The prediction of NICU admission and identifying influential factors in four different categories leveraging machine learning approaches

Abstract

Neonatal mortality is a concerning issue for many families worldwide. With advances in perinatal and neonatal care, neonatal mortality has been markedly reduced. While the Neonatal Intensive Care Unit (NICU) significantly prevents neonatal mortality in many cases, this study utilizes various Machine Learning (ML) models to predict NICU admission based on a real-world dataset of pregnant women to address and improve the raised issue. The dataset features are categorized into four categories: demographics, pregnancy, neonatal, and delivery factors. Six ML models analyzed features in each and all categories (complete case): Support Vector Machine (SVM), Decision Tree (DT), Gaussian Process (GP), Multilayer Perceptron (MLP), as well as ensemble models Random Forest (RF) and Bagging. We assessed the performance of each model in predicting NICU admission based on different categories and complete cases. Then, a Majority Vote (MV) approach determines the most influential category. As a result, neonatal factors emerge as the most predictive for NICU admission. The accuracy and AUC (Area Under the Curve) values for MVs in each category are as follows: demographic (accuracy: 0.78, AUC: 0.78), pregnancy (accuracy: 0.91, AUC: 0.91), neonatal (accuracy: 0.96, AUC: 0.96), and delivery (accuracy: 0.91, AUC: 0.91). Finally, utilizing the TOPSIS method based on evaluation metrics, the best model for each category and the complete case is identified: Bagging for demographic and pregnancy factors, RF for neonatal factors, DT for delivery factors, and SVM for complete case. The outcomes could enhance neonatal care and optimize resource allocation in NICUs on time.