What question did this study set out to answer?

To identify the most effective machine learning model for predicting low birth weight in singleton pregnancies.

April 16, 2026Open Access

Development and Validation of Machine Learning Models for Predicting Low Birth Weight in Singleton Pregnancies in Vietnam

Key Points

To identify the most effective machine learning model for predicting low birth weight in singleton pregnancies.
Conducted a prospective cross-sectional study among women delivering at a hospital in Vietnam.
Employed feature selection using the Boruta algorithm and addressed data imbalance with SMOTE.
Trained seven ML algorithms, including logistic regression and random forest, using five-fold cross-validation.
Assessed model performance using area under the curve (AUC), accuracy, sensitivity, and specificity.
Applied SHAP analysis to interpret feature importance.
Prevalence of low birth weight was found to be 8.7%.
The random forest model achieved the highest AUC of 0.778 and accuracy of 0.871.
Logistic regression had the highest sensitivity at 0.581.
Key predictors of low birth weight included preterm birth, primiparity, and absence of gestational diabetes.

Abstract

Objective: This study compared seven machine learning (ML) algorithms to identify the most effective model for predicting low birth weight (LBW) in singleton pregnancies. The primary goal was to develop a high-accuracy screening tool to support clinical decision-making and early intervention. Methods: A prospective cross-sectional study was conducted among women delivering at the Women and Children Hospital of An Giang, Vietnam. Feature selection was performed using the Boruta algorithm, and data imbalance was addressed with the Synthetic Minority Over-sampling Technique (SMOTE). Seven ML algorithms - logistic regression (LR), random forest (RF), support vector machine, k-nearest neighbor, naïve Bayes, artificial neural network, and XGBoost (Seattle, WA: University of Washington) - were trained using five-fold cross-validation. Model performance was assessed using area under the curve (AUC), accuracy, sensitivity, and specificity, while SHapley Additive exPlanations (SHAP) analysis was applied to interpret feature importance and explain the final model. Results: Among 1,838 women with singleton pregnancies (1,678 non-LBW and 160 LBW), the prevalence of LBW was 8.7% (95% CI: 7.5-10.0%). Of the seven ML models evaluated, the RF model achieved the highest overall performance, with an AUC of 0.778, an accuracy of 0.871, and a specificity of 0.909. LR demonstrated the highest sensitivity (0.581). SHAP analysis of the RF model identified preterm birth as the most important predictor of LBW, followed by primiparity, absence of gestational diabetes, abnormal cardiotocography (CTG) findings, pre-eclampsia, and a prior history of LBW. Conclusion: ML-based prediction of LBW enables early identification of high-risk pregnancies and enables timely preventive strategies. In this study, the RF model demonstrated the best predictive performance, with key predictors including preterm birth, primiparity (first-born status), absence of gestational diabetes, abnormal CTG finding, pre-eclampsia, and prior history of LBW. Early identification combined with appropriate perinatal and neonatal care may reduce infant mortality and severe morbidity.

Mark Helpful

Bookmark

Relay

View Full Paper