A novel training method to preserve generalization of RBPNN classifiers applied to ECG signals diagnosis.

Abstract

In this paper a novel training technique is proposed to offer an efficient solution for neural network training in non-trivial and critical applications such as the diagnosis of health threatening illness. The presented technique aims to enhance the generalization capability of a neural network while preserving its sensitivity and precision. The implemented method has been devised in order to slowly increase, during training, the generalization capabilities of a Radial Basis Probabilistic Neural Network classifier, as well as preventing it from over-generalization and the consequent lack of resulting classification performances. The developed method was tested on Electrocardiograms. These latter are generally considered non-trivial both due to the difficulty to recognize some anomalous heart activities, and due to the intermittent nature of abnormal beat occurrences. The implemented training method obtained satisfactory performances, sensitivity and precision while showing high generalization capabilities.