Machine learning-based prediction of clinical pregnancy in infertile women undergoing in vitro fertilization

Infertility is a public health issue that cannot be ignored. We investigated the value of routine clinical data for building prediction models in infertile women undergoing IVF. The 6,076 included women were randomly assigned to a training set and an internal validation set in a 7:3 ratio. Candidate variables were first screened using univariate analysis and then further selected with the Boruta algorithm. Seven machine-learning models were subsequently constructed and compared. Their performance was examined from the perspectives of discrimination, classification, calibration, and clinical utility. Among the included women, 3,056 achieved clinical pregnancy, while 3,020 did not. Boruta feature selection retained age, AMH, AFC, day of embryo transfer, and FSH as confirmed predictors. In the internal validation set, XGBoost performed best overall, yielding an AUC of 0.606 (95% CI: 0.582–0.632), a sensitivity of 0.701, an F1 score of 0.619, and a Brier score of 0.241. SHAP further indicated that age contributed most strongly to model prediction. XGBoost showed the best overall performance among the evaluated models, but its discriminatory ability remained limited. Further model development and external validation are required before clinical application can be considered.