A comprehensive first-trimester predictive model for preeclampsia based on multi-indicators and machine learning: A retrospective single-center study

Abstract

Preeclampsia (PE) is a severe, pregnancy-specific disorder that significantly contributes to maternal and perinatal morbidity and mortality. Its unpredictable onset after 20 weeks of gestation underscores the critical need for effective early prediction and intervention. This study aimed to develop a comprehensive predictive model for PE using a wide array of maternal, biophysical, biochemical, and hematological indicators from the 1st trimester. This retrospective study included 100 pregnant individuals with singleton gestations (50 PE, 50 controls). Various early pregnancy indicators, including hematological, biochemical, inflammatory, angiogenic, and biophysical markers, were collected. Least absolute shrinkage and selection operator regression was used for feature selection. Subsequently, 7 different machine learning (ML) algorithms were employed for model development. Model performance was evaluated using receiver operating characteristic curves. An independent external validation cohort of 70 participants (35 PE, 35 controls) was used to confirm the model’s generalizability. Baseline characteristics showed significantly higher early pregnancy systolic blood pressure and diastolic blood pressure in the PE group (P < .001). Early pregnancy indicator comparisons revealed the PE group had significantly higher median white blood cell count, neutrophil count, monocyte count, and C-reactive protein (CRP) levels, and lower median hemoglobin and hematocrit. Derived indices like the neutrophil-to-lymphocyte ratio (NLR) were significantly higher (P < .001). Crucially, placental growth factor (PlGF) levels were significantly lower (P < .001), while uterine artery pulsatility index (UtAPI) was significantly higher (P < .001). Least absolute shrinkage and selection operator regression identified 12 key predictive features, including PlGF, UtAPI, CRP, and NLR. Among the ML models, the neural network model demonstrated the highest predictive performance, with an area under the curve of 0.917. The model maintained strong performance (area under the curve = 0.838) in external validation. SHapley Additive exPlanations analysis confirmed PlGF, UtAPI, CRP, and NLR as the most influential features. We developed a robust predictive model for PE based on early pregnancy biomarkers and ML techniques. The neural network model demonstrated superior discriminative ability in both internal and external validation cohorts. Early identification of high-risk pregnancies using this model could facilitate timely interventions, such as low-dose aspirin, potentially improving maternal and fetal outcomes. Further multi-center prospective studies are warranted to validate the model on a broader scale.