Abstract TU212: Prediction of Early-Onset Preeclampsia Using Deep Learning: A Systematic Review of Clinical and Imaging Models

Abstract

Background: Early-onset preeclampsia is a leading cause of maternal and perinatal morbidity and mortality. Deep learning (DL) offers a promising approach for early prediction, but a systematic assessment of its performance is needed. Objective: This systematic review aims to synthesize evidence from DL models for predicting early-onset preeclampsia using clinical and imaging data. Methods: We conducted a systematic review following the PRISMA 2020 guidelines. A comprehensive search of five electronic databases (PubMed, Embase, LILACS, Scopus, and Web of Science) was performed on June 11, 2025. Before study selection and data extraction, two reviewers, trained in a pilot session, conducted independent and blinded reviews. The risk of bias was assessed using the PROBAST tool. Data on study design, population, model type, input features, validation strategy, and performance metrics were extracted. Descriptive statistics and percentages were calculated to summarize key characteristics. Results: From a total of 15 included studies, sample sizes ranged from 100 to 360,943 participants. Descriptive analyses showed that 53.3% (n=8) of studies used DL exclusively, while 46.7% (n=7) combined DL with traditional machine learning. Clinical registries were the primary data source (93.3%, n=14). The majority of models (73.3%, n=11) integrated maternal characteristics, mean arterial pressure, and biomarkers such as PlGF. Imaging data, including electrocardiogram and retinal fundus images, were utilized in only 26.7% (n=4) of studies. Internal validation alone was reported in 80.0% (n=12) of studies, while both internal and external validation were performed in 20.0% (n=3). Model performance varied, with area under the curve (AUC) values ranging from 0.57 to 0.98. The highest performance (AUC 0.98) was achieved by a model using electrocardiogram data. Risk of bias assessment indicated a low overall risk for 80.0% (n=12) of studies, with the analysis domain being the most frequent source of high risk. Conclusions: DL models demonstrate significant potential for the early prediction of early-onset preeclampsia. However, limited external validation, variability in performance, and underutilization of imaging data underscore the necessity for standardized, prospective validation in diverse cohorts prior to clinical implementation.