Predicting adverse maternal outcomes using antenatal and peripartum predictors: A multicenter machine learning study in low- and middle-income countries

OBJECTIVES: To develop and validate machine learning models for composite maternal outcomes using antenatal care and peripartum predictors across eight low- and middle-income countries (LMIC) sites. STUDY DESIGN: Retrospective longitudinal predictive modeling study using prospectively collected registry data. METHODS: We analyzed 2014-2019 data from the Maternal and Newborn Health Registry (MNHR). Two models were developed: one using antenatal predictors and another adding peripartum predictors. Three tree-based ensemble algorithms were evaluated. Models were recalibrated by isotonic regression and assessed using calibration curves and decision curve analysis. Sensitivity analyses examined the contribution of intrapartum management variables and the exclusion of hospitalization from the composite outcome; leave-one-site-out (LOSO) validation assessed cross-site transportability. RESULTS: The dataset included 334,553 women, and the composite adverse maternal outcome occurred in 12.8% of the temporal test set. XGBoost was selected for subsequent analyses. The antenatal model achieved an AUC-ROC of 0.691 (95% CI 0.673-0.707) and the model with peripartum predictors 0.817 (95% CI 0.799-0.832), a difference of 0.126 (95% CI 0.123-0.130), of which 82% persisted after removing mode of delivery. After isotonic recalibration, both models were well calibrated (integrated calibration index ≤ 0.006), with unchanged discrimination, and both yielded greater net benefit than treat-all and treat-none strategies. Performance was heterogeneous across sites, with peripartum predictors improving discrimination in every site under LOSO. CONCLUSIONS: The peripartum model may support peripartum surveillance and postpartum maternal risk stratification, but both models should undergo local evaluation and calibration before implementation.