Machine learning models to predict low birth weight in the Kenyan Coast

Low birth weight (LBW, < 2500 g) is associated with poor outcomes across multiple domains of child development. Predicting high-risk pregnancies for LBW will therefore enable timely intervention. We aimed to develop and internally validate machine learning models to predict LBW in the Kenyan coast and determine the top predictors of LBW. We developed Machine Learning (ML) predictive models using data collected from pregnant women who delivered at Kilifi County Hospital between 2011 and 2019. Model training was conducted with logistic regression, random forest, extreme gradient boosting (Xgboost) and Tabular Prior-data Fitted Network (TabPFN). We used inverse probability of treatment weighting to adjust for potential confounding arising from differences in antenatal care (ANC) visits (> 4 vs. \(\:\le\:\) 4). Hyperparameter search done using 5-fold cross validation and guided using Bayesian optimization. Internal validation was done using a temporal split. Discrimination was assessed using area under the curve (AUC), calibration using calibration plots and overall performance with the Brier score. Approximately 17% of the 25,699 newborns included in the study had LBW. The TabPFN model showed an AUC of 0.693 (95% CI: 0.686–0.701), with performance comparable to logistic regression (AUC 0.688, 95% CI: 0.673–0.704). Across the four models, gestational age at first ANC, having a multiple pregnancy, mother’s age, history of high blood pressure during pregnancy and history of pregnancy complications were among the most common top predictors of LBW. Our findings indicate that logistic regression performed comparably to the best-performing ML models in predicting the risk of LBW. If effectively integrated into antenatal care workflows, these prediction models could enhance existing public health interventions by supporting risk stratification and targeted maternal care. However, further external and prospective validation, as well as cost-benefit analyses across diverse populations and settings, is needed before the models can be implemented in clinical practice.