(1046) Prediction of Severe Preeclampsia Antepartum: Machine Learning Algorithms including Social Vulnerability Data

Multi-center analysis of patients who received prenatal care from January 2018 to July 2025 at Northwell Health throughout the New York metropolitan area. Data collected prior to twenty weeks of pregnancy was used for the prediction models and was sourced from the electronic health record system. Three statistical learning algorithms to predict severe preeclampsia were created: (1) weighted gradient-boosted tree model using XGBoost, (2) stacking ensemble including above model, a random forest and a regularized logistic-regression and (3) a logistic regression baseline. Variables considered in the model include maternal characteristics, obstetric and medical history, and social vulnerability data.

Results:
A total of 156,693 patients were used to train the algorithm, of whom 2.54% were diagnosed with severe preeclampsia during their pregnancies. The algorithms were tested on 17,411 patients. The algorithm thresholds were optimized to F1-scores to balance precision and recall with an F1-score of 0.105-0.187. The prediction model yielded an area under the curve of 0.631-0.749, a sensitivity of 15-24%, specificity of 92-98%, and false-positive rate of 1.5-7.4%.  The factors most predictive of preeclampsia, which differed by model, included prior history of preeclampsia, past medical history of hypertension, and early diagnosis of diabetes in pregnancy (prior to 20 weeks).

Conclusion:
Machine learning algorithms yielded high predictive performance for severe preeclampsia risk from data routinely collected early in pregnancy. Our model has predictive accuracy above that which is currently utilized. Additionally, specifically predicting the severe subtype of antepartum preeclampsia can assist in improved and personalized counseling at the first prenatal visit, without additional workup.