To determine whether transfer-order subgroup modeling improves prediction, explanation, and pre-transfer counseling compared with pooled overall-cohort modeling.
Approach:
Model Development: Developed and compared logistic regression, support vector machine, random forest, XGBoost, LightGBM, and CatBoost models in three cohorts: overall, first-transfer, and second-transfer subgroups.
Feature Templates: Used feature templates T6, T8, and T10 with 6, 8, and 10 variables, respectively, for model evaluation.
Performance Evaluation: Model performance was evaluated using stratified 10-fold cross-validation and summarized binary metrics at ROC-derived Youden thresholds.
Interpretation Techniques: Utilized Shapley additive explanations (SHAP) and logistic regression-based interpretive nomograms for transparent presentation of results.
Key Findings:
CatBoost + T6 achieved an area under the curve (AUC) of 0.704 for the overall cohort and 0.812 for the first-transfer subgroup, indicating predictive performance.
CatBoost + T10 achieved an AUC of 0.825 for the second retained transfer-record subgroup, reflecting its effectiveness in this context.
Interpretation:
Transfer-order subgroup modeling improved model-population matching and provided more interpretable structures for pre-transfer live-birth assessment after freeze-all FET cycles.
Limitations:
Study conducted at a single center, which may limit generalizability.
Retrospective design may introduce biases in data collection and outcome assessment.
Conclusion:
Transfer-order subgroup modeling may enhance predictive accuracy and interpretation for live-birth outcomes in freeze-all FET cycles.
