Comparison of the prediction of extremely low birth weight neonatal mortality by regression analysis and by neural networks

Abstract

Aims: To compare the prediction of mortality in individual extremely low birth weight (ELBW) neonates by regression analysis and by artificial neural networks.

Study design: A database of 23 variables on 810 ELBW neonates admitted to a tertiary care center was divided into training, validation, and test sets. Logistic regression and neural network models were developed on the training set, validated, and outcome (mortality) predicted on the test set. Stepwise regression identified significant variables in the full set. Regression models and neural networks were then tested using data sets with only the identified significant variables, and then with variables excluded one at a time.

Results: The area under the curve (AUC) of receiver operating characteristic (ROC) curves for neural networks and regression was similar (AUC 0.87+/-0.03; p=0.31). Birthweight or gestational age and the 5-min Apgar score contributed most to AUC.

Conclusions: Both neural networks and regression analysis predicted mortality with reasonable accuracy. For both models, analyzing selected variables was superior to full data set analysis. We speculate neural networks may not be superior to regression when no clear non-linear relationships exist.