Development and internal temporal validation of an explainable XGBoost model for predicting clinical pregnancy after frozen embryo transfer

Abstract

Key message An explainable machine learning model was developed using comprehensive clinical variables to predict clinical pregnancy outcomes following frozen embryo transfer. Compared with conventional statistical models, the extreme gradient boosting–based model showed numerically higher discrimination and provided interpretable insights through SHapley Additive exPlanations (SHAP) analysis, supporting individualized decision-making in reproductive medicine. Research question Can an interpretable XGBoost-based model leveraging multi-feature clinical data better predict pregnancy outcomes after frozen embryo transfer (FET) compared to conventional logistic regression? Design This retrospective single-center study included 1,319 patients who underwent FET. Participants were temporally split into a training cohort (n=1,013) and a temporal validation cohort (n=306) according to the date of embryo transfer. Pre-transfer demographic, hormonal, treatment, and ultrasonographic variables were collected. An XGBoost model was developed and compared with logistic regression by assessing discrimination (area under the ROC curve [AUC] and area under the precision–recall curve [AUPRC]), calibration (calibration curves, Brier score, intercept, and slope), and clinical utility using decision curve analysis (DCA). Model interpretability was evaluated using SHapley Additive exPlanations (SHAP). Results The XGBoost model achieved an AUC of 0.854 (95% CI 0.832–0.877) in training and 0.722 (95% CI 0.666–0.779) in temporal validation, with corresponding AUPRC values of 0.840 and 0.707. Calibration in the temporal validation cohort was acceptable (Brier score 0.212; calibration intercept 0.056; slope 0.965). DCA indicated a positive net benefit across clinically relevant thresholds, with high sensitivity at a representative probability threshold of 0.30. SHAP visualizations provided feature-level and individual-level explanations. Conclusion Using single-center data, we developed and internally, temporally validated an explainable XGBoost model for predicting clinical pregnancy after FET. The model delivers accurate, interpretable predictions that may support personalized decision-making and facilitate the clinical integration of artificial intelligence into assisted reproductive medicine.