Preeclampsia early risk stratification based on a multiparametric machine learning model and routinely collected clinical data

Introduction. Preeclampsia (PE) remains one of the leading causes of maternal and perinatal morbidity and mortality, while most cases are still diagnosed at the stage of clinically overt disease. Complex prediction algorithms incorporating biochemical biomarkers and Doppler velocimetry demonstrate high accuracy but are poorly suited for large-scale screening in resource-limited settings. Aim: to develop, internally and externally validate mathematical models for predicting PE risk at gestational age of ≤ 16 weeks based on routine electronic health records (EНRs) data and machine learning methods. Materials and Methods . A retrospective cohort study was conducted using de-identified EНRs of pregnant women from eight regions of the Russian Federation spanning 2010–2025. The analytical dataset included 19,955 visits at gestational age ≤ 16 weeks. The composite outcome comprised PE, eclampsia and HELLP syndrome identified by ICD-10 codes. A broad spectrum of clinical, medical history and anthropometric variables was evaluated as potential predictors. Models (logistic regression, gradient boosting, Random Forest, Extra Trees) were trained with adjustment for class imbalance; feature selection was based on SHAP values (SHapley Additive exPlanations indices). Internal performance was assessed on a held-out test set, and independent external validation was performed on a subsample from healthcare facilities of the Republic of Karelia (n = 918). Results. The final Extra Trees model including 35 clinically interpretable predictors achieved a ROC-AUC (Receiver Operating Characteristic curve; Area Under Curve) of 0.871 (95 % confidence interval (CI) = 0.811–0.923) and 0.862 (95 % CI = 0.833– 0.892) in internal and external validation set, respectively. At a probability threshold of 0.04, sensitivity in the external cohort was 0.886, specificity was 0.631, and negative predictive value (NPV) exceeded 0.99. Probability calibration was moderate (mean absolute calibration error was 0.245–24.5 percentage points). The strongest contributors to PE risk were chronic hypertension, history of PE, blood pressure parameters, antiphospholipid syndrome and diabetes mellitus. Conclusion . The Extra Trees model developed on routinely collected EНRs data demonstrates acceptable discriminative ability, high sensitivity and very high NPV and may be considered as a screening tool for early PE risk stratification, provided local calibration assessment and further clinical evaluation.