Prognosis with neural networks using statistically based feature sets

Abstract

The authors report on several techniques for feature selection utilized in the development of a prognostic tool of predicting recovery for patients with head trauma injuries. The database was examined for features, which were extracted using statistical techniques. ANN (artificial neural network) models were built based on the feature selection of the statistical techniques. These models were trained and tested. Results showed that the ability of the ANN to generalize was dependent on three factors: method of data representation, number of outcome classes, and specific features in the data set. The ANN architecture was kept constant for all the cases. Of the statistical techniques used, the backward selection applied to RA (regression analysis) and stepwise selection applied to LDA (linear disciminant analysis) feature models yielded the best generalizations.<>