Abstract 1385 Medical Informatics Platform, Saturday, 5/1

INTRODUCTION: Mortality is disproportionately high in ELBW neonates, and the interaction and effects of the known possible predisposing factors on an individual neonate are uncertain. Neural networks are non-parametric pattern recognition techniques that can recognize complex non-linear relationships between independent and dependent variables. HYPOTHESIS: Neural networks are superior to regression analysis in predicting mortality in individual ELBW neonates. METHOD: A database of 23 variables (full set) available within the first six hours of life on 810 ELBW neonates admitted to the NICU (1990 to 1996) was randomly divided into training (n=502), validation (n=59), and test sets (n=249). The training set was used to develop the logistic regression equation and train the neural network models (four layer backpropagation) with mortality as the outcome variable. The models were validated and outcome was predicted on the test set. Forward stepwise regression was then performed on the full set to identify the significant variables (Birth weight, Gestational age, Race, 5 minute Apgar score, Antenatal steroids, Multiple births, Respiratory distress syndrome). The regression model and neural network were then trained, validated, and tested using data sets with only the significant variables, and then with variables excluded one at a time. The techniques were compared by analysis of the receiver operating characteristic (ROC) curves. RESULTS(Table): The area under the curve (AUC) for neural networks was not different from the AUC for regression using the same data (p<0.3). The AUC for analysis based on the significant variables was higher than the AUC based on the full data set for both models (p<0.005). Birthweight or gestational age and the 5 minute Apgar score contributed most to the AUC, and the contribution of individual variables was similar for both models. Specificity, positive predictive value (PPV), and negative predictive value (NPV) were identical for both models based on significant variables at 80% sensitivity (85% specificity, 72% PPV, 90% NPV). CONCLUSIONS: Neural networks were comparable to, but not superior to regression analysis. Analysis based only on significant variables was superior to analysis based on the full data set, both for neural networks and regression analysis. SPECULATION: It is possible that neural networks may not be superior to regression analysis in problems where no clear non-linear relationships exist. The relatively high AUC, sensitivity, and specificity suggest prediction of mortality risk in ELBW neonates is feasible using selected significant variables with either regression or neural networks.

Table 1 No caption available
Full size table