Development and validation of interpretable machine learning models for predicting cumulative live birth rates in assisted reproductive technology: a SHAP-based analysis

Objective To develop and validate interpretable machine-learning models predicting cumulative live birth rate (CLBR) in ART patients. Methods In this single-center retrospective cohort (January 2019–June 2024), 7,773 oocyte retrieval cycles were analysed. XGBoost gain ranking, applied to all screened cycles with native handling of missing values, selected eight predictors that were complete in every cycle, so no cycle was excluded for missing data. Six algorithms were compared by stratified 10-fold cross-validation (n=5,441) and an independent test set (n=2,332). We further performed temporal validation (2019–2022 vs 2023–2024), benchmarked against a McLernon-style logistic regression and against a gradient-boosting model restricted to pre-treatment variables, and assessed age-stratified calibration, discrimination (AUC), decision-curve analysis (DCA), and SHAP interpretability. Results Gradient boosting showed the best discrimination on the random-split test set (AUC 0.762; 95% CI, 0.743–0.781; Brier 0.185). Under temporal validation, performance remained robust (AUC 0.712; 95% CI, 0.686–0.737). The full model significantly exceeded the logistic-regression benchmark (0.761 vs 0.691; DeLong p<0.0001; NRI 0.629, 95% CI 0.550–0.712; IDI 0.105, 95% CI 0.090–0.121). However, a gradient-boosting model restricted to the same pre-treatment variables did not outperform that benchmark (0.693 vs 0.691; p=0.72), indicating that the gain was attributable to the inclusion of embryological information rather than to the learning algorithm itself. DCA demonstrated positive net benefit across thresholds 0.01–0.94. SHAP identified female age (mean |SHAP| 0.494) and embryo quality (0.451) as dominant predictors, with an inflection at 38–40 years. Conclusions This interpretable framework delivers robust, well-calibrated CLBR prediction that remains stable under temporal validation while preserving transparency via SHAP. Its advantage over an established pre-treatment benchmark derives from incorporating embryological information rather than from the machine learning algorithm: precise estimation of cumulative live birth requires embryological data that become available only after oocyte retrieval and embryo development, rather than depending on which algorithm is used to model that information.