Machine learning-based prediction of adverse pregnancy outcomes in antiphospholipid syndrome using pregnancy antibody levels.

Background: Antiphospholipid syndrome (APS) is a major immune-related disorder that leads to adverse pregnancy outcomes (APO), including recurrent miscarriage, placental abruption, preterm birth, and fetal growth restriction. Antiphospholipid antibodies (aPLs), particularly anticardiolipin antibodies (aCL), anti-β2-glycoprotein I antibodies (aβ2GP1), and lupus anticoagulant (LA), are considered key biomarkers for APS and are closely associated with adverse pregnancy outcomes. This is a prospective observational cohort study to use machine learning model to predict adverse pregnancy outcomes in APS patients using early pregnancy aPL levels and clinical features.

Methods: This prospective study began data collection and follow-up for APS patients undergoing pregnancy monitoring in January 2023, and all data collection and follow-up were completed by January 2025. The samples were divided into the APO group and non-APO group. Multivariable logistic regression and ridge regression were used to identify independent predictive factors for adverse pregnancy outcomes. Six machine learning models were developed: Light Gradient Boosting Machine (LGBM), CatBoost, Extreme Gradient Boosting (XGBoost), Logistic Regression (LR), Random Forest (RF), and Multi-Layer Perceptron (MLP). The performance of these models was evaluated using the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, and F1 score. The best-performing model was further explained using Shapley Additive Explanations (SHAP) analysis. Additionally, decision curve analysis (DCA) was performed to assess the clinical utility of the models.

Results: A total of 708 patients were included. Ridge regression analysis identified aβ2GP1, LA1/LA2, aCL, gestational week at termination, age at first miscarriage, age, BMI during pregnancy, use of medication, >3 adverse pregnancies, 1-2 adverse pregnancies, preeclampsia, and natural miscarriage as significant predictors. Among the six models, the XGBoost model performed the best for predicting adverse pregnancy outcomes (AUROC = 0.864). Decision curve analysis (DCA) further confirmed the superiority of the XGBoost model, and feature importance analysis revealed that aβ2GP1 levels were the most important variable among the 12 factors.

Conclusion: This study demonstrated that the XGBoost model, integrating aPL levels and clinical features, offers an effective approach to predicting adverse pregnancy outcomes in APS patients. The model enables clinicians to quickly and accurately identify high-risk pregnancies, providing valuable support for personalized clinical interventions and treatments.