Feature reduction and RBF in classifiers based on ANN
Giovanni Bortolan, S. Fusaro · 2002
The application of pruning techniques on artificial neural networks (ANN) and fuzzy pre-conditioning are investigated in the specific problem of the diagnostic classification of 12-lead electrocardiograms (ECG). For this study a large validated ECG database has been employed. A "small size" features space is obtained from the original one reducing it through pruning techniques. In addition, the reduced input space is characterized in terms of a set of linguistic variables by a layer of Radial Basis Functions (RBF) which performs a fuzzy pre-processing or a data abstraction step. The indices used for the validation of the different networks are: the total accuracy, the mean sensitivity and the mean specificity. Different experiments are discussed in detail, pointing out the main characteristics of the resulting architecture. The combination of these techniques has shown satisfiable performances.