Machine learning versus traditional formulas for fetal weight estimation: An international multicenter study evaluating prediction accuracy across birth weight percentiles

Abstract

Objective: To assess whether machine learning (ML) offers improved birth weight prediction accuracy, since despite numerous models, the Hadlock formula remains the clinical standard. Methods: A multicenter retrospective study analyzed data from 9674 singleton pregnancies with estimated fetal weight (EFW) within 7 days of delivery. ML models—Linear Regression, Decision Tree, Random Forest, LightGBM, XGBoost, and Neural Networks—were trained using ultrasound and maternal features. Performance was measured by mean absolute percentage error (MAPE), root mean squared error (RMSE), mean absolute error (MAE), accuracy, precision, recall, and F1-score for percentile categories. Results: LightGBM and XGBoost outperformed Hadlock in overall weight estimation (MAPE ~0.065; RMSE ~252; MAE ~190). For birth weight percentiles (<3rd, <10th, >90th, >97th), ML showed marginal or comparable improvement. LightGBM had higher accuracy and F1 for extreme percentiles, whereas Hadlock showed slightly better recall in some cases. Conclusion: ML models, especially LightGBM and XGBoost, enhanced overall weight prediction but offered limited gains in identifying percentile-based risk. The Hadlock formula remains a strong tool for categorizing at-risk fetuses.