Stage-Specific Models for Predicting Subsequent Pregnancy Loss in Women with Recurrent Spontaneous Abortion

To develop and internally validate stage-specific models for predicting subsequent pregnancy loss in women with recurrent spontaneous abortion (RSA), and to evaluate changes in discriminatory performance associated with the sequential incorporation of sequentially available clinical, biochemical, hormonal, and environmental information.

In this single-center prospective cohort study, we developed and internally validated three stage-specific models for subsequent pregnancy outcomes. The first-stage model included baseline demographic and reproductive variables; the second-stage model further incorporated preconception serum 25-hydroxyvitamin D (25(OH)D); and the third-stage model additionally included gestational age, prenatal hormonal indicators, and pregnancy-course ambient PM_2.5 and O₃ exposure metrics. Logistic regression (LR), random forest (RF), and extreme gradient boosting (XGBoost) models were developed and internally validated using repeated 10-fold cross-validation. Model performance was assessed using ROC-AUC, PR-AUC, Brier score, sensitivity, specificity, and F1 score.

Among 470 women with RSA who subsequently conceived, 86 (18.3%) experienced pregnancy loss again. In RF and XGBoost models, discrimination improved most clearly after incorporating gestational age, prenatal hormonal indicators, and ambient PM_2.5 and O₃ exposure metrics, whereas the addition of preconception 25(OH)D alone provided little incremental value. In internal validation, the RF-based third-stage model achieved the highest discrimination, with a ROC-AUC of 0.935 (95% CI: 0.874~0.996). The XGBoost-based third-stage model also demonstrated improved discrimination, with a ROC-AUC of 0.870 (95% CI: 0.801~0.938).

A stage-specific prediction framework integrating sequentially available clinical, biochemical, and pregnancy-course environmental information improved discrimination for subsequent pregnancy loss among women with RSA, particularly in tree-based models. These findings provide preliminary evidence that progressively incorporating information from different stages of the subsequent pregnancy may improve risk discrimination beyond baseline characteristics alone. However, these findings were derived from internal validation within a single-center cohort, and the third-stage model was limited by a reduced sample size, a small number of pregnancy-loss events, reliance on complete-case analysis, and the use of outcome-anchored environmental exposure metrics that were not available at a predefined prediction time. External validation in larger, independent cohorts and prospective evaluation using exposure windows aligned with a predefined prediction time are needed to determine the robustness and clinical utility of the framework.