JMIR Publications

ABSTRACT

Background:

Low birth weight (LBW), defined as less than 2,500 grams at birth, is a persistent public health issue associated with increased risks of neonatal mortality, long-term developmental challenges, and chronic diseases. Alabama consistently reports higher-than-average LBW rates, emphasizing the need for effective early prediction and intervention strategies.

Objective:

This study aimed to develop an explainable machine learning (ML) framework to predict LBW using large-scale predelivery data, address class imbalance with multiple data-balancing techniques, and generate personalized risk assessments and intervention strategies.

Methods:

We used 514,294 Alabama birth records from 2014 to 2022, with 165 pre-delivery features. Seven balancing techniques were applied: Random OverSampling, Random UnderSampling, SMOTE, ADASYN, Tomek Links, SMOTE-Tomek, and class weight adjustment. Seven ML models were evaluated: Logistic Regression (LR), K-Nearest Neighbors (k-NN), Random Forest (RF), AdaBoost, XGBoost, and Multilayer Perceptron (MLP). Bayesian hyperparameter optimization and stratified train-test splits were used. SHAP (Shapley Additive Explanations) provided global and individual feature importance, and counterfactual sensitivity analysis simulated risk interventions.

Results:

The MLP model with Random UnderSampling achieved the best performance (recall = 0.7779). Key predictors included prenatal visits, pregnancy weight gain, pre-pregnancy weight, and gestational hypertension. Socioeconomic and behavioral variables (e.g., smoking, education, marital status), along with racial disparities, also contributed significantly to LBW risk. Individual-level SHAP and counterfactual analysis demonstrated how modifiable risk factors could lower LBW probabilities.

Conclusions:

This framework combines predictive accuracy, interpretability, and personalization, offering a valuable tool for early, equitable LBW risk identification and intervention.