Prediction of Early-onset Preeclampsia Using Deep Learning: A Scoping Review of Clinical and Imaging Models.

Abstract

This scoping review aimed to map the available evidence on deep learning (DL) models for predicting early-onset preeclampsia (EOPE), focusing on both clinical and imaging-based approaches. A comprehensive search of five databases identified 15 eligible studies published between 2018 and 2025. DL and machine learning (ML) models demonstrated heterogeneous performance, with area under the curve values ranging from 0.57 to 0.92 across different prediction windows. Classical ML algorithms, including random forest, gradient boosting, and XGBoost, often outperformed deep neural networks, while multimodal models integrating clinical, biochemical, and imaging features achieved the highest discriminatory capacity. Despite promising results, only a minority of studies included external validation, and performance commonly declined when applied to independent datasets. Nearly half of the studies exhibited a low risk of bias, but substantial methodological limitations persisted, particularly related to participant selection, predictor handling, and analytical procedures. DL imaging-based models showed potential for non-invasive prediction, although evidence remains limited. Overall, current DL approaches offer valuable opportunities for earlier identification of EOPE risk, but their clinical applicability is hindered by issues of generalizability, interpretability, and validation. Further prospective, multicenter studies and transparent reporting are essential to support clinical translation and ensure safe, equitable implementation of DL-based prediction tools.