Coagulation dysfunction as an independent predictor of early pregnancy loss in women with polycystic ovary syndrome: a machine learning-based retrospective cohort study.
Current studies on adverse pregnancy outcomes in polycystic ovary syndrome (PCOS) have primarily focused on metabolic and endocrine abnormalities, whereas systematic investigations addressing live birth outcomes and coagulation dysfunction remain limited. In addition, there is a lack of comprehensive tools for early pregnancy risk stratification. This study aimed to explore the associations between early pregnancy indicators and early pregnancy loss (EPL) in women with PCOS, identify key risk factors, and subsequently develop a clinically interpretable predictive model to facilitate early risk identification and individualized management.
In this single-center retrospective cohort study, 2,290 women with PCOS who achieved natural singleton pregnancies were included. Participants were randomly divided into a training set (n=1,832) and a validation set (n=458). Demographic, endocrine-metabolic, and coagulation parameters were collected. Independent predictors were identified by multivariable logistic regression. Predictive models were constructed using logistic regression, random forest, and XGBoost algorithms, with model interpretability evaluated using Shapley Additive Explanations (SHAP). Furthermore, a temporally independent external cohort was introduced for external validation to assess the generalizability of the model.
Key predictors included D-dimer, fibrin degradation products (FDP), body mass index (BMI), anti-Müllerian hormone (AMH), and activated partial thromboplastin time (APTT). Elevated D-dimer and FDP levels were associated with increased EPL risk, suggesting enhanced coagulation activation and impaired microcirculation at the maternal-fetal interface. The XGBoost model showed the best performance, achieving an AUC of 0.926 (95% CI: 0.902-0.950). The model also maintained stable discriminative ability in temporal external validation (AUC = 0.705), although a certain degree of calibration drift was observed.
Coagulation abnormalities during early pregnancy are closely associated with the occurrence of EPL in women with PCOS and may serve as independent risk factors. Machine learning models integrating multidimensional clinical indicators can effectively enable early identification of high-risk populations. Temporal external validation further suggested a certain degree of model generalizability, although additional optimization and validation are still required. These findings may provide valuable support for risk stratification and intervention decision-making during early pregnancy.
In this single-center retrospective cohort study, 2,290 women with PCOS who achieved natural singleton pregnancies were included. Participants were randomly divided into a training set (n=1,832) and a validation set (n=458). Demographic, endocrine-metabolic, and coagulation parameters were collected. Independent predictors were identified by multivariable logistic regression. Predictive models were constructed using logistic regression, random forest, and XGBoost algorithms, with model interpretability evaluated using Shapley Additive Explanations (SHAP). Furthermore, a temporally independent external cohort was introduced for external validation to assess the generalizability of the model.
Key predictors included D-dimer, fibrin degradation products (FDP), body mass index (BMI), anti-Müllerian hormone (AMH), and activated partial thromboplastin time (APTT). Elevated D-dimer and FDP levels were associated with increased EPL risk, suggesting enhanced coagulation activation and impaired microcirculation at the maternal-fetal interface. The XGBoost model showed the best performance, achieving an AUC of 0.926 (95% CI: 0.902-0.950). The model also maintained stable discriminative ability in temporal external validation (AUC = 0.705), although a certain degree of calibration drift was observed.
Coagulation abnormalities during early pregnancy are closely associated with the occurrence of EPL in women with PCOS and may serve as independent risk factors. Machine learning models integrating multidimensional clinical indicators can effectively enable early identification of high-risk populations. Temporal external validation further suggested a certain degree of model generalizability, although additional optimization and validation are still required. These findings may provide valuable support for risk stratification and intervention decision-making during early pregnancy.