The study tested machine learning on routine health records in 442,760 Kaiser Permanente Northern California pregnancies, 2011 to 2021. Separate XGBoost models predicted seven complications at six-week points, using only data recorded before each point. At 12 weeks, AUCs were 0.800 for gestational diabetes, 0.789 for gestational hypertension and 0.740 for pre-eclampsia. At 24 weeks, AUCs for preterm birth, stillbirth, placental abruption and small for gestational age were 0.622 to 0.683. Flagging the top 10% by gestational diabetes risk found over 40% of cases; lifestyle counselling was modelled as cost-effective. Kaiser models applied to 6105 All of Us pregnancies did not consistently match locally trained models, even after partial retraining.
Why it is interesting: Most prior models used specialised data or unrepresentative cohorts, whereas this audit uses only routinely collected records. It shows which complications such data can predict, and that models need local tuning in a new health system.
Cite this item
BibTeX
Free sign-up. A digest of the new papers twice a week, grouped by domain.