Prioritising parameters to minimise blood transfusions for postpartum haemorrhage: A machine learning study using national population data

Abstract Background Postpartum haemorrhage (PPH) accounts for approximately 27% of maternal deaths worldwide and often requires blood transfusion, along with its associated risks. In resource‐rich settings like South Korea, transfusion disparities persist, highlighting the need for predictive tools to optimise patient blood management (PBM). Objective To apply machine learning (ML) to rank predictors of transfusion in PPH using national claims data. Methods This was a retrospective cohort study of 123 763 women with PPH (2017–2019) from the National Health Insurance Service (NHIS) database (97% population coverage) in South Korea. The dependent variable was transfusion (yes/no). We used 48 independent variables (demographics, obstetric status, medications and comorbidities) derived from ICD‐10/ATC codes. The random forest model (100 trees, Gini standard impurity criterion) was trained with 80% of the data and validated with 20% of the data after random split (area under the receiver operating characteristic curve [AUC] is a model performance metric). Permutation importance rankings were used for predictors; Shapley Additive Explanation (SHAP) values assessed directional associations and interactions. Results The transfusion rate was 6.7% (8308/123 763). The model achieved an AUC of 77.0% (95% confidence interval: 76.9–77.2). The top 10 predictors by variable importance were coagulopathy (0.311), placenta accreta (0.146), placenta previa (0.129), caesarean section (0.048), uterine atony (0.045), preterm delivery (0.041), age (0.037), vaginal laceration (0.029), iron use (0.025) and socioeconomic status (0.020). SHAP analysis confirmed a positive effect (e.g., maximum SHAP for placenta accreta = 0.2582), and the synergistic interaction between placenta accreta and coagulopathy amplified the risk. Conclusion Coagulopathy and placental abnormalities are the most important drivers of transfusion in PPH. Explanatory ML enhances risk stratification beyond traditional models, informing PBM protocols to curb abuse and inequity. Potential integration into NHIS systems could protect maternal outcomes globally.