To develop and validate stage-specific machine learning models for predicting cumulative live birth rates in women with diminished ovarian reserve (DOR) undergoing embryo transfer.
Approach:
Model Development: Constructed models at three decision junctures: baseline (pre-treatment), post-stimulation (trigger day), and pre-transfer using Random Forest feature selection and six tree-based algorithms.
Cohort and Data: Analyzed 1,234 cycles among women with DOR (AMH ≤1.1 ng/mL) at a single tertiary center, with a training set of 863 and a test set of 371.
Class-imbalance Mitigation: Evaluated strategies to address class imbalance due to low cumulative live birth prevalence (22.8%).
Model Interpretability: Used Shapley Additive Explanations (SHAP) to assess feature contributions in clinically meaningful units.
Key Findings:
The CatBoost algorithm achieved the highest test-set discrimination across stages.
Baseline models yielded an AUC of 0.759 for cumulative live birth prediction.
Post-stimulation markers provided negligible incremental value (Stage 2 AUC 0.755).
Embryological parameters at pre-transfer improved accuracy (Stage 3 AUC 0.793).
SHAP analysis indicated female age and embryo quality were the most important predictors.
Interpretation:
Embryological parameters significantly enhance cumulative live birth prediction in DOR populations, while ovarian response markers add minimal value.
Limitations:
Study conducted at a single center, which may limit generalizability.
Retrospective design may introduce biases.
Conclusion:
Explicit class-imbalance mitigation and interpretable model frameworks are essential for clinically meaningful predictions in reproductive medicine.
