Data source: the INSPIRE trial
Data from the prospective, parallel group, randomized interventional study evaluating the short-term prediction of preeclampsia/eclampsia (INSPIRE) trial was used for the purpose of this research . The INSPIRE trial was conducted in the UK that aimed to evaluate the use of sFlt-1/PIGF ratio in women presenting with suspected preeclampsia and its effect on PE-related hospitalisation within 24 h of the test, within 7 days, or by delivery as the primary outcome. The study was conducted from June 2015 to April 2017 at the John Radcliffe Hospital, Oxford, UK—a tertiary referral centre with a preeclampsia prevalence of 2.9%.
The study enrolled 370 pregnant women 186 reveal trial arm (standard clinical management plus revealing biomarker results) versus 184 non-reveal trial arm (standard clinical management) aged 18 years or above, with singleton pregnancies between 24+0 and 37+0 weeks of gestation with a clinical suspicion of preeclampsia. Women with pre-existing diagnosed preeclampsia/eclampsia were excluded from the trial. Suspicion of preeclampsia was defined by a new onset elevated blood pressure or worsening of pre-existing hypertension or new-onset proteinuria or worsening of pre-existing proteinuria or new-onset headache, visual disturbance, oedema or right upper quadrant pain, or any other clinical suspicion of preeclampsia .
Overall, there were 85 women with PE until delivery and 42 had PE within 7 days of screening. The study found that there was no difference in preeclampsia-related admissions within 24 h of the test between trial arms (sixty patients were admitted in the intervention group (reveal trial arm) and 48 in the comparator group (non-reveal trial arm).
Primary outcome and candidate predictors
The primary outcome in this study was the onset of PE within 7 days of the initial biomarker test as defined in the INSPIRE trial . After informed consent, study participants had standard clinical assessment and additional blood sample for biomarker measurement were collected and centrifuged within 1 h of collection. The sFlt-1 and PIGF values were then measured using the fully automated methods (Elecsys® sFlt-1/PIGF) using the Roche e411 analyzer (Roche Diagnostics Limited, Burgess Hill, United Kingdom) .
A logistic regression model for predicting the onset of PE within 7 days of screening was done using biomarkers sFlt-1 (continuous values), PIGF (continuous values), sFlt-1/PIGF ratio (continuous values), or sFlt-1/PIGF ratio as a binary cut-off of 38.
Sample size assessment for the development of a prognostic model
The adequacy of sample size was assessed using the pmsampsize library in R program as recommended by Riley et al. . The minimum events per parameter required for reliably developing a new model that achieved the desired shrinkage factor, R2, and margin of difference was ~ 10 events per variable. Therefore, with 42 outcome events, ~ 3–4 variables could be used for model development.
Multivariable model building
Four multivariable logistic regression models were constructed for the three biomarkers (sFlt-1, PIGF, sFlt-1/PIGF ratio as continuous, and sFlt-1/PIGF ratio cut off at 38). All four models were adjusted for trial arm . Natural log transformation was carried out for continuous values of sFlt-1, PIGF and sFlt-1/PIGF ratio. The sFlt-1/PIGF cut-off used was 38 since it is commonly used by many studies including the INSPIRE trial . The multivariable logistic regression model assumed the log-odds of PE were linearly associated with the biomarkers. This assumption was formally tested by making comparisons with fractional polynomial regression models. Models were compared using the likelihood ratio test and Bayesian Information Criterion.
Model performance and internal validation
The predictive performances of the four developed models were assessed using calibration and discrimination. Model calibration was assessed by calibration-in-the-large (CITL) and calibration slope whereas model discrimination was assessed by Harrell’s concordance statistic (c-statistic) . The best performing models were tested formally by comparing the respective area under the curve (AUC) using the Delong test [23, 25]. Overall model fit was assessed using Pseudo (Nagelkerke’s) R2 and Bayesian Information Criterion (BIC) . The apparent performance measures (model fit performance in the development data) were adjusted for optimism using bootstrapping by drawing 1,000 resamples from the original dataset and calculating adjusted measures of concordance statistic (calculated from the Somers’ D rank correlation value) , calibration slope and calibration-in-the-large. Adjusted coefficients were corrected for optimism using the uniform shrinkage factor i.e., the calibration slope obtained from bootstrapping. The intercept of the optimism-adjusted model was also re-estimated to maintain the overall calibration. The transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guideline was used for model development and reporting [26, 27].