Conventional statistical and machine learning-based prediction of neonatal outcomes in late-onset fetal growth...

To evaluate associations between antenatal fetal biometric and Doppler parameters and neonatal outcomes in late-onset fetal growth restriction (FGR), and to evaluate machine learning models for neonatal outcome prediction.

This retrospective cohort included 137 singleton pregnancies. Continuous missing values were completed using a single iterative chained-equations procedure. Maternal characteristics, fetal biometric measurements, and Doppler parameters, including umbilical artery pulsatility index (UA-PI), middle cerebral artery pulsatility index (MCA-PI), and cerebroplacental ratio (CPR), were analyzed. Associations with neonatal outcomes were assessed using conventional statistical methods, multivariable regression analyses, and machine learning models with repeated 5-fold cross-validation.

NICU admission was associated with higher maternal age and lower gestational age, fetal biometric measurements, estimated fetal weight (EFW), birth weight, Apgar scores, and umbilical cord pH after FDR correction. In the final logistic models, maternal age, previous abortions, and EFW predicted NICU admission; previous cesarean delivery predicted transient tachypnea of the newborn; and maternal age and EFW predicted neonatal sepsis. The birth-weight linear model had the highest explanatory power (adjusted R²=0.662). Among machine learning models, birth weight had the best regression performance (cross-validated R²=0.592), and neonatal sepsis had the highest classification AUC (0.850).

In late-onset FGR, fetal biometry, Doppler findings, and maternal characteristics were significantly associated with neonatal outcomes. Machine learning models showed moderate predictive performance and confirmed findings from conventional statistical analyses. These findings should be interpreted in light of the retrospective study design and the lack of external validation.