External Validation of a Prediction Model Assessing Risk of Delivery in Fetuses with Growth Restriction after Diagnosis of Abnormal Umbilical Artery Doppler

Abstract

Objective We have previously described a model using maternal, antenatal, and ultrasonographic characteristics to assess the risk of delivery within 7 days following diagnosis of abnormal umbilical artery Doppler (UAD) in pregnancies affected by fetal growth restriction (FGR). Therefore, we sought to validate this model in an independent cohort. Study Design Retrospective, single referral center study of liveborn singleton pregnancies from 2016 to 2019 complicated by FGR and abnormal UAD (systolic/diastolic ratio ≥95th percentile for gestational age [GA]). Prediction probabilities were calculated by applying the original model (Model 1) to the current cohort (Brigham and Women's Hospital [BWH] cohort). The variables of this model include GA at first abnormal UAD, severity of first abnormal UAD, oligohydramnios, preeclampsia, and prepregnancy body mass index. Model fit was assessed with area under the curve (AUC). Two alternative models (Models 2 and 3) were created to identify a model with better predictive characteristics than Model 1. The receiver operating characteristics curves were compared using the DeLong test. Results A total of 306 patients were assessed for eligibility, 223 of whom were included in the BWH cohort. Median GA at eligibility was 31.3 weeks, and median interval from eligibility to delivery was 17 days (interquartile range: 3.5–33.5). Eighty-two (37%) patients delivered within 7 days of eligibility. Applying Model 1 to the BWH cohort resulted in an AUC of 0.865. Using the previously determined probability cutoff of 0.493, the model was 62% sensitive and 90% specific in predicting the primary outcome in this independent cohort. Models 2 and 3 did not perform better than Model 1 (p = 0.459). Conclusion A previously described prediction model to predict risk of delivery in patients with FGR and abnormal UAD performed well in an independent cohort. With high specificity, this model could assist in identifying low-risk patients and improve antenatal corticosteroid timing. Key Points