(611) Predicting Severe Hemorrhage in Operative Cesarean Deliveries Using Machine Learning

Maternal hemorrhage necessitating transfusion is a significant complication in cesarean deliveries. This study aims to develop machine learning (ML) classifiers to predict hemorrhage necessitating transfusion in mothers undergoing cesarean delivery.

Study Design:

AdaBoost, LightGBM, XGBoost, RandomForest, and Logistic Regression mdoels were developed. The models were trained and tested on data from the American College of Surgeons National Surgical Quality Improvement Program database from years 2009 to 2021. Included patients underwent any cesarean delivery identified by specific Current Procedural Terminology (CPT) codes: 59510, 59514, 59515, 59618, 59620, and 59622 with complete data for the outcome. Cases with missing data on the primary outcome were excluded. Training data included preoperative blood biomarkers, clinical history, and sociodemographic information. All models were evaluated using 5-fold cross-validation to ensure robustness. The training data included preoperative clinical and sociodemographic variables for a large cohort of obstetric patients. The primary outcome was severe hemorrhage requiring transfusion.

Results:

Of the total 43,713 patients, 1,425 (3.3%) experienced the outcome of maternal transfusion. AdaBoost and RandomForest models demonstrated the highest accuracy and AUC ROC scores among the evaluated models. Specifically, the AdaBoost model achieved an AUC ROC of 0.72 (95% CI: 0.72-0.72) and an accuracy of 96.52% (95% CI: 96.34%-96.71%). The RandomForest model showed similar performance with an AUC ROC of 0.71 (95% CI: 0.71-0.71) and an accuracy of 96.59% (95% CI: 96.45%-96.73%).

Conclusion:

AdaBoost and RandomForest models show the highest potential for predicting severe hemorrhage requiring transfusion in operative cesarean deliveries. These models can assist in early identification and intervention, potentially improving patient outcomes.