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

Abstract Background and objective Low birth weight (LBW) is an important adverse neonatal outcome and a public health concern. This study aimed to develop and validate an LBW risk prediction model based on machine learning algorithms using routine clinical indicators from early and mid-pregnancy. Methods This single-center retrospective cohort study included 4,498 singleton pregnant women from a tertiary hospital in Urumqi (February 2014 – December 2024), among whom 579 (12.9%) had LBW. The dataset was randomly split into training (70%) and test (30%) sets using stratified sampling. After variable screening via LASSO regression, five models—logistic regression, random forest, XGBoost, gradient boosting machine (GBM), and neural network—were developed and compared. Model performance was evaluated using area under the ROC curve (AUC), calibration curves, decision curve analysis (DCA), and additional metrics including PR-AUC and Brier score. Internal validation was performed using bootstrap optimism correction. Results LASSO regression selected 10 key predictors. Among the five models, GBM achieved the highest numerical performance, with an AUC of 0.869 (95% CI: 0.837–0.902), a sensitivity of 0.751, and a specificity of 0.866 on the test set. However, DeLong’s test showed no statistically significant difference among GBM, XGBoost, and random forest. Therefore, GBM should be regarded as a selected top-performing model rather than a definitively superior one. The calibration curve demonstrated acceptable agreement (calibration intercept: −0.302, slope: 1.135, Brier score: 0.138), and decision curve analysis suggested positive net benefit across clinically relevant threshold probabilities. Feature importance analysis identified hemoglobin, fasting blood glucose, and diastolic blood pressure as the top three predictors in the GBM model. Conclusions The GBM-based prediction model using routine early- to mid- pregnancy clinical indicators demonstrated strong internal discrimination, acceptable internal calibration, and potential clinical usefulness within the study cohort. However, these findings reflect internal performance only and should not be interpreted as evidence of clinical readiness. Further prospective, multi-center external validation is essential before any clinical application. The model may serve as a candidate tool to support early risk stratification and individualized prenatal care, but should be considered an adjunct to, not a replacement for, clinical judgment.