Can AI Beat Ultrasound in Predicting LGA Complications?

Accurate LGA neonate prediction remains crucial for preventing intrapartum complications such as shoulder dystocia and birth trauma. Therefore, obstetric clinicians frequently rely on late third-trimester ultrasound scans to guide labor management. However, researchers continuously debate whether complex artificial intelligence outperforms standard statistical methods. A large retrospective cohort study evaluated whether sophisticated machine learning algorithms truly improve prenatal risk assessment. Consequently, the findings provide clear guidance for everyday obstetric care.

Evaluating AI Against Logistic Regression in LGA Neonate Prediction

The researchers analyzed 21,743 singleton pregnancies undergoing routine ultrasound scans between 35 and 37 weeks. Furthermore, the team tested multiple machine learning architectures, including XGBoost, random forest, elastic net, and support vector machines. They directly compared these models against standard logistic regression. In addition, the investigators examined both birth weight over the 90th percentile and adverse perinatal outcomes. Interestingly, both traditional logistic regression and machine learning achieved an identical AUC of 0.888 for predicting large neonates. Similarly, both approaches yielded an identical AUC of 0.859 for predicting adverse perinatal outcomes. Thus, machine learning failed to show any superior predictive advantage over conventional statistical tools.

The Primacy of Estimated Fetal Weight in Clinical Practice

Explainable AI analyses, specifically SHapley Additive exPlanations, revealed that estimated fetal weight (EFW) and abdominal circumference centiles drove risk predictions. Moreover, EFW centile alone achieved AUC values ranging between 0.809 and 0.878 across all examined outcomes. In contrast, incorporating extensive maternal demographics and Doppler indices offered minimal incremental diagnostic value. As a result, fetal biometry alone captured virtually all clinically actionable information. Clinicians can confidently rely on standard ultrasound biometry rather than requiring complex computational infrastructure. Hence, standard late-pregnancy fetal biometric assessment remains the most reliable strategy for labor planning.

Frequently Asked Questions

Q1: Did machine learning improve the prediction of LGA neonates over logistic regression?

No, machine learning models achieved the exact same discriminatory performance as standard logistic regression, with an identical AUC of 0.888.

Q2: What was the strongest predictor of adverse perinatal outcomes in large fetuses?

Estimated fetal weight centile served as the dominant single predictor, providing robust risk discrimination across all perinatal outcomes.

References

1. Lopian M et al. Machine learning for prediction of large-for-gestational-age neonate with adverse outcome at routine ultrasound examination at 35-37 weeks. Ultrasound Obstet Gynecol. 2026 Sep undefined. doi: 10.1002/uog.70317. PMID: 42669845.
2. Ben-Haroush A et al. Machine learning-based prediction of large-for-gestational-age neonates in diabetic and non-diabetic pregnancies. Int J Gynaecol Obstet. 2026. doi: 10.1002/ijgo.70751. PMID: 41424415.
3. Illanes SE et al. Machine learning to improve the prediction of Large for Gestational Age (LGA) neonates: a cohort study. BMC Pregnancy Childbirth. 2026;26:9411. doi: 10.1186/s12884-026-09411-8. PMID: 42316044.