Machine learning–based prediction of early pregnancy outcomes in women with recurrent pregnancy loss: a retrospective cohort study

We aimed to develop and internally evaluate models for recorded miscarriage in women with recurrent pregnancy loss (RPL), comparing XGBoost with logistic regression. This retrospective study used 1,848 records from women with at least two prior pregnancy losses in the Sun Yat-sen Medical Big Data Platform during 2015–2023. Predictor measurements were restricted to 13 weeks after the last menstrual period. A fixed stratified split allocated 1,293 observations to training and 555 to testing. Twenty-four candidate predictors were fixed after prior exploration. Within each fitting subset, multiple imputation by chained equations generated ten completed datasets, followed by variance inflation factor (VIF) screening using the maximum VIF across imputations and a retention threshold below 5. The workflow was incorporated into five-fold cross-validation and 500 bootstrap refits. Performance was evaluated using out-of-fold (OOF) and held-out predictions, calibration, decision curves, and training-derived classification thresholds. XGBoost predictions were interpreted using SHapley Additive exPlanations (SHAP). There were 1,121 miscarriages (60.7%) and 727 recorded deliveries (39.3%). All 24 candidates passed VIF screening at every fitting stage. OOF AUCs were 0.856 for both models. Testing AUCs were 0.853 (95% CI 0.820–0.885) for logistic regression and 0.843 (0.810–0.875) for XGBoost; the paired difference was 0.0103 (− 0.0052 to 0.0253). Testing Brier scores were 0.152 and 0.160, respectively. At their separate training-derived thresholds, XGBoost had higher sensitivity (0.754 vs. 0.620), whereas logistic regression had higher specificity (0.885 vs. 0.794). Gestational sac size, hCG, progesterone, age, and hemoglobin were leading SHAP contributors. Both models showed good internal discrimination without clear evidence of a difference in testing AUC. XGBoost underestimated average miscarriage risk. The same-source holdout provides internal evaluation only.