Machine learning-based prediction of large-for-gestational-age neonates in diabetic and non-diabetic pregnancies

Objective: This study determines whether a machine-learning model integrating sonographic biometry with maternal clinical parameters improves prediction of large-for-gestational-age (LGA) compared with Hadlock's EFW formula. Methods: We conducted a retrospective cohort study including all singleton live births at ≥32 gestational weeks at a tertiary medical center. Predictors comprised biparietal diameter, abdominal circumference, femur length, maternal demographics and anthropometrics, obstetric history, chronic and gestational morbidity, and glucose values from screening and diagnosis. A CatBoost gradient-boosting model estimated the probability of LGA (birthweight ≥90th percentile). Performance was compared with Hadlock's EFW using area under the curve (AUC) and detection at a fixed 10% false-positive rate. A prespecified subgroup analysis evaluated pregnancies with pregestational or gestational diabetes. Performance was assessed with fivefold cross-validation; calibration and utility were examined by decision curve analysis. Results: Among 31 531 parturients, 18.17% delivered an LGA neonate. The model achieved an AUC of 0.946 (95% confidence interval [CI], 0.938–0.955), significantly outperforming Hadlock's EFW (AUC 0.867; 95% CI, 0.854–0.881; P = 0.01) and yielding a higher detection rate at a 10% false-positive rate (79% vs. 63%). The most influential contributors were abdominal circumference, gestational age at delivery, fetal sex, and maternal age. In 3871 diabetic pregnancies, among whom 24% delivered LGA, performance remained high (AUC 0.890; 95% CI, 0.847–0.918) and exceeded Hadlock's formula (AUC 0.820; 95% CI, 0.772–0.863; P = 0.02). Conclusion: A predictive algorithm, incorporating sonographic and non-sonographic features, as developed here, achieved superior accuracy compared to the traditional EFW formula in predicting LGA neonates, in both general and diabetic pregnant populations.