Prediction of Preeclampsia Using Machine Learning: A Systematic Review

Preeclampsia is one of the leading causes of maternal and perinatal morbidity and mortality. Early prediction is the need of the hour so that interventions like aspirin prophylaxis can be started. Nowadays, machine learning (ML) is increasingly being used to predict the disease and its prognosis. This review explores the methodologies, predictors, and performance of ML models for preeclampsia prediction, emphasizing their comparative advantages, challenges, and clinical applicability. We conducted a systematic search of databases including PubMed, Cochrane, and Scopus for studies published in the last 10 years using terms such as "preeclampsia", "risk factors", "machine learning", "artificial intelligence", and "deep learning". Words and phrases were selected using MeSH, a controlled vocabulary. Appropriate articles were selected using Boolean operators "OR" and "AND". The database search yielded 325 records. After removing duplicates and non-English articles, and completing a title and abstract search 55 research articles were assessed for eligibility of which 11 were included in this review. The risk of bias was found to be high in three of the studies and low in the rest. Clinicodemographic characteristics, laboratory reports, Doppler ultrasound, and some innovative ones like genotypic data and fundal images were predictors used to train ML models. More than ten different ML models were used in the 11 studies from diverse countries like the United States, the United Kingdom, China, and Korea. The area under the curve varied from 0.76 to 0.97. ML algorithms such as extreme gradient boosting (XGBoost), random forest, and neural networks consistently demonstrated superior predictive accuracy Non-interpretable or black box ML models may not find clinical application on ethical grounds. The future of preeclampsia prediction using ML lies in balancing model performance with interpretability. Human oversight remains indispensable in implementing and interpreting these models to achieve better maternal outcomes. Further research and validation across diverse populations are critical to establishing the universal applicability of these promising ML-based approaches.