Using an artificial neural network to predict necrotizing enterocolitis in premature infants

Abstract

Except for degree of prematurity, risk factors for the development of necrotizing enterocolitis (NEC) in very low birth weight (VLBW) infant have not been consistently identified. In addition, fear of NEC determines the majority of VLBW infant feeding regimens in the first postnatal month. About 10-12% of infants weighing less than 1500 grams at birth will develop NEC and about one-third of them will die from the disease. Improved identification of preterm infants at risk for NEC could allow improved infant feeding to focus on growth and nutrition for infants at low-risk of NEC. The objective of this study was to develop an algorithm using artificial neural networks (ANN) to predict prematurely born infants at highest risk of NEC. The majority of ANN's considered optimal used small numbers of variables: 54% used a single variable, 30% used 2 variables, 12% used 3 variables and only 4% used 4 or 5 variables to predict NEC. Sixty-eight percent of the variables were selected first and 79% were selected as second variable at least once. Small for gestational age (SGA) and being artificially ventilated (ventilation: yes/no) were chosen first and second most often among all 57 variables. ANNs as predictive tools provide a first indication for the relative importance of the 57 variables in final decision-making. ©2009 IEEE.