Protocol for the external validation of machine learning models for risk prediction in early-onset preeclampsia: a multicenter prospective cohort study.

backgroundEarly-onset preeclampsia (eoPE), defined as preeclampsia occurring before 34 weeks of gestation, remains one of the leading causes of severe maternal morbidity, iatrogenic prematurity, and maternal mortality worldwide. Its heterogeneous and rapidly evolving clinical course makes the prediction of adverse outcomes particularly challenging. Although machine learning (ML)-based models have shown promise in stratifying eoPE risk and supporting clinical decision-making, robust prospective external validation is still lacking. The present study aims to address this gap through a multicenter validation of three ML models previously developed by our group to: (1) predict maternal complications, (2) estimate the likelihood of delivery within ≤ 7 days of diagnosis, and (3) identify the risk of hemolysis, elevated liver enzymes, and low platelets (HELLP) syndrome or placental abruption.

methodsThis prospective multicenter cohort study (2024-2027) recruits pregnant women diagnosed with eoPE after informed consent. Clinical, analytical, and ultrasound data, including all variables used in the models, are recorded from early pregnancy to delivery. The first 100 cases will undergo quantitative external validation of three ML models and a semi-quantitative evaluation by clinicians, after which these cases will be used to retrain the models. The updated models will then be prospectively applied to the next 100 participants. Model predictions and physician´s assessment of the clinical utility of such predictions are generated through a secure web application accessible only after delivery. DISCUSSION: This study addresses a critical gap in the prospective validation of ML-based tools for eoPE, emphasizing methodological rigor, data quality, and external validation in a multicenter setting. By integrating quantitative performance assessment with clinician-centered evaluation, the project seeks to ensure both statistical robustness and real-world clinical relevance. If successfully validated, these models could support more individualized and timely decision-making in eoPE management.