Machine learning-augmented biomarkers in mid-pregnancy Down syndrome screening improve prediction of small-for-gestational-age infants.

Abstract

Background: Adverse fetal growth outcomes (AFGO), primarily characterized by small-for-gestational age (SGA), large-for-gestational age (LGA), low birth weight (LBW) neonates, and macrosomia (Mac), present substantial challenges in early prediction. This study aims to 1) establish a predictive probability for AFGO using routine biochemical markers from prenatal Down syndrome screening, and 2) evaluate the performance of machine learning-based prediction models that incorporate these biomarkers and maternal characteristics for AFGO identification.

Methods: A retrospective analysis was conducted on 2533 singleton deliveries from 2015 to 2017, with available data on early second-trimester biomarkers [α-fetoprotein (AFP), free β-human chorionic gonadotropin (fβ-hCG), and unconjugated estriol (uE3)], as well as pregnancy outcomes.

Results: Serum uE3 demonstrated higher predictive performance for AFGO compared to fβ-hCG or AFP alone, with higher area under the curve (AUC) values in receiver operating characteristic (ROC) analyses (SGA: 0.626 vs. 0.501/0.500; LGA: 0.557 vs. 0.502/0.537; LBW: 0.614 vs. 0.543/0.559; Mac: 0.546 vs. 0.532/0.519). To improve AFGO prediction, we developed four machine learning-based models. Gradient boosting machine (GBM) and generalized linear model (GLM) models demonstrated optimal performance for SGA prediction, achieving AUC values of 0.873 and 0.706, respectively, in the training set (n = 1782, SGA 143), and 0.717 and 0.739 in the test set (n = 751, SGA 68).

Conclusion: Serum uE3 is superior to fβ-hCG and AFP in predicting AFGO. GBM and GLM models significantly enhance SGA prediction performance, highlighting the potential of integrating routine prenatal screening biomarkers with machine learning for early identification of AFGO.