Using Fuzzy Clustering Technique for Function Approximation to Approximate ECG Signals

Abstract

Radial Basis Function Neural Networks (RBFNN) has been applied successfully to solve function approximation problems. In the design of an RBFNN, it is required a first initialization step for the centers of the RBFs. Clustering algorithms have been used to initialize the centers, but these algorithms were not designed for this task but rather for classification problems. The initialization of the centers is a very important issue that affects significantly the approximation accuracy. Because of this, the CFA (Clustering for Function Approximation) algorithm has been developed to make a better placement of the centers. This algorithm performed very good in comparison with other clustering algorithms used for this task. But it still may be improved taking into account different aspects, such as the way the partition of the input data is done, the complex migration process, the algorithm’s speed, the existence of some parameters that need to be set in a concrete order to obtain good solutions, and the convergence guaranty. In this paper, it is proposed an improved version of this algorithm that solves some problems that affected its predecessor. The approximation of ECG signals is not straightforward since it has low and high frequency components in different intervals of a heart stroke. Furthermore, each interval (P wave, the QRS complex, T wave) is related with the behaviour of specific parts of the heart. The new algorithm has been tested using the ECG signal as the target function to be approximated obtaining very small approximation errors when it is compared with the traditional clustering technique that were used for the centers initialization task. The approximation of the ECG signal can be useful in the diagnosis of certain diseases such as Paroxysmal Atrial Fibrillation (PAF).