An Interpretable Longitudinal Preeclampsia Risk Prediction Using Machine Learning.

Area of Science:

• Obstetrics and Gynecology
• Medical Informatics
• Artificial Intelligence in Healthcare

Background:

• Preeclampsia affects 2-8% of pregnancies and causes up to 26% of maternal deaths.
• Current predictive tools for preeclampsia fail to identify up to 66% of affected patients.
• There is a critical need for improved tools to longitudinally predict preeclampsia risk.

Purpose of the Study:

• To develop and validate a novel tool for the longitudinal prediction of preeclampsia risk.
• To compare the performance of various machine learning and deep learning models for preeclampsia prediction.
• To investigate the explainability and variable relationships within the predictive models.

Main Methods:

• Retrospective study of 120,752 patients across eight hospitals.
• Utilized sociodemographic, clinical, family history, laboratory, and vital signs data.
• Developed and compared linear regression, random forest, xgboost, and deep neural networks across eight gestational time points.

Main Results:

• Preeclampsia incidence was 5.7% in the study population.
• Model performance (AUC) ranged from 0.73-0.91, with external validation.
• The novel model identified 48.6% more at-risk patients compared to first-trimester standard care.

Conclusions:

• A novel, longitudinal preeclampsia prediction tool demonstrates high predictive power using routine clinical data.
• The tool enables early identification and personalized risk assessment throughout pregnancy.
• Implementation in electronic health records can improve clinical decision support and perinatal outcomes.