RESULTS
Of the 309 samples from 253 pregnancies with an sFlt-1/PlGF ratio of 38 or below, none delivered with early-onset pre-eclampsia in the subsequent week and none of the pregnancies developed early-onset pre-eclampsia. Conversely, 42.3% (52 of 123) of pregnancies with an sFlt-1/PlGF ratio above 38 delivered with early-onset pre-eclampsia. However, only 22.1% (47 of 213 samples from 123 pregnancies) were diagnosed with early-onset pre-eclampsia leading to delivery within one week. Note that five pregnancies that delivered with early-onset pre-eclampsia did not have any blood test during the last week of pregnancy. Table S1 (supporting information) shows the epidemiological and clinical characteristics of the population and samples, respectively, by study group. NT-proBNP did not vary significantly during the considered gestational weeks in pregnancies that did not develop pre-eclampsia (Pearson correlation= -0.092 p= 0.149) (see Figure S2). In contrast, sFlt-1, PlGF and sFlt-1/PlGF ratio changed with gestational age. Equations describing sFlt-1, PlGF and sFlt-1/PlGF ratio medians per gestational week are described in supporting information (see Table S2).
As the assessed endpoint is a subrogated marker of severity, we compared this endpoint with the definition of severe pre-eclampsia in our sample (13). We found that 97.9% (46/47) of pregnancies with early-onset pre-eclampsia leading to delivery within one week were cases of severe pre-eclampsia. However, the assessed endpoint was observed in only 48.9% (46/94) cases of severe pre-eclampsia. Therefore, we conclude that the assessment of early-onset pre-eclampsia leading to delivery within one week is a more restrictive criterion of severe pre-eclampsia than the ACOG definition of severe pre-eclampsia.
When assessing individual marker prediction performances, we observed a significantly lower estimate of the AUC of the model based on gestational age and PlGF MoM than the AUC obtained with the model that includes gestational age and sFlt-1 MoM (p< 0.001) (Figure 2).
Model development
We developed two types of linear mixed model. One type including the raw marker values and another considering the gestational age-corrected markers. Marker values were logarithmized to overcome the skewness in the data.
Addition of NT-proBNP to sFlt-1/PlGF ratio
We compared the prediction ability of the raw value marker models from the respective ROC curves (Figure 3, left panel). The estimate of the AUC of the model that includes gestational age, sFlt-1/PlGF ratio and NT-proBNP was significantly greater (DeLong test, p= 0.013) than the estimate of the AUC of the model without NT-proBNP.
Use of sFlt-1/PlGF ratio
PlGF MoM was excluded from the MoM transformed marker model during its construction due to low prediction ability. We did not consider including sFlt-1/PlGF ratio MoM in the models as the inclusion of PlGF MoM was non-informative.
The estimate of the AUC of the model that combines gestational age, sFlt-1 MoM and NT-proBNP was significantly greater (p= 0.031) than the estimate of the AUC of the model without NT-proBNP (Figure 3, right panel). Therefore, the selected model for the prognostic prediction tool included gestational age, sFlt-1 MoM and NT-proBNP. Description and predictions of the selected model for each possible value are summarized in Table S3 and Figure S1 (supporting information), respectively.
There were no significant differences between the AUC of the raw value marker model that combines gestational age, sFlt-1/PlGF ratio and NT-proBNP and the gestational age-adjusted model that includes gestational age and sFlt-1 MoM and NT-proBNP (p= 0.648).
Subgroup analysis
Prediction ability of the model that combines gestational age and sFlt-1 MoM and NT-proBNP did not differ from the model without NT-proBNP in pregnancies with intrauterine growth restriction (p= 0.200) or chronic hypertension (p= 0.361).
Model performance
The area under the AUC for early-onset pre-eclampsia diagnosis leading to delivery within one week was 0.882 (95% CI 0.822-0.934) for the model that combines gestational age, sFlt-1 MoM and NT-proBNP and 0.826 (95% CI 0.752-0.892) for the model that combines sFlt-1/PlGF ratio raw values and gestational age model (P = 0.044).
At a 5% false positive rate cut-off level the model that combines gestational age, sFlt-1 MoM and NT-proBNP reached a detection rate of 59.6%, which was significantly greater (p= 0.001) than the model without NT-proBNP (31.9%). At this cut-off level, the model with NT-proBNP resulted positive in 16.9% of the sample and the likelihood ratio of a positive test was 12.4 (95% CI: 6.0-25.3). In other words, the odds of a developing the event is increased twelvefold when the prognostic prediction tool result is positive.
At the sFlt-1/PlGF ratio 655 cut-off the detection ratio was 31.9% (19.1-47.1) with false positive rate of 4.2% (1.7-8.5), predicting early-onset pre-eclampsia diagnosis leading to delivery within one week. With the same false positive rate, the detection rate with the prognostic prediction tool was 53.2% (38.1-67.9) (P=0.03).
Globally, if we compare the application of the criteria based on the sFlt-1/PlGF ratio cut-off value of 38 with the application of the prognostic assessment tool (to pregnancies with an sFlt-1/PlGF ratio above 38), the latter reduced the false positive rate from 34.9% to 1.7% increasing positive predictive value from 22.1% to 77.8%, at the expense of including 33.9% inconclusive valid results (Table 1).