Predicting length-of-stay in preterm neonates

Abstract

In neonatology, the early prediction of length-of-stay (LOS) may help in decision making. We retrospectively studied the accuracy of two LOS prediction models, namely a multiple linear regression model (MR) and an artificial neural network (ANN). Preterm neonares (n = 2144) were randomly assigned to a training-and-test (75%), or validation patient set (25%). A total of 40 first-day-of-life items (input data) and the date of discharge (output data) were routinely available. Training-and-test set data were used to identify input items with impact on LOS (input variables) using MR analysis to establish a MR prediction model and to train and test an ANN on those selected variables. Fed with validation set data, predicted LOS obtained from MR and ANN was compared individually with actual LOS. Predicted and-actual LOS were highly correlated (for MR, r = 0.85 to 0.90; for ANN, r = 0.87 to 0.92). Conclusion Even first-day-of-life data may contain substantial information with which to predict individual length-of-stay.