(969) Prediction of adverse neonatal outcomes at the start of the second stage of labor

Objective: To develop a risk stratification for unexpected adverse neonatal outcomes at the start of the second stage of labor.

Study Design: This is a retrospective cohort study of individuals at term who reached the second stage or labor and delivered at two hospitals between July 1, 2016, and June 30, 2020. The outcome was unexpected newborn complication, modeled after the Joint Commission metric. Predictor variables included only information that would be known at the start of the second stage, including maternal history, pregnancy characteristics, vital signs, and patterns observed in the nursing documentation of the fetal heart rate tracing (full list below Table). Data were divided into a training and testing sets. Multiple machine-learning variable selection techniques were compared, in addition to standard logistic regression modeling. Discrimination and calibration were assessed between models in the testing set. In the final model (chosen after comparing discrimination and considering parsimony), sensitivity, specificity, negative and positive predictive values were calculated for those in the top 10% and 5% of predicted probabilities in the testing set to better assess model performance.

Results: 19,091 patients were included: 12,994 (68%) in the training and 6,097 (32%) in the testing sets. The prevalence of the outcome in both groups was 4.3%. Random forest, extreme gradient boosting, and Latent Dirichlet Allocation modeling approaches did not result in improved discrimination compared to the logistic regression model. The overall discrimination (AUC 0.65 (95%CI 0.62-0.69)) and calibration were poor, attributed to the rarity of the outcome. However, when examining performance among individuals predicted to be at the highest risk, the model identified patients with >2-3x increased risk of adverse neonatal outcomes.

Conclusion: A model constructed using information available at the start of the second stage can be used stratify patient risk for adverse neonatal events. Future work should focus on how this information could be presented and used in clinical practice to improve neonatal outcomes.