Maternal characteristics include age, height, weight, smoking status, race, method of conception, low-dose aspirin intake, heparin prophylaxis, and risk factors for PE and other PD-related disorders according to the National Institute for Health and Care Excellence (NICE) guidelines (13) were collected from the medical records. During the study period, PE was screened according to the NICE risk-factor guidelines criteria. Those women with ≥ 1 high risk factor or ≥ 2 moderate ones were considered at high risk of PE, and a recommendation of prophylaxis with aspirin was made. However, not all women were evaluated at first in our center so the screening protocol may have differed. Furthermore, there were some women with low molecular heparin treatment, either in the context of thrombophilia, systemic erythematosus lupus, or assisted reproduction techniques. Uterine artery evaluation at 20 weeks was recorded and centiles were calculated (14). Gestational age (GA) was estimated according to the American College of Obstetricians and Gynecologists, that is, reliable last menstrual period was corrected by the crown-rump length before 14 + 0 weeks or biparietal diameter from 14 + 0 weeks to 21 + 6 weeks, when a significant discrepancy of more than 7 days or more than 10 days was found, respectively (15).

Preeclampsia was defined as the presence of both hypertension and proteinuria, according to the National High Blood Pressure Education Program Working Group on High Blood Pressure in Pregnancy (16). In the clinical setting, patients were managed as having PE even in the absence of proteinuria when other severity criteria were met and indications for delivery followed current recommendations (17).

At diagnosis, a complete blood count, biochemistry including angiogenesis biomarkers, and spot protein/creatinine ratio were measured. The sFlt-1 and PlGF concentrations (picograms per milliliter) were determined using an automated assay system (Cobas^® 6000 e701 module, Roche Diagnostics, Penzberg, Germany). The sFlt-1/PlGF ratio was expressed in absolute values, and the obstetricians involved were aware of the results of the sFlt-1/PlGF ratio. Values ≤ 38 were considered to rule out PE ≥ 85 as “aid in diagnosis,” and > 655 as a high risk of the need to deliver within 48 h (18). This information was not intently used to indicate delivery, although its knowledge could have influenced the clinicians in the interpretation of the severity of PE-related symptoms that leads to decisions about the continuation of the pregnancy.

Corticoids for fetal maturity (betamethasone 12 mg/day for 2 days) were administered before 34 + 6 weeks if fetal viability was reached [provided that the GA was ≥ 24 + 0 weeks in normally grown fetuses or 26 + 0 weeks and estimated fetal weight (EFW) ≥ 500 g in growth restricted fetuses]. A repeated cycle of corticoids was considered after 1 week of the administration of the first cycle if the GA remained below 34 + 6 weeks. Magnesium sulfate was indicated in the presence of severity features or regardless of maternal status when there was a risk of imminent delivery < 32 weeks.

Expectant management was initially attempted for fetal interest whenever the GA was ≥ 24 weeks and was discussed with the parents at earlier weeks. Severe complications that indicated immediate delivery irrespectively of GA (19) were pulmonary edema, refractory hypertension (uncontrolled blood pressure despite two antihypertensive medications at maximum doses), HELLP syndrome, abruptio placentae [defined as placental detachment prior to delivery of the fetus, including the identification of a retroplacental hematoma or evidence of blood clot in at least 20% of its surface based on clinical data provided by the attending obstetrician at delivery (20)], renal failure (oliguria < 500 mL/24 h, creatinine > 1.2 mg/dl), and neurological deficit (persisting visual alterations, stupor, clonus, or eclampsia). Delivery was also recommended after 34 + 0 weeks in the presence of severe features of PE and in any PE case after 37 + 0 weeks.

Fetal assessment including a detailed anatomical scan and growth evaluation was undertaken at our placental dysfunction consult within 48 h of eoPE diagnosis. Fetal weight was estimated (21) and centiles customized (22) to maternal and fetal characteristics. Fetal growth restriction (FGR) was diagnosed according to a stage-based classification (23): stage I was considered in cases with normal fetal Doppler or abnormal but with antegrade umbilical artery (UA) flow; stage II was those with absent-end diastolic UA flow; stage III those with reversed end-diastolic UA flow or ductus venosus PI > 95th centile, and finally stage IV was limited to cases with a reversed a-wave on the ductus venosus or spontaneous decelerations on the CTG. Whenever anterograde flow in UA was present, biweekly monitoring (fetal Doppler including interrogation of the ductus venosus plus conventional cardiotocography) was planned, and vaginal delivery (in the absence of other contraindications) was recommended after 37 weeks. If the absent end-diastolic UA flow was detected, subsequent follow-up controls were performed every 48–72 h, and elective cesarean section was indicated at 34 weeks. When reverse end-diastolic UA flow or ductus venosus PI > 95th centile was found, hospitalization and daily monitoring were carried out until elective cesarean section at 30 weeks. Whenever a reverse a-wave flow in the ductus venosus or spontaneous decelerations in the cardiotocography were noted, elective cesarean section was indicated.

Perinatal data included date and reason for delivery and perinatal mortality.

All data were recorded in a database created on the Research Electronic Data Capture (REDCap) tool (24) hosted by the “imas12” research institute.

Continuous variables were expressed in mean (SD) or median (interquartile range) when non-normally distributed. Categorical variables were expressed in percentage (%). Univariate comparisons between the cases in which delivery occurred within 7 days after eoPE diagnosis and those that did not were performed using the t-test or Mann–Whitney U-test for continuous variables and the chi-square or Fisher’s exact test for categorical variables. Two-sided p < 0.05 was considered statistically significant. Statistical package STATA, version 14.1 (TX, United States: StataCorp LP) was used for this analysis.

First, we developed a predictive model of the need for delivery within 7 days of diagnosis (model D), considering this as the window of the effect of antenatal corticosteroids for fetal maturation (26). Second, we created a model to calculate the risk of developing HELLP syndrome or abruptio placentae at any point after eoPE diagnosis (model HA), as these are the most acute and harder to predict complications.

In order to cover different resource availability scenarios, we developed a reduced version of these models using only demographic characteristics of the patient (D1, HA1) and an extended version adding information available at diagnosis of eoPE (D2, HA2). Accordingly, variables were structured into two datasets: those prior to the onset of PE (baseline) and those evaluated at the moment of diagnosis (diagnosis).

Three preprocessing steps were performed. There were very few missing data in the baseline and “at diagnosis” variables and given the pattern they were considered missing at random. However, there were some variables and some patients with a high number of missing values, especially in the follow-up variables, since this depends on the time a patient participates in the clinical study. First, missing values were imputed using the MissForest imputation technique. This is an imputation method suitable for both categorical and numerical data. It performs an iterative imputation by training a Random Forest model followed by a prediction of the missing values in an iterative process (27). This technique was applied for variables with less than 20% missing data and patients with less than 27% missing data in order to avoid synthetic deviation of the statistical distribution. The remaining variables and patients were directly removed from the dataset. To apply the technique, the Iterative Imputer tool of scikit-learn was used with the Random Forest classifier (28). Treatment of missing values is shown as Supplementary Material 1. Subsequently, nominal variables were categorized and represented by one hot vector. Finally, numerical and ordinal nominal variables were normalized using the Min–MaxScaler tool of the scikit-learn library, which scales the data within a 0–1 range.

The available data present a high dimensionality, and some of them may be redundant or uninformative, which can negatively affect the performance of the classification models. Variable selection (feature selection) methods allow us to obtain the most relevant sets of variables, optimizing the development of machine learning models. We have used a genetic algorithm, a heuristic optimization technique that simulates the natural process of evolution and performs a bio-inspired exploration of a large space of solutions to find the best combination of features, something unfeasible with traditional feature selection techniques in high-dimensionality problems. A mono-objective genetic algorithm (PyWinEA python library)¹ was used since the main interest was minimizing the number of characteristics. During the process, variables are selected based on a series of previously fixed parameters and a fitness function that needs to be optimized. This function is based on a supervised classification model, using a specific evaluation metric (29). In our case, we tried support vector machine (SVM), K-nearest neighbor (KNN) algorithm, Gaussian Naïve Bayes (GNB), and decision tree (DT) models and selected them relying on the F1-score metric.

Finally, once selected, the best advanced models provided by the genetic algorithm (D2, HA2), we created an interface using Streamlit open-source app framework in Python language where we exported the models to generate a calculator that evaluates the risks as a function of the input variables.