657: Improving the prediction of shoulder dystocia using artificial intelligence – a novel approach

Abstract

The American College of Obstetricians and Gynecologists (ACOG) recommends consideration of elective cesarean section (CS) to prevent shoulder dystocia (ShD) for women without diabetes carrying fetuses with an estimated fetal weight (EFW) of at least 5,000g and for women with diabetes carrying fetuses with an EFW of at least 4,500g. However, most ShD cases occur in nondiabetic women with normal-sized infants. Our aim was to analyze the potential of artificial intelligence (AI) to predict ShD. A cohort of 53,754 singleton deliveries between 2011 to 2018 from a large tertiary university-affiliated hospital was analyzed. All ShD cases during the study period were identified, and comprehensive prenatal, obstetric and neonatal data from ShD cases were compared with 4:1 controls of singleton vaginal deliveries from the same institution. The AI model incorporated maternal factors (age, height, weight at delivery, smoking, diabetes, insulin treatment), obstetric history (parity, prior CS and prior ShD), fetal sonographic parameters within 2 wks of delivery (BPD, AC, HC, FL, and EFW) and neonatal factors (birth weight and sex). To develop a ShD prediction model, we split the data to training and test sets (80%-20%), and fit a generalized linear model with Lasso regularization on the training set. We repeated the learning with 100 random splits of the data. 239 (0.44%) cases of ShD occurred during the study period. The mean newborn weight among ShD cases was significantly higher than controls (3857g vs 3230g respectively, P < 0.01); however, there was a wide overlap between groups (Fig 1). The AI model identified 12 variables to include in the prediction of ShD with different predictive strengths, some of them protective (Fig 2). Incorporating all factors, the model was able to prevent 41.7% of ShD cases using a 10% elective CS rate cutoff. This CS rate is equivalent to a universal 4000g elective CS recommendation, although doing so would result in prevention of significantly less (35.9%, P<0.01) ShD cases compared with the AI model. We present the first robust study analyzing the potential of AI for prediction of ShD. The model was superior to traditional clinical parameters and therefore we plan to validate the clinical performance of the model using additional datasets with different population characteristics.View Large Image Figure ViewerDownload Hi-res image Download (PPT)