Development and internal validation of prediction models for maternal adverse pregnancy outcomes in HELLP syndrome: a retrospective cohort study in China

Introduction HELLP syndrome is uncommon but can cause serious maternal adverse pregnancy outcomes (APOs). We investigated whether routinely available clinical and laboratory variables could support outcome-specific risk prediction. Methods We included all women with HELLP syndrome treated at Fujian Maternity and Child Health Hospital from January 2013 to June 2023. Five outcomes were modeled: placental abruption (PAB), posterior reversible encephalopathy syndrome (PRES), postpartum hemorrhage (PPH), disseminated intravascular coagulation (DIC), and a severe maternal composite outcome. The primary analysis used repeated nested cross-validation in the full cohort. Missing-data filtering and imputation, predictor screening, model-family selection, hyperparameter tuning, and threshold estimation were performed within each outer training fold so that the held-out fold evaluated the entire modeling process. Candidate families were ridge logistic regression, elastic-net logistic regression, and random forest. A paired sensitivity analysis using multiple imputation by chained equations (MICE) was conducted to assess the effect of replacing mean/mode imputation. Results Among 272 women, PAB occurred in 33 (12.13%), PRES in 23 (8.46%), PPH in 19 (6.99%), DIC in 9 (3.31%), and the severe composite outcome in 25 (9.19%). The corresponding AUCs were 0.715, 0.677, 0.833, 0.907, and 0.742. Ridge logistic regression was selected in 62%, 74%, 56%, and 62% of the outer folds for PAB, PRES, PPH, and the severe composite outcome, respectively. For DIC, the single-predictor logistic regression fallback was selected in 85% of the outer folds. In the paired analysis, MICE changed AUC by -0.002 to 0.009 and AUPRC by -0.005 to 0.002. PAB, PRES, and PPH did not meet the prespecified analysis criteria. DIC and the composite outcome met these criteria, although both estimates were limited by sparse events and the lack of external validation. Conclusions Although performance was not uniformly strong and external validation is still needed, routinely available data showed potential for early risk stratification. In particular, the severe composite outcome provided a clinically meaningful way to combine individually rare but serious complications and may help identify women who warrant closer surveillance.