1148 Racial disparities in cesarean delivery rates: can a machine learning model reduce these biases?

Abstract

Racial Disparities in Cesarean Delivery Rates: Can a Machine Learning Model Reduce These Biases? Variation exists in both rates of and indication for cesarean delivery (CD) by race which may be due to provider bias. As machine learning (ML) models trained on provider behaviors become more common, it is critical to ensure that these biases are not being replicated and amplified. The objective of this study is to identify whether there are differences in CD rate by race in our total delivery population and whether these persist in a supervised ML model. 39,452 deliveries were performed with this distribution: 21,360 White, 6,976 LatinX, 5,048 Asian, 3,733 Black, and 2,335 Other. Actual rates of CD differed by race: White subjects had the lowest rate (17.8%), Black subjects the highest rate (26.5%), p < 0.001 (Figure 1). As expected, the ML model recommended lower CD rates for all races (Table 1); however, the model recommendations reduced the disparity for some races (White, Black, Other) while exacerbating it for others (Asian, Hispanic). Consistent with the literature, significant variability exists by race in the CD rate. Despite excluding race as a predictor, a ML model trained on low CD rate physicians demonstrated persistent differences in recommendation by race. This suggests that low CD rate physicians also treat patients differently based on race, and consideration should be given to mitigation strategies in creating prescriptive ML models.View Large Image Figure ViewerDownload Hi-res image Download (PPT)