Design of hybrid architectures based on neural classifier and RBF pre-processing for ECG analysis

Abstract

A dynamic hybrid architecture is designed for electrocardiogram (ECG) data analysis, combining the fuzzy with the connectionist approach. The data abstraction is performed by a layer of Radial Basis Function (RBF) units and the upcoming classification is carried out by a classical two-layer feedforward neural network. The role of the RBF parameters is investigated, by using different strategies in designing, initializing, and training the RBF pre-processing layer. Generally a more detailed description of the input space by means of a larger number of RBF units does not grant dramatic improvements. An untrained RBF layer allows a compact meaningful description of the input space with performance slightly worse than those of a multilayer feedforward neural network. Other structures with trainable RBF parameters show only a slight improvement of the performance while potentially loosing the interpretability of the RBF layer. The proposed architecture is tested on a real problem in the medical field: the diagnostic classification of ECGs. Several experiments are performed, changing architecture, training strategy, and initial conditions, in order to point out their influence on the overall performance. For the evaluation a large clinically validated ECG database is employed. Some particular configurations have shown a significant improvement with respect to classical methodologies such as statistical classifiers.