Gradient Boosting Model Predicts Low Birth Weight in Retrospective Singleton Study

In a single-center retrospective cohort study evaluating routine clinical data, researchers developed and internally validated machine learning models designed to predict the risk of low birth weight among singleton pregnancies during early and mid-pregnancy. Among five distinct algorithmic architectures tested, a gradient boosting machine achieved the highest numerical discriminative performance, recording an area under the receiver operating characteristic curve of 0.869. However, formal statistical comparison using DeLong’s test demonstrated no statistically significant difference in discriminative accuracy between the gradient boosting machine, extreme gradient boosting, and random forest models. As a consequence, the study authors emphasized that the gradient boosting machine should be recognized strictly as a selected top-performing model within the analyzed cohort rather than an inherently or definitively superior predictive tool.

Low birth weight represents a critical adverse neonatal outcome and an enduring public health challenge worldwide, making early risk identification a major clinical objective for prenatal care systems. To investigate whether routine prenatal indicators could reliably forecast this outcome, the study analyzed health records from 4,498 singleton pregnant women receiving medical care at a tertiary hospital in Urumqi between February 2014 and December 2024. Within this single-center study population, 579 pregnant women, corresponding to an overall incidence of 12.9 percent, delivered infants with low birth weight. To prepare the clinical data for supervised machine learning and unbiased validation, the investigators divided the patient dataset into a training subset containing 70 percent of the cohort and an independent test subset containing the remaining 30 percent, using stratified sampling to preserve the proportion of low birth weight cases across both subsets.

To identify the most informative clinical parameters from routine early- and mid-pregnancy examinations, the investigators conducted initial feature selection using least absolute shrinkage and selection operator, or LASSO, regression. The LASSO screening procedure successfully identified 10 key predictors from the available prenatal variables, filtering out redundant indicators to optimize algorithmic training. Following this feature selection step, five separate predictive modeling approaches were developed and compared: traditional multivariable logistic regression, random forest, extreme gradient boosting (XGBoost), a gradient boosting machine (GBM), and an artificial neural network. Each algorithm was trained on the identical 10-variable predictor set derived from the training data, allowing for a standardized head-to-head comparison of their predictive capabilities in identifying low birth weight risk.

When evaluated on the independent test dataset, the gradient boosting machine demonstrated the strongest numerical discrimination, attaining an area under the receiver operating characteristic curve of 0.869, with a 95 percent confidence interval ranging from 0.837 to 0.902. At this level of performance, the model demonstrated a sensitivity of 0.751 and a specificity of 0.866 in classifying low birth weight cases. Nevertheless, statistical evaluation via DeLong’s test indicated that the discriminative differences between the gradient boosting machine, XGBoost, and random forest were not statistically significant, demonstrating that multiple decision-tree ensemble methods achieved comparable classification accuracy. To better understand the biological and clinical drivers behind the gradient boosting model’s predictions, the authors conducted feature importance analysis, which identified maternal hemoglobin, fasting blood glucose, and diastolic blood pressure as the top three most influential predictive variables.

In addition to receiver operating characteristic metrics, the investigators thoroughly assessed model calibration, overall predictive precision, and clinical utility across the patient cohort. Internal validation was systematically conducted using bootstrap optimism correction to evaluate the stability of model performance. The gradient boosting model’s calibration curve revealed acceptable agreement between predicted probabilities and observed clinical outcomes, evidenced by a calibration intercept of −0.302, a calibration slope of 1.135, and a Brier score of 0.138, supplemented by precision-recall area under the curve metrics. Furthermore, decision curve analysis showed that the gradient boosting model produced a positive net clinical benefit across a broad spectrum of clinically relevant threshold probabilities, suggesting that the model provides actionable risk stratification without generating excess false-positive interventions within the study cohort.

Despite achieving strong internal discrimination and acceptable calibration, the researchers explicitly cautioned against viewing the model as currently suitable for direct clinical implementation. Because the findings are derived entirely from a single-center retrospective cohort at one tertiary hospital, the reported performance metrics reflect internal cohort characteristics only and cannot be interpreted as evidence of general clinical readiness. The authors highlighted that extensive prospective, multi-center external validation remains an indispensable requirement before any algorithmic tool can be applied in routine obstetrical practice. Ultimately, the authors concluded that while the gradient boosting framework offers a promising candidate tool to assist in early risk stratification and individualized prenatal care, it must function solely as a supportive clinical adjunct rather than a substitute for professional clinical judgment.

Subject of Research: Machine learning prediction models for low birth weight in singleton pregnancies

Article Title: A machine learning-based clinical decision support tool for predicting low birth weight: a retrospective cohort study

Article References: Ma, J., Li, F., Zhang, Y., Chen, Z., Zhang, Y., Zhao, X., Ku, T., Chang, M., Du, X., Zheng, Y., Wang, L., & Guo, G. (2026). A machine learning-based clinical decision support tool for predicting low birth weight: a retrospective cohort study. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03896-1

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03896-1

Keywords: low birth weight, gradient boosting machine, machine learning, retrospective cohort study, prenatal care, singleton pregnancy, gestational clinical indicators, LASSO regression, XGBoost, random forest, fasting blood glucose, maternal hemoglobin, diastolic blood pressure, decision curve analysis, bootstrap optimism correction, Brier score, DeLong test