Leveraging Machine Learning to Enhance Maternal and Fetal Health Outcomes 1. Predicting Postpartum Hemorrhage (PPH) and Transfusion Using Machine Learning Ahmadzia HK et al. (2024) developed multiple ML models, including gradient boosting, to predict PPH (≥1000 mL blood loss) and blood transfusion needs using data from over 185, 000 births in 12 U. S. hospitals (Consortium for Safe Labor). Gradient boosting outperformed others (ROC-AUC 0. 833; PR-AUC 0. 210) when combining antepartum and intrapartum features. Key predictors included delivery mode, oxytocin dose, tocolytic use, anesthesia nurse presence, and hospital type, highlighting clinical and system factors. While combining antepartum and intrapartum data improved accuracy, precision remained modest (~13%), with calibration curves indicating risk overestimation. The study’s strengths are its large multicenter cohort and inclusion of system-level variables; limitations involve older data (2002–2008), imputation methods, and varying transfusion thresholds, suggesting the need for external validation on contemporary datasets and fairness audits considering race, insurance, and hospital characteristics. 2. Admission-Time Prediction of PPH via ML and Statistical Models Venkatesh KK et al. (2020) compared logistic regression variants, random forest, and XGBoost to predict PPH using data from 152, 279 births (10 U. S. sites; 2002–2008). XGBoost achieved the highest performance (C-statistic 0. 93), outperforming random forest (0. 92) and logistic models (0. 87). Important variables included pre-pregnancy/admission BMI, macrosomia, vital signs, trial of labor, gestation number, anemia, and spontaneous labor. PPH incidence ranged from 4. 7–4. 8%, with a notably higher ~15% risk in cesarean deliveries versus ~0. 6% in vaginal births, emphasizing class imbalance and the need for refined thresholding. The study employed temporal and site validation plus decision-curve analysis to assess clinical utility. Limitations include outdated data and potential blood loss measurement errors. Recommendations highlight prospective external validation, recalibration, decision-threshold optimization, and equity auditing. 3.

Predicting Severe Maternal Morbidity (SMM) and Assessing Inequities in Maryland Li Q et al. (2025) utilized linked hospital administrative and American Hospital Association data covering 261, 226 deliveries (2016–2019) in Maryland to predict SMM and examine disparities by race, income, insurance, and language. LASSO models with 18 features attained an AUC ~0. 80, outperforming logistic regression (AUC ~0. 69–0. 71); nonetheless, recall was low due to the rarity of SMM (~76 cases per 10, 000 deliveries). Elevated SMM risk was found among non-Hispanic Black women (risk ratio ≈ 2), low-income residents, publicly insured, and non-English speakers. Using a CDC SMM definition that excluded transfusion and limiting features improved interpretability and convergence while retaining discrimination. Strengths include large linked statewide data and explicit equity analyses. Limitations include low sensitivity for case detection, geographic restriction, and omission of transfusion from SMM definition. Future work should focus on class imbalance strategies, decision-curve analyses aligned to resource allocation, and broader validation. 4. Automated 20-Week Fetal Biometry from Full Ultrasound Scans Venturini L et al. (2025) introduced a fully automated AI pipeline to estimate standard fetal biometric measurements (head circumference, biparietal diameter, abdominal circumference, femur length) by analyzing every frame of 20-week ultrasound videos from the iFIND dataset (7, 309 scans; 1, 457 test scans, ~48 million frames). The system integrates real-time plane classification (SonoNet), per-frame U-Net measurement, and a Bayesian mixture model to aggregate measurements, reject outliers, and produce credible intervals, thus minimizing operator bias. Machine-human measurement differences aligned closely with human inter-observer variability (~95% within range) for most metrics. Aggregating all frames significantly reduced variability compared to single-frame estimations, and test–retest variability was about half the human variability on different ultrasound devices, demonstrating robustness. Bayesian credible intervals provide interpretable confidence and tighten with more data. Strengths include scale, real-time application, repeatability, and uncertainty quantification. Limitations involve limited abnormal fetus representation, variable metric performance (e. g. , transcerebellar diameter), and uncertain generalizability across ultrasound technologies. Recommendations include clinical trials enriched for abnormalities, site-specific model updates, and usability testing for confidence interval display. For prior Research Roundups and the latest updates in global digital health, please follow the provided links.