Prediction of mortality of premature neonates using neural network and logistic regression

Neonatal mortality is one of the important health indicators and mortality prediction is applied for auditing and benchmarking, comparing the outcomes in neonatal intensive care units (NICUs), controlling individual differences in populations in clinical trials and evaluating efficacy. In this research work, we aimed to establish and compare two models (neural network and logistic regression models) for prediction of mortality in premature neonates upon admission to the NICU. This modeling research was conducted based on the information of 1618 neonates for prediction of mortality risk until the 28th day of life. In total, 80% and 20% of the data were considered for training and testing of the designed models, respectively. Finally, we achieved to predict the probability of infant mortality based on the 5th minute after birth data. Modeling was performed with two methods; neural network [multi layer perceptron (MLP) with education of back-propagation (BP)] and logistic regression (binominal form in MATLAB R2016a). The results showed that the MLP (with 60 neurons in the hidden layer) had more acceptable indices compared to logistic regression. While both neural network and logistic regression were able to predict the neonatal mortality risk, the neural network is more effective than logistic regression model in performance comparison.