Understanding risk factors for shoulder dystocia using interpretable machine learning

Abstract

To use interpretable machine learning to identify the most important risk factors for shoulder dystocia (SD). This retrospective study used clinical data from 82,889 singleton vaginal births (24-43 weeks’ gestation) from 20 U.S. hospitals (01/2016 - 12/2021) to train and externally validate a predictive model for SD using 57 features available at the time of delivery plus birthweight (which can be estimated prenatally but is not known until birth). Using bootstrap analysis, an Explainable Boosting Machine (EBM) was trained using data from 59,665 births at 13 hospitals and externally validated using 23,224 births at 7 different hospitals. A logistic regression (LR) model was also trained for comparison. SD occurred in 3.4% of births. An AUC of 0.74 (95% CI 0.73 – 0.75) was obtained for the EBM. Sensitivity for SD was 62%, for a false positive rate (FPR) of 25%. LR yielded an AUC of 0.74 (95% CI 0.73 – 0.75). The most important features for predicting SD were birthweight, maternal height, BMI, cervical dilation at initial exam, and time from admission to full dilation. SD risk increased dramatically and linearly with birthweight and decreased as maternal height increased (Figure 1). Surprisingly, length of 2nd stage was not a strong predictor. Using only the top 5 features gave a slightly higher AUC of 0.75 (95% CI 0.74 – 0.76) and a 65% sensitivity for a 25% FPR, suggesting that eliminating some features might yield a small model with clinically useful SD prediction. AI models trained on these data yielded modest accuracy for predicting shoulder dystocia risk and EBM intelligibility highlighted that consideration of factors in addition to fetal weight, may add to the assessment of risk. EBMs also reinforced the need to accurately estimate fetal weight, as evident through birthweight’s exceptionally large contribution to risk. Our results suggest that a novel, clinically useful model for predicting SD at the time of delivery might be achievable using only a small number of features.View Large Image Figure ViewerDownload Hi-res image Download (PPT)