Development and Evaluation of AI Models for Predicting Low Birth Weight: Insights from NFHS-5 Data

Low birth weight (LBW), defined as birth weight below 2,500 grams, remains a significant public health concern in India, affecting approximately 17.29% of infants according to NFHS-5 data. This study aimed to develop and evaluate artificial intelligence (AI) models for predicting LBW using maternal, socioeconomic, and antenatal care-related factors derived from NFHS-5. A dataset of 662 records was extracted following stringent inclusion criteria, with 530 records allocated for training and 132 for testing. Feature selection was conducted using recursive feature elimination, identifying 24 key maternal predictors, including maternal age, education, BMI, anemia status, antenatal visits, and socioeconomic status. The AutoGluon framework was utilized to build predictive models, incorporating ensemble methods such as CatBoost, LightGBM, Random Forest, Extra Trees, K-Nearest Neighbors, and Neural Networks. The best predictive model was the Neural Network, with an accuracy of 86%, sensitivity 82%, specificity 89%, and AUC-ROC of 0.90. XGBoost and Random Forest were not too far behind with AUC-ROC scores of 0.89 and 0.87 as well. Maternal education, hemoglobin levels and birth intervals were identified as the most relevant predictors of LBW in the analysis of the feature importance. This study highlights the effectiveness of predictive modelling using AI in the detection of different risk factors associated with Low Birth Weight (LBW) and demonstrates its implication in predicting women at high-risk and in administering targeted prevention interventions to prevent LBW. This indicates that machine learning models into maternal healthcare processes could be utilized to conduct better early risk assessment and ensure appropriate precautions are taken in a timely fashion. To enhance the clinical utility, future research should target on external validation and real-world implementation.

Keywords: Low Birth Weight (LBW), Machine Learning, Maternal Risk Factors, NFHS-5, Predictive Modeling, Neural Networks