Predictive Modeling of Infant Birth Weight with Kolmogorov-Arnold Networks

Birth weight is an important indicator of neonatal health. Because it influences immediate and long-term health outcomes. This study aims to predict infant birth weight using advanced Machine Learning (ML) techniques, focusing on four classification models, namely, Random Forest, KolmogorovArnold Networks (KAN), Multi-Layer Perceptron (MLP), and eXtreme Gradient Boosting (XGBoost). A dataset from the Centers for Disease Control and Prevention (CDC)’s Natality Records, containing 108,082 entries and 37 attributes, was utilized to train and evaluate these models. Feature selection was conducted using Mutual Information. It reveals key predictors, such as maternal morbidity, intervals since the last live birth, and maternal health factors. The models were evaluated based on F1-score, recall, precision, accuracy, and Receiver Operating Characteristics-Area Under Curve (ROC-AUC). Results demonstrated that KAN outperformed all other classifiers. It achieves an impressive micro and macro average AUC score of 97%. XGBoost followed closely with an AUC of 95%, while MLP and Random Forest exhibited scores of 94% and 86%, respectively. This study not only emphasizes the efficacy of ML in predicting infant birth weight but also lays the groundwork for further research in enhancing neonatal health outcomes.