Interpretable machine learning prediction of live birth after freeze-all FET cycles across transfer-order subgroups

Transfer-order heterogeneity may affect live-birth prediction after freeze-all FET cycles, but existing prediction studies have rarely modeled first- and second-transfer records separately.

We developed and compared logistic regression (LR), support vector machine, random forest, XGBoost, LightGBM, and CatBoost models in three related single-center analytical cohorts: an overall cohort comprising 990 eligible cycles, a first-transfer subgroup comprising 576 cycles, and a second retained transfer-record subgroup comprising 238 cycles. Feature templates denoted as T6, T8, and T10 refer to retained predictor sets containing 6, 8, and 10 variables, respectively. Model performance was evaluated using stratified 10-fold cross-validation, and binary metrics were summarized at receiver operating characteristic (ROC)-derived Youden thresholds.

CatBoost + T6 was retained for the overall cohort (area under the curve [AUC]=0.704) and the first-transfer subgroup (AUC = 0.812), whereas CatBoost + T10 was retained for the second retained transfer-record subgroup (AUC = 0.825). Parallel LR-based interpretive nomograms were used to support transparent presentation. Shapley additive explanations (SHAP) analysis showed recurrent contributions from female age and ovarian-reserve-related indicators, while the second retained transfer-record subgroup additionally highlighted basal progesterone and interval days.

Transfer-order subgroup modeling improved model-population matching and provided more clinically interpretable structures for pre-transfer live-birth assessment after freeze-all FET cycles.