Beyond the unknown: Leveraging machine learning to predict adverse outcomes in pregnant women with aplastic anemia from a representative nationwide cohort

Abstract

Abstract Introduction Aplastic anemia (AA), a rare disorder characterized by bone marrow failure and pancytopenia, poses exceptionally high risks when it occurs during pregnancy. This condition endangers both the mother and fetus, significantly increasing the likelihood of maternal complications such as hemorrhage and infection, as well as adverse perinatal outcomes such as preterm birth and fetal growth restriction. Consequently, pregnancy with AA demands careful management. However, tools to predict these adverse outcomes in affected pregnant women are currently lacking. Here, we applied a machine learning approach to develop and validate a prediction model for adverse pregnancy outcomes in patients with AA, with the goal of guiding early clinical decision-making and improving their overall health outcomes. Methods This study was registered at Clinicaltrials.gov: NCT07101770. We collected data from 310 pregnant women with AA admitted between January 2000 and December 2024 to 15 tertiary hospitals in China. Adverse pregnancy outcomes included at least one of placental abruption, amniotic fluid embolism, postpartum hemorrhage, postpartum infection, maternal mortality, stillbirths, preterm birth, low birthweight, fetal growth restriction, neonatal intensive care unit admission, or neonatal mortality (BJOG, 2014). Feature selection was performed through least absolute shrinkage and selection operator (LASSO) regression. The reliability of the models was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, F1 score, calibration plots, and decision curve analysis (DCA). The SHapley Additive exPlanation (SHAP) method was used to rank the feature importance and explain the final model. Results Among the 310 patients with AA (median age, 30.2 [27.6-33.9]), 201 from 7 specialized tertiary hospitals composed the derivation cohort (training set), whereas an independent cohort of 109 patients from 8 distinct academic medical centers formed the external validation set. To ensure robust model development, the training set underwent a stratified random split, yielding a model-building subset (136 patients, 67.7%) and a hold-out internal validation subset (65 patients, 32.3%), preserving the distribution of adverse outcomes, including postpartum hemorrhage, placental abruption, fetal growth restriction, and preterm delivery. In this study, anemia was present in 280 patients (90.3%). Overall, 195 patients (62.9%) experienced adverse pregnancy outcomes. Notably, among the subgroup with severe aplastic anemia (SAA, n=8), the rate of adverse pregnancy outcomes rose significantly to 75.0% (6/8). These findings underscored the high-risk nature of this cohort, particularly those with SAA, highlighting the critical need for accurate prediction tools to guide targeted antenatal interventions. The data for the variables evaluated in this study, including demographic and clinical characteristics, laboratory results, and treatment, were obtained from patient electronic medical records. Using multivariable LASSO regression, we selected the top five features for model construction: age, hemoglobin level, platelet count, neutrophil count, and the percentage of lymphocytes. Seven state-of-the-art machine learning algorithms were rigorously trained and tuned. The RF model emerged as optimal, demonstrating good discriminative ability both in internal validation (AUC: 0.765, 95% CI: 0.737–0.851) and, crucially, in external validation (AUC: 0.743, 95% CI: 0.723–0.814), confirming its generalizability across heterogeneous health care settings. Furthermore, calibration plots revealed agreement between the predicted probabilities and observed event rates, indicating reliability across risk strata. DCA indicated that the clinical implementation of the prognostic model could benefit pregnant women with AA. Conclusions To our knowledge, it's the world's largest cohort of pregnant women with AA to date. We demonstrated that the model could predict the risk of adverse pregnancy outcomes in patients with AA. The model will help clinicians identify pregnant women at high risk early and provide a basis for individualized patient treatment plans.