(443) Machine Learning for the Prediction of Neonatal Neuroskeletal Injury Following Shoulder Dystocia

Objective: To use machine learning (ML)-based decision tree modeling for the prediction of neuroskeletal injury in neonates following shoulder dystocia (SD).

Study Design: A multi-center retrospective cohort study including all cases of SD across three tertiary care perinatal centers and members of the New York City Maternal Fetal Medicine research consortium, from 2015-2020. A total of 59 variables were collected including sociodemographic and obstetrical variables and outcomes. Tenfold cross-validated decision tree model (DTM) was then constructed for the prediction of neuroskeletal injury among the cases of SD. Neuroskeletal injury was defined as any of the following: Erb’s palsy, Klumpke’s palsy, facial nerve palsy, humeral or clavicular fractures.

Results: Among the 1,155 cases of SD identified, there were 94 (8.2%) cases of neuromuscular injury. There were 56 cases of Erb’s palsy (59.6%), two facial nerve palsy (2.1%), 20 humeral fractures (21.3%) and 28 clavicular fractures (29.8%) (Twelve of the cases had overlapping injuries). There were no cases of Klumpke’s palsy in our cohort. ML-based, DTM encompassing five variables including head-to-body time, maternal BMI, neonatal birthweight, oxytocin use and type of delivery were able to predict neuroskeletal injury with a with 73.4% sensitivity, 70.3% specificity, 0.76 AUC (95% CI 0.70, 0.81, Figure 1) and 0.28 F1 score. Decisions were observed with head-to-body time >1.67 minutes, maternal BMI >29, type of delivery being spontaneous vaginal delivery and forceps assisted vaginal delivery, use of Oxytocin and neonatal birth weight of >3575g.

Conclusion: It is possible to identify a combination of factors that are associated with SD cases that maybe complicated by neonatal neuroskeletal injury with a moderate level of sensitivity and specificity by using DTM models.