Related Experiment Video
Updated: Aug 5, 2026
Z-Scores for Assessing Ovarian Reserve in Young Patients Undergoing Fertility Preservation
Published on: October 25, 2024
Stage-specific machine learning prediction of cumulative live birth in women with diminished ovarian reserve
Lidan Liu1, Bo Liu1, Qianyi Huang1
1Guangxi Reproductive Medical Center, The First Affiliated Hospital of Guangxi Medical University, Nanning, China.
Background:
Women with diminished ovarian reserve (DOR) experience cumulative live birth (cLBR) rates below 30% following embryo transfer, yet existing prediction tools rely on static baseline parameters and lack interpretability, limiting their clinical utility for counseling about long-term treatment success.
Methods:
We developed and validated stage-specific machine learning models for predicting cumulative live birth per oocyte retrieval cycle-encompassing all fresh and frozen embryo transfers-in 1,234 cycles among women with DOR (AMH ≤1.1 ng/mL) at a single tertiary center. Using Random Forest feature selection and six tree-based algorithms, we constructed models at three decision junctures: baseline (pre-treatment), post-stimulation (trigger day), and pre-transfer. The cohort was randomly split into training (n=863, 70%) and test (n=371, 30%) sets. Class-imbalance mitigation strategies were systematically evaluated given the 22.8% cumulative live birth prevalence. Model interpretability was assessed using Shapley Additive Explanations (SHAP) to quantify feature contributions in clinically meaningful units.
Results:
The CatBoost algorithm consistently achieved the highest test-set discrimination across stages. Baseline models (5 features: female age, male age, AMH, BMI, infertility duration) yielded AUC 0.759 (95% CI 0.704-0.810) for cumulative live birth prediction. Post-stimulation markers conferred negligible incremental value (Stage 2 AUC 0.755, ΔAUC = -0.004). Embryological parameters at pre-transfer substantially enhanced accuracy (Stage 3 AUC 0.793, ΔAUC = +0.034 vs baseline), achieving sensitivity 70.6%, specificity 72.4%, and F1-score 0.536. Algorithms without explicit class-balancing exhibited severely depressed sensitivity (<30%) despite competitive AUCs (0.71-0.77). SHAP analysis revealed that female age (32.8% of total importance) and embryo quality (29.7%) dominated predictions, with non-linear thresholds at age 37 years and clear stratification across embryo grades.
Conclusions:
Embryological parameters substantially enhance cumulative live birth prediction in DOR populations, while ovarian response markers provide minimal added value. Explicit class-imbalance mitigation and interpretable model frameworks are essential for clinically meaningful predictions in reproductive medicine.
