Machine learning of umbilical artery doppler flow improves prognostication in fetal growth restriction

Abstract

Umbilical artery doppler (UAD) is an important tool in deciding appropriate delivery timing in fetal growth restriction (FGR) cases. UAD flow demonstrating absent or reversed end diastolic flow is associated with stillbirth, but early abnormal UAD are not clearly predictive of poor outcomes. Machine learning (ML) is a useful tool in detecting nuances in blood flow data that may more accurately predict outcomes. The purpose of this study is to determine if UAD ML improves FGR prognostication. UAD and clinical data were obtained from FGR affected pregnancies from the Iowa Intergenerational Health Knowledgebase containing data for pregnancies receiving prenatal care at a major midwestern academic medical center since 2009 (N >63000). UAD severity groups were determined (Normal, Elevated S/D [ > 95% percentile], Absent End Diastolic Flow [AEDF], and Reversed EDF [REDF]). The UAD images were optimized for ML analyses. Xception (Google) machine learning framework was used. Regression analyses were performed to identify important clinical covariates. Birth weight decreased with severity of UAD category (p < 0.001). Gestational age at delivery was shown to decrease with increasing severity of IUGR as well (p < 0.001). In line with clinical practices, diagnosis to delivery time (DDT) increased from Normal to Absent End Diastolic Flow, and then decreased for Reversed End Diastolic Flow (P = 0.03). Preeclampsia was associated with the higher categories of UAD severity (P = 0.0007). Poor neonatal outcomes were not more severe for one group, but high in all categories, ranging from 64-82 percent (p = 0.4). Machine learning of the UAD alone (R2 = 0.84) was more associated with DDT than clinical variables alone (R2 = 0.24) and clinical variables with UAD categorization (R2 = 0.75) ML of UAD alone was more associated with the diagnosis to delivery time then other traditional prediction factors in FGR pregnancies. We conclude that ML of UAD images will add additional predictive power to FGR prognostication.