Predictive Ability of The St-1/Plgf Ratio for the Development of Perinatal Complications

Hypertensive disorders of pregnancy (HDP) and fetal growth restriction (FGR) of vascular placental origin are two main pregnancy complications associated with an increased incidence of poor outcomes. These conditions are both characterized by vascular dysfunction and syncytiotrophoblast oxidative stress. sFlt-1/PlGF ratio as an index of oxidative stress could be a valid diagnostic tool for predicting adverse perinatal outcomes in high risk population. We recruited women affected by HDP, classied according to fetal growth or with isolated FGR. Women with an uneventful pregnancy were recruited as control group. Ultrasound data and sera sample were collected at recruitment. Perinatal complications were the main outcome of this diagnostic study. A survival analysis, a logistic regression model with all covariates and a predictive model based on a Random Forest and a Mean Decrease Gini index were performed. oxidative stress, independently from the cause of the placental imbalance.


Introduction
Placental dysfunction (PD) is a pregnancy disorder characterized by suboptimal placental activity, which can lead to a spectrum of conditions such as hypertensive disorders (HDP) and fetal growth restriction (FGR) [1,2]. HDP and FGR associated with poor placental function share a multifactorial pathogenesis [3], often involving placentation vascular malfunctions or other causes of placental oxidative stress [4][5][6][7].
Indeed, these diseases are characterized by an angiogenic imbalance, with an elevation of soluble blocking factors such as soluble fms-like tyrosine kinase 1 (sFlt-1) levels and a decrease in placental growth factor (PlGF) levels [8,9]. Pregnancies complicated by HDP and/or FGR have higher risks for the fetus to develop perinatal complications, due both to the poor placental function typical of these pathologies and both to the numerous iatrogenic premature deliveries to which these conditions lead. [10].
This longitudinal study was performed on patients with pregnancies complicated by HDP, FGR or both of these conditions, in order to evaluate the possible role of the sFlt-1/PlGF ratio to predict perinatal complications.

Study design
This is a longitudinal, observational, multicenter study performed from January 2019 to January 2021.
The study participants were recruited at the High Risk Maternity Mangiagalli Centre, Department of Woman, Child and Neonate, Fondazione IRCCS Ca' Granda, Ospedale Maggiore Policlinico, Milan, and at the Unit of Obstetrics, Department of Obstetrics and Gynecology, MBBM Foundation at San Gerardo Hospital, Monza, Italy.
The objective of the study is to create a predictive model for the development of perinatal complications, using sFlt-1/PlGF values and without knowing the delivery gestational age.
All pregnant women admitted for HDP, FGR, or both, were eligible for the present study. Women with an uneventful pregnancy were recruited as control group. Exclusion criteria were: multiple pregnancies, genetic or structural anomalies of the fetus, maternal infections, maternal age less than 18 years old. All patients provided written, informed consent prior to enrolment. The protocol of the study complies with the European Union's Good Clinical Practice standards and the Declaration of Helsinki, and it was approved by the Milan Area 2 Ethics Committee (MATER n° 71_2020 Fondazione IRCCS) and PREBIO STUDY (San Gerardo Hospital).
Patients were recruited at the time of referral at the outpatient high risk clinics, or at admission at the Maternal Fetal Medicine Wards. The visit included a general and obstetrical physical examination, with the assessment of maternal weight gain and the measurement of blood pressure. The diagnosis of HDP was made according to ISSHP guidelines [11], including both gestational hypertension (GH) and chronic hypertension (CH). Patients were evaluated longitudinally during pregnancy and they all underwent an obstetric ultrasound in order to evaluate fetal well-being by the measurement of the fetal biometry, Amniotic Fluid Index (AFI), and Doppler velocimetry of uterine arteries (mean UtA PI), umbilical arteries (UA PI), middle cerebral artery (MCA PI) and the fetal ductus venosus (PIV) when appropriate. The diagnosis of FGR was made according to the Delphi consensus [12].
Placental biomarkers sFlt-1 and PlGF were measured at recruitment. Venous blood was sampled and collected in tubes containing a separating gel, then the tubes were labeled and centrifuged for 10 minutes, within three hours of collection. sFlt-1 and PlGF were analysed by Roche's Elecsys automated method, immunoassays based on electro-chemoluminescence technology. The limits of detection varied between 10 and 8500 pg/ml for sFlt-1 and between 3 and 10000 pg/ml for PlGF.

Statistical analysis
We chose to use the Mann Whitney U-test for scalar variables and Fisher's exact test for categorical variables. Data are expressed as mean and standard deviation for continuous variables and as absolute and relative frequencies for categorical, respectively.
The survey is divided into an initial interpretative analysis of the test population and a subsequent predictive phase.
The rst step of our study was a Survival Analysis performed in order to evaluate perinatal complications in relation to gestational age at delivery. The event of interest is the gestational age at the onset of complications. Cumulative incidence curves were built using Kaplan Meier estimates and gestational age at delivery was considered as the time variable. We considered the cumulative incidence of perinatal complications distinguished for the four risk categories established on the basis of the sFlt-1/PlGF ratio values at recruitment (low <38, medium 38-85/38-110, high >85/>110 and very high >655/>201 before and after the 34th week of gestation respectively) [13]. Log rank test was used to compare the curves.
The second step of our analysis was a logistic regression model in order to understand whether sFlt-1/PlGF ratio is among the predictors of perinatal complications and how it in uences the outcome. The dependent variable of the model were perinatal complications. The covariates we used were: sFlt-1/PlGF ratio at recruitment, pre-pregnancy BMI, maternal age, aspirin (ASA) or low-molecular-weight heparin (LMWH) during pregnancy, gestational diabetes, mean UtA PI, pre-pregnancy disease, gestational age at recruitment, parity, previous PE/FGR/IUFD, conceive with assisted reproductive technology (ART), and the macro-ethnicity of women divided into North African, Central and South African, Asian, Indian and South American. Odds Ratio were used to interpret the relationship between predictors and perinatal complications, and Wald test was performed in order to assess the signi cance of predictors.
Finally, the third step consisted in a classi cation analysis for which we used a Random Forest (we speci ed 500 trees and 3 randomly selected variables for each split) to create a model capable of predicting the occurrence of these complications. The dataset was divided into two groups: 75% of the data were used to train the model (training set), and 25% of the data were used to test the model (test set).
Furthermore, since the number of pregnancies with perinatal complications was much lower than uncomplicated pregnancies, we adopted the Random Over Sampling Examples (R.O.S.E.) algorithm to balance only the training set. Finally, the model for predicting perinatal complications built on the training set was applied to the test set and its classi cation performances were reported. Mean decrease Gini index was used to evaluate the rank of each variable. Statistical analysis was performed by a statistician using R version 4.0.3 (2020-10-10).

Results
Three hundred and fty patients, of the 359 recruited, were eligible for the analysis of the study. Five patients were excluded during the study for SARS-COV-2 infection during pregnancy, three were delivered at another hospital and lost at follow-up, one was excluded because of fetal malformations discovered after recruitment.
One hundred and ninety-one patients developed HDP, among whom 90 were associated with FGR. One hundred and eighteen pregnant patients suffered of FGR without clinical evidence of hypertension. Fortyone patients, among those referred to Maternal Fetal Medicine clinics, had an uneventful pregnancy. The population was then divided according to the perinatal complications into 90 complicated cases and 260 uneventful cases. Figure 1 shows recruitment, selection population and the prevalence of maternal complications into clinical groups.
Interpretative analysis Table 1 shows maternal demographic and clinical features, delivery data and neonatal outcome of patients included in this analysis. We observed a signi cant higher prevalence of positive medical history, previous obstetric complications, and gestational diabetes in pregnant with perinatal complications. Of interest, the percentage of smokers during pregnancy was higher in the group without complications, even though all women reported a major reduction in the number of daily cigarettes compared with the prepregnancy period. Finally, the group that developed perinatal complications showed signi cantly lower average neonatal weight and gestational age at delivery than the uncomplicated group.  (17) <0.001 Table 2 shows data regarding the biochemical pro le. The sFlt-1/PlGF ratio values have been divided into four risk categories, as suggested by Stepan et al [13]. The mean value of the sFlt-1/PlGF ratio at recruitment in patients developing perinatal complications was much higher than the group without complications. Among women who developed perinatal complications only 18.9% fell into the low-risk pro le, a much lower percentage than the 60.4% of patients without complications (p < 0.001). Even in the medium risk the percentage of patients without perinatal complications is higher, while it is possible to see an increase of the number of cases distributed in the high risk among patients with perinatal complications (p < 0.001).  (14) 0.054 Figure 2 shows the results of the Survival Analysis as the cumulative incidence of perinatal complications in the four risk groups (reported in Table 2) according to sFlt-1/PlGF values at recruitment. Patients with sFlt-1/PlGF ratio within the high and very high risk values developed more perinatal complications at an earlier gestational age than the low and medium risk patients (p <0.0001). Table 3 reports the results of the logistic regression analysis for all variables that might be associated with perinatal complications. Variables signi cantly associated with complications were: mean UtA PI and sFlt-1/PlGF ratio at recruitment, gestational diabetes, pre-pregnancy disease and gestational age at recruitment. This logistic regression analysis showed how a unit increase of the sFlt-1/PlGF ratio induced an increase (OR>1) in perinatal complications of (1.006-1) * 100 = 0.6%.

Predictive analysis
We developed two different predictive models, one including the gestational age at delivery, an extremely strong predictor for the development of complications, and the other one based on gestational age at admission. Then, we compared the performances obtained by these two models ( Table 4). The rst model achieved a sensitivity of 88%, a speci city of 90% and an area under the curve (AUC) of 89%; the second model reached a similar diagnostic accuracy, but unknown the gestational age of delivery. The Random Forest adopted to construct the predictive models also allows us to obtain an index of the importance of each variable included in the model: the Mean Decreased Gini. Results are shown in Figure  3. In the rst predictive model gestational age at delivery is the variable with the highest prediction weight, but when the gestational age at delivery is replaced by the gestational age at recruitment the sFlt-1/PlGF ratio and the mean UtA PI measured at recruitment become the two most important variables for the prediction of perinatal complications.

Discussion
Our analysis shows that sFlt/PlGF ratio measured at the time of recruitment is a very useful parameter for predicting the possible onset of perinatal complications, especially in pregnancies with HDP or FGR.
I. The survival analysis performed in order to evaluate perinatal complications in relation to gestational age at delivery, obtained by dividing sFlt-1/PlGF values into four risk classes according to reported clinical criteria [13], showed that the majority of patients who subsequently developed complications fell into the high and very high risk sFlt-1/PlGF ratio groups, and only a very small percentage into the low and medium risk groups (p <0.001).
II. The logistic regression analysis that included all the variable associated with perinatal complications proved that sFlt-1/PlGF ratio was among the signi cant variable independently associated with perinatal complications, and that for every 10-point increase in the ratio, the risk of developing perinatal complications increased by approximately 6%.
III. After interpreting the predictors of perinatal complications, our goal was to build a model capable of predicting the onset of these. The Random Forest model adopted to assess the predictive value of sFlt-1/PlGF ratio which excluded gestational age at delivery from the possible covariates, proved that this ratio in pregnancies affected by HDP and/or FGR, measured at the time of clinical diagnosis and recruitment, signi cantly predicted the later development of perinatal complications. The AUC of the sFlt-1/PlGF ratio was as high as 85% in predicting perinatal complications. Of interest, the inclusion into the model of gestational age at delivery, obviously related to the onset of perinatal complications, did not signi cantly improved the outcome of the model. According to the Mean Decreased Gini criteria the two most important variables were sFlt-1/PlGF and mean UtA PI at recruitment. This nding further underlies the independent predicting value and strength of sFlt-1/PlGF ratio.
With regard to perinatal complications, most surveys [14][15][16] identi ed different cut-offs for the sFlt-1/PlGF ratio in order to de ne its predictive ability. The study by Bednarek-Jędrzejek M et al. [14] assessed neonatal outcomes by dividing patients into three risk classes: sFlt-1/PlGF ratio <38, between 38 and 85, and >85. The survey showed that perinatal outcomes were signi cantly worse in the group with sFlt-1/PlGF >85, in which there was a lower birth weight, a shorter duration of gestation and a lower pH of cord blood. Similar ndings were observed by Chang YS et al. [15] that used sFlt-1/PlGF values >85 to distinguish high risk from low risk group, and they showed that surviving infants in the high-risk group had a higher incidence of preterm birth, lower birth weight, higher incidence of respiratory distress syndrome and bronchopulmonary dysplasia. Our survey assessed similar correlations in patients belonging to the high and very-high risk sFlt-1/PlGF ratio groups, which had a higher incidence of premature births and thereby developed more perinatal complications at an earlier gestational age.
Finally, according to Zhu X et al.
[16] the sFlt-1/PlGF ratio can be useful to predict birth weight, indeed there is an inversely proportional correlation between placental biomarkers values and neonatal weight.
Also our analysis showed that infants who developed perinatal complications, and who had higher biomarkers values, also had on average a lower birth weight than infants without complications and with normal biomarkers values.
A recent meta-analysis [17] con rmed these ndings on six studies that assessed the sFlt-1/PlGF ratio predictive ability on both maternal and perinatal complications with a sensitivity of 68% (59-75%), speci city 86% (74-93%) and AUC 79% (75-82%). To our knowledge our survey is the rst longitudinal study, using placental biomarkers values sampled at recruitment, able to create a predictive model on the development of isolated perinatal complications at delivery.

Strengths and limitations of the study
The main strength of our study is its reproducibility. In fact, we only used sFlt-1/PlGF ratio continuous values without manipulating the data by the use of thresholds. Additionally, in our analysis the predictive model was constructed using Random Forest, only after dividing our population of 350 patients into a training set and a test set: this means that performances reported in our study always refer to the prediction of complications on data from a set external to the one used to build the model, thus excluding the possibility of over tting that can occur when the model is applied to the data on which it was built.
The adoption of this robust procedure allowed us to obtain a more realistic performance, equivalent to that which would be obtained by recruiting a new cohort from scratch or estimating the risk of a single patient. Finally, to the best of our knowledge, our study is the rst one using placental biomarkers sampled at recruitment to predict the development of perinatal complications at delivery.
Unfortunately, due to real-life clinical setting of the study, the recruitment of patients did not take place within a precise pre-established gestational age window, but at the time of diagnosis of HDP and/or FGR.
The dataset was also collected on population with a predominance of women of South European ancestry and only a minority of different ethnic groups.

Conclusions
Our study proved that in a cohort of patients with HDP and/or FGR the majority of patients who subsequently developed complications fell into the high and very high risk sFlt-1/PlGF ratio groups.  Cumulative incidence curves of perinatal complications in risk groups divided according to sFlt-1/PlGF ratio levels. The y-axis shows the cumulative incidence of complications, while the x-axis shows the gestational age at the time of occurrence of complication. Each step indicates the onset of a new complication.