Prediction and feature selection of low birth weight using machine learning algorithms.

Area of Science:

• Public Health
• Biostatistics
• Machine Learning

Background:

• Low birth weight (LBW), defined as <2,500g, poses significant global health risks, including increased neonatal mortality.
• Identifying LBW predictors is crucial for developing effective interventions and improving infant health outcomes.

Purpose of the Study:

• To identify significant risk factors for LBW using machine learning (ML) approaches.
• To determine the optimal feature selection technique and the best predictive ML model for LBW.

Main Methods:

• Utilized the Boruta algorithm and wrapper method for feature selection.
• Employed Logistic Regression (LR) and ML classifiers (DT, SVM, NB, RF, XGBoost, AdaBoost) for LBW prediction.
• Evaluated model performance using specificity, sensitivity, accuracy, F1 score, and AUC.

Main Results:

• The wrapper method identified key features such as maternal age, education, wealth index, and twin birth.
• Random Forest (RF) demonstrated superior performance in predicting LBW, achieving an accuracy of 85.86% and an F1 score of 0.9243.
• The wrapper method significantly improved the performance of both LR and ML models, with specific models identifying crucial factors like twin birth and cesarean delivery.

Conclusions:

• The wrapper method is the optimal feature selection technique for LBW prediction.
• Machine learning models, particularly Random Forest, outperform traditional methods in predicting LBW.
• Identified risk factors can inform Bangladeshi policymakers to mitigate LBW prevalence.