JMIR Publications

ABSTRACT

Background:

Women with polycystic ovary syndrome (PCOS) undergoing fresh in vitro fertilization /intracytoplasmic sperm injection (IVF/ICSI) embryo transfer carry elevated miscarriage risk. However, individualized risk stratification before pregnancy outcome remains challenging in clinical practice.

Objective:

To develop, externally validate, prospectively validate, and interpret a machine learning model for miscarriage prediction in women with PCOS who achieve clinical pregnancy.

Methods:

This multicenter study included 8,911 clinical pregnancy IVF/ICSI cycles across development (n=7,241), external validation (n=1,170), and prospective validation (n=500) cohorts. Predictors were limited to pre-outcome demographic, endocrine, endometrial, and embryo-transfer variables to avoid data leakage. XGBoost was compared with logistic regression, LASSO logistic regression, K-nearest neighbors, decision tree, random forest, support vector machine with radial basis kernel, gradient boosting machine, neural network, and naive Bayes models. Performance was assessed using discrimination (AUC), precision-recall, calibration, threshold metrics, decision-curve analysis, risk stratification, and SHAP interpretability.

Results:

Miscarriage rates were 13.4%, 12.5%, and 12.4% in development, external, and prospective cohorts. XGBoost achieved the highest internal AUC (0.858; 95% CI 0.833–0.883), with stable performance in external (AUC 0.787) and prospective (AUC 0.763) validation. Risk quartiles showed strong stratification: observed miscarriage rates ranged from 0.6% to 37.3% internally, 2.7% to 31.4% externally, and 0.8% to 28.0% in the prospective cohort. SHAP analysis identified AMH, hCG-day endometrial thickness, basal LH, and BMI as the leading predictors. AMH and testosterone dominated early miscarriage prediction, whereas BMI and basal LH were most predictive for late miscarriage.

Conclusions:

This multicenter XGBoost model using routine pre-outcome variables provides reliable and generalizable miscarriage risk stratification for PCOS patients after fresh IVF/ICSI. The model supports individualized counseling and risk-adapted post-transfer management.