arterial pressure of 96 mm Hg, BMI 30, a family history of pre-eclampsia, and no protective factors, her probability of pre-eclampsia is 39%. Her risk of pre-eclampsia decreases as each risk factor is removed in stepwise fashion; if her mean arterial pressure is 80 mm Hg her probability of pre-eclampsia would be 18%, if her BMI was 24 her probability would be 14%, if she had no family history of preeclampsia her probability would be 8%, and if her birth weight had been 3500 g her probability would be 5%. If she had protective factors, such as a previous early miscarriage with her partner, her risk would be reduced to 2%.
Impact of definition of pre-eclampsia To evaluate the impact of 24 women receiving a diagnosis of pre-eclampsia based on the presence of gestational hypertension combined with multisystem complications, the model was reconstructed defining the cases as pre-eclamptic women with proteinuria (n=162). Most risk factors and protective factors remained with similar odds ratios, except that age, high intake of fruit, and cigarettes were excluded and a sexual relationship of six months or less (odds ratio 1.7, 95% confidence interval 1.05 to 2.7), hyperemesis at 15 weeks (2.0, 1.1 to 3.7), and maternal height (0.87, 0.76 to 1.0) per 5 cm increase) were included.
Reproducibility of prediction model To investigate the stability and potential reproducibility of the model (using all candidate predictors) we constructed 10 “best models” that included 10 variables. The risk factors (mean arterial blood pressure, BMI, family history of pre-eclampsia, family history of coronary heart disease (woman’s father), participant’s birth weight) and protective factors (≥12 months to conceive, alcohol used in the first trimester) occurred in all “10 best models.” Of the other variables in our model, six occurred in three to seven of the best models, while cigarettes a day was not selected by the stepwise model fitting procedure.
We have shown systolic blood pressure rather thanmean arterial blood pressure as that requires calculation and, unless incorporated into an algorithm, is not easily applied in a routine clinic setting.
Specialist referral framework for nulliparous women
To better understand potential clinical implications, we developed a framework for specialist referral in a population (n=1000) based on observations in our study population (fig 4). In the first stage, women were referred to a specialist if their post-test probability of pre-eclampsia generated by the model was at least 15%. Among those not referred, women with a systolic blood pressure >120 mm Hg, a BMI ≥30, or a family history of pre-eclampsia underwent a uterine Doppler ultrasound with their fetal anatomy scan at 20 weeks, and those with an abnormal uterine artery resistance index also had a specialist referral. Of the women referred for a specialist opinion, 21% developed preeclampsia and 8% developed preterm pre-eclampsia. The relative risk for developing pre-eclampsia and preterm pre-eclampsia in the specialist referred group compared with standard care group was 5.5 and 12.2, respectively.
Clinical risk prediction for pre-eclampsia in nulliparous women. Statistical methods
Statistical methods
WeusedSAS (version 9.1) for univariable data analysis and to generate a multivariable logistic regression model. We used Student’s t test, Wilcoxon rank sum test, or χ2 test for comparing characteristics in the study population and pregnancy outcomes between women who did and did not develop pre-eclampsia. Stepwise logistic regression was used to determine independent risk factors for pre-eclampsia in both datasets. The order of variable selection was determined by the χ2
statistic for each potential variable and the forward selection step could be followed by removal of variables in one or more backward elimination steps. We calculated receiver operating characteristics curves and determined screening test characteristics at a 25%, 10%, and 5% false positive rate. For internal validation we evaluated the calibration and discrimination (10-fold cross validation) of the model using methods described by Altman et al.30 Calibration was assessed by plotting the observed proportion of events against the predicted probabilities. For the cross validation, participants were stratified by region (New Zealand, Australia, Ireland, and UK), pre-eclampsia status (positive or negative), and gestation (<260 days or ≥260 days) and randomly allocated to one of 10 groups. Tenfold cross validation was then performed, with 90% of the data used to generate a model, and estimation of disease risk was performed in the 10% remaining. These predicted values were then combined across the 10 runs and summarised by the C statistic (AUC). This entire procedure was repeated 10 times. To determine the variables most consistently retained in the prediction models, we generated the 10 “best models,” based on the proportion of variance explained, by calculating all possible logistic regression models retaining 10 variables. We determined the frequency of each variable present in the 10 highest scoring models and identified key risk factors.Wethen calculated the proportion of women with specific combinations of key clinical risk factors and abnormal uterine artery Doppler at 20 weeks’ gestation who developed pre-eclampsia.
Mexican pharmacy online
RESULTS
Of the nulliparous women (n=4961) invited to participate in the study, 3780 (76%) agreed but a further 208 were excluded before or at the 15 week interview. Among the 3572 women recruited into the study, we had data on pregnancy outcome for 3529 (99%). When we compared the women who declined to participate (n=1202) with the 3529 women in the study population there were minor but significant differences in the ethnicity mix (79% v 87% white, 4% v 3% Maori or Pacific Islander, 11% v 7% Asian, 6% v 3% other, P<0.001) and age (mean 28.8 (SD 5.7) v 28.1 (SD 5.8), P=0.001). A further 182 women were excluded relationship of six months or less and uterine artery Doppler waveform indices. Based on clinical risk factors, the mean AUC from the ten 10-fold cross validations was 0.71 (SE 0.002). The AUC for the proposed model based on the observations used to create the model was 0.76, indicating a bias in the C statistic of about 5%. The addition of 20 week uterine artery Doppler indices did not improve performance based on the study population (internally validated AUC 0.71 (SE 0.003)). Figure 3 shows that the model has a reasonable level of calibration, but there is an indication that, at the higher probabilities for preeclampsia, it might underestimate cases.
Clinical risk prediction for pre-eclampsia in nulliparous women. Part 4
Among the 39 candidate predictors, data were complete in 32, missing in <1% for six variables, and missing in 6% for participant’s birth weight. We imputed missing continuous data (n=4) with expectation maximization and used the median for three variables unrelated to other observed data. The expectation maximisation algorithm was implemented in the validation we evaluated the calibration and discrimination (10-fold cross validation) of the model using methods described by Altman et al.30 Calibration was assessed by plotting the observed proportion of events against the predicted probabilities. For the cross validation, participants were stratified by region (New Zealand, Australia, Ireland, and UK), pre-eclampsia status (positive or negative), and gestation (<260 days or ≥260 days) and randomly allocated to one of 10 groups. Tenfold cross validation was then performed, with 90% of the data used to generate a model, and estimation of disease risk was performed in the 10% remaining. These predicted values were then combined across the 10 runs and summarised by the C statistic (AUC). This entire procedure was repeated 10 times. To determine the variables most consistently retained in the prediction models, we generated the 10 “best models,” based on the proportion of variance explained, by calculating all possible logistic regression models retaining 10 variables. We determined the frequency of each variable present in the 10 highest scoring models and identified key risk factors.Wethen calculated the proportion of women with specific combinations of key clinical risk factors and abnormal uterine artery Doppler at 20 weeks’ gestation who developed pre-eclampsia. RESULTS Of the nulliparous women (n=4961) invited to participate in the study, 3780 (76%) agreed but a further 208 were excluded before or at the 15 week interview.
Among the 3572 women recruited into the study, we had data on pregnancy outcome for 3529 (99%). When we compared the women who declined to participate (n=1202) with the 3529 women in the study population there were minor but significant differences in the ethnicity mix (79% v 87% white, 4% v 3% Maori or Pacific Islander, 11% v 7% Asian, 6% v 3% other, P<0.001) and age (mean 28.8 (SD 5.7) v 28.1 (SD 5.8), P=0.001). A further 182 women were excluded “mix” package in R, version 2.9.1.To evaluate its imputation, we used a permutation technique on the complete dataset. For each variable, we systematically removed each data point and imputed the “missing” value using expectation maximisation. We calculated the ratio of mean absolute error between imputed and original values to the mean value for that variable. The mean ratio for the variables imputed with expectation maximisation was 10.8%.
Clinical risk prediction for pre-eclampsia in nulliparous women. Other definitions
The estimated date of delivery was calculated as follows: if the woman was certain of the date of her last menstrual period (LMP), the estimated date of delivery was adjusted only if a scan at <16 weeks’ gestation found a difference of seven or more days between the scan gestation and that calculated by the LMPor a scan at 20±1 weeks found a difference of 10 or more days. If her date was uncertain, scan dates were used to calculate the estimated date of delivery. Preterm pre-eclampsia was pre-eclampsia resulting in delivery before 37+0 weeks’ gestation. Small for gestational age was defined as a birth weight below the 10th customised centile, adjusted formaternal height, bookingweight, ethnicity, and delivery gestation and infant’s sex.
Power calculation The number of women required to be screened was based on achieving suitable screening test characteristics and precise estimates of their values. Given a pretest probability (prevalence) of pre-eclampsia of 5%, then a post-test probability of 30% or greater would make this a useful test, based on current clinical practice. The algorithm must therefore have sufficient ability such that it is unlikely the post-test probability will fall below 0.30 (30%) for pre-eclampsia. This can be attained, with a power of 80%, in a cohort of 3000 if the true positive likelihood ratio of the screening test is 9.2 to 10.0. Given a prevalence of 5%, if we observe a sensitivity of 90% this cohort size will give a 95% confidence interval for this sensitivity of 84.0 to 94.3, and a specificity of 91%.
Datasets
We used two datasets to construct the predictive models for pre-eclampsia. The first comprised clinical variables obtained at 16-16 weeks’ gestation and the second comprised clinical data at 14-16 weeks combined with variables from the 19-21 week ultrasound scan. Of the 933 original and derived variables recorded, we excluded variables added after recruitment commenced (n=76), paternal variables (n=48), variables not applicable to prediction of pre-eclampsia (n=246), variables with more than 10% missing data (clinical laboratory data and work variables not applicable to women not working, n=27), and 402 variables with P>0.10 on univariable comparison of women with and without pre-eclampsia. Of the remaining 134 variables, we selected 38 as candidate predictors on the following criteria: known potential risk factors for pre-eclampsia, ease of collection in the clinical setting, and potential applicability to future populations (see table A in appendix 1 on bmj.com for a full list of variables). With this approach the only established “risk” factor not included in the candidate predictors was cigarette smoking, as in our dataset this was not associated with pre-eclampsia. We added the number of cigarettes smoked a day at 15 weeks’ gestation as a candidate predictor, giving a total of 39 variables for the multivariable analysis. Variables were not included as candidate predictors because of colinearity (n=61), a low cell count (<5) in the χ2 test (n=11), lack of a consistent relation with pre-eclampsia in literature (n=4), or not readily applicable to a future obstetric population (n=20).