Employing Artificial Neural Network Technique to Predict NICU Utilization by Maternal Clinical Data at Admission of Delivery

We used a neural network data mining technique to develop, validate and compare models for prediction of neonatal intensive care unit (NICU) admission and length of stay (LOS), using readily-available clinical information. Data from the Consortium on Safe Labor (CSL), included hundreds of maternal and neonatal variables captured from electronic medical records of routine obstetric care for 233,844 births (228,668 deliveries from 2002-2008). These 233,844 births resulted in 32,253 NICU admissions, with the following LOS distribution: 0-4d (n=14,413), 5-14d (n=8,206), >=15d (n=9,634) . An artificial neural network (ANN) prototype was used to create a non-linear model yielding an area under the receiver operating curve (ROC) for the NICU admission of 0.67. Prenatal history of threatened pre-term birth, gestational age at delivery, and estimated fetal weight at admission were the top factors from the importance analysis. Prediction performance was improved by inclusion of the following early-neonatal variables: actual birth weight, Apgar scores, birth injury, congenital anomaly, and neonatal fever, increasing the area under the ROC to 0.73. Of note, for those neonates admitted to the NICU, a model using the same pre-delivery data correctly predicted NICU LOS category in 61.5% of the observations. Therefore, maternal demographic and clinical characteristics, available at the time of admission can predict NICU utilization, potentially improving resource allocation and planning and suggesting the potential for improved obstetric clinical decision support systems that utilize medical information technology.