Imputation of missing values by integrating artificial neural networks and case-based reasoning

The implementation of a neural network and case-based reasoning hybrid system to impute missing values was successful. The Canadian Neonatal Network database of 19427 patient cases was used to test this approach. The connection weights of a linear neural network were extracted and used as the match weights in the case-based reasoner to find the closest-matching cases. The missing values in the queries were replaced with the means of the matched cases. A neural network with the weight-elimination cost function and the log-sensitivity stopping criterion was used to develop a new neonatal mortality model that identified the most influential risk factors for predicting mortality in the neonatal intensive care unit. The resultant model classified the patients equally well or better than the statistically-based models in the literature.